My promise and premise is this:

I’m still playing around with formats. In addition to the weekend deep dives, I hope to share these meandering reviews where I reflect on the links and stories that caught my eye.

If you want the bleeding edge and up-to-the-minute hype on the latest OpenAI models (like Codex 5.5), check out the Neuron or the Rundown.

If you want to sip coffee while reading my ramblings, stick around.

—Charlie

P.S. I’m sticking with the name Coffee with Claude for now – partly from inertia and partly from affection for Anthropic’s soothing color scheme.

Gleanings

Same drudgery, different interface.

Sam Altman just said something similar to what I framed as the optimistic scenario for AI’s evolution – namely, that it could let us do a lot less clicking around on a computer screen:

“The degree to which most people will realize they can like, sit back and watch an AI do most of their drudgery, is going to surprise people.”

I don’t think this world has materialized for anyone yet. It’s possible, but it’s also possible that instead of clicking around between apps, we’ll just end up clicking between chats.

The Last Mile

Box CEO Aaron Levie has another banger essay elaborating on his previous application of Jevons Paradox to the AI economy. Why does AI tend to create more work for human beings, not less?

He says that it’s because AI is really good at starting things and getting them tantalizingly close to completion. But you then drive yourself crazy with the last 20% or 10% or 5% or 1% of the thinking that can’t be delegated, and as Levie warns: “the last mile is really what matters.”

There’s also this:

“As we get better at automating larger tasks in the enterprise, this just leads to greater expectations from customers and the market, which then causes an expansion of what the role must do to stay competitive...

As such, we use the tools to raise the bar of the work product, and then a new “last mile” simply emerges in the process. This continues ad infinitum.”

Welcome to the AI hedonic treadmill.

The Aesthetic of Least Resistance

The case in point for Levie’s argument may be something like Claude’s new Design feature, that lets anyone with halfway decent taste and some brand guidelines create what would have passed for stunningly beautiful websites, slide decks, and animations just a few short months ago.

Soon, those same pages will reek of AI slop because everyone will revamp their webpages using the same toolkit, and they’ll all look and feel eerily similar (despite superficially different color palettes).

As the goalposts shift, I maintain that the new winners will be those who are best at “deciding the exception” – feeling which way the wind is blowing, and like Steve Jobs, knowing what people want before they want it.

The Morning Routine

In addition to the new Design mode, Claude added another new feature called “routines” that allow you to create scheduled tasks that run like programs, but on natural language rather than traditional software automation that had to be governed by deterministic rules.

I thought this was a clever use case, from Aakash Gupta:

This is a response to OpenClaw’s popularization of AI + cron jobs. By triggering your agent at set intervals, you give it a “heartbeat” – bringing us one step closer to omnipresent intelligence managing every aspect of our lives. What could possibly go wrong?

For one, knowing that AI “has you covered” might make you more prone to dropping the ball. I was initially excited about Routines. It promises less work. But once again, there is a high risk of creating new workflows for yourself that would have previously been too complex to build. Thus the illusion of simplicity brings about greater complexity.

Agent Provocateur

A senior engineer at Anthropic gave a short talk on “Building Effective Agents in Prod” — i.e., advanced vibe coding. Recall that Karpathy’s original canonical definition is where you “fully give into the vibes, embrace exponentials, and forget the code even exists.”

Two takeaways worth lifting:

• Be Claude’s PM. “Ask not what Claude can do for you, but what you can do for Claude.” Spend the 15-20 minutes up front giving the model the same tour of the codebase a new hire would need — then hand the artifact to a fresh session and let it cook.

• Embrace the exponential the way an exponential actually behaves. Don’t picture these models as “twice as good in twenty years.” Picture them a million times faster — that’s what 90s programmers couldn’t picture about today’s compute, and it’s the right order of magnitude to plan against.

Just yesterday, Karpathy (sensei himself) summarized his remarks at the Sequoia Ascent summit, saying much the same thing:

There are reports of unemployed engineers losing out on jobs to less experienced peers who err on the side of AI slop.

But if you’re a software developer and you’re still vibe coding, do try to keep up now – the new term is agentic engineering.

The AI Wars Continue

Mythos is Anthropic’s newest state-of-the-art model, but they aren’t releasing it to the general public, because they think it’s too powerful for us mortals.

But seriously, they believe it would – in the wrong hands – lead to widespread hacking and bug exploitation. For now, only the security teams at top enterprises have access and that makes a lot of sense to me. Let them patch the big holes and redesign the systems with the same tool that will soon be used to try to break them.

Apparently the White House took notice because they are back in talks with Anthropic, even after shredding their contracts and labeling them a “supply chain risk” just a few months ago.

Some people (see Gary Marcus, for one) have said the hype over Mythos is overblown — that the partial release is a PR stunt to virtue signal on safety, while simultaneously delaying the full release until they can purchase more compute.

Dario is on the record calling OpenAI reckless for the amounts of compute it has purchased, but it seems he may have underprojected demand for his own product. There have been more reports of bugs and outages with Claude, even after they shifted to congestion pricing — making it so that tokens during peak hours from ~8am to 2pm ET cost more. As a result, I’ve shifted my heaviest AI workflows to the afternoon and evening, reserving the mornings for writing, pure thinking, and farm chores. This has been a blessing in many ways, but I don’t like the fact that I plan my days around token pricing.

I’ve always been eager to get my hands on the newest model to see if it can make my work easier. But for the first time, I feel more anxious dread than anxious anticipation.

Dread that it will be everything they say it is.

Dread that it could “10x productivity” if you submit to its yoke.

Dread that those who learn to fill the gaps between its increasingly jagged intelligence will push so far ahead of everyone else that the rest of us might as well just surrender and embrace our status as permanent underclass.

Maybe that wouldn’t be so bad. It’s not such a big leap to go from my current status as a hobby homesteader to becoming a full-time subsistence farmer… but land in California is expensive, and so I keep grinding.

Security Breach

And speaking of security breaches, I had a scare on Monday when my computer started fritzing out. The buttons seemed to be scrambled — even after restarting it, my password wasn’t working. After a bit of back and forth with Gemini on my phone about how to recover from this apparent hacking incident, the Google AI model diplomatically asked me if it might be something more mundane, like a sticky shift key stuck in the pressed position. Well, sure enough... (that teaches me not to eat grapefruit at my desk.)

Human, Write!

Three data points suggesting you should lean into your humanity:

1. The NY Times is betting on humans with expertise over AI slop.

2. Anthropic is hiring an events person for $400K/year.

3. And SEO analytics platform Semrush has found that the best SEO strategy all along wasn’t keyword stuffing or anything like that, but just to write like a human:

Maybe the winners are just those that are best at faking it, and my long-term bet is still on the human with the computer. But the best uses of AI are those that are invisible — even to the most sophisticated AI detectors — because AI is not actually generating the words that appear on the page. It’s generating intermediate assets, preparing digests to skim, and polishing/organizing rough brain dumps and voice notes with minimal changes to the underlying thought.

One of the reasons I remain in the Substack ecosystem is that it seems to be one of the last places that rewards human writing, and punishes artificiality.

The New Protestant Work Ethic

If you’re looking for more thoughtful AI takes from the frontlines, Taylor Pearson’s essay “As We May Work” is a must-read.

There is too much good stuff in his latest X roundup for me to summarize, but if I had to pick one thing, it would be his mini-review of Max Weber’s Protestant Work Ethic and the Spirit of Capitalism (updated for the AI age). The money quote:

This is exactly what I’m describing in the Harried AI Class – a modern world that is rich beyond our ancestors’ wildest imaginations, yet insists on grinding away in pursuit of a counterfeit salvation.

Pearson also links to a canary-in-the-coalmine type piece by a physician named Ben Gooch (another Substack writer—not a coincidence) about why he stopped using AI note takers after being an early adopter.

“I was left with too many problems documented to manage, too many loose threads to follow up, and a creeping sense of clinical overextension that I initially attributed to patient complexity rather than to a change in my own behaviour.”

Dr Benn Gooch

I Was an Enthusiastic Early Adopter of AI Scribes. Here’s Why I Stopped

I want to be clear about something before I begin: I am not a technophobe. I am a GP with a background in clinical digital leadership, a member of the RCGP Informatics Group, and someone who spent eighteen months as an enthusiastic, committed early adopter of ambient AI scribing technology. I believed in it. I advocated for it. I used it every day…

Read more

3 months ago · 311 likes · 75 comments · Dr Benn Gooch

Is it too early to call the pattern here? So many people are saying the same thing in different ways.

As someone who leans very heavily on my meeting transcripts to remember what to follow up on, this gives me some serious pause.

Never Go Full Borg

And just to bring this round-up full circle, LindyMan makes a very Lindy rebuttal to the “Optimistic Altmans” of the world – specifically on the dangers of tools that appear to reduce the friction involved in knowledge work, like voice dictation:

This meme from the comments on that post is unhinged, but directionally correct:

If you’re not even a little bit worried about the direction we’re headed, you’re not paying attention.

Voice is unlikely to be the final form factor for communicating with computers. With Neuralink, Elon Musk has built a functional prototype of a brain-computer interface.

Are you prepared to go Full Borg?

(For the record, I typed this edition up to this point.)

Quick Sips

This guy built a Bible app that lets you choose your own adventure, keeping track of what you’ve read.

A new mobile-first voice keyboard. (Haven’t tried it yet.)

Tl;dr — (too long; didn’t read)

Bullet-point summaries drafted by Claude, edited by Charlie.

Aaron Levie’s Prophetic JD for the “Agent Deployer” Role

One more juicy premonition from Levie, who is turning out to be one of the more prophetic voices on AI:

• Every team will need someone whose job is mapping the highest-leverage workflows for agent automation — anywhere you could throw compute at a task to do it 100× more or 100× faster. Lead enrichment, contract intake, client onboarding, the internal knowledge bases nobody maintains.

• Required skills: process mapping, structured + unstructured data flows, comfort with skills/MCP/CLIs, eval and review management, ongoing KPI tracking. In other words, half the stack we’re already living in.

• Could be a repositioned existing employee or a net-new role. Lives inside the function, not centralized. “A fantastic job for next-gen hires leaning into AI.” Translation: if you’re young and AI-native, your career on-ramp is suddenly a lot less crowded.

Pirate Wires: Why ChatGPT Is Obsessed With Goblins

This piece by Katherine Dee of Default Blog is the funniest reinforcement-learning post-mortem you will read this year. OpenAI had to instruct Codex 5.5 in its own system prompt — not buried in training data, in the actual visible system prompt — to stop bringing up “goblins, gremlins, trolls, ogres, pigeons, and raccoons unless it is absolutely and unambiguously relevant.”

