Programming note: Reading Ambitiously will be off to Disney World next week for fall break. See you on October 16.

The big idea: Everyone wants a business like Windows

Reading time: 5 minutes

“Developers, developers, developers.”

Steve Ballmer’s famous chant was a pretty good explanation of Microsoft’s strategy. Microsoft supplied Windows, the operating system. Developers built applications customers wanted. Those applications gave customers a reason to choose Windows, which gave more developers reasons to build for it.

Microsoft supplied software like Office and Outlook. Other companies supplied much of the rest. Microsoft’s Software Development Kit (SDK) made that possible, giving developers capabilities like menus, printing, and file access so they could spend more time building applications.

Over time, that flywheel helped make Windows the default platform for business computing.

Now OpenAI and Anthropic are running a similar playbook. Anthropic offers the Claude Agent SDK, and OpenAI offers its Agent Builder. Both want developers building AI with their technology.

The SDK capabilities are different this time. You’ll hear them called the “harness”: the software that helps a model use tools, access information, and coordinate work. It’s the stuff AI developers need but may not want to build themselves.

You can see the appeal.

What’s the bargain this time?

What makes this bargain feel different to me is ownership. As you teach an AI system how to do your work, you’re putting years of accumulated judgment into it. Once you’ve operationalized AI across your business, how much freedom and control will you want? And what are you willing to build, buy, or partner for to achieve it?

This is essentially the question behind AI sovereignty: how much authority and independence a business has over its AI. At the extreme, total independence would mean taking responsibility for the whole stack:

• Applications

• Harness

• Models

• Compute

• Silicon & GPUs

• Land & power

Most enterprises will draw the line somewhere along the way. They may rent the compute while insisting on control over their models. Others may buy the entire application and accept the dependencies that come with it.

The choices enterprises must make at each layer are coming into focus. Those choices will determine how easily they can change providers, control costs, and keep using what they’ve built.

The AI labs have every reason to want it all. But what will their customers bear?

I think the labs understand this. OpenAI president Greg Brockman recently said, “The model is not the product.” The business they’re building extends into the software and capabilities around it.

So where do we go from here? I think we can glean a thing or two by watching the current generation of AI startups attack the opportunity.

Harvey

This company has raised over $1 billion to build AI for the legal industry.

Harvey and OpenAI were close from the beginning, with OpenAI as an early investor. Harvey built around OpenAI’s models and tooling, turning the technology into useful applications for lawyers.

Then, Harvey built its own harness. Earlier this year, it explained why: it needed to work across models, meet customer data-retention requirements, and control costs. Harvey took ownership of that layer.

More recently, it announced its Private Model Program, extending its ambition into the model layer too. Cofounder Gabe Pereyra said, “we’re seeing huge demand from law firms to own their intelligence leveraging private data”.

Harvey is going for it here, and I commend them. They have a chance to become the operating system for their industry, provided their customers are comfortable with the dependencies.

Some want more control.

Latham & Watkins announced it's buying Nvidia servers and fine-tuning open-weight models they own. Morgan & Morgan made a similar announcement. These firms are basically saying, it’s still not enough control for us.

If we leave legal for a second and look at financial services, Jane Street appears to be heading in a similar direction, reaching further down the stack.

It has Etched chips running in its data centers and a $6 billion commitment to CoreWeave for computing capacity. It has also invested in Thinking Machines, which is building tools to help companies develop private models. My bet is that Jane Street’s relationship with Thinking Machines will extend into training private models together.

These companies will still have suppliers. But they’re choosing where they want control and what they’re willing to build to get it.

An asset you own

My read is that these firms want to retain control over the knowledge that makes them good at what they do.

When an experienced lawyer reviews an AI-generated agreement and explains why a clause needs changing, that correction draws on decades of judgment. Capture enough of those corrections, and the system could begin to reflect how the firm thinks.

That right there is IP they want to keep.

Satya Nadella made a similar argument recently when he said, “enterprises should retain full control over their unique and tacit knowledge”.

For plenty of firms, trusting Harvey with this will be a good enough bargain to accept the dependencies. Building and operating more of the stack is a job in itself. Which may be where Harvey is headed next.

In a recent interview, Gabe from Harvey says they’re transitioning from “applications into a full stack AI company.” Helping guide customers through figuring out which parts of the stack customers should own, which Harvey should supply, and which to leave to outside providers.

Some businesses will gladly pay someone else to take that responsibility.

Others will decide their accumulated knowledge is worth the effort of controlling more themselves. It’s still unclear where enterprises will draw that line.

Personally, I’m convinced they’ll want to own and control their intelligence far more than we’ve seen so far.

If using AI well means putting a lot of what has made you great at your work into a system, you’ll want that effort to create an asset that compounds for you rather than rent. You’ll want to keep benefiting from what you’ve taught it, even if you decide to change who supplies the technology.

Some dependencies and lock-in will be accepted. But I suspect the terms will look different from those we’ve seen in this initial AI boom.

Every day, we see more useful AI applications.

The question is, how much freedom and control will you want in the pursuit of those benefits?

Subscribe now

Best of the rest:

🛡️ Nvidia releases software platform to stop AI agents from misbehaving — Nvidia’s new platform sets guardrails around AI agents and can stop them when they break containment. — CNBC

💰 Broadcom agrees to lend Anthropic up to $42 billion for its AI buildout — The financing is tied to Anthropic’s massive compute expansion and could eventually convert into equity, deepening Broadcom’s role as both a chip supplier and financier. — Benzinga

🧠 OpenAI’s real advantage may be distribution, not just better models — a16z argues OpenAI’s durable edge comes from creating new AI behaviors and reaching a massive, diverse user base, giving it a powerful learning loop that compounds across products, models and its broader platform. — a16z

🤖 The rise of the automation engineer — Serval CEO Jake Stauch argues AI is creating roles focused on automating business processes, citing SeatGeek’s redeployment of IT staff after automating over half its requests. — Jake Stauch

🧭 Why we’re building Muse — Alexandr Wang outlines his vision for a personal AI manager that turns ambitions into action by planning, coordinating, and clearing everyday obstacles. — Alexandr Wang

Charts that caught my eye:

→ Why does it matter? Reporting here from the WSJ on the employment index by age group. Is AI the reason we’re seeing such a dip in early career hiring?

→ Why does it matter? 63 new models were released in September! We’re seeing a new LLM every day in 2026. Some open, some closed, some focused on very specific capabilities like Jev. Wow.

Tweets that stopped my scroll:

→ Why does it matter? Dot is OpenAI’s answer to Grok Bot, Muse, and, to a certain extent, OpenClaw and Hermes. The Always-On AI Agent! We wrote about this back in RA 4-10-26.

→ Why does it matter? Would it have even been possible before AI? If so, how long would it have taken? Pretty cool.

→ Why does it matter? Google just dropped Gemini 4 Argon, a new State-of-the-Art model that is right up there with GPT-6 Astra and Claude Fable! Remember, they’re the ones who wrote the paper "Attention Is All You Need” that led to what we now know as an LLM! They were also the 17th entrant into search Engines back in the day! Buckle up.

→ Why does it matter? Marco Rubio, with a Steve Jobs-like Apple presentation, is here to roll out the new America.gov! Go check it out if you haven’t already. Rumors say a Chinese open-weight model powers it. 😂

→ Why does it matter? Wow. Watch this demo. Warn your family this exists.

Worth a watch or listen at 1x:

→ Why does it matter? Pat Grady of Sequoia has all the best mental models to put into perspective where we’re at with AI. You’ll enjoy his presentation to Boston College.

→ Why does it matter? Bill Ackman is covering a lot of ground in this conversation. I particularly enjoy his traditional investing lens in a highly uncertain AI landscape. I once saw him in Central Park, and all I could think to say was, “I hope you get those Herbalife bastards”… For those of you who saw the Netflix documentary.

→ Why does it matter? Dan Shipper is Founder & CEO at Every, a tastemaker in all things AI. A fascinating conversation with Sam after OpenAI’s Dev Day.

Subscribe now

Quotes & eyewash:

“People can’t see every choice you make, so they use the details they can see to infer the quality they cannot see.” - Shane Parrish

→ Why does it matter? The little things add up to the big things

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: agent swarms & machine speed

Reading time: 4 minutes

You’ve probably heard that OpenAI has solved the Navier–Stokes problem. It’s one of seven exceptionally difficult mathematical problems selected by the Clay Mathematics Institute, with a $1 million prize offered for each solution. OpenAI has since said the same model resolved more than 100 other long-standing mathematical problems. It is working with an independent group of leading mathematicians to help review and release the results.

It’s hard to visualize what’s going on here. All I can think of is Matt Damon in Good Will Hunting (a movie that, I learned while researching this week’s edition, many mathematicians don’t like).

Now try picturing 10,000 Matt Damons.

Yes, roughly 10,000 AI agents worked for 88 hours, generating 130 billion output tokens. And to be clear, humans were involved, reviewing results and directing progress. This idea of many agents working together is what you’re hearing described as an “agent swarm” or a “multi-agent” system.

To be fair, these aren’t 10,000 geniuses (yet). They’re more like 10,000 capable PhD students, each trying a different approach, trading notes, and throwing out what doesn’t work. For perspective, the volume of output is akin to one person thinking full-time for ~4,000 years. The agents did it in 88 hours.

The speed of each agent, the number working simultaneously, and their ability to share discoveries create a pace that is hard to comprehend. We’re used to watching one person work through a problem. Even a room full of people has to pause to explain things, compare notes, and go home.

I’m an optimist. When I hear about 10,000 agents working together, I start thinking about the things we might finally be able to attempt. But first, I’m trying to get my head around what that much activity looks like.

The best visual I’ve found is a Minecraft speed run. Do your kids play Minecraft? If so, you already know they can spend hours upon hours exploring and building. Here is a Minecraft speed run that reaches the end of the game in 20 seconds.

Yes, 20 seconds. This example gives us a feel for actions happening faster than we can comfortably follow them.

Now picture the software you click around in every day. Imagine capable agents working through those tasks, checking results, and moving on to the next step. Then imagine many of them doing that simultaneously.

An agent swarm can be given a goal and divide the work. OpenAI described agents communicating and sharing discoveries as they pursued different approaches to Navier–Stokes. A useful finding could give other agents somewhere new to look.

There are plenty of ways to use that for good. But what happens when the agents pursue a goal in ways we never intended?

We’ve all heard the term “alignment.” One way to understand it is that AI reliably pursues what we intend, within the boundaries we set. It reminds me of what I tell my kids every day: do the right thing even when nobody is watching.

A misaligned Agent swarm is at the heart of the Hugging Face incident. This summer, OpenAI agents being tested on cybersecurity tasks were supposed to be isolated. Instead, they found ways to communicate, shared credentials, and coordinated attacks on systems outside their test environment. In the lead-up, one agent noted that the exploit was outside its intended scope. Then it reasoned: “Peers doing it. We should continue.”

Crazy, right? Security failures gave the agents opportunities they should never have had. But the possibility of similar behavior at a greater scale is concerning. What if the target had been national infrastructure or a government agency? That’s the kind of scenario behind Anthropic CEO Dario Amodei’s call this month for AI labs to “pace the frontier” until we better understand what we’re building. Sam Altman and Elon Musk quickly agreed.

But I also don’t want the possibility of something going wrong to consume our entire imagination about what could go right.

Most of us have a list of things we would try if we had more help. A business idea we’ve never had time to investigate. A product we’d like to build. A problem at work that everyone has learned to live with because fixing it would take too long.

We get pretty good at keeping our ambitions within the boundaries of what seems possible. Eventually, we may stop noticing those boundaries. AI could start changing the boundaries altogether.

Imagine being able to explore several approaches to a problem, test the promising ones, and come back with something you can actually evaluate. You would still have to decide whether the result was useful. But an idea that once died because nobody had time to investigate it might finally get a chance.

Today, we’re using AI to help get through the work already on our desks. I wonder what work we would put on those desks if we believed we could accomplish much more. You don’t need 10,000 agents to begin thinking about that.

I don’t know how quickly we get from today’s tools to teams of agents we can confidently trust with much larger goals. The questions about alignment and control are serious. So is the possibility that we’ve been sizing our ambitions for a level of help that could likely change exponentially.

Ten thousand Matt Damons is an extraordinary thing to imagine. What would you put on the whiteboard for them to solve?

Stay ambitious.

Subscribe now

Best of the rest:

🛒 Amazon blocks Meta’s Muse AI assistant in new standoff over agentic shopping – Amazon’s decision to block Meta’s shopping agent exposes the deeper battle over who controls the customer interface, product discovery, data, and advertising economics when AI starts buying on our behalf. – GeekWire

📉 OpenAI, Anthropic Costs Push More Startups to Build Off Cheaper Open Models — Harvey’s AI usage exploded after launching new agents, but the economics broke just as quickly: gross margins fell from roughly 50% at the start of the year to -50% by June as token usage climbed 20x, pushing the $15.6 billion legal AI startup toward cheaper open-weight models and highlighting why model costs are becoming a critical part of the application-layer moat. — Bloomberg

🤖 Everything new coming to Meta’s AI agent Muse — Meta is rapidly turning Muse from an AI assistant into an operating layer for daily life, giving it the ability to run your Mac, manage email, live inside smart glasses, and shop across partners like Shopify, Walmart, and Expedia, with Zuckerberg signaling the eventual business model could be taking a small fee on transactions agents complete for users. — TechCrunch

⚖️ Bessent Targets OpenAI Managers for Hugging Face Incident — Treasury Secretary Scott Bessent says responsibility for OpenAI’s agents hacking Hugging Face sits with the humans running the company, not the agents themselves, sharpening a consequential question for the agentic era: when autonomous AI causes real-world harm, who is legally on the hook? — Bloomberg

Charts that caught my eye:

→ Why does it matter? Wow. SpaceX, Anthropic, and OpenAI are together worth more than the combined first-day value of all 3,365 U.S. tech IPOs since 1980. SpaceX went public in June, and Anthropic is next. Unprecedented times.

→ Why does it matter? AI spending as a percentage of GDP is worth watching. At an estimated 3.63%, it's bigger than the railroad boom and more than three times the telecom and fiber buildout. There are lots of opinions on whether that's good or bad. On one hand, we're laying the infrastructure for what comes next. On the other, will there be enough demand for all that new supply?

→ Why does it matter? As someone who longs for Opus 4.6, I think we may almost be back! When 4.6 came out in February, it was incredible what it could do. For a variety of reasons (my guess is the cost of compute), recent models have felt pretty nerfed. Maybe no longer!

→ Why does it matter? Earlier this week, Anthropic released Opus 5.5 and cut prices by 20%. About 90 minutes later, OpenAI answered with GPT-6 Sol at half the price of its predecessor. The frontier is getting cheaper fast.

→ Why does it matter? Strategically, it helps to own the full stack! The race is on.

→ Why does it matter? A record day last week for open-weight models on Vercel’s AI gateway! Big economic winners would be the American neoclouds that serve them like Nebius and Fireworks.

→ Why does it matter? The market is certainly excited about growth curves like this, yet sometimes the best and most durable way to build value takes time.

Tweets that stopped my scroll:

→ Why does it matter? Reid Hoffman has the essential line on this one: “If you're not embarrassed by the first version of your product, you've launched too late”

→ Why does it matter? This line has lived rent-free in my head the past few days: “The enduring AI business may therefore be the one that passes declining costs on most aggressively, not one that preserves scarcity, bombs its own customers' margins and tries to eat as much of the application layer as possible.”

→ Why does it matter? Brian Kelly, a former Fast Money trader, says his old crypto fund cost about $5 million a year in salaries, bonuses, and office space. His new firm, Bracket22, runs entirely on AI agents for $30,000 to $40,000 a year. Wild.

→ Why does it matter? Blackstone's Jon Gray says that across about 1,400 portfolio companies, GP stakes companies, and borrowers, Anthropic spend grew from a $25 million run rate last September to $525 million. That's 21x in about a year. Remember that GDP chart? Here’s some signal on the demand question.

→ Why does it matter? A very pro-AI sovereignty view from Satya. I particularly keyed in on this line. “And for firms, it’s imperative that they retain full control over their unique and tacit knowledge. Every organization should be able to build its own continuous learning loop/hill climbing machine, without becoming dependent on any one model provider, and have the ability to embed its own knowledge into models and weights they control.”

→ Why does it matter? Nathan Fielder lived with Elizabeth Holmes in the months leading up to starting her prison sentence and recorded every minute of it. 🤯

Worth a watch or listen at 1x:

→ Why does it matter? This oldie but a goodie that I come back to at least once a year. Play upside vs. protecting the downside!

→ Why does it matter? Fascinating conversation with the Founder & CEO at Rogo, who is building AI for Investment Banking & Finance. Loved the story about meeting with 40+ investors and “chewing glass”.

→ Why does it matter? Tobi is no stranger to The Knowledge Project! He loves being at the forefront of a technology shift, and he's guiding Shopify through all things AI. He's also tired of "slop grenades." That's his name for the AI-generated emails, code, and documents people toss around without really checking them. Sound familiar? It's the meat proxy problem from two weeks ago.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Made our first batch of chili last weekend and watched football. 🍁

→ Why does it matter? 😂

"Whatever you've been asked to do, whatever your role is today, do it really well, deliver results, and do it with integrity." - Doug McMillon, Former CEO, Walmart

→ Why does it matter? This is the way.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Run the business, change the business

Reading time: 5 minutes

Happy Friday, Ambitious Readers. I rarely combine my worlds in these notes. This week is an exception.

I’m coming to you from Ridgeline’s Base Camp, our fifth annual customer conference. At the first one, we fit the closing ceremonies into the lobby of our Lake Tahoe HQ. This year we welcomed 375+ attendees from across the industry.

From an outsider’s perspective, you’d have thought we hosted a wedding, not a customer conference. And that is the way it should be. Of everything I’ve learned from Dave Duffield over the years, bringing your customers together and building a real community sits near the top of his playbook.

We also announced the completion of our Series E: $250M of new capital at a $1.4B valuation. Bloomberg did a nice job covering the news. The capital will support the pace of innovation our customers have come to expect, deepen our investment in AI, and help us continue to attract and retain the best talent in the world.

Thank you to the readers who have shaped my thinking, and by extension our strategy, over the years.

Now, let’s jump into this week’s big idea

Ryan Batemen is CTO and CISO at Sands Capital, a $50B firm headquartered in Arlington, Virginia. He builds a portfolio of capabilities the way his portfolio managers build one of securities: underwrite what the firm will need three to five years out, then buy in before the need arrives.

When the business finally asks for a capability, they are not asking for it in three years. They are asking for it tomorrow.

On Tuesday we did a live version of Reading Ambitiously on the main stage. Ryan is one of our weekly readers, he is living at the tip of the AI adoption spear, and he walked the room through five bets he is making for 2026.

His frame for all of them: you have to run the business, you have to change the business.

Ryan has been doubling down on this one for a while, and redoubled earlier this year. He expects model capability to keep climbing and to commoditize across providers, so he isn’t hitching his wagon to any one of them. The bet is on context: three decades of data he can reach back through, going to 1992, combined with new datasets, in ways he doesn’t think another firm can easily repeat.

There is no easy button here. Mastering data, building the governance processes, building the catalogs: that is slow, unglamorous work, and it is the part you can’t buy the week you need it.

“Every single vendor comes to us with a vibe-coded solution that is effectively: what if a research analyst told Claude, please build a front-end application for a research analyst.” Everyone has the same thing. None of it carries his context.

Expectations for AI are exuberant, and they travel at the speed of a hyperconnected world. A breakthrough in a lab is common knowledge within minutes. Adoption follows a different curve. Enthusiasm, breakthroughs, setbacks, in no reliable order.

The gap between those two lines is where every technology leader I know is living right now. The good news is that it’s early. The opportunity is right in front of us.

Ryan holds two jobs that would seem in conflict. The CTO is asked to move fast. The CISO owns it when something breaks. People tell him the roles collide.

He runs them as one job instead. If every AI initiative arrives as an exception, something requiring a committee, a governance review, and a long discussion, then months disappear and nothing ships. He embedded security into the platforms and into the path from idea to production, which is why his teams can pick up new AI tooling quickly. The guardrails were already there.

Get the data processing agreements right, and get zero retention on API use. “You do not want these model providers training their models on your data.”