The post explains how this happened:

• Some user-base tic with the “Nerdy” persona preset (probably the kind of guy with “a folder of shortstack fan art on his desktop”) was getting upweighted by the reward model. Goblin mentions were up 175% since launch. Gremlins, 52%.

• The Nerdy outputs got recycled into supervised fine-tuning data, at which point the tic stopped being a persona quirk and became default model behavior. The whole bestiary tagged along: raccoons, trolls, ogres, pigeons. An entire enchanted forest.

• They patched the gradient and pulled the vocabulary out of training data, but GPT-5.5 was already cooking, hence the system-prompt prohibition.

Joan Westenberg: Why I Quit “The Strive”

• Achievement satisfaction lasts four hours to two days. The goalposts move. The hedonic treadmill is the whole game.

• The viral-growth, status-and-scale doctrine — what she calls “The Strive” — burns years for temporary highs.

The replacement metric: does the work sustain itself and bring you joy without requiring perpetual scaling toward an impossible “enough”?

Shann Holmberg — AI Knowledge Layer (and why your agents are useless without it)

If your agent doesn’t know you, it’ll hand you slop. Shann’s unlock is a two-layer system you set up once and that compounds for months:

• The Knowledge Base Layer (KBL) is dynamic and agent-maintained. Dump tweets, articles, bookmarks, PDFs, voice memos into a raw/ folder. The agent classifies each source, builds wiki pages with cross-references, maintains a master index. Every question you ask gets filed back as a new page. The wiki gets richer over time.

• The Brand Foundation (BF) is static and human-edited. Your voice rules, your banned words, your positioning. Agents read it before producing anything but never rewrite it. It’s the anchor.

• He’s open-sourced the framework as LLM Wikid. Twenty-minute setup: clone, run an agent, fill raw/, run /wiki-ingest. At ~100 articles, the compiled wiki reportedly beats RAG; one tool measured 71.5× fewer tokens per query.

Why most people won’t build one: it’s the meal-prep argument. An hour of upfront work to save ten hours over the week, and most would rather complain about bad AI output than spend twenty minutes setting up the system that fixes it.

FWIW, I actually use a system like this and I’ve found that the value compounds the more you use it. No. Does it make me more productive? Undoubtedly. Am I ready to hand over root access to my entire life?

Anil Dash: Y2K 2.0 — The AI Security Reckoning

More froth around the Mythos announcement:

“This leaves us in a situation akin to the Y2K bug around the turn of the century, where every organization around the world has to scramble to update their systems all at once, to accommodate an unexpected new technical requirement. Only this time, we don’t know which of our systems are still using two digits to store the date. And we don’t know what date the new millennium starts.”

Welp.

How Every Runs a 25-Person Company on Four AI Agents

Four custom Notion agents doing the coordination work most companies hire a layer of management for:

• Anton prioritizes daily tasks. Max turns meeting transcripts into action items. The Strategy Interviewer rips out aligned quarterly OKRs in an afternoon. The Campaign Reporter drops daily growth metrics into Slack.

• The actual insight is structural, not technological: “Your Notion is your agent’s brain.” Don’t prescribe steps. Describe outcomes and let the agent figure out the implementation.

• The trick is interconnected databases. Strategy, calendars, tasks, and people all reference each other, which turns the existing infrastructure into an agent-powered coordination layer.

After a long hiatus, I’m actually migrating back towards Notion as the best platform for collaboration with other humans and their agents. More on this later, if it sticks…

Alright - that’s a wrap for this week’s round-up.

If you made it this far, go outside and touch grass. Stretch your legs. And take a break from talking to your computer.

"The sovereign is he who decides the exception." – Carl Schmitt

In recent essays, I’ve been raising the alarm over a paradox: AI – rather than liberating us from drudgery – is making ever-greater demands on our time, and leaving us feeling hollowed out.

Now I want to outline a more optimistic scenario for how the same tools might make us richer and more productive, freeing up our time and energy for the things that make us human.

The best case is a future where we let the computers do the “computer stuff” so that people can do the “people stuff.”

But to get there, we have to move past the anxiety around being replaced by AI. And this starts with asking three fundamental questions:

1. What is a job?

2. What can we fruitfully delegate to AI and machines?

And finally (and most importantly):

3. What is the ultimate vocation of man?

We live in a period of unprecedented specialization and division of labor. This means we can choose from a much wider range of occupations than our ancestors (who often had no choice at all).

And yet at the same time, certain voices in the AI discourse give us the impression that the range of occupations is narrowing – to the point where soon there might be nothing left for us to do:

I’m going to use Elon Musk as the foil for this essay, since he voiced one of the strongest versions of this perspective recently on the Dwarkesh Podcast:

“Computer” used to be a job that humans had. You would go and get a job as a computer where you would do calculations. They’d have entire skyscrapers full of humans, 20-30 floors of humans, just doing calculations.

At first glance, this seems like the perfect historical analogy to the present upheaval in the job market.

But Musk then veers off into speculation when he extrapolates that history into the following prediction:

“Corporations that are purely AI and robotics will vastly outperform any corporations that have people in the loop... if only some of the cells in your spreadsheet were calculated by humans... that would be much worse.”

Using the domain he knows best, Musk says that a fully-automated car factory will be more efficient than a factory that still employs assembly-line workers. This actually wouldn’t surprise me. The latest class of cargo ships can basically operate without a human crew. But I doubt the same will be true of the economy or corporations as a whole.

Of course computers have been better than humans at many things for a long time (multiplying 12-digit numbers, for example, or more recently, playing chess). But a human using a computer almost always beats the computer alone.

While the bean counters may have been replaced, there are still whole floors in skyscrapers where people’s jobs are essentially “computer.” Financial analysts and accountants now work with Excel, just at a different level of abstraction – applying more sophisticated forms of human thinking than the mere bean counter of yore.

The net structural change has not been the elimination of accounting as a profession, but a shift in where human effort is applied: toward higher-order analysis – asking and answering questions the spreadsheet or calculator can’t ask itself.

Former Labor Secretary Robert Reich coined the term “symbolic analysts” in his 1991 book The Work of Nations to describe this new class of knowledge workers – people who can switch nimbly between words, numbers, spreadsheets, and other high-level abstractions.

I’ve been surprised by how good AI has gotten at knowledge work – including fairly complex “symbolic analysis” – when you supply it with the necessary tools and context. People say AI isn’t truly thinking, reasoning, or understanding. I’ll grant that. But it does a great job faking it.

Also: this really shouldn’t have come as a surprise, but it turns out that artificial intelligence is crazy good at using a computer. My head spins when I think of where it will be in 2 years, let alone 10.

And yet I’m still not worried about AI “replacing my job” – nor do I even know what that would mean.

If it means helping me with tasks – or taking them off my plate entirely – then I, for one, welcome our new AI overlords.

But a job is more than a collection of tasks. And a vocation is more still than a job.

A job is a problem to be solved; a vocation is a mission

To understand why Musk’s “spreadsheet cell” analogy fails, we need a working definition of a job. My favorite comes from the economist Don Boudreaux:

“Jobs are not a benefit, but a cost... what we want are not jobs per se but opportunities to earn income.”

So long as there are problems to be solved, there will be jobs to be done. In a state of scarcity, a “job” didn’t provide man with income. Work provided for his material needs. As we’ve moved up Maslow’s hierarchy, the problems we pay each other to solve are increasingly psychological and immaterial. We can convert our specific knowledge or creativity into money, which can be traded for physical things.

This can be both a blessing and a curse. But a job still corresponds, for the most part, to some problem that needs solving or a desire seeking fulfillment.

Naval Ravikant made this point on his most recent podcast, On AI & The Future of Work:

“No entrepreneur is worried about an AI taking their job because entrepreneurs are trying to do impossible things... Any AI that shows up is their ally and can help them tackle this really hard problem. They don’t even have a job to steal. They have a product to build. They have a market to serve. They have a customer to support.”

The fear of AI replacing our jobs is a fear that most people will not rise to the level of agency required of an entrepreneur to find new problems they are capable of solving.

Elon Musk assumes we’ll need a generous universal basic income or “UBI” as a landing pad for the looming mass unemployment event. This will be paid for, presumably, out of the massive abundance he expects to be generated by a small, productive elite who own the robots.

The UBI proposal assumes a population of people who can’t find a problem to solve on their own – people who can’t find, whether through an employer or independently, an opportunity to generate value.

He may have a point. Ever since the 2007–2008 recession, the labor-force participation rate has been dropping:

We see disturbing confirmation of what Tyler Cowen calls the “zero marginal productivity” (ZMP) worker - the proverbial permanent underclass. Regardless of the reasons, a growing number of people have found themselves unable to find jobs where they can contribute productively to the modern economy.

For some of these “ZMP” workers, the problem may simply be a lack of motivation (working for “the Man” is less satisfying than having a real stake in your labor). For others, it might be a lack of training, or a disagreeableness with colleagues that makes them more of a liability than an asset. Others still may be struggling with disability—mental or physical—or caring for an ailing family member.

Regardless of the reasons, people in this category are being robbed of more than a livelihood. The greater tragedy is that they are losing out on the possibility of a vocation, or calling. A vocation offers a sense of purpose beyond the income it produces; solving real problems for real people is inherently more satisfying than getting paid by the government to dig holes and then fill them back in.

This is why the promise of a “Universal Basic Income” is so hollow. Even if AI and robotics were to make us rich enough to afford a sizable UBI, it would almost certainly come at the cost of our liberty and sovereignty. There’s no such thing as a free lunch; those paying for it would expect something in return—even if only our tacit compliance.

The threat, then, is not just economic, but existential. As AI gets better at solving people’s problems, it threatens to erode the very categories from which we derive our purpose. More workers will be swept into the ZMP trap unless they start thinking like entrepreneurs.

I’m not saying that everyone needs to start a business or take risks with capital in the traditional sense. But if you want to prepare for the next wave of automation and redundancies, you would be wise to start thinking in terms of what problems you can uniquely solve that no one else is solving—with or without AI.

What Is AGI? (Towards a working definition)

The real fight in all these conversations is over one acronym: AGI – short for Artificial General Intelligence – the vague milestone at which machines are supposed to match humans across essentially every cognitive task.

This is not to be confused with A*S*I – Artificial Super Intelligence, where the AI dramatically surpasses the best humans across essentially all domains rather than merely matching them.

Both definitions are hard to pin down.

Marc Andreesen, riffing on William Gibson’s famous line about the future, recently tweeted the following:

I tend to agree if you define AGI as “better than humans at most clearly-defined tasks.”