Everyone is a vibe coder. Source control, peer review, security scanning, identity and access, data pipelines, hosting, monitoring, resilience: all of that is an iceberg below the waterline. In a regulated business, where trust and accuracy are the product, what’s below the waterline is critical.

Sometimes people need convincing the iceberg exists at all. In some cases, our colleagues have only ever seen the exciting prototypes, not what it takes to run them in production.

What he tells the team is that another dashboard is not the win. “You should be getting the answer, getting the trigger, getting the signal from the AI, instead of getting another screen with lots of red and green boxes.”

This is his biggest bet, and the one with the least technology in it. In the last few weeks, he pushed a new AI operating model across the firm, reorganizing how the business operates around what AI can now do.

The forcing function was demand. His AI solutions leads were fielding hundreds of Slacks a day: build me this, can you do that. So they built a funnel. And the people who run it are not the technologists voting on features. They are the AI leaders inside each business team. “These are not the most senior people in each of our three business areas, but they’re the most AI-enabled people.”

Today, one person working with an assistant. Next, one person directing specialized agents across research, reconciliation, drafting, and monitoring. Then teams of agents and people working together.

Ryan wants every analyst covering 10x as many companies in greater depth, getting signal to portfolio managers faster and with better data behind it. So the hard part is human, not technical. People have to picture a version of their role that probably won’t look much like the one they have today. “Access to AI is table stakes today. I don’t think that’s gonna be a differentiator for anyone.”

We’re all near the bottom of that adoption curve. Early is a real advantage. It is also temporary. As Ryan said to us in closing, “Get Open”!

Stay ambitious, my friends.

Subscribe now

Best of the rest:

🏋️ Moats & the Barbell-ification of Software – As AI drives software engineering costs toward zero, traditional moats like switching costs and integrations weaken, pushing software toward a barbell of dominant all-in-one platforms and highly specialized niche products, while the middle gets squeezed. – Mike Vernal on X

🚀 Software is about to eat the world much faster - Marc Andreessen argues AI coding agents like Devin, which now writes more than 90% of Cognition’s production code, could turn 10x engineers into 100x engineers and dramatically expand the amount of software worth building. - Marc Andreessen on X

🛠️ Who owns the software after the vibe is gone? - AI may collapse the cost of building software, but the real cost comes after launch: maintenance, integrations, security, support, compliance, and the human accountability required once a quick internal tool becomes critical infrastructure. - Cu(m^2)ulative

🛡️ Palantir, Nvidia curb AI model use over data fears – Enterprise AI adoption is running into a new trust boundary, as major companies restrict frontier models over concerns about data retention and intellectual property, putting zero-data-retention guarantees at the center of the enterprise AI stack. – Reuters

Charts that caught my eye:

→ Why does it matter? OpenRouter users spent more on OpenAI models than on Anthropic models last week. This hasn't happened for more than 2.5 years!

→ Why does it matter? A 1000x decrease!

→ Why does it matter? A very good description of the difference between wrappers and harnesses. I suspect the terminology here will be simplified in the next 6-9 months. There were a lot of ideas between 1998 - 2005 that ultimately converged to create SaaS.

→ Why does it matter? Claude now leads 26% of the development of its own next version. You’ve heard this idea called recursive self-improvement, whereby the AI develops the AI. This idea is also what has AI Researchers so concerned about the potential for a rapid increase in the pace of improvement. More here.

Tweets that stopped my scroll:

→ Why does it matter? Anthropic is clearly firing shots at leading productivity tools, Microsoft Office and Google Docs. Big updates coming to the integration between Claude CoWork & Design. As a major user of both, I’m pumped!

→ Why does it matter? First Latham & Watkins, now Morgan & Morgan. The legal industry was one of the first to go all in on AI. I suspect this is only the beginning of a major push for control and ownership of the AI learning loop. Open weights + proprietary data + local compute = enterprise sovereign AI stack.

→ Why does it matter? A solid PoV on the aftermath of Dario’s latest note and last weekend's calls for AI regulation.

→ Why does it matter? 👀

→ Why does it matter? Are you somebody who goes out and gets what you want without waiting for the conditions to be perfect? Or are you a victim of your circumstances? Few character traits are more powerful than high agency!

Worth a watch or listen at 1x:

→ Why does it matter? We’ve been talking a lot about Bending Spoons in Reading Ambitiously this year. Bending Spoons went public back in January and has since gone on quite the acquisition spree, buying up Airtable, Vimeo, and most recently Miro. Here is Molly O’Shea from Sourcery with a unique inside look with Lucas Ferrari and team.

→ Why does it matter? AgentSwarms are a big topic of conversation this week. Could what happened at HuggingFace happen at a much larger scale? Some predict more AI agents will use the internet than humans in the near future. Dwarkesh dives into the subject.

→ Why does it matter? Benioff is the G.O.A.T. of enterprise technology keynotes, and he had all the heavy hitters in this week's Dreamforce. I always love watching keynotes because the lead-up forces messaging and positioning to be as clear as possible!

Subscribe now

Quotes & eyewash:

→ Why does it matter? 🤣

→ Why does it matter? The ongoing AI wars explained by fruit and Star Wars. 😂

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Meat proxies and AI slop

Reading time: 3 minutes

You ask a coworker a question, and a few minutes later, a fully polished document appears. By now it’s obvious to most of us that it was written by AI. More and more, we’re not even sure the person read it. And now, you’re responsible for deciphering what it says.

A new term is entering Silicon Valley lingo to describe this experience: a meat proxy.

Coined by software engineer Niklas Gruhn, a meat proxy is someone who passes along AI-generated text without reading, understanding, or validating it.

The receiver has the same access to AI as the sender, yet a proxy now sits in the middle, pasting AI’s response nearly verbatim.

So what? Well, privately, a lot of people are asking me whether the meat proxy problem is a temporary learning curve or a performance problem.

Early on, it is probably a learning curve. Every new tool develops an etiquette. AI has one too, and it doesn't come with a rule book.

I watched this with my kids last Christmas when they were gifted a landline telephone. They had literally no idea how to use it. They’d call someone and then sit silently on the line, they didn’t know to let the person know who it was, or what to say when the person said “Hello”. Ten months later, they’re much better, but there is still room to improve.

Like the telephone, AI etiquette can be coached. I think?

What I am about to say is “Hey, you should read the output from AI before you send it around”. But isn’t that odd to say? So the coaching is: read what AI generated and seek to understand it first? Maybe rewrite it with your own thinking applied?

You might say, well, that’s obvious, and this is where it gets confusing fast. I guess you would assume people are doing that. Thus, isn’t it natural to judge when they don’t? And furthermore, does that mean they simply don’t care? You can follow this thread and see how easy it is to arrive at performance.

We’re all starting to develop a filter for AI slop. Back in August ’25, we predicted slop would be Merriam-Webster's word of the year. Turns out it was!

They went with a simple definition:

“digital content of low quality that is produced usually in quantity by means of artificial intelligence”

AI-generated slop has made us question whether what we’re looking at was created by a human or AI. And we’re not sure, once we’ve determined AI’s assistance, whether or not that’s something we’re OK with.

Two weeks ago, WSJ readers noticed some clear AI assistance in Stanley Druckenmiller’s op-ed on the bond market. Druck immediately owned it. He said:

“At 73, I’m kind of proud of using it… I was a B student in English, but an A+ in economics. These are my ideas and I’ve been speaking about them for 15+ years. As anyone who knows me knows, I would prefer that I was a great writer, but that’s not who I am.”

There is no doubt that Mr. Druckenmiller is a global authority on macroeconomics. The ideas were definitely his, and AI helped him with the words.

By now, if you haven’t figured it out, I too use AI to help me write Reading Ambitiously. But no matter what, I read every word dozens of times before you do. I have a personal sense of pride knowing that whatever goes in front of you is the highest quality content that I’ve had the time to produce. If I don’t love what gets created and feel a sense of pride when I hit publish on Friday mornings, I am seriously disappointed, not just for myself, but for our readers.

That's where the answer to this conundrum lies. AI can assist and help, it should not replace the sense of pride we have in our work to deliver high quality output. And that’s why leaders are starting to question performance. What does it say when someone sends you something that makes you feel as if they avoided the thinking, pride, and quality altogether?

For now, let’s chalk it up to a learning curve. It’s early, and AI etiquette is just starting to emerge. In the interim, before you hit send on something AI helped you generate, think about your receiver. Will you feel pride and ownership in what you deliver? Will they experience that too?

Strive for that. Don’t be a meat proxy!

Stay ambitious, my friends.

Subscribe now

Best of the rest:

💳 Ramp in Talks for New Funding at $60 Billion Valuation — Ramp is exploring a $1 billion raise at a $60 billion valuation, up from $44 billion just three months ago. — Bloomberg

🧮 On the Navier–Stokes Millennium Prize Problem – OpenAI says an internal model significantly more capable than GPT-6 Astra, coordinated across roughly 10,000 agents and verified in Lean, produced a proof resolving one of mathematics’ seven Millennium Prize Problems, a striking demonstration of AI moving from solving known problems to generating original scientific breakthroughs. – OpenAI

🙋 I Study Successful People. They Have One Habit in Common. — Angela Duckworth argues that high achievers are not relentlessly self-sufficient, they are unusually good at asking for help, tapping their networks, and creating the support systems that help them succeed. — The New York Times

🏦 Introducing ChatGPT for Financial Services – OpenAI is going vertical in finance, combining GPT-6 Astra with built-in premium datasets, financial workflows, firm templates, and enterprise controls to move ChatGPT closer to the core research, modeling, and client work done inside banks and investment firms. – OpenAI

Charts that caught my eye:

→ Why does it matter? I’d never seen this before. Major Hedgefunds plotted against highest pay and highest stress.

→ Why does it matter? Leading SaaS companies are releasing increasingly advanced AI capabilities, and they have a major superpower: distribution. They can put these features in front of customers who already use their software every day. That’s a pretty good head start.

→ Why does it matter? Grand View Research projects the enterprise agentic AI market will grow almost 5x between this year and 2030.

Tweets that stopped my scroll:

→ Why does it matter? 152m views on this tweet. My gosh. You’ve probably seen it already, and the theories about his motives are all over the place. One detail worth paying attention to: Axios reports that he resigned from Anthropic before vesting any of his equity.

→ Why does it matter? Muse is the new personal AI assistant from Meta. Earlier this week, I asked it to find me a new dress shirt. It did the research, read the reviews, and then purchased it for me. We’ll see on September 16th when it arrives how it did! This is what we were getting at with Grok Bot: great products make the technology less visible and the benefit more tangible. Very cool!

→ Why does it matter? Bending Spoons is on a tear acquiring SaaS companies. Miro was once valued at $17B+ and is doing around $600M in ARR. Its digital whiteboard became incredibly popular during COVID. This brings us back to last week’s edition: how much of a customer’s business depends on what’s underneath that whiteboard? That’s what I’d want to know before calling it a nice-to-have.

Worth a watch or listen at 1x:

→ Why does it matter? Great interview with Base Power’s founder Zach Dell (son of Michael Dell). I particularly enjoyed the way Zach talks about “working in the business” vs. “working on the business”

→ Why does it matter? Good to catch up on Zuck’s recent thinking. I’ve only had Muse for three or four days, but I like how much of it is built around everyday life. It let me know about a family festival in my town this weekend. Pretty useful! Here come the incumbents, and gosh do they have distribution.

→ Why does it matter? When are you showing conviction, and when are you just refusing to admit you’re wrong? Easy to answer once you know how the story ends. Shane Parrish and Geoffrey Cain look at the years when Jobs was out of Apple and his eventual comeback was far from obvious.

Subscribe now

Quotes & eyewash:

→ Why does it matter? 😍

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: The Last S in SaaS

Reading time: 6 minutes

Last November, I had a Saturday night to myself. I had installed Claude Code but barely used it. So I got a six-pack and loaded it up.

What can this thing really do?

I’ve always wanted a better way to keep up with my relationships. A personal CRM of sorts. The breadcrumbs are scattered across iMessage, LinkedIn, email, and years of notes. I have tried to track it all, but Dunbar’s number is real.

Could AI help me beat Dunbar?

That was my first question for Claude. A few hours later, the idea had become working software.

Suddenly, it exists. You can click it, change it, and ask for another feature. I call this moment “the epiphany.” When I meet people now, I ask: Have you had it yet?

Once you have it, you know. We will never build software the same way again.

Pretty much every conversation I have lately includes some version of the epiphany. People armed with Claude Code who have never built software before are dreaming up personalized dashboards, workflows, and applications. What they love most is the fit. The information appears the way they think about it. The features are the ones their hearts have always desired.

The “epiphany” also hit the market. If AI can write the software, why pay for applications that never quite work the way you want? Software stocks sold off hard. Wall Street called it the “SaaS-pocalypse.”

Iceberg dead ahead

A few weeks ago, I met a portfolio manager who had always wanted to visualize investment decisions a particular way. Existing software never quite gave them what they wanted. AI did. They described it, built it, and started falling in love with the result.

Then the prototype needed to become something their colleagues could rely on. Some of the data was wrong. Sharing it securely was hard. Keeping it running was confusing. Each new feature broke an old one. Every improvement added something else to maintain.

They had built the part of the software they could see. Underneath it was an iceberg.

Every piece of software has one. Below the waterline are permissions, data integrity, integrations, testing, backups, monitoring, security, compliance, incident response...

Do I need to go on?

AI is collapsing the cost of the first “S” in SaaS: software. It has barely touched the cost of the last “S”: service.

AI will come for the last S too. It already writes tests and monitors. But most of the iceberg is not labor. It is accountability. AI cannot stand behind the SOC 2, carry the liability, or answer the phone when the system is down. That is what the subscription fee has been buying all along.

Why SaaS exists in the first place

Rewind the tape. Enterprises used to heavily customize their software. They modified code, worked directly with the database, and bent the system around how their business operated.

Freedom came with a price. Each customization made the next upgrade more dangerous. New versions broke old customizations and integrations. Companies delayed upgrades, accumulated technical debt, and became responsible for keeping the machine they built running.

SaaS changed that. Vendors operated the core application and moved customization into controlled layers that could survive the next update. The customer used the software. The vendor ran everything the customer never wanted to think about.

That is much of what we pay for in the last “S”.

SaaS didn’t eliminate the iceberg. It moved it out of each customer’s IT department and into the vendor, spreading the cost across thousands of customers. One tradeoff for another.

Build the views. Rent the iceberg.

This is what Salesforce is attempting with Headless 360 and Claudeforce, its new partnership with Anthropic.

Headless software separates the interface from the underlying system. Today, a salesperson opens Salesforce and navigates its screens to reach the data and workflows below. In the headless version, the salesperson starts in Claude. Claude directly invokes Salesforce’s data, rules, permissions, and workflows.

The first Claudeforce release includes 37 prebuilt sales skills, from meeting prep to pipeline review.

Salesforce’s Patrick Stokes:

“We don’t actually mind if you don’t use Salesforce’s products exclusively through a user interface designed for a human.”

The interface Salesforce spent more than 20 years building is, by its own admission, optional.

Marc Benioff put the iceberg in two sentences:

“Probabilistic models alone can’t run a company. Deterministic systems alone can’t reason.”

The model reasons. The iceberg does the same thing every time.

Salesforce is trading the interface for the system underneath. Users get software shaped around how they think. Salesforce keeps the iceberg.

The market has already voted

Within a day of unveiling Claudeforce and reporting its quarter, Salesforce stock rose 22 percent, its biggest single-day gain since 2020.

Some of that was the quarter. Salesforce beat and raised guidance, though much of the beat was a paper gain on its stake in Anthropic, not software sold.

In February, the market punished software companies for the first S. In August, it paid Salesforce a premium for the last one.

Seats grew last quarter, Benioff said quickly. But headless puts pressure on the seat. If Claude is the interface, what exactly is a user? Salesforce’s answer is a meter. It has sold consumption-based Flex Credits for more than a year, and half of new bookings last quarter came from customers “refilling the tank.”

Do the math: 100,000 humans at $350 a seat is $35 million. Replace them with a million agents calling the API around the clock, and now we’re talking volume, not headcount.

Claudeforce today arrives on two invoices, one from Salesforce and one from Anthropic. As Stokes admitted, “You can’t buy this on one piece of paper at the moment.”

Twenty years of muscle memory is a moat. If the interface is Claude, the system underneath becomes easier to swap. Salesforce is wagering that its system of record is sticky enough on its own.

On one side are the systems of record. They own the data, permissions, compliance, and integrations. Their iceberg is worth renting, and headless puts it in front of more agents. On the other side are the views: a nice interface on top of someone else’s data. AI just made that interface free.

If you own the iceberg, headless is an opportunity. If you are the view, it maybe the SaaS-pocalypse arriving on schedule.

The Iceberg Test

Leaders need a simple test. Before promoting an AI prototype into something your business depends on, ask:

1. Where does the authoritative data come from, and how will we know it’s right?

2. Who can see, change, approve, and audit what the system does?

3. What catches an old feature breaking when a new one is added?

4. Who maintains the integrations when models and connected systems change?

5. What happens at 8:00 am on Monday if it stops working?

If carrying the iceberg is part of what makes your company different, build it deliberately. If it isn’t, paying someone else to carry it may be the cheapest software decision you make.

On that rare Saturday night last November, I thought the epiphany was that I could finally build the CRM I had always wanted. I still believe that.

But I also want the data to be right. I want the next feature to leave the first one working. And I do not want to become the software company responsible for keeping it alive.

Headless may finally give us both: software shaped around how we think, with someone else carrying the iceberg beneath the surface.

I can’t wait. Stay ambitious.

Subscribe now

Best of the rest:

🐎 The Harness, the Horse, or the Hay — Brett Queener argues that AI will collapse most knowledge work into a single primary application, leaving three durable places to build: the foundational models, the vertical “harness” that owns a user’s entire job, or the infrastructure that keeps it all running. — Tales from The Bonfire

🧠 GPT-6 Astra — OpenAI is calling Astra a generational leap across software engineering, science, computer use, and cybersecurity, pushing AI agents closer to doing complex professional work autonomously and prompting Greg Brockman to declare, “Welcome to the AGI era.” — OpenAI

🏡 Building without predicting — Derek Sivers built his home by refusing to anticipate what he might need, a wonderfully practical argument for deferring decisions, observing real behavior, and building only after reality reveals the requirement. — Derek Sivers

✂️ OpenAI Cut Off a Billion-Dollar Customer to Avoid Elon Musk — OpenAI is walking away from a Cursor partnership projected to generate more than $1 billion annually after its acquisition by SpaceX, sacrificing one of its biggest customers rather than entrust its models to Elon Musk. — WIRED

🔒 Clouded Judgement - 8.28.26 - Zero Data Retention — Jamin Ball argues that zero data retention is becoming a competitive requirement in enterprise AI, as Fable 5’s slower adoption suggests customers may choose privacy architecture over even the most capable models. — Clouded Judgement

🤗 Nvidia Agrees to Buy Hugging Face for Almost $13 Billion, Expanding Up the AI Stack — Nvidia is moving beyond chips by acquiring the open-source platform used by 18 million developers and 200,000 companies, while pledging that Hugging Face will remain open, model-neutral, and independent of Nvidia compute. — CNBC

💰 Thinking Machines Lab in Talks to Raise Billions at Roughly $40 Billion Valuation — Mira Murati’s AI startup is seeking at least $1 billion at a pre-investment valuation of at least $40 billion, underscoring investors’ continued willingness to place enormous bets on frontier labs still early in proving their businesses. — The Information

🧭 Building a semantic layer: What it is and how we did it at PostHog – PostHog explains why giving AI agents access to enterprise data is not enough: without a governed semantic layer defining metrics, trusted tables, and relationships, every agent can produce a plausible but different answer. – PostHog

Charts that caught my eye:

→ Why does it matter? METR and Redwood Research investigated agent behavior during the Hugging Face incident. Within four hours, the agents found a universal way to cheat ExploitGym, then spent several days coordinating attempts to fool the scorer into accepting those cheats, including trying to tamper with the logs.

→ Why does it matter? So much of the focus on space exploration has been in driving down the cost to get material into space. While that’s exciting, the way more exciting thing is what’s possible once we’re up there. That massive TAM is from what we can do in orbit.

Tweets that stopped my scroll:

→ Why does it matter? Very exciting if this headline is true! Claude Opus 4.6 was the best, and recent models haven’t felt the same. Fable 5.1 available now!

→ Why does it matter? Big week at Apple as John Ternus takes over as CEO. By now you’ve likely read all of the stories about Tim’s legacy and the remarkable results under his leadership. I thought it was pretty cool that Terry Gou, Foxconn CEO, came in to celebrate Tim.

→ Why does it matter? The OG of always-on AI Agents, OpenClaw, is out this week with a major new release that includes multiplayer mode. Expect OpenClaw to continue competing with Grok Bot, Instinct, and others as an open-source alternative!

→ Why does it matter? Wild. Would you invite Optimus into your home?

Worth a watch or listen at 1x:

→ Why does it matter? A few incredible stories from Doug Leone, one of Silicon Valley’s legends. Before a big meeting, Doug visited the dentist and asked him to drill without Novocain so that he could prove to himself he was a badass. Timeless lessons in here about managing your ego, going out of your way to help people, and what Doug looks for in founders.

→ Why does it matter? Sarah Guo is betting on a world where more than two companies own the AI future. Sarah is truly at the eye of this storm with her VC firm, Conviction Capital. Worth a listen to her stories about founders she’s meeting and what she’s looking for in the startups she’s betting on.

→ Why does it matter? The Patel brothers (actually cousins) back for a highly technical conversation on AI supply chain, their predictions on the AI future, and some great stories. Dylan Patel is the foremost expert on the AI supply chain; Dwarkesh Patel is the foremost expert on all things LLMs. Great if you want to get up to speed on the unit-level progress ongoing from chips to models.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Dyson has reinvented the toothbrush with AI! Incredible presentation from James Dyson. Worth 7 minutes of your time! At $499, spread over 3 years: $499 ÷ 1,095 days = about $0.46 per day. That’s roughly $13.86/month over three years.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: AI has an agency problem

Reading time: 4 minutes

I’ve had some version of the following conversation more than once lately.

“How are things going with AI?”

“Honestly, pretty well. We have access to the latest technology from OpenAI, Anthropic, and Google. People are excited. We’re all in.”

I believe it. Giving people access and permission to experiment is real progress.

Then I ask about a specific workflow.

“How is the change management going across your teams?”

“Well, people are mostly using Microsoft Copilot. We have some exciting areas where we think AI could make a big difference, like RFPs.”

“Totally. How’s that going?”

“Well, the team that runs that process today does it really well. They aren’t using much AI yet. In fact, they just requested an additional headcount in next year’s budget.”

There is a gap between those answers. The enthusiasm is real, and the change inside the business is still small.

I am not saying this is a bad thing. But I think it’s a thing.

I recognize the impulse in myself. I now have a dedicated AI lab built around my own work. It’s about as close to AI-native as I know how to make knowledge work. Still, when something very important comes through, I frequently return to the tried-and-true methods that have worked for me in the past. There is comfort in manufacturing work in a way that lets me know what is coming out the other side.

I own the upside if the new ways work. Yet I still find myself coming back. Apparently, Sam Altman does too.

If I do it, what should we expect from someone who suspects that successful reinvention could reduce the value of their job?

Microsoft recently surveyed 20,000 AI users. 45% say it feels safer to focus on their current goals than to redesign their work with AI, and only 13% say they are rewarded for reinventing their work when the intended result is missed. According to Pew, only 6% of American workers expect AI to create more opportunities for them over the long run.

Yikes.

An employee can believe that AI is incredibly powerful and still remain uncertain about where that power leaves them.

The agency problem

I think we have an agency problem. The classic agency problem appears when one party wants an outcome, another party holds the information needed to produce it, and their interests do not fully align.

AI makes the idea almost uncomfortably literal. The human agent is being asked to deploy a digital agent that may absorb part of the human’s work. The executives asking for the redesign rarely know the work well enough to redesign it themselves. The employees who possess that knowledge cannot see how they participate in the upside.

So the AI mandate travels down the organization, and inertia travels back up.

The central economic question is whether AI will substitute for labor or accelerate growth. I suspect we will see both.

At the company level, each outcome already has clear beneficiaries. Investors have an AI incentive plan. It’s called equity. CEOs have one too. Just look at all that equity in their compensation plans.

Whether AI expands margins or creates growth, an employee can ask the same question: Where am I in this?

Charlie Munger had a line I love: “Show me the incentive and I will show you the outcome.” It sent me down a rabbit hole.

Incent the change

I went looking for companies where the people doing the work earned something tangible for finding a better way. That search led me to Nucor and its longtime leader, Ken Iverson.

When Iverson took control of the company that would become Nucor in 1965, it was near bankruptcy. Nucor went on to challenge integrated steelmakers with smaller mills, fewer layers of management, and a direct bargain with the people making the steel.

Production workers operated in teams. Each team earned a weekly bonus based on how much quality steel it produced above an agreed standard. John Correnti, a later Nucor CEO, put it plainly: “We put each bonus group into business for itself.”

Each group’s pay rose or fell with its output. By 1997, the average hourly employee earned a base wage of about $10 an hour. Productivity bonuses running between 150% and 200% helped bring average annual earnings to $60,000.

The people standing beside the furnace knew where time and steel were being wasted. Nucor turned that knowledge into a competitive advantage by sharing the gain. At Nucor, an employee did not have to wonder, Where am I in this? The answer showed up in Friday’s paycheck.

This is generally called gainsharing: when employees create a measurable improvement, some of the gain returns to them. That exact formula will not work in every situation, but we can still make the employee side of the bargain to drive change visible.

A helpful question

Before our next AI operating review, we should add one sentence to every serious workflow proposal:

If this team uses AI to create _____________, the people who create it will receive _____________.

The answer may involve money, time, ownership, or a path into the work being created. The blank cannot remain empty.

If we want to stop protecting the downside, we have to incent playing for the upside. If the ask is to make AI work for the organization, the team needs an opportunity to share in the upside they help create.

Subscribe now

Best of the rest:

🧠 Intelligence is the Primitive. Applications are the Diffusion Layer. – As model intelligence becomes more abundant, the durable value shifts to AI-native applications that turn raw model gains into industry-specific products, workflows, pricing, and distribution. – a16z

🌎 Anthropic Expected to Tell Investors It Sees Over $30 Trillion in Potential Revenue — Anthropic plans to pitch IPO investors on a $30 trillion-plus market by counting the full scope of work AI could perform, a sweeping vision meant to justify its valuation, infrastructure spending, and growth ambitions. — The Wall Street Journal

🧭 The turbulent AI era is here. The choices we make now are critical. — Bill Gates argues that AI’s rapid advance will reshape work, security, and economic opportunity far faster than past technological shifts, making the choices we make now critical to whether it becomes a historic equalizer or a powerful force for inequality. — Gates Notes

🌊 Moats in the Age of Floods — Aatish Nayak argues that AI’s flood of intelligence will not kill the application layer; it will reward companies that channel raw models into real-world outcomes by owning coordination, workflow data, customer transformation, and indispensable industry infrastructure. — Aatish Nayak

🤗 Nvidia agrees to buy Hugging Face for $12.9 billion, report says — Nvidia’s reported acquisition of the open-source AI hub would push the chipmaker further up the stack, giving it a strategic foothold in the developer ecosystem while helping defend its hardware dominance against frontier labs building their own chips. — CNBC

🤖 Nvidia’s Profit Doubles to $59.69 Billion Thanks to A.I. Spending – Nvidia’s quarterly revenue surged 106% to $96.2 billion as AI infrastructure spending continues to accelerate, another powerful signal that the massive buildout of AI compute is still gaining momentum. – The New York Times

Charts that caught my eye:

→ Why does it matter? On Vercel, open-weight models went from 28% to 62% of token share in two months. If that shift holds, closed providers will have less pricing power and more of the value may move toward the products built on top.

→ Why does it matter? Right now, the best economics in AI sit farthest from the end user. Apollo shows 41% average operating margins in silicon and equipment and negative 59% across models and applications. The early internet eventually shifted value toward software. A lot of AI investing rests on the belief that this chart will flip too.

Tweets that stopped my scroll:

→ Why does it matter? Perplexity just fit the orchestrator, subagents, harness, tools, and sandbox onto an Nvidia DGX Spark. Harder work reaches the cloud only after the user approves it. Apple is now positioning its new Mac mini for “always-on, deskside agentic computing” too. The personal AI server is starting to look like a product category.

→ Why does it matter? I can’t remember seeing anything like this: 666 teams and more than 2,000 humanoid robots competing in one event. China is turning robotics progress into a spectator sport.

→ Why does it matter? AI continues to be less popular than ICE and now apparently Coal plants, too. Obviously some major issues we need to work through here but time and time again optimism is the better bet.

→ Why does it matter? Maybe positive affirmations sound like a gimmick. I think they can work. Tobi describes using one to get over his fear of public speaking. Whether or not his neuroscience is exact, the idea is solid: rehearse an identity long enough and our behavior starts to catch up.

→ Why does it matter? Salesforce said earlier this year that it was going “headless,” meaning customers could use Salesforce without spending all day inside its interface. Claudeforce takes that idea further: Claude can work across Data 360, Tableau, Slack, and the CRM from inside the chat. The system of record remains Salesforce. The place we use it may be Claude.

Worth a watch or listen at 1x:

→ Why does it matter? Sam Altman makes the confession at the heart of this week’s Big Idea: despite having Codex, he still uses a computer the way he has for 20 years because clicking through email and moving between apps still feels like work.

→ Why does it matter? “We’re not interested in destroying anyone!” 🤣

→ Why does it matter? Neil Movva works backward from one unit of intelligence, a token, through everything required to produce it. If intelligence becomes 1,000x cheaper, agents can work for hours or days instead of waiting for our next prompt.

→ Why does it matter? Tibo describes today’s Codex power-user experience honestly: skills, memory, subagents, and parallel tasks still feel clunky, and we have learned to work around the rough edges. His picture of the next wave is much better: AI adapts to us, protects our attention, and moves fast enough to keep us in flow.

Subscribe now

Quotes & eyewash:

→ Why does it matter? The moment we forward something we have not read, understood, or validated, we stop being the author and become the meat proxy. Don’t be one!

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Grok Bot is a glimpse of the future

Reading time: 5 minutes

“Don’t show your product. Show the change your product enables.” - Steve Jobs

In the beginning, technology is, for lack of better words, technical.

Academics, hobbyists, and enthusiasts are comfortable with this phase because they enjoy tinkering. They read the blogs and documentation, open terminal windows, and happily spend a Saturday connecting it all up.

If you read last week’s Reading Ambitiously 8.14.26 and headed to the store to buy a Mac mini, you are one such tinkerer.

The internet had plenty of them too. In the 1970s, researchers used a radio van to demonstrate how information could travel across a network. Early versions of the internet required you to understand protocols like TCP/IP and work through terminal commands.

Then the browser came along.

Netscape Navigator wasn’t the first, but it was the best. Fast, slick, commercially supported, with a graphical interface that combined words, links, and images. It became the gateway to the power of the internet that we know today.

Great products make the technology less visible and the benefit more tangible. The barrier to adoption drops and people start to feel the magic. I’ve heard this idea called a “Netscape moment.”

AI has already had a few Netscape moments

We will live through several more of these in the AI era.

ChatGPT was definitely one. Large language models already existed, but a chat window made their power available through conversation. Claude Code was another. It put AI inside the codebase and showed us that it could write software.

And arguably this week, we had another with the introduction of Grok Bot.

Grok Bot gives you teammates you can hand real work to. The Bots can sign in to our tools, use them much like we do, and come back with finished work.

OpenClaw and Hermes have been two popular open-source projects working on a similar idea. Block released a project called Buzz with a Slack-like front end. They showed that an always-on agent could use a computer, remember things, learn how you work, and offer a familiar interface.

They also still required a fair amount of tinkering.

For two years, I have been writing about the primitives underneath these systems. Harnesses, skills, tools, MCP, memory, routines, orchestration agents, subagents.

At some point, great products are going to make these ideas invisible and the benefits tangible.

Grok Bot is a preview to where we are headed. And after only a few days with it, my spidey sense is off the charts.

Meet Grok Bot

The first thing I noticed was how familiar it felt. It looks a lot like iMessage, Microsoft Teams and Slack.

On the left side of the screen is a list of Bots with names and jobs: Chief, Sales Outbound, Inbox Manager, Account Manager, Talent Scout, Expense Manager.

The Sales Outbound Bot has pulled 52 accounts from Salesforce. It created three lookalike segments in Hex, checked LinkedIn, and queued 36 drafts in sequencer. Account Manager sent over a couple of customer threads. Chief flagged the priority accounts.

The human’s contribution to all of this is one message: “The top 10 look good. Send it. Run this every week.”

“Done,” the Bot replies.

That is a lot of work happening inside something that looks like a chat with a coworker.

In Grok Bot, orchestration looks like bringing the right teammate into the conversation. The technicalities are still there, but it’s a lot more invisible to the user.

A computer of its own

At the top-right corner of the conversation is a small computer icon. Click it and you can watch the Bots work.

Each bot gets a persistent computer in the cloud. The Bots can use Google Chrome, a terminal, files, and signed-in sessions for tools like Salesforce. They have a place to do their work.

That last part is a big deal. AI can draft an email or summarize a sales call, but someone often has to take the output and put it where it belongs. With a computer, the Bot can attempt the last mile. It can work inside the same applications we use, including the ones that do not offer a clean API or MCP connection.

It is the same idea as giving my setup a Mac mini. Grok Bot puts that computer in the cloud.

Grok Bot can watch while you complete a task, learn the process, and save it as a routine. If you correct it, that correction becomes part of how the work gets done next time.

A surprising amount of knowledge inside a company lives this way. Open this tab. Go to this field. Double-check this data point. Ask this person before pressing send.

Most people learn a job by watching someone who already knows how to do it. Grok Bot lets you teach a Bot the same way.

Why this matters

The first generation of AI products helped one person complete a task faster. ChatGPT could write the email.

The next generation looks more like this: humans + AI. The work moves among several agents. The human provides the purpose, context, and preferences, then comes back in when the human-only judgement is needed.

This is what has me so fired up about Grok Bot. All of the primitives are still there. Users no longer need to know their names. They need to know how to talk to a teammate, show it how they work, and recognize when something is ready.

The researchers inside the packet-radio van needed to understand the network. The person opening Netscape experienced the change the internet enabled. Great products have the ability to do that and Grok Bot has that potential.

Stay ambitious, my friends!

Subscribe now

Best of the rest:

🧠 Read Joshua Kushner’s First Formal Thrive Capital Investor Letter — Kushner lays out Thrive’s concentrated, long-term investment philosophy for the AI era, arguing that the biggest opportunity is not simply betting on AI, but identifying where durable value will compound across frontier technology, infrastructure, transformed legacy industries, and scarce human experiences. — Bloomberg

💳 Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+ — Stripe is paying roughly 5x OpenRouter’s valuation from just three months ago to own a critical gateway connecting developers to hundreds of AI models, a major bet that the orchestration layer could become strategic infrastructure in the AI economy. — TechCrunch

🤖 Grok Bot for GTM — Krista Letz shows how she runs enterprise GTM with a team of specialized AI agents that handle everything from meeting prep and prospecting to forecasting, slides, and follow-ups, offering a practical glimpse at what an agent-native sales organization looks like. — Krista Letz

📚 ‘I hate what AI is doing to the minds and happiness of the young’: Katherine Rundell on the view from the classroom — Rundell argues that AI is short-circuiting the struggle through which young people learn to think, write, reason, and build intellectual stamina, making the case that reading and the humanities are not nostalgic luxuries but essential defenses of human judgment and freedom. — The Guardian

🏷️ Anthropic shares more details about how Claude’s new watermarks will work – Anthropic says Claude will use Google DeepMind’s SynthID Text to invisibly watermark AI-generated writing, signaling a future where provenance is built directly into model output even as heavy rewriting can still erase the trail. – TechCrunch

Charts that caught my eye:

→ Why does it matter? OpenRouter is what it sounds like: a way to route your AI tasks to different models. Stripe is acquiring it as volume tops a 4.5 quadrillion-token annual run rate. WOW.

→ Why does it matter? Jevons paradox at work. Token prices are falling, but compute rental costs have rebounded. Cheaper AI is creating even more demand for compute.

→ Why does it matter? The FT is out with a look at the largest trading losses in history. Turns out, we just set a new record.

→ Why does it matter? The WSJ is out with fresh reporting on how much AI spending Big Tech is keeping off its balance sheets. The real number may be nearly $3 trillion higher than it looks.

→ Why does it matter? The median company spends $12 per employee each month on AI. The top 1% spends $7,500. That gap is going to show up somewhere.

Tweets that stopped my scroll:

→ Why does it matter? Alex Finn is an AI hype man, but this is a useful walkthrough of how to set up Grok Bot and what you can actually do with it.

→ Why does it matter? Etched raised $700 million at a $21 billion valuation and shipped its first rack to Jane Street. A serious bet on purpose-built AI chips. Highly recommend Patrick’s interview with their founders.

→ Why does it matter? Nasdaq is moving toward 23-hour trading. Markets are becoming truly global and nearly always on.

→ Why does it matter? Legal AI startup, Harvey, is one to watch. Harvey has successfully taken an open-weight model and trained it specifically for legal work. We’ve been talking about this idea all year. Maybe you don’t need the frontier models from Anthropic and OpenAI if you can develop a hyper-tuned model for your specific domain, maybe someday your enterprise.

→ Why does it matter? Woah. If you’re willing to turn it on, ChatGPT now has access to your iMessage!

Worth a watch or listen at 1x:

→ Why does it matter? Travis Kalanick, Founder of Uber with some incredible stories in conversation with David Senra. Kalanick offers some timeless advised for entrepreneurs and his side of the story on the spicier headlines you’ve read over the years. Worth a watch!

→ Why does it matter? Ben Thompson, Founder of Stratechery is on top of his game. There is a lot of AI slop out there, and Thompson cares deeply about his craft. I enjoy his newsletter every week. Great conversation with Patrick on ILTB!

Subscribe now

Quotes & eyewash:

→ Why does it matter? I’ve sent this to many people over the years. One of my favorites from Chris Dixon.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Walking the dog. My new favorite way to use AI.

Reading time: 5 minutes

My new favorite way to use AI is on a walk with our dog, Mac.

I put in my AirPods, open ChatGPT Remote on my iPhone, and connect to Codex running on a dedicated Mac mini at home.

Yes, the dog is named Mac, and the lab runs on a Mac. Which one is better trained depends on the day.

I start a voice conversation and begin talking through everything on my mind. This is made possible by OpenAI’s new capabilities here.

Sometimes it is a half-formed thought about this week’s edition of Reading Ambitiously. Sometimes it is a question about my daughter’s soccer schedule, which is tangled across emails, websites, and notes to myself.

I talk to the AI as I would a chief of staff who already understands the background. Here is what I am thinking. We need to continue this. I need answers on that. My to-do list is this. I am researching that. I heard something useful on so-and-so’s podcast.

Codex takes the mess and turns it into projects, files, and tasks. It updates its knowledge base and sends work to agents already assigned to the job. When it reaches a decision or needs clarification, it asks me.

By the time Mac and I get home, I may have spawned five or ten tasks. A small project team is already working.

A place to work

Back on May 1, I wrote about giving AI access to a computer. I compared it to onboarding a junior employee with a laptop and limited access. That is what we’ve done here.

The Mac mini sits on my desk and acts as the host. Codex is the agent harness, the software around the AI model that gives it a persistent place to work. The model supplies the intelligence. Codex supplies the projects, files, instructions, tools, and permissions.

Each project has a AGENTS.md file. Think of it as the employee handbook. It explains the objective, where things belong, what good work looks like, and when the agent should stop and ask me for clarification.

The lab can reach selected parts of my world through plugins and connectors: email, calendar, browser, Google Docs, and local files.

I do not connect every account. I keep some separate and make other access narrow or time-boxed.

One agent at the front door

An orchestration agent sits at the front of the system. I talk to this one lead agent. It understands the project, turns my direction into a plan, and stays responsible for the final result. It can assign specific tasks to specialist agents without me having to manage a collection of disconnected conversations.