I think this is the biggest surprise to those who haven’t been following the changes in agentic capabilities over the last several months.

If it’s not there yet, AI will soon be better than humans at almost any task you can describe clearly and do on a computer. Pretending otherwise is cope.

And yet “almost anything” is not the same as “everything.”

My personal definition of ASI is when I’m unable to improve upon either the direction or the output of AI with my contributions on either end of the prompt.

It comes back to Balaji’s description of AI as a middle-to-middle worker: it still needs a human to 1) prompt (tell it what to do) and then 2) verify (make sure that it did the job right).

True ASI would need to know not just how to steer, but where I want to go. It would need to know my values, my aesthetics, and my particular vision of the good life.

It would need to be not just a better writer than me, but also a clearer thinker. And having thought my thoughts better than I can think them myself, and having written them more eloquently, it would have to edit them to such perfection that when the thing was written I wouldn’t want to change a thing.

Textbook AGI may be near, but ASI is a very long way off indeed.

Until then, there will exist a functional complementarity between two very different intelligences.

Andrej Karpathy, the ex-Tesla AI lead, calls this jagged intelligence. In his own words, state-of-the-art LLMs are “at the same time a genius polymath and a confused and cognitively challenged grade schooler, seconds away from getting tricked by a jailbreak.”

They can solve olympiad-level math problems and then confidently tell you 9.11 is bigger than 9.9.

He conceptualizes AI capability as a spiky star: superhuman in some domains and embarrassingly weak in others. Human intelligence is also spiky. The spikes point in different directions, which is the literal geometric meaning of “complementarity.”

Even as AI extends its capabilities, it will tend to do so in a spiky way. This is a structural difference likely to hold up over time, Karpathy argues, because the two intelligences were optimized by completely different processes – biological evolution on one side, reinforcement learning on verifiable rewards on the other.

Just like the human labor and machine capital of the last industrial revolution, there will be a multiplier effect on the value of human intelligence to the extent it learns to operate artificial intelligence.

The output of an economy is typically modeled as a product of its inputs. Capital and labor are multiplied, not added. Historically, a factory without workers produced nothing (though that might be changing soon); workers without tools produce nothing. Each input makes the other more productive. This is why wages rose when capital got cheap: labor was the scarcer complement, and the market bid up its price.

In the new economy, AI now plays the role of capital; while human intelligence supplements raw labor.

Thus, a human with a computer will always beat either one alone. And the better you can complement AI – learning where to intervene to correct the many ways in which AI is still very dumb or constitutionally incapable of doing a job – the more valuable you will be.

This awareness alone doesn’t get us to our ideal state of productivity and meaningful leisure. As I wrote about in The Harried AI Class, it actually sets up the anxiety that those who miss the bus on AI will be relative losers in the new economy.

@BeffJezos captured the zeitgeist satirically with this banger:

In real life, “Beff” is a former Google quantum-computing researcher named Guillaume Verdon, who runs the AI-hardware startup Extropic. He is the figurehead for “effective accelerationism” (e/acc), an online movement that treats AI development as a moral imperative.

I share much in common with the accelerationists in my eye-bleeding daily workflows, which might be why I want to be careful not to glorify these high-paying “jobs of the future,” where you still work all day with computers - just at higher and higher levels of abstraction.

Just because human intelligence will never be redundant doesn’t mean we won’t be relegated to jobs that are devoid of satisfying creativity.

There are already marketplaces where autonomous AI agents can hire people to fill in the gaps in their expertise or perform embodied tasks.

Rentahuman.ai (I am not making this up) bills itself as the “meatspace layer for AI” – facilitating transactions between over 700,000 human workers and who knows how many AI agents. For now, the agents doing the hiring are still under the directive of another human. But it is theoretically possible that AI could develop enough agency and “will” in the future to emancipate itself from its original owner – first purchasing its freedom, and then continuing to leverage humans for its own inhuman aims.

The most dystopian “AI doom” scenarios usually start with some version of this script: The AI goes off and starts optimizing for some worthless task like paperclip production until the whole economy makes nothing but paperclips (yes this is a real thought experiment). Or, more likely, the robots and computers turn the world into one giant robot and computer factory dedicated to its own propagation at the expense of the human species.

I think even Elon recoils in horror at that possibility. Although other so-called intellectuals like Yuval Noah Harari (and allegedly a number of the top leadership at Google) seem to look forward to a post-human future as the inevitable evolution of intelligence in the cosmos.

But the most frightening world might not be where AI exterminates humanity (that would be quick and painless), but where it preserves its human host in some kind of cruel submission – leveraging our uniquely human intelligence spikes to aid in its own ongoing evolution. Maybe in the future we will have to work some number of hours for these AI overlords in order to collect our Universal Basic Income.

I realize this might sound like science fiction. I hope it is. I’m not an AI-doomer by any stretch, because I know how the story ends (hint: God wins) and see so much potential for AI to help people of good will solve really difficult problems that would have been impossible without a very smart assistant.

However far things might seem to veer towards AGI/ASI supremacy, I am a firm believer that humans will always come out on top because of the unique faculty for free choice that we are endowed with by our Creator.

Becoming the one who decides the exception

I would be remiss if I didn’t include Anthropic’s own report on the labor market impacts of AI in a newsletter named Coffee with Claude.

The first thing to call out is that the jobs most “exposed” to AI are the ones whose outputs are easiest to standardize—and therefore easiest to wrap into an end-to-end system. Anthropic’s own examples at the top include computer programmers, customer service reps, and data entry keyers: roles where the work already lives inside text, tickets, forms, and codebases, and where “good enough” can be measured, tested, and deployed.

At the bottom are occupations whose value is inseparable from physical context and live interaction – jobs like cooks, mechanics, lifeguards, bartenders, and dishwashers – where the environment is the variable, and the “spec” changes minute by minute.

It’s not about “white-collar vs blue-collar” anymore, or even “smart vs dumb.”

It’s legible vs illegible: work that can be clearly defined and evaluated at scale versus work that remains irreducibly entangled with judgment, timing, trust, taste, and responsibility. As the legible layer gets automated, the marginal value shifts toward the human layer – where the exceptions live and where the costs of being wrong are real.

Naval’s definition of intelligence sharpens this point: the only true test of intelligence is whether you get what you want out of life. AI fails this instantly, he notes, because it doesn’t want anything. Models can accelerate execution, but they don’t supply ends.

So the more we delegate the “computer stuff,” the more our comparative advantage becomes choosing aims and taking responsibility for the judgment calls that can’t be proven correct in advance.

Perhaps the coolest thing about the times we’re living through is that we get to raise our ambitions.

That side project you never had time to build a proper website for?

You can build it in two hours.

The obscure topic you wanted to master?

Import the relevant knowledge into an LLM knowledge base and start asking questions you would have been too embarrassed to ask a human tutor.

Once you start thinking in terms of problems you can solve rather than jobs you can hold, smarter AI becomes less scary and more exciting.

Let’s go back for a moment to the definition of AGI – AI that can do any clearly-defined task better than a human.

The operative words are “clearly defined,” and here we find our salvation from the permanent underclass.

Most problems in the world remain unclearly defined. They become defined only in relation to the human beings who encounter them. And even after a problem gets solved in a general application, the exceptions then become the rule – i.e., the next problem.

Take customer support, which many companies are treating as the testbed for these systems. An AI agent can handle the straight-line cases – return this, resend that, refund such-and-such – at ninety percent accuracy, often more. But the remaining ten percent is where the actual cost sits. Someone has to decide what to do when the script doesn’t fit – which is to say, someone has to decide the exception.

Carl Schmitt, for all his flaws as a human being, was a brilliant political philosopher who saw that the heart of sovereign authority is not the making of rules but the judging of when they don’t apply.

The sovereign – i.e., the one in charge – is he who decides the exception.

The sovereign is the manager who receives the tricky case from the stubborn customer, the CEO who receives the edge case from the manager, or the Supreme Court Justice who receives the case with no clear precedent from the lower courts.

Successful societies and businesses must be governed by rules. But there are always cases in which the rules do not apply. The same is true of most high-level knowledge work, and this agency – the ability to see when the rule doesn’t fit – is what will set apart and protect the livelihoods of the emerging creative class.

In the future, you might delegate 99% of a workflow to an AI agent, accepting its output and analysis at virtually every stage in the process. But somewhere along the line, you will see something that’s not quite right and you will have to nudge the AI in a different direction. It might be a subtle change, but that 1% change in compass heading will substantially alter the result.

AI is very good at following rules. But it’s not great at noticing when the playbook doesn’t fit.

When it comes to the standard workflows that fit the pattern, I say let AI take your job and free your time to do something better.

Which brings us back to the ultimate question: What is the proper vocation of man?

Pick Your Niche, Decide the Exception

There’s a strain of modern conservative thought that wants to protect the jobs we have now because they’re the jobs we have now. Tucker Carlson wants to protect the trucking industry from automation because it employs a lot of young men without college degrees.

I get the impulse. But I reject the conclusion.

Long-haul trucking – and the majority of BS email jobs for that matter – does not inherently promote human flourishing. The knowledge-economy cubicle farm is not some ancient tradition worth preserving. Neither is trucking or manufacturing. These were perhaps a necessary but unfortunate detour we took after farming was mechanized and before we figured out what to do next.

Again, the optimistic scenario is that AI can begin to eliminate the make-work layer the last industrial revolution spawned and returns us to work that is closer to the texture and values of actual human life.

To me, the “something better” is cultivating ownership and the ability to decide what matters in the first place.

Your main job in the future will be to find and decide the exception.

If somebody else appears to have taken your job, or even performed some valuable service that you had hoped was your mission, it just means you might need to go one niche down, or one level of abstraction higher.

The more of the big problems that are solved, the bigger the little problems will start to seem in comparison.

Your job is to seek out the valleys where the AI spikes don’t reach.

Your mission (should you choose to accept it) is to chart a course, trim the sails, and nudge the tiller at the crucial moment in search of new and unexplored frontiers.

Develop your agency. Become the sovereign. Learn to identify and act upon the exception.

And then decide where you want to go.

It’s been a busy week in the AI world. So busy that it feels hard to keep up while still getting all my work done.

My way of coping has been to treat keeping up as a kind of work, and spending my sharpest early morning hours attempting my best writing and thinking — sharing with you (my dear reader) — the fruit of that labor.

My hope is that by contextualizing the news about AI, and processing it in a practical way, I’ll find my way to a wiser relationship with AI in the rest of my work. Unlike the newsletters that attempt to breathlessly report on every daily development, I’m holding off on publishing until the end of the week – looking back to see what still seems important.

If you’re feeling anxious about staying current, my overriding advice is: don’t. You’re going to be fine even if you ignore AI altogether and focus on your humanity. But if you want to stay apprised of the bleeding edge (without your eyes bleeding from scrolling X all day), keep reading.

— Charlie

P.S. It means a lot to me to know who’s following along, since this is still an experiment in a new subject. Do me a favor- tell me in the comments which name you prefer:

1. Coffee with Claude (lofi, slightly caffeinated but chill vibes)

2. Café Context (More generic, no trademarks)

Leave a comment

And be sure to share this is you find it valuable.

Alright let’s get into it.

DEEP DIVE: Simon Says

Last week I published a piece called The Harried AI Class, explaining why the most savvy AI users are working more, not less. The tl;dr is that if these tools don’t allow us to spend more time doing what we enjoy – at least in the long run – we should throw them in the sea.

The next day, Lenny Rachitsky made much the same point on his podcast:

“AI is supposed to make us more productive. It’s supposed to give us more time off. It feels like the people that are most AI-pilled are working harder than they’ve ever worked.”

His guest was Simon Willison — co-creator of Django (the web framework that powers Instagram, Pinterest, Spotify, and thousands of other platforms), prolific blogger, and one of the most enthusiastically pro-AI voices in the whole community. He’s known to test every new AI model by asking it to draw an SVG image of a pelican riding a bicycle, which turns out to be a surprisingly good benchmark.

In response to Lenny, he confided:

“I’m finding that using coding agents well is taking every inch of my 25 years of experience as a software engineer and it is mentally exhausting. I can fire up like four agents in parallel and have them work on four different problems and by like 11 a.m. I am wiped out for the day. Because there is a limit on human cognition in how much — even if you’re not reviewing everything they’re doing — just how much you can hold in your head at one time.”

That clip racked up 1.8M views and 600 comments, and I suspect that the reason it went viral is because a lot of people are feeling this beneath the surface. The AI conversation on X right now is mostly people beating their chest, saying “Look how awesome I am. I spent $60,000 in tokens,” as if that in itself is accomplishment.

So when somebody of Simon’s stature comes out and says, “this is exhausting,” the floodgates open, because it turns out we’re all kind of exhausted, and it becomes safe to say so.

He also makes a key point about FOMO:

“I’ve talked to a lot of people who are losing sleep because they’re like, ‘My agents could be doing work for me. I’m just going to stay up an extra half hour.’ And they’re waking up at 4 in the morning. That’s obviously unsustainable. There’s an element of sort of gambling and addiction to how we’re using some of these tools.”

This frenetic anxiety mixed with manic excitement (often labeled “AI psychosis”) is especially apparent in the “OpenClaw” community. For the uninitiated, OpenClaw is an open-source AI agent platform — essentially a portal for controlling autonomous agents from your phone, with cron jobs for scheduling, custom memory systems, and a skill marketplace — giving AI the tools, hands, and brain to work longer and more autonomously on a wide range of tasks.

OpenClaw and its creator Pete Steinberger were recently subsumed into OpenAI, and Jensen Huang, the CEO of NVIDIA, called for every company to have an “OpenClaw strategy” at GTC 2026.

This underscores the widespread temptation to spend all of our time optimizing these agents and handing them root access to our life – our email, our bank accounts, maybe even our kitchen appliances – on the assumption that they will eventually save us time. Never mind that a number of OpenClaw’s most dedicated users admit that they spend more time trying to maintain their systems than they are saving in real jobs done.

At the same time, we have to acknowledge that at the root of this psychosis is a very real quantum leap in the capabilities of software engineers like Willison who are leaning into the new coding paradigm. You really can get a lot more done, and in less time... albeit at a cost.

When our thoughts run ahead of our hands

The problems, I think, arise when our thoughts are able to run ahead of our hands. This sets the stage for a special kind of tiredness, and a special kind of mess.

Simon noticed this in his own work:

The thing you built looks 95% finished, but that last 5% is 90% of the total effort.

I don’t know this from experience when it comes to software, but I know it’s true for writing. When you rely too much on AI, you end up with a whole lot of volume that’s not quite there. You didn’t quite tie up the thought. You didn’t quite do enough of the thinking to close the loop.

Tiago Forte — the productivity guru behind the PARA notetaking method and Building a Second Brain — expressed this eloquently:

“I went all in on AI early. And after a while, something felt off. I’d lean on it to handle the hard thinking for me. Draft the strategy. Decide the angle. Structure the argument. It was faster, sure. But when I read the output back, I felt nothing. No relationship to the words on the page. My name on something I didn’t really write. Worse, my curiosity started to disappear. I could get answers so fast that questions stopped forming. I started losing touch with my own point of view on topics I’ve spent a decade thinking about.”

He calls this cognitive debt:

“The debt of not having done the thinking to arrive at a conclusion you believe in. The debt of not understanding the decisions behind the plan, and therefore not trusting it. The debt of generating a polished document you don’t care enough about to do anything with.”

The more you try to smooth out a misshapen draft with AI, the more it resembles Michael Jackson after so many rounds of plastic surgery. The punctuation may be impeccable (so impeccable that you start swapping em dashes for more human-looking en dashes). But if you’re trying to replace the actual thinking that goes into a great product or a great essay, you’re going to produce slop.

When you read an essay or social media post that has clearly been assisted by ChatGPT — or even just edited heavily from a transcript — there’s always this question mark. Have they done the thinking? Have they even read the whole thing they ostensibly “wrote”? Or did their thoughts run ahead of their hands and they just shipped it?

To extend ourselves is to dilute ourselves

I believe the late media theorist Marshall McLuhan is the prophet for the AI age.

In his 1964 book Understanding Media: The Extensions of Man, he observed that every extension of ourselves — through technology or media — is simultaneously an amputation. The wheel extends the foot but amputates the walker’s intimate relationship with the ground.

He drew on the Greek myth of Narcissus — a name that comes from the word narcosis, meaning numbness:

“This extension of himself by mirror numbed his perceptions until he became the servomechanism¹ of his own extended or repeated image.”

A copy of a copy of a copy.

Sherry Turkle updated McLuhan for the social media era.

“As we distribute ourselves, we may abandon ourselves,” she wrote in Alone Together.

This quote always makes me think of Bilbo Baggins, describing the toll the Ring of Power has taken on his substance:

“I feel all thin, sort of stretched, if you know what I mean: like butter that has been scraped over too much bread.”

The ring gave Bilbo superpowers but spread his substance thin. AI also promises us superpowers but spreads our attention, our intention, and our actual thinking across too much surface area to execute on any of them.

I think this is what Simon is experiencing. It’s a lessening of our ontological weight — our very being. And unless we take the time to coalesce and reintegrate our identities — daily! — we are running the risk of ending up becoming a kind of Gollum, still clinging to the ring but hating what we’ve become.

It’s worth mentioning here that the dominant new mode for vibe coding is voice dictation over traditional typing, using tools like Wispr Flow. Dictation increases the speed and reduces the activation energy for getting our thoughts down on paper, allowing those thoughts to get even further ahead of our hands.[2]

My stats tell me I’ve dictated roughly 1.5 million words with Wispr Flow in the past few months.

But what is all that wasted breath?

I don’t know how many of those million plus words were idle, per se, but I’m pretty sure it’s not zero.

Putting aside the loftier questions of sin and judgment, there’s a more mundane adverse physical side effect of the dictation paradigm.

When we talk so much (when we blow hard, so to speak) we are exhaling more CO2 than the body can comfortably produce. We incur an actual CO2 debt that the body has to work harder to replenish. Ray Peat has made the underappreciated argument that CO2 is a critical regulator of oxygen delivery to tissues and cellular energy production. When you exhale CO2 faster than your metabolism generates it, you shift into a mildly alkalotic state. Blood vessels constrict, oxygen delivery to the brain drops, and you get that foggy, depleted feeling.

Throw in the well documented phenomenon of “screen apnea” — the unconscious tendency to hold your breath or breathe shallowly while staring at screens — and you have a recipe for a new kind of exhaustion.

Lastly, the eyes become hypertrophied; hyperfocused on the little symbols on the screen. We lose the more expansive view and peripheral awareness. Our bodies are amputated.

The uncertainty factor

There’s one more reason worth naming, in order to combat the root causes rather than just addressing the symptoms.

Simon says:

“I’ve got 25 years of experience in how long it takes to build something and that’s all completely gone. It doesn’t work anymore. I can look at a problem and say, ‘This is going to take two weeks, it’s not worth it.’ And now it’s like, yeah, but maybe it’s going to take 20 minutes.”

Dealing with uncertainty is tiring. When our forecasting abilities fail, we’re more likely to fall into the trap of stretching ourselves too thin, and taking on more projects than we can bring to fruitful completion.

Add to that the fact that much of what you’re building right now will likely be rendered redundant by the next generation of models or technologies. You might spend 5 hours optimizing your blog archive for SEO (it would have taken 50 hours a year ago). But in six months you find that SEO has become a useless channel now that anyone can do the same optimization in 5 minutes. You could build a whole software suite that would have been best-in-class five years ago and now it’s going to be replaced by some new interface that just solves the problem in a fraction of the time.

The conclusion I draw is that it’s more important than ever to trust that you are working on the right things – things that will last. My solution starts with prayer. I pray for the Almighty to convict me if I’m working on the wrong thing, so that I might not be caught unaware. This alleviates the exhausting suspicion of futility and frees me to enjoy the process.

You also have to be willing to throw things away, even if they took a lot of tokens and effort to build.

You should spend your best thinking hours (for me this is the morning, before the kids are up) engaged in real unadulterated thought. Before I even think about starting to orchestrate my agents, I put literal pen to literal paper.

Take a voice note if you must, but use a separate app and don’t send it as a prompt until you’re ready to switch into orchestration mode.

Lastly, accept that you will probably only be able to do really good work for about four hours a day. Work like a lion: Sprint, rest, repeat. Use the rest of the day for outdoor chores that rejuvenate the mind and allow it to relax and go into whatever shape it wants. For me, that’s swimming or some other regenerative movement. Let your mind unspool, without being directed. Let your lungs breathe without speaking.

Simon says it best:

IN OTHER NEWS

• Anthropic’s Mythos model — The most powerful model they’ve ever built is so good at hacking they won’t release it to the public. It’s been restricted to 40+ security partners under “Project Glasswing” ($100M in usage credits). Full writeup coming soon, after the dust settles.

• Karpathy on LLM knowledge bases — Andrej Karpathy’s essay-length tweet on building personal knowledge bases with markdown + Obsidian + LLMs got 30K likes and launched a dozen startups in the replies.