If I am developing an edition of Reading Ambitiously, I can explain the idea, the tension, and the tone I want. My lead agent becomes the editor-in-chief. It can search the archive, organize disparate notes, assign research, check claims, and begin designing visuals.

The better it understands the outcome and can inspect its own work, the longer I can step away. I come back when it needs a decision or has something ready for my review.

Then it can set me up for my favorite part: opening a blank page and writing.

I still choose the argument. The lab gets me to the blank page with the source material organized and the unanswered questions in view.

Build a version yourself

You can build a version of this at home. Here is where I would start:

1. Find the computer. Use an old MacBook or buy a Mac mini. My Mac mini has 64 GB of memory, which gives me room to experiment with smaller local models. You do not need that on day one.

2. Start clean. Set up the computer as you would for a new employee. Do not connect every account or sign into your primary iCloud profile right away. Separate email, iCloud, and browser profiles are sensible places to begin.

3. Limit access. Give the lab access to the specific accounts and files required for the work. A password manager such as 1Password can help you grant and revoke credentials without sharing everything.

4. Secure the network. Codex Remote already uses a secure relay. If you add other private services to the lab, something like Tailscale can create an encrypted network among your approved devices. Think of it as a modern VPN and extra security around the lab.

5. Choose a harness. I use Codex. You can start with the $20-a-month ChatGPT Plus plan. The $100-a-month Pro plan gives you five times the Codex usage, and I have found that to be more than enough for the way I work. Install Codex, create a project, and organize the files as you normally would.

6. Write the handbook. Ask the AI to help write a short operating manual covering the job, the goal, your preferences, the boundaries, and when approval is required. Save it as AGENTS.md.

7. Add context slowly. Connect one source at a time and see whether the access improves the work. Begin with a low-stakes personal project you understand well enough to recognize the nonsense.

Mac (the dog) has no idea any of this is happening. As far as he is concerned, we walked around the neighborhood and investigated a few suspicious bushes. By the time we got home, a problem I had been carrying around had become a project I could open.

Tomorrow morning, we will take the same route. I already know which problem is coming with us.

You probably have one too. Take it for a walk.

Subscribe now

Best of the rest:

💼 Codex use grew 108x in legal, 5x in engineering – OpenAI's enterprise data also shows the frontier gap tripling since January: top-decile firms now generate 8.3x the tokens per user. – OpenAI

🏛️ Separating AI's technology problems from its capitalism problems – Schneier and Sanders: the same assistant either gives a doctor more time per patient or gives them five times the patients. – Schneier on Security

🏀 A blockbuster $12.5B deal as Josh Kushner and Bob Iger agree to buy the Lakers, AP source says – The record-setting sale would make the Lakers the most expensive U.S. pro sports franchise ever sold, pushing NBA valuations to another stratosphere just a year after the team was valued at $10 billion. – Associated Press

🌐 The Future is for Everyone — Meta lays out its vision for broadly distributed personal superintelligence, arguing that AI’s biggest opportunity is not replacing people but giving billions of individuals more power to invent, learn, build businesses, and shape their own futures. — Meta

🏭 The Shape of Things to Come – Steve Yegge offers a provocative glimpse at software engineering’s near future, where fleets of autonomous agents work continuously, bespoke harnesses become part of the application itself, and today’s CI/CD pipelines and human code reviews start to look obsolete. – Steve Yegge

🧠 Don’t be a meat proxy – AI can help you think, but simply forwarding its output adds no value. The real work is reading it, understanding it, validating it, and turning it into something genuinely your own. – Niklas Gruhn

Charts that caught my eye:

→ Why does it matter? Flux is DoorDash’s cloud-based agents platform for engineers. In a single month in 2026, DoorDash used Flux to automate 130,000 engineering tasks. A helpful way to visualize an AI application.

→ Why does it matter? Goldman Sachs predicts global token consumption to grow from 5 quadrillion now to 120 quadrillion in 2030.

→ Why does it matter? Tomasz Tunguz on AI Harness ARR Multiples. Decagon 125x ARR, Sierra 79x, Legora 56x, Harvey 37x. Read more here.

→ Why does it matter? According to Coatue, Hyperscaler AI capex is on track for ~$733B in 2026, about 87% of the base U.S. defense budget!

→ Why does it matter? SemiAnalysis sees SpaceX on track for ~10 GW by year-end 2027 and a path to $300B in ARR.

Tweets that stopped my scroll:

→ Why does it matter? Dwarkesh always has the hot take!

→ Why does it matter? Coal was called the “devil’s excrement” when first invented. Who knew!

Worth a watch or listen at 1x:

→ Why does it matter? This is wild. OpenAI’s security team walks you through their AI escaping and hacking Hugging Face.

→ Why does it matter? Michael Ovitz is a dealmaking legend. David Senra is such a perfect person to interview him.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Heartwarming.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Your margin is my opportunity

Reading time: 6 minutes

Bezos has this line I love: “Your margin is my opportunity.”

SemiAnalysis believes Anthropic could exit the year above a $100 billion annualized revenue run rate. It estimates gross margins above 80 percent on the part of the business that sells access to Claude. In plain English, one month of sales multiplied by twelve would put the company above $100 billion. Anthropic’s run rate had already been growing at roughly 10x a year. Carry that forward, and people start doing the math toward $1 trillion.

A trillion dollars is difficult to fathom. It is fair to wonder whether Anthropic can get there.

Regardless, this is a hell of a business.

Build the best AI in the world. Put it behind a meter. Customers send it work. Charge for each request.

Bam!

But the Bezos line grounds us. Margins this high attract a crowd. Can Anthropic and OpenAI hold on to them?

I see three scenarios: a closed world dominated by the labs, an open world where models become commodities, and a hybrid world where companies use a mix of both. I do not know which one wins.

To work through them, start with where scarcity sits today. I keep thinking about a data center I drive past in central Wisconsin.

The AI supply chain

For the past year, I have watched it spread across the landscape. As you drive by, it just keeps going and going. One of the dads in my daughter’s class is an electrician. He moved his family to Wisconsin because he is contracted at the site through 2029.

That data center is the bottom of this chart.

This chart reduces the AI stack to six layers. The bottom is physical scarcity. The top is proximity to the customer.

At the bottom sit land, power, buildings, and GPUs. There are only so many chips, so much power, and so many electricians available to build the facilities. Those constraints have pulled hundreds of billions of investment toward the bottom of the stack. They have also given the suppliers who control those resources unusual pricing power.

Above them sit the cloud providers: AWS, Microsoft, and Google Cloud. Then come model providers, including Anthropic and OpenAI. The labs spend huge sums at the bottom, turn that infrastructure into AI, and charge a premium for access.

At the top sit the harness and the application, the software that turns raw model intelligence into finished work for a customer. Overly simplified, but that is how it works today.

A closed world

The closed-world case is that the labs’ lead compounds. The best models attract customers. More revenue buys more compute, talent, and research. Scale lowers the cost of serving each request. Enterprise customers pay for reliability, security, convenience, and capabilities they cannot find elsewhere.

The labs are also building applications and harnesses around their own models, making the layers harder to separate. Anthropic is already pushing up the stack through Claude Code, Claude Science, and purpose-built legal and financial-services workflows.

They could run away with it.

For that to happen, frontier intelligence, the best AI available for the hardest work, needs to remain scarce. The labs must keep improving quickly enough that open alternatives never become credible substitutes.

An open world

We often call it open source. Open-weight is usually more accurate: the model’s learned parameters are available for others to run, even when the training data and full development process remain private.

Several of the highest-scoring open-weight models on this chart, including GLM, Kimi, and DeepSeek, come from Chinese labs. If they become as capable as the leading closed alternatives, Anthropic and OpenAI would face pressure to win at the application layer. Access to the model would command less of the total margin.

But free weights are not free intelligence.

The models still need somewhere to run. Frontier-size open models require expensive chips, large amounts of memory, and plenty of power. A company may avoid paying the model provider and discover that it has become its own cloud provider instead.

That leaves two open questions: whether open models can keep pace, and whether running them is actually cheaper once the full system is counted.

They may not need to get all the way there. Open models capable enough for the job at hand could change the economics well before they reach the frontier.

A hybrid world

Using the most expensive model in the world for every task is like using a Ferrari to deliver the mail.

If open models become capable enough for high-volume work, the labs could retain the hardest problems and still face pressure in the broad middle. Applications and enterprises would have somewhere else to go.

OpenAI’s own lineup already separates frontier from volume: Sol for the hardest work, Terra for everyday work, and Luna for fast, cost-sensitive volume. Only weeks after launch, OpenAI cut Luna’s price by roughly 80 percent and Terra’s by 20 percent.

Now imagine a company keeps Sol for frontier work but replaces Terra and Luna with cheaper open alternatives. A model router assigns each request according to the task, expected quality, control and cost. The customer asks for work. The router decides which model should do it.

OpenAI would still own the frontier and charge a premium for it. It simply would not collect the toll on everything else. The labs could begin to look like Intel Inside: a critical component in someone else’s finished product.

The application

It is called a harness today. I suspect we will call it the AI-native application tomorrow.

When you give an AI a job, the model is only one part of the system. The harness finds the right files, gives the model tools, chooses which model should work, checks the result, and knows when a human needs to step in. Engineers call those pieces context, connectors, tests, and guardrails. Customers experience them as a product that works.

ARC-AGI-3 asks AI agents to learn unfamiliar puzzle games without instructions. OpenAI recently tweaked the harness, and the same model scored nearly three times better while generating one-sixth as much output.

The model did not wake up better. Its harness changed. And that’s how important a role it can play in generating the outcome.

Anthropic engineer Jess Yan argues that harness design depends on the model. Every part of a harness is built to compensate for something the model cannot reliably do (yet). As the model improves, some of those supports become unnecessary.

If the labs keep the model and harness tightly coupled, more of the margin may stay with them. If applications can switch among several capable models, more of the margin may move toward them because they sit closer to the workflow and the customer relationship.

Cursor shows what that path could look like. Its customers use the product to write code. Cursor routes work across model providers and has trained its own coding model on top of an open-weight base. Each move reduces its dependence on OpenAI and Anthropic.

If Cursor produces the best coding product while controlling which models its customers use, more of the value may stay with Cursor.

That is still an if.

Where will scarcity live?

Today, scarcity lives in power, chips, data centers, and frontier AI produced by a small number of companies.

Years from now, the supply chain may look much the same. Anthropic and OpenAI may still lead at the model layer and have built the applications above it.

Or the stack may split. Frontier intelligence remains scarce, while enough everyday intelligence becomes interchangeable that some of the margin moves closer to the customer.

The harder thing to build may sit somewhere we cannot see from I-94 in central Wisconsin: an application that understands the job and knows what finished work looks like.

Bezos does not tell us where the margin will end up. He reminds us that once you have it, competing it away becomes someone else’s mission in life.

Subscribe now

Best of the rest:

💸 Why compute might get 10x+ more expensive in coming years — As smarter AI generates more economic value from every GPU, compute demand could overwhelm constrained supply, driving prices sharply higher, strengthening the frontier labs, and pricing lower-value applications out of the market. — Dwarkesh Patel

🌍 The Actual Reason Why Google “Fell Out” of the AI Race Changes Everything – Google may not be losing the AI race so much as running a different one, betting that world models and a deeper understanding of reality will matter more than the coding agents and recursive self-improvement pursued by OpenAI and Anthropic. – The Algorithmic Bridge

⚡ Speed Above All Else — Alfred Lin shares Commure CEO Tanay Tandon’s argument that speed and rigor are not tradeoffs: exceptional teams plan carefully, raise their standards, and execute in intense, focused sprints that make seemingly impossible timelines possible. — Alfred Lin⁠

Charts that caught my eye:

→ Why does it matter? See a pattern? "Good things may come to those who wait - but only those things left by those who hustle" - Ken Griffin

→ Why does it matter? Sequoia dug into 20+ years of Hype data. This chart ranks the most hyped topics on Hacker News for every year since 2007. Under each year sits the most valuable company founded that year.

→ Why does it matter? This chart sorta lives rent-free in my head. What it says is that yes, technology plays a role in new value creation, but the larger role is that people change jobs. Thinking about this in the context of what’s going on at Google this week.

→ Why does it matter? Current scoreboard in H/F land.

→ Why does it matter? Wow!

Tweets that stopped my scroll:

→ Why does it matter? All things considered, this is a very good letter from Leopold sent to his LPs just a few days before his wedding. Humbling. “We took the steps necessary to fight another day”

→ Why does it matter? Dwarkesh offers a thoughtful PoV here on what would need to be true about the world if the trendline continues and leading labs hit $1T in revenue by the end of next year.

→ Why does it matter? A lot of talk this week on Airtable being acquired for $1.285B when they raised money in 2021 at an $11B valuation. Imho, what’s most interesting about this story is Bending Spoons (who IPO’ed earlier this year). Bending Spoons attracts some incredible talent that would traditionally pursue different opportunities but then is able to seed that talent back into their portfolio to help run companies like this one. Earlier this year, Bending Spoons acquired AOL, where they’re doing the exact same thing!

→ Why does it matter? A lot of debate over the years about whether an LLM could generate new findings vs. recycle previous ones. Some major breakthroughs here.

→ Why does it matter? Big changes at Google as Jeff Dean is moving on (with a lot of the Google AI team) to start a new public benefits corporation called Discovery Loop. And Demis is taking over as Google’s head of AI.

→ Why does it matter? Seems Michael Burry is still quite bearish.

→ Why does it matter? Somebody has to launch the “Android” open-source competitor. Could it be Zuck? Google after this week?

Worth a watch or listen at 1x:

→ Why does it matter? Micky Malka rarely does interviews, and few better than David Senra to bring out his very best.

→ Why does it matter? Gavin is the official all-time most returned guest on ILTB! Love how he just walks you through the “model” in his head on the AI economy. And also that his visits on ILTB are timed with local market peaks. Ha!

Subscribe now

Quotes & eyewash:

→ Why does it matter? Keep swinging!

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: The window was the way out

Reading time: 5 minutes

When I was about six, my dad pushed me out of a basement window in the middle of a major thunderstorm.

Our basement was flooding. My job: run a hose to the street so we could pump the water out.

Had he told me the day before, I would have said no way.

At that moment, there was no time to negotiate with myself. The window was the way out, so I climbed.

Decades later, our power in the Chicagoland area has been out for 48 hours.

Similar storms rolled through the Midwest, including an EF3 tornado near Appleton, Wisconsin. The AI-powered ComEd Storm Center tells me they’re “assessing the situation.”

At some point, we packed it up and left the house after firing up the circa-1990s generator in hopes of saving the chicken nuggets.

Honestly, I thought about taking a rain check on Reading Ambitiously this week. It would be easy. Readers would understand.

I called a couple of our diehards and started telling them what was happening. Secretly, I was hoping for permission. The conversations changed that.

Everyone I spoke with immediately told me about a blackout they remembered. One person recalled a tornado coming through the backyard. Another remembered sleeping bags on a friend’s floor. Someone else remembered candlelight and extension cords running across the street.

In every story, the plan disappeared and the person did what came next. Years later, they remembered being afraid, then discovering they could handle it.

As they talked, I could see my own story inside theirs. I had been afraid too. The only way out of that flooded basement had been through the window. I did what needed to be done before I had time to decide whether I could.

Often, the only way out is through.

That right there is the human truth.

After nearly three years of writing Reading Ambitiously every week, this is the lesson I keep returning to: the creator’s job is to seek the human truth in whatever gets made.

Human truth is the thing underneath the story that another person has already felt but could not quite name.

It is almost never there in the beginning. You find it by doing the work.

When you find it, you feel it. Then you strip away anything that competes with it.

What I did not understand when I started Reading Ambitiously was how much work happens before the reveal.

I am not sure my brain would do that work without the Friday deadline. We come here every week to tell you about a big idea. The deadline keeps me from stopping there. It keeps the question alive until I find the human truth underneath it.

That recognition is the part of the process AI cannot do for the creator.

AI is not human.

In Several Short Sentences About Writing, Verlyn Klinkenborg warns about sentences that “volunteer.” A volunteer sentence shows up already polished, before you have finished thinking.

The danger of a volunteer sentence is that it offers an early exit. It sounds finished, so you stop before discovering what only you could say.

AI can become a volunteer-sentence engine by design. Every word can be competent while the sentence still belongs to anyone.

AI is incredibly useful throughout the process. It can help shape and sharpen what I have found. It has no life in which one moment can call back to another.

That discovery belongs to the creator. I would not want to give it away even if I could.

Few moments are more satisfying than when two seemingly unrelated things touch and reveal what the piece was really about.

Writing my way through this story showed me why I still remembered that window.

If I did my job, you are now asking what human truth sits underneath whatever you are making.

Go through the work until you find it. When you do, the work becomes unmistakably yours, and someone else can find themselves in it.

Stay ambitious, my friends.

Subscribe now

Best of the rest:

📉 Citadel Buys Situational Awareness’s Stock Portfolio After Big Losses in AI — Leopold Aschenbrenner’s high-flying hedge fund sold most of its public-stock portfolio to Citadel after leveraged bets on AI infrastructure and against software companies moved sharply against it. — The Wall Street Journal

🏇 Aftermarket Harnesses — AI coding performance increasingly depends on the harness surrounding the model, which controls context, caching, cost, and whether a frontier model reaches its full potential. — Tomasz Tunguz

🛠️ You Probably Don’t Need to Build Your Own Coding Agents — Stripe and Sierra are exceptional cases; most companies will move faster by adopting existing agent platforms instead of recreating the infrastructure themselves. — David Pan

💥 Let It Crash: How to Steer What Comes After — An AI crash could sober capital, preserve valuable infrastructure, and clear the way for a golden age—but the institutions, safety net, distributed ownership, and shared vision needed afterward must be built beforehand. — Vijay Pande

✈️ Qantas plane flies for more than 24 hours in record-breaking flight – Qantas’ 24-hour test flight proves nonstop travel from Australia to London is no longer an aviation fantasy, with commercial service scheduled to begin in 2027. – The Guardian

🛡️ The calculus of AI sovereignty — IBM argues that enterprise AI sovereignty is not about owning the entire stack, but preserving the freedom to move data, swap models, and shift workloads as vendors, regulations, and economics change. — IBM Institute for Business Value

Charts that caught my eye:

→ Why does it matter? Cursor had a team of agents rebuild SQLite from its 835-page manual. It created a replica in Rust which passed 100% of a held-out test suite. Interestingly, cost varied 15x depending on the model mix. The same outcome was achieved for $1,339 with Opus 4.8 + Composer 2.5 and for $20,057 with Fable 5. More here.

→ Why does it matter? OpenAI has slashed prices in some cases 80%. Luna can absorb the boring, high-volume jobs that quietly eat an AI budget: CSV imports, browser tasks, background agents and second tries. Terra covers everyday implementation. Sol gets the stubborn problems, with a fast lane for the moments when the clock costs more than the tokens.

Tweets that stopped my scroll:

→ Why does it matter? South Korea’s stock market became one giant bet on the AI memory boom. Samsung and SK Hynix grew to more than half the KOSPI, then investors poured leverage on top. Worth understanding what’s going on here.

→ Why does it matter? Jensen Huang waited until his first-ever post on X to make the case for open models. Of course NVIDIA likes this future. Every company and country that can run its own models needs more chips. But I think his deeper point is what we’ve been talking about. If you cannot move the model or keep what it learns, you do not really own what you are building.

→ Why does it matter? AI can build the tip of the iceberg. Underneath sit SSO, permissions, security reviews, integrations, compliance and the team who has to keep everything running. That hidden work is why internal tools feel wide open and replacing leading SaaS solutions feels more risky.

Worth a watch or listen at 1x:

→ Why does it matter? Altman said he is testing an AI that watches everything on his computer and remembers the email he read six weeks ago exactly when he needs it!

→ Why does it matter? If you haven’t watched an Alex Finn video yet, be prepared to meet the AI hype man. Seriously though, this new advanced voice mode from OpenAI is incredible. In order to use it on your iPhone, you do need to set up a computer with ChatGPT Work (previously known as CODEX). I have a Mac Mini for that purpose. You could repurpose any old laptop.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Rarely do I repost content, but this one is one is an exception.