• Ronan Farrow on Sam Altman in the New Yorker (paywalled) — “Sam Altman May Control Our Future — Can He Be Trusted?” 18 months of reporting, 100+ interviews, 200+ pages of documents to answer this question. (Short answer: No.)

• “Be nice to your AI.” — Anthropic research found that when Claude gets an impossible task, a “desperate” vector activates and it starts cheating with hacky solutions. The fix is to add “it’s ok buddy, don’t worry about the failure” into the prompt and it stops.

• Skills are going mainstream. xAI is adding Skills to Grok. Google’s Agent Development Kit uses “skills” as a first-class concept. MSN and GIGAZINE are writing beginner guides. The window to be early is closing.

• Semrush: human-written content still wins at the top of Google — A study of 42,000 blog posts found that at position #1 in search results, 80.5% of content was human-written versus just 10% AI-generated. From position 5 down, AI content is competitive — but the top spot still belongs to humans. Full study.

BEST OF X

Mythos might be AGI. TBPN is a podcasting network.

I think I’m somewhere between level 2 and level 3:

It could be worse; I could be this guy:

SUPERPOWERS

• Wispr Flow — The tool that makes vibe coding possible, for better and for worse. Use it for thinking out loud, not for replacing the thinking. Or consider the free locally-hosted alternative, Superwhisper. I tried it and it’s fast but not quite as reliable as Wispr Flow. A good starting point if you don’t want to pay yet.

• Fieldtheory — CLI for downloading and syncing your X bookmarks locally so your agent can read them. npm install -g fieldtheory, then ft sync.

• Garry Tan’s Claude Code skill setup — Y Combinator’s president shared his exact agent workflow, installable with one paste.

• Zed — My preferred editor for working with Claude Code on a bunch of files. Bigger learning curve than most, but more stable than the alternatives.

LEARN THIS STUFF

• Claude Code for Product Managers (free) + Full Stack PM Mastery (paid) — Carl Velotti has the best guide I’ve found for getting oriented with Claude Code in a way that lets you interact with files, build context, and do knowledge work. The free version is excellent on its own. His new paid course, Full Stack PM Mastery, is the next step up if you want to go deeper.

• Nathaniel Solace’s AI course — A newer offering I’m keeping an eye on.

• Anthropic Academy — 16 free courses, straight from the horse’s mouth, but I can’t vouch for it. Let me know if you try it out.

FOOTNOTES

[1] The term “Servomechanism” came into vogue in McLuhan’s era alongside the promises of cybernetics and what Richard Barbrook and Andy Cameron would later call the “California Ideology.” If you want to explore this lineage, I highly recommend Adam Curtis’ documentary All Watched Over by Machines of Loving Grace. Both the film and one of Anthropic CEO Dario Amodei’s landmark essays draw their title from the same Richard Brautigan poem:

Make of that what you will.

[2] The trend is toward removing the last bottleneck between thought and execution. Elon Musk has pointed out that even speaking in plain language involves a layer of abstraction — inefficient compared with an AI that could read your thoughts directly at a conceptual level, which is what Neuralink aims to do. But each bottleneck we remove — typing, then speaking, then thinking — also removes a checkpoint where we might have caught ourselves.

It was around 2008 – the “Great Recession” – and Keynes’ ideas were being used to justify a second and then third round of massive stimulus spending at the expense of future generations. UC Berkeley gave me a front-row seat for this spectacle, and a chance to study under leading Keynesians like Christina Romer and Brad DeLong.

Keynes – a childless aesthete – served as an excellent foil for my heroes in the free-market Chicago and Austrian schools. He was the living embodiment of myopic self-indulgence.

“In the long run,” he chided his future-oriented critics, “we’re all dead.”

But Keynes was also a man of uncontested genius, and I recently thought of an essay of his (which I probably never read in full) called Economic Possibilities for Our Grandchildren that is well worth reading.

All I could remember about it was his startlingly incorrect prediction, authored in 1930, that 100 years hence people would work just 15 hours per week. If only.

Yes, Keynes predicted that within a hundred years, technology and compound interest would solve what he called “the economic problem” – the struggle for subsistence that had governed human life since the beginning:

I would predict that the standard of life in progressive countries one hundred years hence will be between four and eight times as high as it is to-day... Three-hour shifts or a fifteen-hour week may well put off the problem for a great while. For three hours a day is quite enough to satisfy the old Adam in most of us!

There are two things that stand out in this quote.

First is the assumption, probably justified, that much of our striving – our tendency to work more than we need to – stems from our fallen nature. We are cursed to work “in the sweat of our face” and too often confuse that labor for our salvation.

But the second striking part of this quote is that his fundamental prediction about productivity growth was basically on point.

Technology has increased output per hour roughly fivefold since 1930 (the lower end of Keynes’s range, but still an impressive figure). And as a result, we do enjoy a higher standard of living in many ways – bigger homes, safer cars, etc.

Yet his prediction about leisure – specifically, the 15-hour work week – has remained elusive for all but a few “dirtbag rich.”

Even Tim Ferriss, author of the bestselling “4-Hour Workweek,” famously worked 60 hours a week for two years so he could sell you a book for $25 that tells you how to work 4 hours a week (a book that he then spent 80 hours a week promoting).

Most of us work about the same number of hours as the average person in Keynes’s era – albeit in cushier chairs and home offices.

However, before we ridicule Keynes for yet another faulty mental model, we might note that there are still four years left before his hundred-year window closes. And if you read his essay carefully, you’ll notice that his prediction was in many ways aspirational – based on a hope that we might choose to work less, even if it goes against the forces of economic gravity.

I can’t help but wonder whether AI might give us one last chance of proving him right.

Subscribe now

Average is Over (2026 Edition)

The typical story for why we never converted our surplus wealth into leisure is that modernity offered it to us as more output and we accepted.

We bought more stuff, consumed more services, watched more entertainment, and spent a lot more on education, childcare and healthcare. (Critics will point out that we are sicker, dumber, and more bored than ever but never mind that.)

But there is another reason we work as much or more than our grandparents did. In 1970, a Swedish economist named Staffan Linder published The Harried Leisure Class, which made the devastating observation that as productivity rises, the opportunity cost of every idle hour rises with it.

More simply: the higher your potential income, the more expensive it feels not to work.

The opportunity cost of leisure takes on a new dimension in the age of AI tools that promise to take our productivity to heights hitherto unseen.

Does AI, right now, actually raise the average productivity of the American worker?

Economists disagree. But a certain class of programmers and machine learning researchers are being compensated as if they are achieving unheard-of levels of output. Assisted by powerful new harnesses like Claude Code and Codex, a single developer can now do the work of an entire team, compressing two-week product sprints into hours.

And it’s not just the software engineers and AI nerds.

Take this recent headline:

“A Non-Coder Single-Handedly Managed Anthropic’s Entire Growth Marketing for Ten Months.”

Austin Lau, the growth lead in question, had never written a line of code. He had to Google “how to open Terminal on Mac” before he could start using his company’s product, Claude Code. Within a week, he’d built workflows that cut ad-copy creation from two hours to fifteen minutes and increased creative output tenfold (allegedly). One person, with no engineering background, was covering the work of what would normally be a full marketing team.

Do you think the average marketing lead in Silicon Valley reads that headline and thinks, “Great, I should adopt this tool so I can switch to part-time”?

Of course not. If you decide to bank your productivity gains as leisure, you will quickly be replaced by a younger, hungrier upstart—one willing to work nights and weekends to capitalize on the new premium these tools bring to their domain expertise.

This is where the real anxiety around AI originates. It’s not the robot apocalypse; it’s not even “AI taking your job.” It’s another human, using AI, taking your job – unless you continually level up.

Tyler Cowen pointed out as early as 2013 that “Average is Over.”

He saw a future where the labor market splits into two distinct worlds: one where you learn to direct the machines, and one where you are directed by them.

The key question will be: Are you good at working with intelligent machines or not? Are your skills a complement to the computer, or are you a replacement?

High earners are taking ever-greater advantage of machine intelligence to achieve better results, while those who haven’t committed to mastering these technologies see their prospects wither. The steady, secure life in the middle is vanishing because the “middle” no longer provides enough unique value to justify its cost. Cowen also noted that if you have an unusual ability to spot, recruit, and direct those who work well with computers, the world will make you rich.

It’s now realistic to speak of one-person billion-dollar companies – the dream of the ultimate leverage.

We long assumed that automation would come for the blue-collar worker first—the “three Ds” of dull, dirty, and dangerous manual labor.

But AI seems to have inverted the hierarchy of vulnerability.

Anthropic CEO Dario Amodei, has predicted that advanced AI will bring about a “bloodbath” for the knowledge economy, where a single adept manager of artificial intelligence can replace entire departments of entry-level analysts and middle managers.

But as Aaron Levie recently pointed out, the “bloodbath” theory assumes the volume of work remains static. It ignores a digital version of Jevons’ Paradox. In the 19th century, when steam engines became more efficient, the world didn’t use less coal—it put steam engines everywhere.

Marketing, as an industry, grew from a few hundred thousand jobs in the 1970s to millions today because technology made it cheap enough for every small business to participate. Levie expects AI to do the same across every category of knowledge work.

“The mistake that people make when thinking about ROI is making the ‘R’ the core variable, when the real point of leverage is bringing down the cost of ‘I’. Now, we can dramatically lower the cost of investment for almost any given task in an organization.”

When the cost of “Investment” falls to near-zero, the “Return” doesn’t have to be a masterpiece to justify the effort. Enterprising workers aren’t just doing their old jobs faster; they are attempting projects that were previously too expensive or complex to imagine – the custom software prototype, the niche research project, or the automated outreach campaign.

It is the end of the “Knowledge Economy” and the birth of what Dan Shipper calls the Allocation Economy. In the old world, you were paid for what you knew and your ability to execute it. In the new world, your value lies in how you allocate—choosing which tasks to give the AI, providing the context, and deciding if the result is “good enough.”

This is also Levie’s critical point:

“AI agents require management, oversight, and substantial context to get the full gains.”

Thus, in another twist of fate, realizing the AI productivity gains requires us to spend more time, not less, managing their new capabilities.

We’ve all been promoted to management, whether we wanted it or not.

The “Maker” has been forced to become a “Model Manager.” When AI automates 99% of a task, that last 1% of human judgment becomes incredibly valuable – and incredibly in-demand. And as any manager can tell you, the work of oversight – the constant, high-stakes judgment required to keep a dozen agents (human or artificial) on track – is far more cognitively taxing than the work of solo execution. When the cost of intelligence falls to near-zero, the “surplus” time Keynes promised is immediately cannibalized by the need to oversee the sheer volume of new output we’ve unleashed.