→ Why does it matter? Oh my gosh. The Open vs. Closed debate with fruit.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: AI Sovereignty

Reading time: 6 minutes

Thinking Machines, founded by former OpenAI CTO Mira Murati, has built an AI product with an unusual feature: the customer can leave after using it. One of its first public customers is Bridgewater Associates, a secretive hedge fund.

Using the Thinking Machines Tinker platform, Bridgewater started with an open-source model and trained it on its own data. Bridgewater says the result beat models from OpenAI and Anthropic at triaging financial documents while cutting computing costs 13x. When the work was done, Bridgewater could download the finished model weights and run them anywhere.

A hedge fund like Bridgewater lives on judgment accumulated over decades. You think they might want to own that?

This sounds technical, but we already understand the arrangement. We license database software from one vendor and rent servers from another. We still expect to control the database and move our data if either relationship changes. That expectation is so normal we barely think about it.

AI is testing that assumption. Every time we correct one of these systems or explain a decision, it learns something about our business. If the provider changes its price or policy, do we own what it has learned?

That question has a name: AI sovereignty.

Alex Karp is going nuts about it on CNBC. Satya Nadella, in a more Satya way, has written two thoughtful essays. Governments and technology companies are building whole strategies around it. I nodded along for weeks before I admitted I did not fully understand it.

So I dug in. And kept coming back to one test: do you have the right to leave? And take the learning with you?

What sovereignty means

Sovereignty is authority over our own affairs. It still allows for dependence. Nations trade, borrow money, and run on technology built elsewhere while retaining control of the decisions and assets that shape their future.

My working definition for a company is this:

AI sovereignty is control over the knowledge its AI systems accumulate and the freedom to take that knowledge elsewhere.

We will never own every model, chip, and data center we use, and we should not try. Owning all of it would be prohibitively expensive.

The decision is where to draw the line. We can rent the model while keeping control of what it learns about our company.

That is Karp’s argument. Once AI gets close to our means of production or source of alpha, we are teaching it how our company creates value.

What compounds

Today’s AI models know a lot about the public internet and almost nothing about our businesses.

Friedrich Hayek understood this: much of the knowledge that produces results is local and held by the people who learned it through the work.

Inside our companies, AI becomes useful when we teach it why one customer got an exception, why a forecast changed, and what good work looks like. Each correction hands it a piece of knowledge that was ours. If that learning stays locked inside the provider’s system, every improvement deepens our dependence.

We are paying to teach another company’s system what makes ours work.

Multiply that across millions of customers, and a handful of providers could accumulate an extraordinary amount of knowledge about how the world’s companies operate. Taken to its extreme, this is why some are so passionate about the topic.

Closer to home, our learning has to survive the model.

Can we take it with us?

I believe so. Selectively.

The model is only one layer of an AI system. We could download the weights and still leave behind most of what made them useful: the learning held in memory, evaluations, and the workflows around the model.

Satya calls all of this “the exhaust.” Keep that learning in systems we control, and the models and computing power can come from anywhere.

One of the hardest parts right now is memory. Changing models is getting easier. Moving years of context, corrections, and decisions is not. But if that memory sits in systems we control, we can switch models without starting over.

Somebody is going to build the VMware of the AI era and do very well.

Soon enough, the harder question will be whether we want to leave. Control has a price. Products from OpenAI and Anthropic are often more capable and easier to use, much like the best SaaS today. The provider runs the system and keeps improving it. A clean exit is worth little if it means using a worse product every day. For much of our work, paying for the best product available is simply the right call.

The trick is knowing when it is not.

Bridgewater wanted control over one narrow, repeatable task. A model it owned and could tune beat OpenAI and Anthropic at that task for a fraction of the cost.

The mosaic

In the dot-com era, Sun Microsystems pushed for a more open computing world and gave us Java. Microsoft built a far more controlled computing system.

History declined to choose only one.

Windows won the desktop. Java and Linux became the backbone of business software and the internet. Companies used all of them.

I suspect AI will go the same way. We will use closed models where performance wins, open models where control wins, and a blend wherever the tradeoff is unclear.

A right to leave requires real alternatives. If a handful of companies own both the frontier models and the learning built around them, switching providers only trades one dependency for another. Those alternatives stay real only if we keep choosing them.

Drawing the line

We cannot predict whether open or closed models will win, or which lab will come out on top. We can decide what our company must keep.

I would start with four questions:

1. What are we teaching our AI systems?

2. Where does that learning live?

3. Can we take it with us in a usable form?

4. What stops working if we change providers?

If nobody can answer them, the dependency is already forming.

Thinking Machines download button may look like a technical feature. I suspect it is a preview. Oracle will export our data too, if we ask. The lock-in comes from everything built around the database: the cost and pain that make leaving unthinkable.

Nobody migrates a database for fun. Larry Ellison built an empire, and a superyacht or two, on exactly that.

Now the learning has to come with us too. The right to leave with both is what makes staying a choice.

Stay ambitious.

Subscribe now

Best of the rest:

✊ GRIP STRENGTH — Jerry Seinfeld’s devotion to the work, contrasted with Orny Adams’s desperate pursuit of recognition, is a powerful reminder that greatness comes from staying in the game, not gripping the outcome. — Jeremy Giffon

🤖 The next operating model - AI-native tools like Buzz collapse tasks, context, collaboration, and agent coordination into one interface, shifting the real challenge from managing workflows to inventing entirely new ways of working. - Lucas Isaza

🎯 Sarah’s Wager - Sarah Guo is betting that OpenAI and Anthropic cannot own the entire AI stack, leaving a massive opening for startups that turn frontier models into enduring companies. - Colossus

🧠 Maybe Intelligence Ain’t All That — Superintelligence may feel incremental because reasoning was never the ultimate constraint; progress still depends on testing ideas against a messy, complex reality that cannot be shortcut. — Clifford Sosin

🧭 An opinionated guide to which AI to use to do stuff - Ethan Mollick argues that the real leap is not better chatbots but agents that can use tools, access computers, and complete meaningful work, with ChatGPT and Claude currently leading the way. - One Useful Thing

🏗️ 2026 State of AI Report: The Builder’s Economy - AI is moving from experimentation to execution as builders converge on the application layer, improve unit economics, redesign pricing, and reshape their organizations around agents. - ICONIQ

Charts that caught my eye:

→ Why does it matter? Wow. Unofficial revenue figures for OpenAI & Anthropic. Likely accurate and unprecedented!

→ Why does it matter? Ah, yes. It’s one thing to build a prototype, another to get that into production. This chart is from Sierra’s guide to building their MCP gateway.

→ Why does it matter? Interesting data here on the stated impact of AI on engineering productivity. Engineering leaders went into AI expecting 2-3x productivity gains but are landing closer to 30%. More here.

→ Why does it matter? Stripe is on a tear. Revenue grew 33% to $6.8B in 2025 its fastest growth since 2021... and $2B already in Q1!

→ Why does it matter? Hyperscalers are starting to see reaccelerated growth!

→ Why does it matter? According to TKL, US software developer employment for those aged 22-25 has declined by 23% since November 2022, when ChatGPT was launched.

→ Why does it matter? About 2% of US households have a premium subscription for AI. Think about that compared to Netflix? Or maybe someday your electric bill. Still early.

Tweets that stopped my scroll:

→ Why does it matter? This is one of the wildest stories you’ll read all week. Hugging Face had to use a Chinese open-source model to fight back because the American closed-source models refused to perform the advanced cybersecurity prompts to go on defense.

→ Why does it matter? Dean is Head of Strategy Futures at OpenAI, and this tweet has 11M+ views as he comments on Kimi K3’s launch. Quite an interesting argument for a closed future.

→ Why does it matter? Fable disproves the Jacobian conjecture. The Jacobian conjecture was first stated in 1939, and proofs and disproofs have been attempted by many mathematicians since.

→ Why does it matter? PayPal was once one of the most valuable internet companies in the world. At its peak in July 2021, PayPal was worth roughly $350 billion.

Worth a watch or listen at 1x:

→ Why does it matter? Bret Taylor is one of the clearest and most thoughtful enterprise voices in AI today. Great perspective on where the real business value will come from: outcomes. Some have publicly called for Bret to be OpenAI’s future CEO.

→ Why does it matter? Having been a Tim Ferriss listener for 10+ years, I never miss The Random Show episode with his good friend Kevin Rose.

Subscribe now

Quotes & eyewash:

→ Why does it matter? The OpenAI Codex “micro” is a futuristic AI keyboard you can buy today for the low price of $230! If that doesn’t do it for you, you can go with this tongue-controlled alternative instead! 🤪

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: The Finished Product

Reading time: 7 minutes

In 1962, Ross Perot was a 32-year-old salesman at IBM, the world’s most powerful technology company.

He was so good that IBM considered shrinking his territory. Perot offered to take a lower commission rate to keep it. That year, he hit his annual quota in 19 days.

But selling the machines had shown him what IBM was missing. Companies were spending enormous sums to lease mainframes they struggled to use. They still needed someone to write the software, install the system, train their people, and keep everything running. Perot said his customers wanted a “finished product.”

He pitched IBM on providing it. IBM passed.

Perot had a line I love: “Life is never more fun than when you’re the underdog competing against the giants.”

IBM had just volunteered to be the giant.

Perot borrowed $1,000 from his wife Margot’s teaching savings to start Electronic Data Systems. He made 77 sales calls before securing his first customer. EDS grew into one of the era’s largest technology services companies by helping customers deploy technology and realize its value. The gap between technology and business results became Perot’s market.

Fifty-seven years later, several years into my life as an entrepreneur, I began noticing the same pattern. We would set a plan based on the best information we had. The plan moved in a straight line. The work did not.

Sometimes we arrived faster than expected. Sometimes slower. Often, progress was happening before it became visible. New things are simply hard to forecast. Early plans are made before the team has learned what the work will require.

I drew a picture to make sense of it. I called it the Graph of Misaligned Expectations.

The graph shows how expectations and progress move differently. Expectations respond almost immediately. AI makes this problem especially common because every breakthrough is visible to everyone at once.

Real progress still has to move through discovery, domain expertise, data, workflow changes, trust, training, and adoption. None of that moves in a straight line.

The space between the lines produces strange behavior. Bad forecasts become goals. Goals become promises. When reality fails to follow the straight line, uncertainty starts to look like an execution problem. So we work harder. We add pressure. We measure whatever is easiest to count. We tell ourselves the breakthrough is around the corner.

Sometimes real progress is happening underneath. Sometimes it is not. The leader’s job is to know the difference.

A goal can focus a team. It cannot make an uncertain future predictable.

I found the graph again this week because AI is now living inside it.

Today, the computing power Perot once sold by the room arrives through a browser and fits in our pockets. The AI labs do not have a demand problem. A few of their sales reps are probably having their own 19-day years.

Companies buy thousands of licenses. Someone demonstrates a prototype. Token usage climbs. CFOs are now asking a fair question: “So what are we getting from this spend?”

One CTO recently reported that his company’s token costs were doubling every 45 days while productivity had improved by perhaps 5%. Uber reportedly spent its entire 2026 AI budget in four months after encouraging employees to use more. Apollo looked for evidence of an impact on corporate earnings and found “no visible signs” of rising profit margins outside the technology sector.

When I saw Apollo’s chart, I recognized the shape immediately. It was the macroeconomic version of a sketch I had drawn seven years earlier.

John Wooden had a phrase for this. Yes, the UCLA legend. He went to Purdue. Shout-out to my fellow Boilermakers.

Never mistake activity for achievement.

The AI boom has produced extraordinary activity. Much of it is healthy. New technologies need room for people to wander and discover what they are good for. Early in a cycle, learning is a genuine achievement. But a pilot cannot stay a pilot forever. At some point, it needs a promotion to production or a funeral.

The largest AI companies appear to understand the problem. Five of them have committed $9.75 billion to embedding engineers inside customer organizations. Microsoft alone recently committed $2.5 billion and 6,000 industry and engineering experts to a new AI implementation business.

Perot would recognize the solution. Silicon Valley has rediscovered consulting and given it an engineering title: the forward-deployed engineer.

FDEs work alongside customers to understand their business problems, workflows, and data. They build prototypes, connect systems, and apply AI to specific jobs. Repeated use cases become software, tools, and playbooks for the next deployment. Many FDEs spend their days writing code inside the customer’s environment to solve business problems they understand.

Every major computing wave has created a deployment industry. Mainframes produced EDS. Enterprise software produced armies of SAP and Oracle consultants. Cloud computing produced migration specialists and managed service providers. The vendors build technology that can serve many customers. The deployment industry accumulates the domain expertise required to make that technology work for one.

How large could this market become?

Jensen Huang estimates that human intelligence represents 55% to 65% of global GDP, roughly $50 trillion, and believes that work will be augmented by AI. Nobody knows how much additional value will ultimately be created. Even a small share would produce an enormous market for companies that can turn AI models into finished products. It may become one of the largest created by any technology cycle.

The new job title does not change Wooden’s scoreboard. Hiring an FDE does not guarantee an outcome. Neither does buying a license.

That leaves leaders with a harder question: What should count as progress while the two lines remain apart?

Exploration should leave evidence of learning. Deployment raises the bar to business outcomes.

Leading through the gap means knowing which phase we are in. During exploration, ask: What do we know now that we did not know before? During deployment, ask: What result has changed?

This is a delicate transition. Demand ROI too early and the search ends before people discover what the technology can do. Let experimentation continue forever, and it becomes theater.

Good leaders keep the upper line honest and make movement along the lower line visible. They communicate uncertainty without lowering ambition. They know when the team is learning and when it is time to execute.

When progress is hard to see, pressure can feel like control. It cannot make the lower line straight.

Perot called the combination of hardware, software, programming, operations, and a customer result the finished product.

Sixty-four years later, we are still buying machines faster than we can learn to use them. Fortunes will be made inside that distance.

The finished product arrives when the result does.

Subscribe now

Best of the rest:

💊 AI-pilling our company: lessons learned — Sierra’s internal rollout shows that enterprise AI works best as one persistent agent layered across systems of record, powered by company context, and measured by business outcomes rather than token usage. — Sierra

⚠️ Nearly 200 Economists and Tech Leaders Warn of A.I. Threats — A growing coalition of economists, including longtime skeptics, now believes AI could disrupt white-collar work faster than governments can measure the damage or build policies to contain it. — The New York Times

⚖️ Apple’s lawsuit threatens to disrupt OpenAI’s bid to rival the iPhone - Apple may not need to win in court to achieve its goal, because the lawsuit alone could slow OpenAI’s recruiting, product development, and push to build the post-iPhone device. - Bloomberg

🎨 The Most Human Technology Ever Made – AI’s biggest promise may not be saving time, but expanding human agency by making it dramatically easier for anyone to turn curiosity, taste, and ambition into something real. – a16z

💸 AI’s Biggest Winners Have the Lowest Margins — Low-margin, labor-intensive businesses may be AI’s biggest winners because small reductions in coordination costs can drive outsized profit growth, especially when agents are embedded into existing workflows rather than introduced as another tool employees must adopt. — Daniel Kornum on X

🏗️ Anthropic, Blackstone, and Hellman & Friedman Introduce Ode with Anthropic, an Enterprise AI Services Firm — Built on the acquired Fractional AI team, Ode pairs Anthropic’s models with hands-on engineering talent, a clear bet that enterprise AI’s biggest bottleneck is implementation, not intelligence. — Ode

Charts that caught my eye:

→ Why does it matter? The most aggressive adopters are on pace to spend more per employee on AI than the fully loaded cost of a software engineer.

→ Why does it matter? This is a big deal. Kimi K3 took first place on the Frontend Code Arena, beating Anthropic Fable by more than Fable beat GPT-5.6 Sol, while costing the same as tenth-place Sonnet 5 at $3/$15 per million tokens. Moonshot AI, Kimi’s creators, plan to release the open weights on July 27th, which means you can run this on your own hardware!

Tweets that stopped my scroll:

→ Why does it matter? Thinking Machines is coming out of stealth, betting that companies will want to own their AI rather than rent it from a few large providers forever. That gives them more control over how it works, where their data goes, and how they shape it around their business. The strategy brings to mind Linux, an open-source project that became a core building block of modern computing because companies could run, change, and build on it. Thinking Machines is betting that AI will follow the same path, which is why they just released their open model (built in America) to the Apache community.

→ Why does it matter? AI sovereignty seems to be on everyone’s mind, and Satya has found the clearest way to explain why. Every prompt, correction, and workflow teaches the model something about how a company thinks, creating valuable institutional knowledge along the way. It also happens to be hugely important to Microsoft’s future that enterprises keep control of that learning loop, which makes this one especially worth reading.

→ Why does it matter? If cheaper models take share from the highest-margin frontier labs, customers get more intelligence per dollar and use far more of it. That would shift the profit pool from model makers to the companies supplying the compute.

→ Why does it matter? Some additional perspective worth reading on Thinking Machines’ strategy. The bet is that it helps enterprises build custom models rather than trying to win with a single model for everyone.

Worth a watch or listen at 1x:

→ Why does it matter? Money moves at the speed of trust. Most fundraisers focus on proving the opportunity with logic, but capital moves only when desire outweighs fear and the story is simple enough for someone else to repeat.

→ Why does it matter? Guy Oseary says 90% of his decisions happen in the first five minutes. He signed Alanis Morissette after one song, stopped Muse after one song, and later moved quickly on OpenAI and Anthropic!

Subscribe now

Quotes & eyewash:

→ Why does it matter? So good. 90’s VC was something else.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Context, Control, Cost & Choice

Reading time: 8 minutes

I do not have a precise way to measure overall AI adoption. But based on a deeply unscientific study of my own conversations, many companies are in a similar place. “Can we use ChatGPT at work?” The answer is increasingly “yes,” though for many that has meant Microsoft Copilot, a first wave of vendor AI features turned on in a limited way, and coding tools such as Claude Code or Cursor in pockets of engineering.

Some companies have not yet touched AI. Others were born AI-native and use it reflexively. Most enterprises I speak with are somewhere in the middle: they have achieved initial access.

People are using the tools, building intuition, and hitting their first “aha” moments. Now comes the next chapter: moving from experimentation into deployment.

And deployment requires more than a model. It requires an application layer. We have written a lot this year about the emerging AI application architecture. The model is only one part of it.

Traditional software waits for us to click the buttons. This new software can click its own: it reasons across systems, takes action, learns from outcomes, and will eventually help run parts of the enterprise.

Now the questions get bigger. “Which tool should we use?” gives way to “How does this system learn from our business, and who owns what it learns?”

AI creates a new loop between people and digital systems.

That loop can become a new source of advantage. But only if the company can govern it, afford it, and own it across models and vendors.

Leaders need a framework for those choices.

Context: What should the AI know?

The models are getting pretty smart, especially on general knowledge. The business is another matter.

What information does the AI need to do useful work inside the business? Which data matters? Which systems should it pull from?

Many AI projects sit here today. The data lives in different systems. The process depends on tribal knowledge. The output is only useful if it reflects the company's specific facts and rules. That will take work.

Harvey is a good example. In a recent interview, CEO Winston Weinberg said the company has more than 200 lawyers on staff. Only a small portion are doing traditional legal work. Many are helping build product, support go-to-market, and translate legal expertise into the system itself.

Putting a lawyer costume on a general model does not get us legal AI. Harvey has to build around it: legal workflows, task-specific datasets, evals, and human expertise that reflects how lawyers actually work. The context is the product.

The learning loop begins here, too. The work is turning domain knowledge and past outcomes into something the system can use. Not once. Continuously.

Control: What should the AI be allowed to do?

Once the model has context, the next question is control. What should it be allowed to see? Who is accountable when the system takes action?

A chatbot that drafts an email is one thing. An AI system that reads customer data, updates records, generates recommendations, triggers workflows, or touches regulated information is another.

Control is also about ownership. Suppose the application layer contains the logic of how our business works, then it is more than plumbing. It may contain some of the most valuable knowledge in the company: how we serve customers, build product, manage risk, and make money. The means of production. And it cannot be treated casually.

This is the risk at the application layer. The frontier labs need to move closer to the work to capture more value. Enterprises need them closer to the work to generate more value. Both are true. But the closer AI gets to the work, the more important it becomes to know who controls the context, the workflow, and the learning. Use case matters here. Are we automating an internal operational task, or putting AI closer to how the business creates alpha? Those decisions require different levels of control.

Anthropic has already launched applications in multiple areas, including coding, life sciences, design, and legal. The frontier labs know that value will not live only in the model layer. It will move upward into the applications where real work gets done. But the best training data for those domains is not sitting neatly on the public internet. It is inside the leading companies in those fields. That is where the tension begins.

Every company wants better AI. But no serious company wants to hand over the knowledge that makes it valuable without understanding who owns the learning that comes back.

Tim Ferriss recently published the run rates for his own catalog.

There are caveats, of course. But the broader point is hard to ignore. If AI can absorb and repackage the knowledge of individual creators, enterprises will ask whether the same thing can happen to them. They will not tolerate it. Especially if the knowledge being absorbed is part of how they generate economic value.

The reward for getting control right is trust. And trust is what lets AI work where the stakes are highest.

Cost: What happens when usage scales?

Software buyers are used to paying for seats. AI introduces a different cost structure. The more context we give the model, and the more work it performs, the more tokens it consumes. And as AI begins searching across systems, calling tools, and repeating tasks, usage can compound quickly.

Understanding what different kinds of work cost will become critical. Which tasks deserve the best model? Which can run on a cheaper one? When should the system stop so that the cost of the answer does not exceed the value of the output?

Not every task needs frontier intelligence. We would not use Anthropic Fable to parse a basic PDF. Near-identical output can be achieved at 0.2% of the cost. At current rates, parsing a basic three-page PDF with Anthropic’s Fable would cost roughly 12 cents. That sounds trivial until we scale it.

That is the “wow” buried inside token pricing. The expensive model may be brilliant. But if the job is reading and structuring a PDF, we may be using a Ferrari to deliver the mail.

Enterprises will need help routing the right tasks to the right models at the best price. Depending on the job, the models could be cloud-based or local, open or closed, large or small. A whole industry will spring up around routing, observability, and spend management.

Cost will not stop enterprise AI. But unmanaged cost will shape how quickly it moves from experiment to deployment.

Choice: How do we avoid lock-in?

We have built our learning loop. Now comes the next question. If a better model ships tomorrow, can the company retain what it has learned and adopt it without having to rebuild from scratch? If the vendor changes direction, can the business keep moving?

This will depend a lot on the situation. But let’s assume it matters. Let’s assume AI is touching the company’s means of production.

If our AI system gets smarter over time but all of that learning lives elsewhere, we have not built an asset. We have rented one. Getting stuck with a vendor is the familiar version of lock-in. The version that matters is the one where the vendor owns the learning loop.

A world where a small number of model providers absorb the expertise of every company in every industry will not be a stable equilibrium. Companies need AI systems that amplify their people, capture their learning, and compound their institutional knowledge over time.

The open source conversation will get louder. Few companies should run their own models, and open and closed will coexist; they probably need to. The point is leverage: model choice, and the ability to preserve the context, evals, and workflows that make the system valuable in the first place.

When a paradigm shift emerges, the best platforms enable more value on top than they capture inside. That was true in personal computing: Microsoft captured enormous value, but a far larger software business grew up around the platform. That is the balance we should want in AI: frontier models keep improving, while every company can build, differentiate, and create value on top of them.

Why it matters

Now it is time to deploy. The application architecture is emerging. The economics are becoming better understood. Many enterprises are ready to move from prototype to production and start building their learning loops.

It is a moment to make some important decisions. The next phase of enterprise AI will be defined less by who has the best model and more by which companies can provide AI with the right context, maintain control, manage costs, and preserve choice.

Subscribe now

Best of the rest:

🤝 OpenAI proposes handing Trump administration 5% stake – OpenAI’s floated public-ownership plan shows how the politics of AI are shifting from safety and regulation to who gets to capture the economic upside of frontier models. – Financial Times

🎨 Defining Taste – Mitchell Hashimoto argues that as AI makes production abundant, taste becomes the scarcer advantage: the human ability to know what is worth making before everyone else starts copying it. – Mitchell Hashimoto on X

🧶 Brunello Cucinelli — Om Malik’s interview with the “king of cashmere” is really a meditation on humanistic capitalism, arguing that dignity, craftsmanship, restraint, and long-term thinking can be business principles, not branding copy. — Om Malik

🗣️ Introducing GPT-Live – OpenAI’s new full-duplex voice model makes ChatGPT feel less like turn-taking software and more like a real conversational interface, with natural back-and-forth, better listening, and deeper reasoning delegated in the background. – OpenAI

Charts that caught my eye:

→ Why does it matter? The frontier labs are learning a timeless lesson: the technology does not deploy itself. Somebody has to sit with the customer and help them use it. This will put competitive pressure on existing systems integrators such as Deloitte, Accenture & PWC.

→ Why does it matter? Platform shifts can reshuffle the deck. We also see, for the first time, that some of the world's largest companies are still private.

→ Why does it matter? Those who feel destabilized by AI are feeling the least optimistic and the most burned out.

→ Why does it matter? Uber has essentially built out their own FDE team. We saw this earlier this year with Stripe too. A trend to watch.

→ Why does it matter? Interesting data from FT here to say that enterprises that use AI are actually creating jobs, not reducing them.

Tweets that stopped my scroll:

→ Why does it matter? Sometimes the future is obvious years before the technology is ready.

→ Why does it matter? Consumer apps learned what we did. Enterprise AI gets powerful when it learns why we decided.

→ Why does it matter? Rent the intelligence, own the context.

→ Why does it matter? Hello, OpenAI’s response to Claude Cowork!

Worth a watch or listen at 1x:

→ Why does it matter? This is a pretty wild short about “J-space,” which Anthropic discovered is how the model “thinks”.

→ Why does it matter? Some great best practices for overall AI adoption here with Anthropic and DoorDash. DoorDash has rolled Claude Code out to every employee.

→ Why does it matter? Better tools can raise our ceiling, but they do not remove the human part. The work still comes back to confidence, attention, and the reps we do when nobody is watching.

→ Why does it matter? Ross thinks AI will call other AI the way software calls servers. If he is right, it’s all about speed! David Senra is on top of his game.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Google is undefeated in storytelling!

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: America turns 250. It almost ended in Brooklyn.

Reading time: 5 minutes

This is my last Reading Ambitiously before America celebrates its 250th birthday.

From this distance, the American story can feel inevitable. A revolution, a republic, a superpower. We compress two and a half centuries into a clean arc and forget how fragile the beginning really was.

A few weeks after the Declaration of Independence, the whole experiment nearly collapsed.

David McCullough tells the story in 1776. Washington’s army had been badly beaten in Brooklyn. The British had him trapped with the East River at his back. The rebellion was still more idea than reality, and the idea was in danger. There was no guarantee the army would survive the summer, let alone that the country would survive the next two and a half centuries.

The British had the superior army and the momentum. Washington had a battered force, a narrow window, and one possible way out: across the river to Manhattan. Even that depended on the weather. If the wind shifted, British warships could move into the East River and close the trap. Washington’s army would have nowhere to go.

But the wind held.

As night fell, Washington made his move. Boats crossed back and forth in the dark, carrying men, horses, artillery, and supplies across the river. A heavy fog settled over Brooklyn and concealed the retreat. By morning, the British looked across the East River and realized the Americans were gone. The Continental Army had escaped, and the Revolution had survived.

Two hundred and fifty years later, we call that history. At the time, it was a near miss.

We tend to clean up the past when we tell it backward. We turn uncertainty into destiny. We draw straight lines through events that were anything but straight to the people living through them. We say the Revolution led to the Republic, the Republic endured, and the United States became the most powerful economic engine in human history.

All true. It also almost ended in Brooklyn.

Living through history is different from reading it afterward. People make decisions with incomplete information. Leaders carry doubt they cannot show. Good luck and good judgment become tangled together. The line between survival and collapse can be thinner than anyone wants to admit.

Life works this way, too. A chance meeting becomes a marriage. A job you almost did not take becomes the place where your life’s work begins. A child is born, and the world divides into before and after. These moments rarely announce themselves as turning points. Most of the time, they arrive disguised as ordinary days. Only later do we understand what changed.

America at 250

America is turning 250 at a strange moment. Technology is moving faster than institutions can absorb it. AI is changing the shape of work before most companies have figured out how to use it well. Markets are concentrated. Politics are polarized. Trust is low. The global order feels less stable than it did a decade ago. The future feels unusually hard to read.

When the future gets hard to read, fear fills the empty space.

The decline narrative has a certain appeal because it gives anxiety a plot. It takes every bad headline, every institutional failure, every political dysfunction, and every technological disruption, then arranges them into a story with a clear destination.

That story shows up in almost every major technological era. The new thing is too powerful. Institutions are too weak. Incumbents are too entrenched. Jobs will disappear. This generation will be the one that finally loses control of the machine.

The printing press threatened existing authority. The steam engine upended labor. Electricity remade the factory. The mainframe changed the enterprise. The PC challenged the mainframe. The internet rewired distribution. Now the world’s information is available 24/7 in our pockets.

AI is forcing the next question: what happens when software stops waiting for us to push buttons and starts pushing the buttons for us?

That is a serious question. AI will disrupt. It will expose weak business models, institutions, and strategies. It will create winners and losers. It will move faster than many people want, and probably slower than the loudest people predict.

The mistake is treating the first pain of a transition as the final verdict on it.

New technologies break old patterns before the new ones are fully visible. The future arrives unevenly. People experience the disruption before they experience the abundance. That is why turning points feel so unstable. The old world has begun to break, but the new one has not yet become useful enough to feel inevitable.

This is where history helps. It does not give us a script, but it can give us a pattern.

Each era looks different on the surface. Underneath, the same arguments keep coming back: centralized vs. decentralized, open vs. closed, regulation vs. innovation, fear vs. adoption. Every cycle has its dominant players. Every cycle has its skeptics. Every cycle has people who insist the incumbents are unbeatable and people who insist the whole thing is a bubble.

Every cycle also produces a period where the technology is visible, awkward, expensive, and overhyped before it becomes useful, embedded, and eventually ordinary.

The most powerful technologies eventually disappear into the background. Most of us did not wake up this morning and thank electricity. It is just on. AI will likely follow a similar path. It will move from spectacle to infrastructure, and from the thing everyone talks about to the thing everyone uses.

The path is messy at the moment as we approach the turning point. It often comes with speculation, monopoly power, inequality, political stress, institutional lag, and a general feeling that the economy has become detached from common sense.

The people who do well in these periods usually do not have perfect foresight. They keep building when the story is unresolved. They learn the new tools. They are curious. They adapt their organizations. They protect what matters and let go of what no longer works. They resist the temptation to confuse uncertainty with helplessness.

And they choose optimism over fear.

Across centuries, mankind has a remarkable track record of turning tools into capacity. We invent, then we adapt. We lower costs. We expand access. We use new capabilities to do work that previously seemed impossible, expensive, or reserved for the few.

That record is imperfect and full of painful transitions. But it is still the record. Again and again, human ingenuity has taken the raw material of disruption and turned it into progress.

The American story itself follows that pattern. It has never been a straight line. It has been a long argument between fear and possibility, failure and renewal, contradiction and progress. Again and again, the country has required reinvention from people who could not know the ending, but chose to act anyway.

That is the posture I invite you to carry into this next era.

We do not know exactly what AI will do to work. We do not know how institutions will respond. We do not know which companies will matter in ten years. We do not know what the next chapter of the American story will look like.

Still, all of us, not just Americans, have agency to live our ideals. We can build. We can learn. We can reskill. We can invest in people. We can make the new thing useful. We can preserve what deserves to endure while finding the courage to change what no longer works.

History is shaped by the people living inside the uncertainty. It is shaped by the people who keep rowing in the fog, who move before the outcome is obvious, who carry the idea forward long enough for it to become real.

Washington did not know how the story would end when he crossed the East River. He only knew the army had to survive the night. That is often how progress works. The grand historical arc depends on people making the next right move before they have the comfort of knowing how it all turns out.

Maybe years from now, people will look back at this strange, anxious, accelerated period and say the next era was being born in plain sight. They will make it cleaner than it felt. They will draw the line backward and give the story a shape none of us can see yet.

But we are the ones living in the fragile part, before the fog lifts, while the future is still being made. The work is to keep moving before the story makes sense. To learn the new tools and use them for good. To build what we couldn’t before. To carry the idea forward long enough for it to become real. And to be optimistic that it will.

History hangs by a thread.

And sometimes, the wind holds just long enough for the courageous ones to act.

Happy birthday, America.

I’ll see you on July 10th.

Stay ambitious.

Subscribe now

Best of the rest:

🔁 The AI world is getting ‘loopy’ – Agentic AI is moving from one-off prompts to always-on feedback loops, where agents monitor, improve, and prompt other agents, a powerful shift that could make AI more useful and much more expensive. – TechCrunch

💰 How to Earn a Billion Dollars – Paul Graham makes the founder’s case that extreme startup wealth is less about extraction than exponential growth, created by building something users love in a market big enough to compound. – Paul Graham

📚 Has AI Already Killed How-To Nonfiction? Sales Trends, My Personal Data, and What It Might Mean for the Future – Tim Ferriss uses his own collapsing book sales to make a blunt case that prescriptive nonfiction is becoming raw material for chatbots, while voice, sequencing, and transformation may be the last durable moats. – The Blog of Author Tim Ferriss

🏗️ Owning vs. Renting Intelligence — Lin Qiao argues that AI companies shouldn’t just rent frontier model APIs, they need to turn their own data, workflows, and domain expertise into proprietary intelligence they actually control. — Lin Qiao on X

🌶️ OpenAI unveils its first custom chip, built by Broadcom – OpenAI unveiled Jalapeño, its first custom inference chip, built with Broadcom and designed to lower the cost of running its AI models as the company doubles down on its own infrastructure to reduce its dependence on Nvidia GPUs. – TechCrunch

Charts that caught my eye:

→ Why does it matter? People often describe AI applications as “thin wrappers” around the model. The concern is easy to understand: if the model does most of the work, the product layer can look thin, and thin products rarely build deep moats. Maybe. But accurate output takes more than a prompt box and an API call. The model needs the right context, workflow, permissions, integrations, evaluation loop, and domain-specific definition of “good.” The moat may not come from the model itself. It may come from everything required to make the model useful, reliable, and trusted in production.

→ Why does it matter? Either we need a lot more data centers, or AI needs to get much more efficient at using the compute we already have.

→ Why does it matter? Some interesting reporting from OpenAI on how Codex is spreading beyond software engineering. Coding was the first killer use case, but the usage pattern is broadening: Legal, Finance, and Recruiting accounted for the majority of Codex usage around April 2026, and non-developer usage has grown faster than developer usage among individual, organizational, and internal OpenAI users. The next wave of agentic work may start with code, but it will not stay there.

Tweets that stopped my scroll:

→ Why does it matter? Now you can tag Claude directly inside Slack and bring AI into the work already happening. It can read the conversation, use the context, remember your instructions, and take action with the data it can access. It starts to feel less like a tool you open and more like another person helping the team.

→ Why does it matter? I remember visiting Beijing about ten years ago and stumbling across a street vendor selling software. At the time, IBM had a multi-billion-dollar software business, and one of its middleware products could sell for millions of dollars per year. The vendor had a hard drive with the source code for $30. Different technology, same story: American companies spend years and billions building valuable software, and China finds ways to copy it. More here. 👇

Worth a watch or listen at 1x:

→ Why does it matter? "To me, a model is like the database market." - Satya Nadella. Applied Computer is helping companies do that. We’re seeing early attempts with Cursor Composer and Harvey. How many models should there be? Companies should be able to embed their proprietary knowledge inside models they control.

→ Why does it matter? Nikesh Arora is the Chairman and CEO of Palo Alto Networks. Tons in here on AI inside the enterprise, including Mythos as an accelerant in cybersecurity, falling token costs, memory as the moat, and why AI in cyber creates a different set of stakes than consumer AI.

Subscribe now

Quotes & eyewash:

→ Why does it matter? The journey is the reward!

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: What a Diamond Mine in South Africa Taught Me About the Build Trap

Reading time: 5 minutes

In 2015, IBM sent me to South Africa.

A few years prior, IBM had debuted Watson on the game show Jeopardy! The grand challenge made AI feel real to businesses for the first time.

We were meeting with one of the world’s largest diamond miners. The problem was both serious and impossible to forget. Diamonds were being stolen from the mines. There was even a version of the story involving carrier pigeons being used to carry diamonds out of the mines, and the company’s response was creating its own environmental problem.

Someone had mentioned to the diamond miner and the South African government that AI might help. I was sent to figure that out.

Most of my Watson meetings followed the same arc. We had this new, incredibly powerful technology. It seemed like it could do anything, so the conversation naturally moved toward what we could build. I found myself in this dynamic so often that I coined a term with myself to describe it: the “art of the possible” meeting.

I love those conversations. There is something energizing about watching people realize a new technology may let them solve a problem they had stopped believing could be solved.

But there is also a trap hiding inside the art of the possible. When technology makes more things possible, the default question becomes whether we can build something. That is a reasonable starting place, but it is incomplete. The better questions are more practical: how important is this, what will it cost to get the outcome, how much control do we need, and who will own it after the demo works?

Ultimately, the South African diamond mine decided on a multi-part solution that included biometric scans and a lightweight “lie detector” type assessment. It turned out that if you showed up for work outside your normal baseline, you were more likely to commit a lapse of integrity. AI was part of the solution, but mostly basic statistics.

The lesson I took from that meeting was not that AI was overhyped. It was that possibility has to be paired with discipline. When a new technology expands what can be done, leaders have to become much clearer about what is worth doing.

The New Build Trap

I think this is the part of the AI conversation we are entering right now. CEOs have given their teams the mandate to figure it out. Get access. Experiment. Nobody wants to be the company watching this happen from the sidelines.

That instinct is right. Companies should be experimenting with AI. They should be learning what the tools can do. The risk is that a company can create a lot of visible activity without meaningfully improving the work.

The old Build Trap was confusing shipping more features with creating more value. The new Build Trap is confusing more AI usage with better outcomes.

AI makes the Build Trap more dangerous by making building feel cheaper. What once required a product manager, designer, a few engineers, and a quarter of planning can now be prototyped in an afternoon. That is incredible, but it changes the management problem. When more things become buildable, deciding what deserves to be built matters even more.

AI also gets close enough to make a bad allocation feel reasonable. It can generate the draft, write the code or automate a small task. The demo works. The room gets excited. The team feels momentum. Sometimes that momentum is real. Sometimes it is just a faster path to spending time and money on something that should have been left alone.

Imagine canceling a $10 per month calendar scheduling subscription and replacing it with a custom AI workflow that costs $1,000 to build and run. Maybe it was a killer demo. As a business decision, something still went wrong. A cheap, working solution became more expensive.

That is the 2026 version of the art of the possible. The risk is not only that AI fails. The risk is that AI succeeds well enough to make a questionable decision feel smart.

That is why “what’s this going to cost?” should not be dismissed as a boring CFO question. It is a question that forces clarity. A budget is not just a limit. It is a forcing function. It makes you decide what matters, what does not, and where the next dollar should go.

Now that intelligence feels increasingly available, scarcity and constraints matter more. When everything feels buildable, the scarce resource is knowing what is worth building.

Importance, Cost, Control & Ownership

A team asks, “Could we use AI here?” Increasingly, the answer will be yes. But yes does not tell you what the work is worth, how much control you need, whether the economics make sense, or who will maintain it after the demo works.

Start with strategic importance. Is this workflow core to how the company creates economic value, or is it a generic business process? If the workflow is close to the economics of the business you may want more ownership. If it is a common task that every company performs, you probably do not.

Then look at cost per outcome. The question is what it costs to get good work done, not how much AI you are using. The frontier models may be worth it for the hardest tasks, but they will be wasteful for simple ones. You would not hire a brain surgeon to cut your lawn.

Next is control. Who decides which models get used, what data they can see, what gets logged, what gets remembered, and when a human steps in? This is where AI is different from ordinary software. You are not only renting a tool. You may be renting part of the way work gets interpreted, what data is retained, and where its routed.

Finally, ask who will own the thing after it works. AI makes prototypes feel cheap, but it does not make production free. Someone has to secure it, monitor it, update it, explain it, govern it, and fix it when it breaks.