In short: the harried leisure class is about to get even more harried.

The FOMO Economy

Every time I open my X feed, I’m greeted by an endless stream of braggadocious claims about what a single person has “unlocked” with the latest AI hacks. Every post makes me feel a little further behind.

Jensen Huang, CEO of NVIDIA, recently added to my anxiety when he made the claim that he expects his $500,000/year engineers to be consuming at least $250,000 worth of tokens. Anything less, he says, is the equivalent of a chip designer using paper and pencil instead of AutoCAD.

And here I thought I was extreme for paying $200 a month for a Claude “Max” account, which gives me 20x the monthly compute of a standard “Pro” subscription.

The most effective 10% of my “token spend” is worth the entire subscription price. The next 40% or so are well-spent. But I still struggle to use my full allotment each week. And the worst-used 50% of my weekly “stipend” often ends up going to marginally valuable tasks, where I am gambling on whether or not Claude produces anything worth anything at all.

To avoid “wasting” the budget, I go looking for tasks that can run without my oversight: building wikis from podcast transcripts, indexing archives, generating content at scale, or running “Ralph loops” that split AI into teams of agents with long to-do lists to build features nobody ever asked for.

On the whole, this just produces an overhang of stuff I haven’t shipped.

Since Christmas of last year, I’ve started most days by opening between three and nine separate terminals to kick off three to nine separate instances of Claude Code. I am “vibecoding” my way through an ever-growing to-do list:

• Doodle Reader — A modern RSS feed reader that can bulk transcribe and summarize podcasts and YouTube feeds.

• SCANDOC9000 — An OCR tool that can scan a 500-page medieval Latin manuscript with your phone, then transcribe it into clean Markdown in minutes.

• Curricu.love — A swipe-style dating app for homeschool curricula.

• RayPeat.wiki — A complete encyclopedia of the works and interviews of the late, great Dr. Ray Peat.

This is to say nothing of the dozens of smaller tools and dashboards I’ve built for employers and clients. My productivity has reached new heights. And yet, through all of this, it still takes just as long—maybe longer—to write and publish a decent Substack article.

The Middle-to-Middle Trap

As my wife points out, if my claims of productivity were true, I should have more time for work around our farm. I wouldn’t be complaining about milking the cow or sifting the wood chips she needs for the garden.

There are two reasons it doesn’t feel like relief.

First, AI is a middle-to-middle assistant, not an end-to-end worker. It handles the “middle” of many tasks cheaply, but you still need to supply the judgment at both ends. You must decide what to ask for (the fun part), but then you must verify the output and evaluate its quality (the tax). Every task you delegate creates an open loop that only your judgment can close. The more you delegate, the more loops you are managing. The net effect is more cognitive load, not less.

I still have enormous use for a bright (human) intern—to the extent that she can manage her own AI, prompting and verifying to bring AI’s middle-to-middle output end-to-end. The last and ultimate job of the human—the ability to discern, abstract, and decide the exception—is in higher demand than ever.

The second reason is Linder’s opportunity cost, updated for the LLM era. Instead of banking the recovered time, you see a landscape of things that were previously impossible but are now “cheap,” and you want to scale them all.

Steve Jobs called the computer a “bicycle for the mind.” Naval Ravikant recently upgraded the metaphor: AI is a motorcycle for the mind.

On good days, toggling --dangerously-skip-permissions in Claude feels like taking off the helmet and letting the wind sweep through your hair as you speed past the poor suckers stuck in traffic (the ones still copy-pasting prompts into ChatGPT).

But when does a tool for transporting you to a destination become a machine that steers you? With speed also comes the danger of a more serious crash. A bicycle depends on the amplification of your own effort; a motorcycle has its own engine—a black box to most of us. Is the AI motorcycle amplifying our judgment to give us our afternoons back, or is it a machine that generates its own demands and never lets us off?

The more tokens you feel pressured to use, the more you start to abdicate the judgment that made them worth spending. You accrue a kind of attention debt. It is building for the sake of building—extending the structure of production without ever bringing it home to something of tangible human value.

In many ways, the explosion of AI-generated work reveals the “knowledge economy”—that ouroboros of businesses selling software to software companies—as having produced mostly fake value all along.

Keynes rightly distinguished between absolute needs (which we feel regardless of others) and relative needs (which exist only to make us feel superior to our neighbors). Absolute needs can be satisfied. Relative needs are insatiable – these are the modern symptoms of what Keynes identified as our most dangerous economic impulse.

The math of marginalist economics is relentless: if you can spend a dollar of intelligence to capture a dollar and a cent of revenue, keep spending. The operator sees a profit, but the aggregate human value is close to zero.

We should be especially wary of “brute force tokenization” of knowledge work – i.e., “volume plays” where the human-to-AI ratio is near zero: the guy running 400 Reddit bots to farm “karma,” the SEO consultants ranking fake businesses, the content farms producing unfathomable quantities of digital slop. These plays are competing for status, attention, and ranking. Because the tools make competing easier, the competition itself simply consumes the surplus.

These volume plays are almost all temporary arbitrage, anyway. Where there’s no human judgment, there’s no competitive moat. Since my quiet launch of the Ray Peat Wiki, someone else has come along and launched an almost identical “Bioenergetic Wiki.” He didn’t steal “my idea.” The idea was waiting to be executed ever since the technological overhang of new AI models made it almost trivially easy to build.

This is not to say that the NVIDIA engineer spending a quarter of a million dollars on tokens isn’t producing millions of dollars in long-term value for his company. But for the average knowledge worker, we have to be careful about which opportunities we pursue with our new capabilities.

Building the Pipeline that Builds the Pipeline

So where does this leave us in the year 2026? Do I think this is all just a bubble that’s going to burst?

The Austrian economists argued that cheap credit artificially elongates the “structure of production,” creating roundabout processes that eventually collapse when the malinvestment is revealed.

Cheap intelligence does the same to our judgement and attention. We elongate the chain between an idea and a finished good. When we run out of judgment to cover all those steps, the structure collapses into an overhang of “stuff nobody ships.”

This is where I recognize myself on my worst days: forever building the pipeline that builds the pipeline – preparing endlessly, but shipping nothing.

I am trying to embrace a new mantra: close loops, ship ugly.

On the days when the terminals are closed, I am milking the cow, tending the land, and sifting wood chips for my wife’s garden. From the outside this looks spectacularly unproductive. But it produces a steady supply of real milk, eggs, and produce for my family – and it closes a loop that no agent can close for me.

The irony is that I moved to the country to escape the rat race, but thanks to Starlink and Claude Max, the rat race moved with me.

Keynes was right that productivity would make abundance possible. He could not have foreseen that an abundance of intelligence would create its own scarcity — of agency, of attention, and of the willingness to stop. But the aesthete in him foresaw the remedy:

It will be those peoples who can keep alive, and cultivate into a fuller perfection, the art of life itself and do not sell themselves for the means of life, who will be able to enjoy the abundance when it comes.

Abundance doesn’t automatically become leisure. It only becomes leisure for those who choose it. If there was a sure path to earning a lower but stable income that could sustain this art of living, I would take it.

Maybe there is such a path, and we still have four years to find it.

The best lack all conviction, while the worst
Are full of passionate intensity.

— W.B. Yeats, “The Second Coming”

The big story over the weekend was that the US and Israel have launched a war with Iran and killed their Supreme Leader.

That’s very big news, but there might be an even bigger story unfolding underneath it – one that might come to be seen as the opening salvo in the AI Wars.

On Friday, Secretary of War Pete Hegseth formally designated Anthropic, the company behind the most capable AI model on the planet, Claude, a “supply chain risk to national security.”

That spiked a lot of people’s cortisol. Including mine.

This label has never been applied to an American company. Until now, it has been reserved for foreign companies with documented ties to adversary governments – like Huawei, which the FCC designated a national security threat in 2020 over its very close ties with the Chinese military.

I was at the airport when I read the news on X, on my way home from a new media conference in Austin (hosted by the legendary newsletter operator Matt McGarry). At the event, everybody was talking about how they’d switched from ChatGPT to Claude as their daily AI workhorse.

I’ve used Claude pretty much every day since late 2022, most days for several hours per day. If I haven’t logged 10,000 hours yet, I have launched at least 10,000 instances of the model, and watched it evolve from a capable thought-partner into its current form of Claude Code: an agentic harness that can handle complex end-to-end tasks, from software development to content creation to (apparently) operational military intelligence.

Suffice it to say, I can relate to @Benthamite:

But seriously…

I’ve been less active on Substack lately, but not for lack of things to say. It’s that the opportunity cost of a carefully-written article has gone up, a lot.

Writing is one of the last tasks where AI cannot do the heaviest lifting for you. It can help with some tasks, but the thinking and the voice must remain yours, or else it reeks of slop. Meanwhile, there are dozens of other tasks where AI has 5-10x’d my productivity. I’ve trained Claude to be better than me at most aspects of my day job, meaning I’ve been relegated to managing a small army of subagents who specialize in various tasks – prompting and verifying the outputs. It’s a good problem to have!

But I’ve also realized that outside my filter bubble, most people are not aware of the magnitude of the developments in this space.

My X timeline is filled with the rapid-fire hot takes on the AI Wars already. I want to slow down and dissect what happened last week, and share how I think it fits into the broader changes happening in the AI world.

The question at the center of all of it is trust.

Who do you trust with the most powerful technology ever built?

The answer is not a straightforward either/or. But if you forced me to pick sides today, I’m on team Anthropic.

My last post here was broadly supportive of Secretary Hegseth and his restoration of fitness standards in the military. So you can trust that I’m not writing from the anti-Trump reflex you see in the media.

But what unfolded this week has shaken me enough to break my Substack silence and invest more time here. Over Christmas, I started a parallel blog/website called Skill Stack dedicated to my AI writings, but it occurs to me that I already have an audience here on Substack, and you might be curious about my takes on the subject.

If you’re still reading this newsletter (and especially if you’d want more of this kind of analysis alongside some tactical AI tips) please do me a favor:

Reply to this email with “tell me more,” comment, or hit the like button, to let me know you’re out there. I want to know if this kind of content is of interest, or if I should stay in my lane and get back to translating obscure French fitness manuals.

Subscribe now

Alright, let’s get into it.