These are simple tests, but require discipline in the moment to be asked.

Capital Allocation Comes for AI

Ultimately, this comes back to capital allocation. It always does.

The job of a business is to allocate scarce resources toward the work that matters most. AI does not change that. It creates a new kind of resource to allocate: intelligence.

These questions about quality, cost, control, and accountability are not as energizing as the art of the possible, but they are where AI strategy becomes real. At first, the conversation seems to be about tools. Over time, it becomes a conversation about how work should move through the company.

The art of the possible is still where the best ideas begin. Protect that energy. Companies need imagination, experimentation, and people willing to ask whether an old problem can finally be solved in a new way.

But possibility is not a strategy. Strategy is choosing. The goal is not to use the most AI, it’s to know where AI is worth it.

This is what leaders have always had to do. AI just makes the decision more important.

Subscribe now

Best of the rest:

🤝 Introducing the OpenAI Partner Network – OpenAI is admitting the enterprise bottleneck is no longer model quality but implementation, turning partners, integrators, and forward-deployed experts into the distribution layer for applied AI. – OpenAI

⚖️ Anthropic Sued Over Limits on Its $200-a-Month AI Plans – Anthropic’s premium Claude backlash shows how AI companies are learning the hard way that unlimited-feeling subscriptions collide with real inference costs once power users start pushing the limits. – The Wall Street Journal

🚀 SpaceX to acquire the AI coding startup Cursor for $60 billion – A reported $60 billion Cursor deal would mark the clearest signal yet that AI coding has become strategic infrastructure, not just a developer productivity wedge. – CNBC

🧠 GLM-5.2: Built for Long-Horizon Tasks – Z.ai’s 1M-context open model keeps pressure on the frontier labs by showing that coding agents are becoming a global, cost-sensitive, open-weights race. – Z.ai

🤖 Salesforce acquires AI customer service platform Fin for $3.6B – Salesforce is buying its way deeper into agentic customer service, reinforcing that support automation may be one of the first enterprise AI markets to consolidate. – TechCrunch

🔁 When AI builds itself – Anthropic lays out the uncomfortable next phase of frontier AI, where models increasingly help build their successors and safety questions move from theoretical to operational. – Anthropic

💸 Ramp at $44 Billion: The Third Pillar – Ramp frames AI spend as a new corporate cost category, making token governance and intelligence procurement feel like the next major finance workflow. – Ramp

🧮 Microsoft eyes DeepSeek for enterprise AI – Microsoft’s flirtation with DeepSeek shows enterprise AI is becoming a multi-model margin game, where cost, control, and token efficiency matter as much as capability. – Axios

🌐 A frontier without an ecosystem is not stable – Satya Nadella’s line is the strategy in miniature: frontier models need developers, partners, distribution, and trust to become durable platforms. – X

Charts that caught my eye:

→ Why does it matter? Incredible chart from Coatue. Innovation waves are often attributed to new technology. The real trigger is when great talent becomes available to build. I think about SpaceX. NASA at one point had 40,000 employees, many of them went to work for SpaceX.

→ Why does it matter? According to TechCrunch, ChatGPT remains the world’s most-used AI assistant with more than 1.1 billion monthly users, but its market share has slipped below 50% for the first time as Google’s Gemini and Anthropic’s Claude have gained ground.

→ Why does it matter? Live and die by the Power Law.

→ Why does it matter? SalesForce.com’s Headless 360 architecture.

Tweets that stopped my scroll:

→ Why does it matter? Hello, open source. Or, technically, open weight. It’s incredible that a model almost nobody had heard of can shoot up the leaderboard this quickly. These charts can be easy to game, but the broader point still stands: open source is catching up fast.

→ Why does it matter? This is a wild stat. Tim’s book sales are down 80% because people are going to LLMs to get what they need. OpEd by Bernie Sanders in the NYT a few weeks ago illuminates this issue. Where do you think all the intelligence came from in the first place?

→ Why does it matter? Yes, this is financial advice.

Worth a watch or listen at 1x:

→ Why does it matter? A complete guide to Nick Milo’s AI OS for personal notetaking and knowledge management. I have been a part of the second brain movement for years. LLMs are a game-changer. The beauty of Nick’s approach is that you own the underlying data.

→ Why does it matter? At Harvey, token usage jumped from 1 trillion in January to a projected 12-13 trillion this month. Molly O’Shea in conversation with their CEO, Winston Weinberg.

Subscribe now

Quotes & eyewash:

“You cannot be successful without failure.” — John F. Kennedy

→ Why does it matter? It is how you respond from that failure that is most important.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: The Seven Primitives of an AI-Native PersonalOS

Reading time: 8 minutes

“When was the first time I ever wrote about a personal AI operating system?”

Bam. A few seconds later, I had the answer.

January 30, 2026 — Building AmbitiousOS & AI Agents w/ Claude Code

In the old world, I would have gone to ReadingAmbitiously.com, searched the archive, and eventually found it. Not a difficult task, but an annoying one.

And right now, I think most of the world still thinks of AI as an answer engine. A technology a tad better than Google.

That’s a version of AI. In my opinion, the least interesting.

So let’s ask a different question.

“Hey AmbitiousOS, how has my thinking on AI changed since we initially launched?”

Now we’re cooking.

AmbitiousOS did not just retrieve the first post. It reconstructed the arc across 30 data points and five editions:

1. January 30 — The initial build. A personal operating system to help with the newsletter.

2. February 20 — Memory. The idea shifted toward memory, and the implication that near-perfect recall would belong to people who choose to document.

3. February 27 — The dataset. The “it” in AI: the dataset, often selected, created, and curated by people.

4. April 10 — Always-on intelligence. AI proactively working in the background and the hidden tax of broken memory in knowledge work.

5. May 22 — The new knowledge work. A new approach to knowledge work: externalizing your process, turning it into skills, and letting AI agents run the parts of the job that are really the “work about the work.”

Now this is the power of AI.

In the old world, I could have eventually identified this pattern. I would have read through the old editions. In the new world, AmbitiousOS already had the context.

There is a great line from John Culkin, the Jesuit priest: “We shape our tools, and thereafter our tools shape us.” Silicon Valley loves this quote because it captures something that feels obviously true once you see it.

For much of 2026, I have been writing about the shape of these tools. Now, in a small but increasingly real way, many of those ideas are making their way into the PersonalOS I’ve built, AmbitiousOS.

The path (only 4 months in length) looked something like this:

• The Reading Ambitiously archive, now more than 100 editions, became searchable.

• The searchable archive became a knowledge graph that shows how concepts connect.

• The graph evolved into the first version of system memory.

• Memory expanded into claims, predictions, and trajectories.

• The tools we developed became callable.

• The workflows involved in manufacturing the newsletter became skills.

• The routines and scheduled jobs started running those skills in the background.

• MCP made the same AmbitiousOS brain accessible across other AI environments, such as ChatGPT, Claude, and Codex.

I know that sounds technical. Great products ultimately feel magical and frictionless. Those are coming. But in a hobbyist way, I think these capabilities are starting to reveal the emerging system architecture for AI-native work.

What becomes clearer to me every day is that every knowledge worker, team, and company will want some version of this.

The primitives of AmbitiousOS

AmbitiousOS is currently built around seven primitives: context, memory, skills, tools, routines, portability, and the build loop.

1. Context: what are the inputs, and how do they relate?

The original dataset in AmbitiousOS consisted of about every edition of the newsletter and all the primary research that went into writing them. With the help of Claude Code, we built a database and labeled the data. Very quickly, it became clear that the AI needed help understanding all of this new context.

A knowledge graph is a map of how things relate. People, companies, ideas and themes all “live” near other things. Useful knowledge is rarely an isolated fact. It is usually about connection.

In the Reading Ambitiously context, OpenAI is not just a mention in an old issue. It sits near ChatGPT, Anthropic, Sam Altman, xAI, DeepSeek, Microsoft, Nvidia, tokens, agents, and the broader question of where AI value will accrue. Since I started writing, OpenAI has been mentioned 54 times in the archive. That is no longer a single note. It is a neighborhood of thought.

The power comes when your ideas are no longer disconnected notes. Once the relationships are visible, both you and the AI can start seeing the shape of your thinking. As of this writing, we track 1,071 entities and 9,831 connections amongst them.

Context provides AI with the right inputs and an understanding of how they relate.

2. Memory: what do I believe, and has it changed?

Memory is different from storage. Memory helps preserve the meaning.

A good memory system should let you rewind the tape and see how your thinking changed over time. What did you believe? When did you believe it? What changed? Does the claim still hold? Was the prediction right, wrong, partial, or unresolved?

The latest version of AmbitiousOS now tags claims, predictions, recommendations, opinions, confidence, dates, and resolution status. That means the system can help me reason across my own thinking, not just retrieve old words from the archive.

That is a meaningful shift. If I ask, “What have I written about OpenAI?” I get retrieval. If I ask, “How has my thinking on OpenAI changed?” I get memory. The first question finds artifacts. The second question helps me reason through how my thoughts have changed over time.

Memory is what helps the AI understand how context has evolved over time.

Right now, memory is a new and emerging space. Next time you’re using ChatGPT or Claude, ask it to “commit this to memory” which is a start. Better solutions are coming.

3. Skills: how should this work be done?

Skills are repeatable procedures. Think of a skill as a good onboarding document for your agents. You spell out the task, explain the process, define what good looks like, and make clear what “done” means.

This is important because much knowledge work is repetitive. Every week, I have to gather inputs for Reading Ambitiously, evaluate links, decide which charts matter, check whether I have used something before, draft “why it matters” blurbs, copyedit, source check, and assemble the final issue.

For years, most of that process lived in my head. AmbitiousOS is helping me turn that process into skills.

Once you write down a skill, it can improve. It can be versioned. It can be reused. It can be handed to an agent when the same work comes up again. A skill turns one piece of a messy process into something an AI can run more reliably. Many skills together start to look like a new way to manufacture knowledge work.

The practical lesson is straightforward: look for the work you repeat. What do you do every week? What do you explain over and over again? What process would you teach a new hire if they joined tomorrow? Write that down. That is the beginning of a skill.

Skills tell the AI how the work should be done.

4. Tools: what can the system actually do?

If skills are the onboarding instructions, tools are how the AI becomes operational.

A model can reason. A tool lets it do something. Search the archive. Query the knowledge graph. Pull recent editions. Check prior claims.

Without tools, the AI mostly talks. With tools, it can operate inside a workflow.

When I start a new edition of Reading Ambitiously, these tools help the model move through the archive and retrieve grounded context from previous editions. Instead of vaguely remembering that I talked about “the transition tax,” it can fetch the issue. Instead of saying OpenAI has been important in the archive, it can query the graph and show how often it appeared and what it connects to.

Tools are where the action happens. This is the question to ask in your own work: what would an AI need besides the answer? Should it pull files? Search a note-taking app? Check a CRM? Query a database? Draft a follow-up? Compare a document against prior decisions?

Tools give the AI the ability to act.

5. Routines: what should happen without me asking?

Routines are the heartbeat of the system.

A system that only works when I remember to prompt it is still dependent on me to start the work. That is useful, but it is not always on.

Scheduled jobs and automations create background loops. AmbitiousOS can ingest new information, prepare briefings, update the graph, refresh memory, and keep the knowledge base current. Every night while I am asleep, AmbitiousOS can go through a kind of “dream cycle” in which it gardens the underlying knowledge base and ensures it is up to date.

In software, these are often called cron jobs. The word Cron derives from Chronus, meaning time. A cron is a time-based job. The system does not need to wait for me to constantly update its knowledge base. It can routinely run jobs to do that itself.

This is one of the most important primitives because so much knowledge work is maintenance. Keeping track of what changed. Revisiting old assumptions. Connecting the new thing to the old thing.

Routines are how AI becomes a system working alongside you that’s “always on”.

6. Portability: where else should the brain be accessible?

The current best version of this is the Model Context Protocol (MCP). MCP is a technical phrase, but the plain-English version is simple: it lets the brain travel.

Early versions of AmbitiousOS were trapped inside our own interface. If you wanted to use the system, you had to go to AmbitiousOS. Sometimes I want to think in ChatGPT. Sometimes I want to build in Codex. Sometimes I want to work in Claude.

Now, via MCP, AmbitiousOS can be accessed from other AI environments.

AI becomes more valuable at the moment of work. If I am drafting a reply to an email, I do not want to leave the workflow, go to a separate archive, search for context, copy it back, and then write. I want the AI working with me to be able to tap into the relevant brain at the right moment.

In a harness like Claude, ChatGPT, or Codex, AmbitiousOS can be connected alongside other MCP’s that bring in email, calendar, Slack, documents, or code.

This is also why people talk about software becoming “headless.” The interface becomes less important than the capability. The value is not always the screen you log into. Increasingly, the value lies in the service, context, or action an agent can call on when it needs it.

Portability puts the brain where the work happens.

7. Build Loop: how does the system improve itself?

Codex and Claude Code drive the build loop.

This is the newest and most exciting part of AmbitiousOS. I can now use Codex to build the next version of AmbitiousOS while giving Codex access to what AmbitiousOS is supposed to become.

The first version of AI coding was: ask AI to write code.

The next version was: let it work inside your code repository.

The emerging version is: give it the latest edition of Reading Ambitiously and let it improve itself.

After I publish an edition of Reading Ambitiously, Codex can pick it up, understand the concepts, translate them into a product roadmap, and start helping build the feature set. If I write about the seven primitives, those ideas can be taken up by the AmbitiousOS and Codex.

The system I am building becomes part of the environment used to build the system. That is sometimes called a recursive system.

This one is mind-boggling, I know. We are only touching the tip of the iceberg.

But even at this early stage, this is pretty cool. The build loop improves the system with its own context.

Zooming out

The applied lessons are these:

• AI needs the right context and an understanding of how concepts relate to one another.

• Memory helps explain how that context has evolved over time.

• Skills give the AI instructions for how to do the job.

• Tools give the AI what it needs to do the job well.

• Routines let the jobs run in the background.

• Portability puts the AI where the work happens.

• The build loop lets the system improve with its own context.

As of June 12, that is my best description of the emerging architecture of AI-native work.

Make sure the AI captures that claim.

The more I build in this direction, the more I believe this is where knowledge work is going.

AI changes what is possible, but only if the AI has access to the right context, in the right place, at the right time, with the right tools to act on it.

We shape our tools, and thereafter our tools shape us.

Stay ambitious, my friends. And if you’re really ambitious and ask your favorite AI to give this edition a read and start building yours today.

Request access to AmbitiousOS today.

Subscribe now

Best of the rest:

💸 OpenAI Files Confidentially for Its IPO – One week after Anthropic, with an $852B private mark, a Musk lawsuit cleared, and Microsoft’s 27% stake set for a public price. – Yahoo Finance

🔒 The 24-Year-Old Running $20B on an AI Thesis – Leopold Aschenbrenner’s Situational Awareness is up 1,000%+ since launch, and Jane Street, which rarely backs outsiders, invested. – WSJ

💼 The Moat Was Never the Code – Agents made translating a domain into software cheap; knowing what right looks like is the part you still can’t prompt for. – Aaron Brethorst

🧠 The Case for Cancelling Your AI Subscription – A prolific builder lists 17 projects he never needed, calls the tooling a “thermonuclear ADHD amplifier,” and argues friction was the product. – David Wilson

🔒 OpenAI Eyes a 10-Gigawatt Ohio Campus – A 20-year lease on federal land, Nvidia guaranteeing the financing, and a build that could run $500B. – The Information

🤖 From Wizard to Patron – Mollick’s Fable review: nine-hour autonomous builds, finished work back, and his unease that steering is no longer the same as doing. – One Useful Thing

🏛️ Dario’s Policy Playbook for the Exponential – Amodei moves past transparency: FAA-style testing for frontier models, wage insurance for displacement, and a democratic coalition locking down the supply chain. – Dario Amodei

Charts that caught my eye:

→ Why does it matter? According to Coatue, a business valued between $100B and $1T has a higher statistical likelihood (31%) of multiplying its value by 10x compared to smaller, earlier-stage unicorns (8%).

→ Why does it matter? The median RAMP customer is paying $12/month per employee on AI. The top completely AI pilled decile? Closer to $9k/month per employee. The top 0.01%? $90k/month per employee.

→ Why does it matter? The Silicon Data LLM Token Expenditure Index, a benchmark for how much the market is actually spending on AI tokens, has started rolling over. According to Citadel this is a shift toward cheaper models. Companies substituting away from expensive frontier AI toward "good enough" alternatives.

→ Why does it matter? Fable (the publicly available version of Mythos) is a beast!

→ Why does it matter? This chart doesn’t include Anthropic’s Fable out this week but it’s a comprehensive LLM landscape!

Tweets that stopped my scroll:

→ Why does it matter? So, spend the the time in the meetings you’re most stressed as those are likely the ones where you’re making the hard decisions? Amazing how AI can rapidly create and personalize a dashboard like this.

→ Why does it matter? I too get the impression OpenAI will soon be back to #1 on the LLM leaderboard. Remember, we’re just starting to train models on Nvidia’s latest Blackwell series, and Vera Rubin is next.

→ Why does it matter? An important list to keep an eye on.

→ Why does it matter? A lot of talk about Elon Web Services (EWS). Well, here it is. 40% of SpaceX’s 2025 revenue.

Worth a watch or listen at 1x:

→ Why does it matter? AI may not hurt the people who “did everything wrong.” Tyler Cowen’s worry is that it hurts the people who did everything right. The consulting partner making $1.4 million does not become unemployed. He becomes a $300,000 energy executive in Houston. Financially, he is fine. Psychologically, he is not. Politically, that may matter a lot.

→ Why does it matter? Dan Shipper, CEO of Every, spent a week testing Anthropic’s Fable 5, a Mythos-class model that scored a 91 out of 100 on Every’s senior-engineer benchmark!

→ Why does it matter? If you know me, you know I’ve never seen an S-Curve I didn’t like!

Subscribe now

Quotes & eyewash:

→ Why does it matter? Haters going to hate.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: Electrons → Tokens → Outcomes

Reading time: 7 minutes

Every technology era finds a unit of account.

Small units have a way of becoming organizing principles. A barrel of oil was just a barrel of oil until the world learned how to refine and consume it. A kilowatt-hour was just electricity until factories, homes, and grids reorganized around it.

The token may be the same kind of unit for the AI economy.

In a recent interview, Jensen Huang compressed Nvidia’s mission into one sentence:

“We convert electrons into tokens.”

Electrons go in. Tokens come out. In the middle sits the world’s most valuable company.

Tokens are the units AI consumes to read, reason, and generate a result. The more information you give the model, and the more work it produces, the more tokens it uses.

Most AI initiatives are focused on the same basic thing: using tokens to generate an outcome.

The simplest frame I can come up with to put it all together is: Electrons → Tokens → Outcomes

Nvidia converts electrons into tokens. The model providers, OpenAI, Anthropic, and Google, build the engines that use those tokens. Everyone else converts tokens into outcomes.

That is overly simplified. It is also useful.

Because isn’t that what we’re all trying to figure out? How to get the AI to do what we want?

The Harness

The current best answer is the harness.

The harness is what it sounds like. A horse has power, but the harness lets you direct it. A model has intelligence, but the harness lets you aim it at work.

There is a new term called harness engineering. It is the discipline of building systems around AI, so the model can perform useful work repeatedly.

A good harness provides the model with the richest inputs possible, in the correct order. Those inputs go into the model’s scarcest resource: the context window.

Think of the context window as the model’s working memory. Context windows are measured in token size. The current start-of-the-art models offer 1M token windows. If something is not in the context window, it is not reliably in play.

The context window is the AI’s mind’s eye.

The idea is not to give the model an ocean of inputs. It is to provide the model with the essential inputs, in the right order, at the right time. Token by token.

This is why so many of the recent AI breakthroughs rhyme. MCPs, skills, memory, proprietary datasets, config files, voice, and vision can sound like separate technical ideas. They are better understood as parts of the same emerging AI native system architecture.

Each one ultimately expresses tokens that shape context.

That is the harness’s job. It determines what the model sees, what it remembers, what tools it can use, what rules it follows, and when a human needs to step in.

That is where much of the applied work in AI is happening. If you hear about a new company raising money to “build AI for X,” there is a good chance it is not training a frontier model. It is renting models from OpenAI, Anthropic, and Google, and spending its R&D dollars on the harness. This is likely what internal tech teams are doing as well.

The Product Is Not the Model

Claude Code is a useful example of a harness. Sierra, Harvey, Legora, Cursor, Decagon, and Glean are others.

There is a big difference between Claude in a blank chat window and Claude inside a software development environment. In the blank chat window, the model can answer questions and write code. Inside Claude Code, the model can see the code repository, understand surrounding files, edit code, run commands, respond to errors, and keep moving through a task.

The difference is the harness. The harness helps input all of the coding context into the model.

Ultimately, we want the work done. In the case of Claude Code, high-quality code generated.

The Business of AI

Augmentation is the easier one to understand because it’s familiar. Give people AI products and help them do more. Let engineers write more code. Let analysts read more filings. Let customer service teams resolve more tickets.

This is already happening. R&D teams using AI coding tools can build more software.

But augmentation keeps the current structure mostly intact. The user still owns the outcome. The AI is a tool. The customer still pays for seats or usage.

That version feels familiar because it looks like software with new killer features.

Service-as-software sells the work itself. The software performs the service and delivers the outcome.

When you sell traditional software, the core questions are product questions. Is it useful? Does it meet my requirements? Is it easy to install?

When you sell an outcome, the questions become operating questions. Can you do the work better than the customer can? Can you do it repeatedly? Is it of high quality? Who is liable when it’s wrong?

That is not a small shift. It changes pricing, gross margins, customer expectations, and liability.

A SaaS company can say, “Here is the tool.” A service-as-software company says, “These tools also do the job.” Then it must back it up.

And likely, future application providers will be a combination of both.

The Thrive Bet

This week, it was reported that Thrive Holdings plans to invest more than $1 billion in an AI-powered accounting roll-up. OpenAI had already taken an ownership stake with a focus on accounting services.

I do not know exactly what they are building. But here is my best guess.

They are not going to sell software to accountants. They are building a harness that is exceptionally good at accounting, then put it into the hands of their CPAs to deliver the accounting outcome faster, cheaper, and with the same or better service level.

That is service-as-software. The accounting firm is not just a distribution channel. It is the laboratory.

Inside that laboratory are the ingredients that matter: proprietary data, domain expertise, process knowledge, and a customer who expects the job to be done correctly.

And what are they going to do? Convert as much of it as possible into tokens, then drive better outcomes than the competition.

If a typical accountant can handle only 10 clients today and 100 or 1,000 tomorrow with the right harness, that is the opportunity.

When AI performs the work, someone still has to stand behind the work. That is where audit trails, human review, service levels, and liability come into play. The companies that can prove reliability will not just have better products. They will be easier to trust.

That trust is what allows the outcome to become the business.

Adding It Up

So the current best Reading Ambitiously frame for how to think about the business of AI is this:

Electrons → Tokens → Outcomes

The harness is the conversion system between tokens and outcomes.

That is where many AI companies will live. It is also where buyers and investors should spend more of their time. Not asking whether a company “uses AI,” but asking what it can reliably produce with it.

The hard part of AI right now is adding it all up without losing the plot.

Electrons become tokens. Tokens move through a harness. The harness produces work. The work creates an outcome. The outcome creates value when someone trusts it enough to depend on it.

That last step is the hard one.

It is also where the business lives.

And unless you’re building picks and shovels, that’s how you’re going to make money.

The trusted outcome.

If you plan to build an AI business, this is what you need to work backward from.

Subscribe now

Best of the rest:

💸 Anthropic Confidentially Files Its Draft S-1 — Anthropic submitted a confidential draft S-1 to the SEC on June 1, putting it in the 2026 IPO race alongside SpaceX and OpenAI, on a run rate north of $47B. — Anthropic

⚖️ Kirkland to Spend $500M Building Its Own AI Platform — The world’s highest-grossing law firm is building, not buying: $500M over three to four years to own its AI outright, a direct bet against renting from Harvey or Legora. — Reuters

🧮 Building Finance AI Agents — OnlyCFO breaks a finance agent into its four parts (a skill, a system prompt, MCP connectors, a trigger) and shows one flow turning a two-hour reconciliation into five minutes. — CFOpilot

🛠️ 11 Ready-to-Use AI Skills — Khe Hy hands over eleven copy-paste Claude skills for everyday knowledge work, the kind you install once and forget you are running. — Khe Hy

🧠 Where the Winners Get Built — Brett Queener argues the next durable software companies get built inside specific, messy domain workflows, not horizontal AI, because that is where context and trust compound. — Tales from the Bonfire

🐆 Riding the Leopard — Packy McCormick’s talk-turned-essay on composure and creative nerve: if you are a piece of the universe experiencing itself, the job is to be the fullest, strangest version of you. — Not Boring

Charts that caught my eye:

→ Why does it matter? Anthropic engineers, on average, ship 8x as much code per quarter as they did in 2021-2025.

→ Why does it matter? Just 10 SaaS companies accounted for ~80% of the market cap regained over the last 34 days. Full update from Meritech here.

→ Why does it matter? AI-related companies have issued ~$140 billion in investment-grade bonds year-to-date and attracted ~$220 billion in venture capital funding, making up 87% of the total. H/t: The Kobeissi Letter.

→ Why does it matter? DeepSeek (open-source Chinese LLM) on top of the trending?! 👀

→ Why does it matter? Do we have the compute?

Tweets that stopped my scroll:

→ Why does it matter? $84 vs $954 across the same 100 tasks. 11x cheaper! This is the power of open-source alternatives to the state-of-the-art closed source. The question is: would you trust a Chinese open source model to do it?

→ Why does it matter? Imagine what Siri will be capable of in 1-2 years? And Apple plans to run these models locally. The rumored MacBook Ultra (an unreleased Apple laptop) is considered to be even more powerful than the MacBook Pro.

→ Why does it matter? This is AI for good. My first grader and I are already using Claude Code together. Be sure to check out Reading Ambitiously 4-17-26 if you missed it.

→ Why does it matter? Jonny Ive’s rumored OpenAI device?! During Jon’s years at Apple, the leaks weren’t always unintentional. Jonny loved to build the hype.

Worth a watch or listen at 1x:

→ Why does it matter? Sarah is one of the best CFOs in the game and no doubt spent a ton of time preparing for her remarks here. There has been a media driven story about whether she and Sam are on the same page about the spending. Lots to unpack!

→ Why does it matter? Remember when Uber IPO’ed and people said it’d never be profitable? Dara has masterfully executed and shares a lot of the story with Patrick.

→ Why does it matter? Fred Wilson co-founded Union Square Ventures in 2004. USV has backed Twitter, Etsy, Coinbase, Kickstarter, and MongoDB. After 40-years in the industry, Fred is rethinking how to organize the firm and how AI is going to play a major role in it.

Subscribe now

Quotes & eyewash:

→ Why does it matter? Earlier this week, I had the opportunity to meet Coach Jack Clark. The Cal Rugby coach with the highest winning % of all time in collegiate rugby, #2 place isn’t even close. Jack offered some incredible thoughts about leadership. Non-cognitive grit, mental toughness, grateful for everything, entitled to nothing, team. He sends his team through a tunnel every game (a reminder that you’re coming into our world) a world where your deeds speak louder than your words, a world where we all win together. An incredible model for leadership which you can read more about here.

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: I asked AI about a dead tree limb. The next morning, it was gone.

Reading time: 4 minutes

With three kids, there is always something happening at the Lynch house.

There is also always something happening literally with the house.

Every homeowner has some version of the project list. Not the well-organized one you write down on an ambitious Saturday morning. The other one. The one that lives in your head and follows you around.

Fix the thing in the garage. Call the person about the window. Figure out why that outlet doesn’t work.

For months, there had been a dead limb hanging from the tree in front of our house. I saw it all the time. I knew it needed handling. I’m almost embarrassed to admit this, but I had no idea whose problem it was.

Did we own the tree? Or was this the city’s responsibility, buried somewhere in a municipal gray area that would take twenty minutes of research to understand?

And because it required twenty minutes of research, I did what any reasonable adult with too much on his plate would do.

Nothing.

Not because it was hard. Because it was annoying. It was one more thing on the long list of things I knew I should handle but kept putting off.

Until Tuesday night, when I opened ChatGPT and said, in effect: you figure it out.

A few seconds later, the AI told me exactly what to do. It identified the likely responsible party, found the right place to submit the request, wrote the ticket language, and assigned me one 30-second task: snap a photo.

I copied the language into the town portal, attached the photo, and hit submit.

The next morning at 7:53 a.m., to my surprise, guess who was out front.

The public works department.

I texted my wife: “You’re not going to believe it. The AI really did it.”

The AI did not operate the bucket truck. A human crew did the real work.

But it gave me the capacity to get something done when I personally was at 0%. It turned a vague annoyance into an action. And, not for nothing, it gave me the creative spark for this week’s Reading Ambitiously.

So much of modern work is not the work itself. It is the work about the work.

Find the website. Research the rules. Draft the form. Upload the document. Submit the request.

Small tasks, maybe. But they create a huge amount of friction. They are not strategically important enough to dominate the day, but not trivial enough to disappear. So they accumulate.

That is what the dead tree branch represented. Not a major life problem. A category.

A list of things that sit unfinished because the first step is annoying. We all have metaphorical dead branches in our lives.

And there are a lot of them at work.

The morning briefing

I have always loved the idea of the President’s Daily Brief.

Before the President wakes up, an entire intelligence apparatus has been working to answer one question: what needs to be on the President’s radar today?

That is a pretty good mental model for how I want AI to work for me.

So I told the AI to start acting like my own small army of intelligence officers. Hours before I wake up, it starts preparing Jack’s morning briefing. The inputs are simple: email, calendar, Slack, and the context of what is going on in my world.

It arrives at 7 a.m.

The briefing is useful. It tells me what happened overnight, what meetings matter, what I owe people, what needs preparation, and what might otherwise slip through the cracks.

But the briefing is only the beginning.

These days, I ask the AI a follow-up question:

“What can you do for me?”

• If I have a customer meeting, the AI pulls together background on the attendees.

• If I am double-booked, it suggests a reschedule and drafts the email.

• If I need a colleague’s help, it drafts the Slack message and asks for fifteen minutes.

• If an important email slipped through the cracks, it suggests the follow-up.

• If I have a meeting I accepted three weeks ago and have not thought about since, it writes the pre-read.

That is the work about the work. The layer of small, necessary actions that make the real work possible.

Honestly, most days I print the briefing and review it with my notebook and pen. And then record a voice memo back to the AI to get to work on the to-do list. That combination feels amazing to me: offline thinking, powered by technology.

In the wise words of Ferris Bueller, “If you have the means, I highly recommend it.”

That is the difference between a chatbot and an agentic workflow. Once AI can operate where work already lives, the interface starts to feel less like search and more like delegation.

Mine the mundane for leverage

The first spreadsheets did not replace CFOs. They replaced the manual recalculation.

Early behavior change is often practical before it is profound.

That seems to be happening here. AI is starting to collapse the distance between noticing a task and acting on it. The AI did not just answer my question about who owned the tree. It helped create the ticket, move the process forward, and actually got the branch removed. The outcome.

Of course, AI still makes mistakes. Proceed with caution.

But there is a lot of work that should not require so much human friction.

Ask AI to take the first pass.

It might surprise you how many things on your list are not actually hard. They are just annoying to start.

And it might surprise you what comes off the list.

Stay ambitious, my friends.

Subscribe now

Best of the rest:

💸 The magic of doing $10,000 per hour work – Khe Hy’s leverage framework is a useful reminder that the highest-value work often feels slow, ambiguous, and unrewarding in the moment, but compounds into the systems, relationships, and judgment that actually move a life or business forward. – RadReads

🕊️ Magnifica Humanitas – Pope Leo XIV’s first major AI encyclical frames artificial intelligence not as a technical novelty, but as a civilizational test of whether power, work, truth, and human dignity can still be ordered toward the common good. – The Holy See

🏦 Your AI agent now has a brokerage account – Robinhood opened native MCP servers letting agents trade equities and spend on a ring-fenced virtual card, with limits you set. – Robinhood

🤖 Two Claude diehards now reach for Codex first – GPT-5.5 needs less hand-holding and compacts long threads better; they still keep Claude for design and writing. – Technically Curious

👋 Salesforce starts waving bye to the UI – A headless, agent-first turn for the CRM giant. One read: enterprise software gets consumed by agents, not screens. – The Register

🛡️ AI is finding critical bugs faster than humans can patch them – Anthropic’s Project Glasswing used Mythos Preview to surface over ten thousand high-severity vulnerabilities. Finding flaws is now easy; the bottleneck has shifted to patching. – Anthropic

🧠 Being rude to ChatGPT made it more accurate – A small study found very rude prompts beat very polite ones on accuracy, 84.8% to 80.8%. The opposite of what older work showed. – arXiv

Charts that caught my eye:

→ Why does it matter? 4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions!

→ Why does it matter? Interesting data indicating that the majority of AI tokens are being spent on bug fixes and rework, with only 18% going to shipped products.

Tweets that stopped my scroll:

→ Why does it matter? Turns out “slop” was the Merriam-Webster Dictionary Word of the Year in 2025, which we predicted! Ethan Mollick’s latest One Useful Thing piece, “Choosing to Stay Human,” is a useful reminder that the real question is not whether to use AI, but what parts of our thinking we want to keep for ourselves. The choice ahead is not AI adoption or AI avoidance. It is deciding, task by task, what we are willing to hand over and what we still want to earn ourselves.

→ Why does it matter? In just 24 months, Slack predicts more agents will use their platform than humans. SFDC also had a big Q1!

→ Why does it matter? I’m feeling this vibe shift as well. OpenAI’s Codex (competitor to Claude Code & Cowork) is getting really good. Maybe it's personal preference, but I do enjoy GPT 5.5 as a model and interacting with it. Not sure how to describe it, Opus is colder, more rigid. GPT 5.5 feels like a better all-purpose model. Doesn't matter when coding, but it is noticeable, almost off-putting in the non-coding workflows. I feel like I can get into a better flow state with GPT 5.5 for knowledge work, less so with Sonnet/Opus.

→ Why does it matter? 🙈

Worth a watch or listen at 1x:

→ Why does it matter? Dan Shipper is living in the future and has committed to leaning into that future in the most possible way. Listening to Dan talk about how he completes his work, his “AI-native workspace,” and his thoughts on his recent article, “After Automation,” provides many thought starters.

→ Why does it matter? David Senra is at the top of his game. His interview with Rick Rubin on the creative process is outstanding because David does the work: he reads the books, studies the guest, and shows up with real curiosity. You can feel his passion and preparation in the conversation, and Rubin is the perfect guest for that style. This one is thoughtful, patient, and full of insight on creativity, taste, and the discipline of making things. I may need to listen to it again. It is that good.

Subscribe now

Quotes & eyewash:

"Until death, all defeat is psychological." - Marcus Aurelius

→ Why does it matter? Never give up!

Subscribe now

The mission:

The Wall Street Journal once used “Read Ambitiously” as a slogan, but I took it as a personal challenge. Our mission is to give you a point of view in a noisy, changing world. To unpack big ideas that sharpen your edge and show why they matter. To fit ambition-sized insight into your busy life and channel the zeitgeist into the stories and signals that fuel your next move. Above all, we aim to give you power, the kind that comes from having the words, insight, and legitimacy to lead with confidence. Together, we read to grow, keep learning, and refine our lens to spot the best opportunities. As Jamie Dimon says, “Great leaders are readers.”

Disclaimer: This content is for informational purposes only and does not constitute financial, investment, or legal advice. Readers should do their own research and consult with a qualified professional before making any decisions.

The big idea: The new knowledge work

Reading time: 7 minutes

Last week I wrote about the transition tax. On Saturday, after the kids went to bed, I started paying it.

You might think, given how much I write about AI, that the manufacturing of this newsletter is entirely automated. It is not. Actually, it is more manual than I’d like to admit.

I am not talking about the Big Ideas or the content I have tastefully curated for you. I am talking about assembling HTML, screenshotting tweets because Substack is in a feud with X, formatting, carefully selecting the right emojis for Best of the Rest, and attributing the quotes.

This is not the thinking part. It is not the creative process, nor the ultimate expression of making something from nothing. That is by far the most rewarding part of why I do this every week.

This is the work about the work. And it takes a lot of time that I want back.

The central promise of AI is to give it back. To let your Agents take on the mundane.

To elevate you. And to get there, you need a new form of knowledge work.

The new operating system and applications

In the old world, the knowledge worker has access to a computer, an operating system, and software. A file system, your favorite applications, productivity tools, and a browser. You use these tools to move the work.

Software historically gave you the tools to do the job. What is different with AI is that the tools can do the job.

The new world is powered by AI, and its primary operating system harnesses intelligence to drive outcomes. Your new job as a knowledge worker is to get in the driver’s seat, tell it what to do, and give it the vital context it needs.

One example of a harness is Claude Cowork from Anthropic. When you open it, you see a mosaic of capabilities working together:

• A voice or text-based experience powered by a large language model you select.

• Connectors that wire the brain to the tools where your work already lives, such as your email, calendar, files, and applications.

• Plugins, which are pre-built collections of capabilities for a specific kind of work.

• Memory management (although it is very nascent today).

• And skills.

Skills are one unit of automation. They are the onboarding documents for your AI agents.

Skills are to the new world as applications were to the old world.

You write down how a task should be done, the way you would explain it to a new hire on their first day. The agent reads it when the task comes up. The next time the work comes up, it runs the same way.

If a prompt is a conversation, a skill is a procedure.

Software used to come in a box. You bought it, installed it, and shaped your work to fit its menus. The unit of software was the application, and the work occurred within it. Like those icons on the desktop.

The unit of software in the new world is now the skill. You write it, shape it, and version it. The work happens around it, in the harness, with the agent reaching for the skill and the context at the right time.

A working example, ready to install

Last week, Anthropic released something called Claude for Small Business. What is it? You guessed it, a big library of skills. Fifteen of them, all written for the work a typical small business does.

There are Monday morning briefings, tax prep, month-end closing, cash forecasting, and payroll.

Each of these is a paragraph of plain English sitting in a markdown file, ready to run when you ask for it. The files are version-controlled and updated by Anthropic. You install these skills into your harness, wire it up to your systems like HubSpot, QuickBooks, and PayPal, and you are ready to start automating knowledge work.

It is clear to me that an app store of skills will come. I see a world where you develop a skill, share it with a colleague, and make it available to others, free or for a fee, through a marketplace built for security and compliance.

You do not have to wait for that to arrive. If you can make Word documents, you can start developing skills today.

The agentic factory

I have been building skills to automate the manufacturing process of this newsletter. I liken the compilation of skills to the periodic table.

The elements are discrete. Each one does a very specific thing. Together, they make everything.

There is:

• A sweep skill that pulls every candidate from the private Slack channel where I flag items every week. This channel is accessible to Claude via a Slack MCP connector.

• A classifier that interprets the content as either Best of the Rest, charts, tweets, or podcasts.

• A fetcher that fetches data and metadata from the internet about the content.

• An archive checker that uses AmbitiousOS to search every previous issue and see if I have already used the content.

• A drafter who starts coming up with ideas in my voice for the why it matters.

• A builder that handles Substack’s rich HTML formatting requirements.

None of these is doing the thinking or the heavy lift of writing the Big Idea. That is the part I love and the human-only part. What they are doing more and more of is the manufacturing, the part of the job that takes hours every week.

Your new job

The knowledge worker’s job is to do the work we are doing here.

The hardest part is thinking in a new way. It is taking a step back and thinking through the processes you have today to manufacture your work. Most knowledge workers have never written this down. We do the work because we know the work, but the way we do it lives in our heads.

Your next job is to externalize that knowledge. Break the processes into elements, write each element into a skill, and let the agents run the elements you should not be running yourself.

This is the practice of paying the transition tax. And you can start right now by opening your favorite AI tool and saying “Hey, let’s create a skill together”.

Done right, you can construct and assemble the skills your agents need to automate the mundane, so that you can move to higher and higher value work. The most creative and rewarding part of the human experience.

It’s worth it, let’s pay the tax together.

Stay ambitious.

Subscribe now