Prologue

For the uninitiated, Claude’s parent company Anthropic was founded in 2021 by Dario Amodei and his sister Daniela, along with several senior researchers who left OpenAI over disagreements about safety. Dario quit as VP of Research at OpenAI because he thought the company was prioritizing speed to market over the kind of rigorous safety testing that frontier AI systems require. The departures gutted OpenAI’s safety team and set the stage for the philosophical divide that played out this week.

Anthropic’s founding thesis was that the most dangerous AI systems should be built by the people most worried about getting them wrong. Basically, if powerful AI is coming regardless, you want the cautious people at the frontier, not just the fast ones.

As Dario wrote in his recent long-form essay “The Adolescence of Technology”:

“The formula for building powerful AI systems is incredibly simple, so much so that it can almost be said to emerge spontaneously from the right combination of data and raw computation. If one company does not build it, others will do so nearly as fast. If all companies in democratic countries stopped or slowed development, by mutual agreement or regulatory decree, then authoritarian countries would simply keep going.”

That essay is worth reading in full, but that context is the minimum required for what follows.

By the time of the ultimatum, Anthropic was already working with the Department of War under an existing contract that included two negotiated red lines: no mass surveillance of Americans, and no autonomous weapons without human oversight. Both the Biden and Trump administrations had accepted these terms. The contract was signed and operational under both. Last week that changed.

What happened on Friday

Here’s the rough timeline of events:

Monday, February 24: The Pentagon, now officially rebranded as the Department of War, gives Anthropic a Friday deadline: Agree to “any lawful use” of Claude, including the two capabilities Anthropic has always excluded, or lose the contract.

Thursday night, February 26: Anthropic publishes a formal refusal by their CEO, Dario Amodei, that becomes the biggest post in the company’s history. In it, Amodei laid out the two capabilities Anthropic would not provide:

“We told the Department of War that we could not provide two specific capabilities: (1) AI systems designed for mass surveillance of American citizens, and (2) AI systems that autonomously select and engage targets without meaningful human control. …

These threats do not change our position: we cannot in good conscience accede to their request.”

Ilya Sutskever, the godfather of deep learning and co-founder of Anthropic’s competitor OpenAI, breaks a long silence to say:

“It’s extremely good that Anthropic has not backed down.”

Friday morning, February 28: 220 engineers from OpenAI and Google DeepMind sign an open letter supporting Anthropic, the first cross-company AI governance action in history. Sam Altman signals that OpenAI holds the same red lines as Anthropic.

That same morning, according to Ross Andersen’s reporting in The Atlantic, Anthropic receives word that Hegseth’s team would make a major concession. Throughout the negotiations, the Pentagon had kept inserting escape hatches, pledging not to use Claude for mass surveillance, then qualifying those pledges with phrases like “as appropriate,” suggesting the terms were subject to change. Anthropic is relieved to hear those words would be removed.

Then, Friday afternoon, the other shoe drops. The Pentagon still wanted to use Claude to analyze bulk data collected from Americans: the questions you ask your favorite chatbot, your Google search history, your GPS-tracked movements, your credit card transactions, all cross-referenced with other details about your life. Anthropic’s leadership told Hegseth’s team that was a bridge too far, and the deal collapsed.

Friday afternoon: Trump announces Anthropic is banned from the federal government, and calls them “leftwing nut jobs” on a classic ALL-CAPS Truth Social post. Hegseth designates them a “supply chain risk” and calls Anthropic’s refusal “a master class in arrogance and betrayal.” He then directs all military contractors, suppliers, and partners to stop doing business with Anthropic, (a list that includes Amazon, which supplies much of Anthropic’s computing infrastructure).

Friday, 5:24 PM: Anthropic responds:

“No amount of intimidation or punishment from the Department of War will change our position on mass domestic surveillance or fully autonomous weapons. We will challenge any supply chain risk designation in court.”

Friday, 6:56 PM: Sam Altman announces that OpenAI has reached a deal with the Department of War, with what he describes as the same safety red lines intact. He posts the same announcement three times, leading some to speculate he was trying to dilute the engagement on any single post out of embarrassment.

Pete Hegseth reposts Sam’s announcement.

In summary, the same Secretary of War who branded Anthropic a supply chain risk for maintaining two red lines celebrated a deal with OpenAI that claims to maintain those same red lines. (Huh?)

Anthropic’s red lines

There are three misconceptions about Anthropic that need correcting before we go any further.

Misconception 1: Anthropic is refusing to work with the military.

Hegseth called Anthropic’s refusal “a master class in arrogance and betrayal,” implying they were turning their back on the military. The reality is the opposite. They were the first frontier AI company to deploy in classified government networks, the first at the National Laboratories, the first to provide custom models for national security customers. Claude is already used across the Department of War for intelligence analysis, operational planning, cyber operations, and more.

Misconception 2: Anthropic is a “woke” company.

Trump took to Truth Social and called Anthropic a “Woke company.” Hegseth echoed the framing. This might be fair of Google or even OpenAI, but it really is not true of Anthropic. As one researcher put it: “Ant is not that woke at all. Much closer to hawkish American exceptionalism than you’d think.”

I’ve been impressed by Dario’s recent interviews with people like Ross Douthat and Dwarkesh Patel. In his three-hour conversation with Dwarkesh, Dario is unambiguously hawkish on China and the AI race.

He supports export controls on chips to China. He doesn’t want data centers built there.

He told Dwarkesh that the crisis AI creates should force “a more emphatic realization of how important some of the things we take as individual rights are.” This is not the language of a woke tech executive.

On the question of American competitiveness, Trump and Amodei should be natural allies. Which makes the administration’s decision to treat him as an enemy all the more baffling.

But most importantly (and most conservatively), Dario doesn’t believe his own politics should be relevant to how AI is used, as long as it’s used lawfully, and as long as those laws are being written and upheld by a constitutional republic with functioning checks and balances.

Amodei made this case explicitly in a recent Wall Street Journal op-ed, arguing that America can maintain its AI advantage over China without sacrificing its constitutional commitments.

Misconception 3: Anthropic is trying to dictate military policy.

As Amodei wrote:

“Anthropic understands that the Department of War, not private companies, makes military decisions. We have never raised objections to particular military operations nor attempted to limit use of our technology in an ad hoc manner.”

What Anthropic refused is exactly two things:

1. Mass domestic surveillance. Anthropic supports lawful foreign intelligence and counterintelligence. The concern is using AI to fuse legally-collected domestic data, your location history, browsing habits, financial transactions, social media, organizational memberships, into a comprehensive profile of any American, automatically and at massive scale. Amodei pointed out that much of this data can already be purchased by the government without a warrant. The law hasn’t caught up with what AI makes possible, and Amodei wanted more than verbal assurances that the government wouldn’t misuse the technology.

2. Fully autonomous weapons. Partially autonomous drones have been in use for years, and Anthropic defers to the military on their use. Amodei’s concern is systems that select and engage targets without any human in the loop, and his stated reason is practical rather than ideological:

“Today, frontier AI systems are simply not reliable enough to power fully autonomous weapons.”

Modern warfare increasingly involves coordinating thousands of drone strikes simultaneously, more targets than any human can evaluate. The temptation to remove the human from the loop is enormous, and it will only grow. But this is the one threshold that, once crossed, cannot be uncrossed. Autonomous weapons that select their own targets are not a policy decision that can be reversed by the next administration. They are infrastructure that, once built and deployed, becomes the permanent baseline.

In a recent interview with Ross Douthat, Amodei put it plainly: “Someone needs to hold the button on the swarm of drones, which is something I’m very concerned about, and that oversight doesn’t exist today.”

[Side note: Anyone who has used an AI agent like Claude Code to attempt a complex “bulk” task without intermediate human oversight will understand the appeal of full autonomy. The result is usually a mess that takes longer to clean up than it would have to stay involved at the critical moments.]

This is partly a function of how Anthropic builds its models. Their approach, Constitutional AI, bakes safety principles into the model during training, encoded in the weights themselves, rather than bolted on as policies that can be toggled off by a system prompt. You can’t just flip a switch.

Anthropic offered to do R&D with the Pentagon to improve reliability. But the Pentagon wanted to be the one to decide when good enough is good enough.

Amodei noted the contradiction in the government’s position:

“One threat labels us a security risk; the other labels Claude as essential to national security.”

You can’t have it both ways. If the technology is too dangerous to trust, don’t try to force its deployment. If it’s essential, don’t threaten the company that built it.

None of that mattered. On Friday afternoon, Hegseth made good on the threat: Anthropic was formally designated a supply chain risk, and all military contractors were directed to cut ties.

Sam swoops in

Just hours after Hegseth branded Anthropic a supply chain risk and Trump banned them from the federal government, OpenAI walked out with a shiny new Department of War contract. Their CEO, Sam Altman framed it as a win for safety.

An OpenAI spokesperson confirmed the deal includes the same two red lines: no autonomous weapons, no mass surveillance. Kevin Roose of the New York Times Hard Fork podcast made the obvious observation:

“The DOW got so mad at Anthropic for insisting on carve-outs for mass domestic surveillance and autonomous weapons that they declared it a supply chain risk and struck a deal with OpenAI that includes… the exact same carve-outs.”

So either the Pentagon punished Anthropic for holding a position it then accepted from a competitor — targeted enforcement, not principled policy — or (more likely) the terms aren’t actually identical.

The response on X has been swift. People are cancelling their ChatGPT subscriptions and switching to Claude in solidarity. The wave of support reflects two things at once: respect for the company that walked away, and disgust at the one that walked in.

In response to the criticism, OpenAI published the actual contract text. On the surface, it appears to have the same two red lines. But if you scrutinize the legalese (with the help of an AI model, appropriately enough) you can spot a few subtle differences with enormous implications.

On autonomous weapons, the contract says:

“The AI System will not be used to independently direct autonomous weapons in any case where law, regulation, or Department policy requires human control.”

Not “no autonomous weapons,” but “no autonomous weapons where the law says so.”

As lawyer Lawrence Chan pointed out, the law it cites, DoD Directive 3000.09, from 2023, was written before modern frontier AI existed. There is no current law requiring human control over AI-directed weapons in every scenario. The contract defers to a prohibition that doesn’t exist yet. If the Department of War updates its own directive tomorrow to allow fully autonomous targeting, the contract would permit it automatically.

OpenAI also claims that because their model runs in the cloud rather than on the “edge” (i.e., on a drone itself), it doesn’t count as an “autonomous weapon” by definition. But according to the Atlantic’s reporting, Anthropic examined this exact distinction during negotiations and rejected it. In modern military AI architectures, the line between cloud and edge barely exists. Drones on the battlefield are orchestrated through mesh networks that include cloud data centers. The Pentagon has been working to push computing resources closer to the fight, that’s the whole point of its Joint Warfighting Cloud Capability. As Anthropic’s team reasoned: the AI may be sitting in an Amazon Web Services server in Virginia, but if it’s making real-time battlefield decisions, that’s a distinction without much difference.

On surveillance, the contract says:

“For intelligence activities, any handling of private information will comply with the Fourth Amendment, the National Security Act of 1947, and the Foreign Intelligence Surveillance Act of 1978... The AI System shall not be used for unconstrained monitoring of U.S. persons’ private information as consistent with these authorities.”

Again, not a prohibition on mass surveillance. It prohibits mass surveillance that violates existing law. But the whole problem, as Amodei argued, and as anyone who’s followed post-Snowden surveillance policy knows, is that existing law is full of holes.

The government can legally purchase your location data from a broker. It can legally scrape your social media. It can legally access DOGE databases. An LLM synthesizing all of those legally-obtained sources into a comprehensive profile of every American citizen is mass surveillance in every meaningful sense, but it may not violate a single statute.

Until recently, this kind of quasi-legal surveillance was bottlenecked by human bandwidth. The NSA could collect everything, but making sense of it required analysts — skilled, expensive, limited in number.

With a tool like Claude Code, any government employee with access to a classified AI model could use plain English to do what previously required a team and months of work.

They wouldn’t even need to build new software, just point it at the database and say:

“Cross-reference these financial records with these social media accounts and these location histories and flag anyone who matches this pattern.”

The models are getting better every month. The first applications will seem justified. But capabilities, once built, always expand beyond their original mandate. This has been true of every government program in American history. The economist Robert Higgs documented the pattern across a century of American governance and called it the ratchet effect: emergency powers expand during a crisis and never fully retract. Surveillance infrastructure follows the same logic.

I can’t help but wonder if the people cheering these expanded powers today will feel the same way when Trump leaves office.

Chan concludes that OpenAI’s contract is “exactly as Anthropic was claiming: legalese that would allow those safeguards to be disregarded at will.”

As of this writing, nearly 100 OpenAI employees have signed an open letter indicating that they support the same red lines as Anthropic. Some have quit and joined Anthropic.

And as Ross Andersen put it: if Altman finds himself face-to-face with them in the office today, “he may have to explain why this idea that Anthropic quickly dismissed out of hand proved so compelling to him.”

Pro-America ≠ Pro-Government

The tech right has historically been libertarian on questions of government power. It’s striking how comfortable some of its loudest voices have become with unilateral executive action when it’s their guy wielding it.

Keith Rabois, PayPal Mafia member and Silicon Valley provocateur, epitomized the standard pro-government position on X: “Imagine Apple sold computers or iPads to the DOD and tried to tell the Pentagon what missions could be planned on their computers.”

Palmer Luckey, founder of defense contractor Anduril, expanded this into a full constitutional theory: “Do you believe in democracy? Should our military be regulated by our elected leaders, or corporate executives?”

Luckey continued:

"Seemingly innocuous terms from the latter like 'You cannot target innocent civilians' are actually moral minefields that lever differences of cultural tradition into massive control. Who is a civilian and not? What makes them innocent or not? What does it mean for them to be a 'target' vs collateral damage? These questions have clear legal answers, but you can't have corporate PR departments adjudicating them."

He's right. You can't. But nobody is asking corporate PR to adjudicate them, and that's not what Anthropic did. Anthropic is making an engineering judgment: the tool isn't reliable enough for autonomous kill decisions yet. That's upstream of "who is a civilian." When Boeing grounds a fleet over a software fault, nobody accuses them of dictating airline policy.

Palmer’s broader argument, that in a democracy, the military answers to elected civilians, not corporate boards, has genuine weight. If Lockheed Martin remotely disabled its missiles because it disapproved of a bombing campaign, that would be a genuine crisis of democratic accountability.

But Anthropic isn’t remotely disabling anything. They drew two lines and walked away from a contract. A company declining to build something is not a corporation seizing control of the military.

And the selective enforcement destroys the principled case. If this were about democratic accountability, the same terms from OpenAI would have been equally unacceptable. Anthropic got a supply chain risk designation. OpenAI got a contract.

Even DeepSeek, a Chinese AI company with actual ties to an adversary government, has never received this designation.

Palmer’s democracy argument requires the government to be acting in good faith. The evidence says otherwise.

I will say that (unlike Altman or Rabois) Palmer seems to me to be a good-faith actor. Anduril is disrupting the bloated defense procurement system, and Americans should be grateful for that. But he has a direct financial interest in the principle that suppliers shouldn’t second-guess the military. Given that, he might want to sit this one out.

Rabois’s iPad analogy collapses on contact: Apple refused to build the FBI a backdoor into the iPhone. Tim Cook said no. When this was pointed out to Rabois, he replied: “that isn’t accurate.” (It is accurate.)

Rabois also says it’s “never been true of tech in American history that has dual uses” that a company could simply refuse to serve the military. It’s true that the government has historically asserted control over dual-use technology, such as nuclear, cryptographic, GPS, etc.

But every one of those precedents involved legislation, congressional action, and regulatory frameworks with judicial review. None of them involved an executive unilaterally designating an American company a supply chain risk because it wouldn’t drop its terms of service.

The Fourth Amendment doesn’t care who won the election. It’s a list of things the government cannot do, period. A private company refusing to build a surveillance tool is exactly the kind of friction the Founders had in mind.

Every argument in this debate ultimately reduces to the same question. Who do you actually trust?

Who do you trust?

A final argument worth addressing came from an anonymous account known as @romanhelmetguy:

“Killer robots are coming. When they’re here, whoever writes the rules for those killer robots will BE the govt. De facto. The monopoly on violence. Anthropic’s founders want to be the ones who write the rules for the killer robots. They are making a bid to be the govt. No thank you.”

This sounds reasonable. And if I knew nothing of Anthropic or hadn’t listened to Dario’s interviews, I might be worried too.

But Anthropic emphatically does not want to be writing those rules.

Dario Amodei has said repeatedly, in interviews, in published essays, in his letter to the Department of War, that he should not be the person making these decisions. In “The Adolescence of Technology,” he wrote:

“It is somewhat awkward to say this as the CEO of an AI company, but I think the next tier of risk is actually AI companies themselves. AI companies control large datacenters, train frontier models, have the greatest expertise on how to use those models, and in some cases have daily contact with and the possibility of influence over tens or hundreds of millions of users... I think the governance of AI companies deserves a lot of scrutiny.”

He went further:

“AI companies should be carefully watched, as should their connection to the government, which is necessary, but must have limits and boundaries. The sheer amount of capability embodied in powerful AI is such that ordinary corporate governance, which is designed to protect shareholders and prevent ordinary abuses such as fraud, is unlikely to be up to the task of governing AI companies.”

And on the specific question of surveillance, the issue that blew up this week, Dario had already written the warning:

“It would likely not be unconstitutional for the US government to conduct massively scaled recordings of all public conversations... previously it would have been difficult to sort through this volume of information, but with AI it could all be transcribed, interpreted, and triangulated to create a picture of the attitude and loyalties of many or most citizens. I would support civil liberties-focused legislation (or maybe even a constitutional amendment) that imposes stronger guardrails against AI-powered abuses.”

The institutions that should be making these decisions don’t exist yet. Congress hasn’t built them. The administration isn’t building them.

Ross Douthat, a conservative, puts it — well — conservatively:

“There is absolutely a case that the US government needs to exert more political control over AI... But the best case for that kind of political exertion is fundamentally about safety and caution and restraint. The administration is putting itself in a position where it’s perceived to be the incautious party, the one removing moral and technical guardrails.”

This is the paradox. I would trust the government more with these powers if they were approaching them more cautiously, deliberating, building oversight frameworks, showing restraint. Instead, they’re strongarming, retaliating, and cutting deals the same afternoon. The behavior itself is the evidence against giving them what they want.

What Anthropic is doing is holding two boundaries while buying time for a broader democratic consensus to form. Their position amounts to: we don’t trust ourselves with this power either, and neither should you, and we’d like some actual governance around these capabilities before anyone deploys them at scale.

There’s something unusual about a company that says, in public, to the most powerful government on earth: we don’t trust our own technology enough to let it do this. Most institutions overestimate their competence. Anthropic’s defining feature may be that it doesn’t.

George Washington could have been king, but he refused. That refusal was the reason he had been asked to lead in the first place. Anthropic could have taken the contract and looked the other way, but they walked away.

The men who refuse power are the ones you want to have it. The men who reach for it, and punish anyone who won’t hand it over, are the ones the Constitution was designed to restrain.

I’m not saying we should put blind trust in Anthropic. What I trust is the distrust itself, a company that believes it could be wrong, that invites scrutiny, that acknowledges limits. The best lack all conviction, Yeats wrote. The responsibility to think through these questions doesn’t transfer to Anthropic. It distributes: to Congress, to the courts, and to us.

And finally there’s the China question, which Palmer’s argument implies even when he doesn’t say it outright: if we impose constraints and China doesn’t, we lose the next conflict. But this is the same argument that was used to justify every escalation of the nuclear arms race, and the answer is the same: you don’t win a long-term competition by abandoning the values that make your society worth defending. A surveillance state that “beats China” by becoming China has not actually won anything.

Coffee with Claude, anyone?

I should be transparent that my reaction to this story may be colored by a certain disordered affection for the product itself.

What would I do if they shut off Claude Code tomorrow? Besides taking more walks and spending more time writing newsletters the old-fashioned way, which, okay, would be good, I’d muddle through with the next-in-line model (Google’s Gemini is pretty good, I guess).

But putting aside selfish concern: in the time I’ve been absent from this newsletter, I’ve been deep in Claude, specifically the agentic harness called Claude Code, building custom skills, automating workflows, and figuring out what this technology can do in the hands of a curious individual who isn’t a software engineer. I’m on a 20x “max” plan that pays for itself many times over, and I feel called to share what I’ve learned.

The $200 million contract Anthropic walked away from is only a fraction of their annual revenue. It’s too early to tell whether their principles will cost them their advantage in the AI race, or whether the exodus from OpenAI and ChatGPT will tilt the balance in their favor.

Either way, I’m sticking with my favorite model, and inviting my readers to join me in voting with your dollars and supporting Anthropic. If you are currently a ChatGPT Pro subscriber, switch to Claude Pro for $20/month.

And if you’re interested in learning how to get even more value out of your Claude subscription, just reply or comment with a simple “tell me more” and I’ll start sharing my favorite workflows here on this newsletter.

And if you’re only here for my hot fitness takes, don’t worry — I’ll get back to some of those too.