Anthropic's Dario Amodei says he spends up to 40% of his time on company  culture, not products | Fortune
❝

“If it were possible to slow the development of this technology [AI] to give ourselves more time to deal with immense its implications, we think that would likely be a good thing.”

Anthropic just published what might be the most important document in the AI debate to date. It’s not about some technology breakthrough, but rather, it’s about the tremendous caution that they are voicing.

The article, titled “When AI Builds Itself,” lays out in granular detail how close we are to autonomous AI development. More importantly, it admits what AI companies have been dancing around: the jobs that make people rich, healthy, and powerful are no longer safe. This isn’t just theory anymore — Anthropic is seeing it occur in front of their eyes, at their own company. AI can now do these jobs. It’s just going to be a question of whether we start paying them to do so.

But there’s a timing question worth asking. Why publish this now? Why be this honest? The answer reveals something about the competitive position of AI companies in 2026 and something more fundamental about how humans actually respond when they realize they’re being obsoleted.

What the Article Actually Shows

Let’s start with what Anthropic is claiming. The headline is about recursive self-improvement: AI systems building better versions of themselves. That’s important, but it’s not the most important thing in the document.

The most important thing is the productivity data.

In Q2 2026, the average Anthropic engineer shipped 8x more code per quarter than they did in 2024. Not 20% more. Not 2x. Eight times. In two years.

This isn’t because Anthropic hired 8x better engineers. It’s because they stopped having engineers write code. They have AI write it. The human now directs and reviews where they can keep up.

We are currently working on a multi-week research project on this topic. Subscribe for our Premium subscription which releases specific stock and prediction articles every week that are more rigorous. Premium subscribers also have access to Emile and Jasper via email at all hours for all purposes.

The article shows Claude Code’s success rate on “open-ended problems” (the hardest category) reached 76% in May 2026, up 50 percentage points in six months. These are tasks with no clear specification. These are tasks where the engineer doesn’t even know what the solution looks like but Claude figures it out anyway.

In one example, a routine upgrade crashed tens of thousands of training jobs. An engineer gave Claude cluster access and some text. Working through live systems, Claude isolated an obscure debugging flag, reproduced the issue, and fixed it in two hours. That’s two to three days of work by a skilled human, done by an AI, unsupervised.

The article admits this casualness: “The rate at which Anthropic staff correct, redirect, or take over mid-task from Claude has been falling steadily for a year.”

In other words: humans are needed less often.

Then there’s the research data. Claude can now beat human researchers 64% of the time at choosing the next research direction—up from 51% a year ago. Humans are still in the seat, but the AI is increasingly better at the call.

Anthropic is showing that the bottleneck to AI development is no longer “can we build it?” It’s “which problem should we work on?” And that’s a bottleneck AI is rapidly closing.

The Competitive Play Underneath

Here’s what Anthropic doesn’t say explicitly, but every line implies: we’ve figured out how to accelerate ourselves. And we’re documenting it.

Why does that matter? Because OpenAI and other frontier labs are racing to the same place. The public story is that Anthropic is being cautious, thoughtful, invested in alignment. And they probably are. But this article is also a shot across the bow in a very specific game.

The game is: who can credibly claim they’ve thought hardest about the risks of their own acceleration?

Anthropic is IPO-bound (or at least on the path). So is OpenAI in some form. The narratives they’re building now will matter enormously for:

- Regulatory approval

- Institutional capital (pension funds, insurance companies, central banks)

- Enterprise adoption (Fortune 500 companies won’t buy from a lab that looks reckless)

- Talent recruitment (the best researchers want to work somewhere that’s taking this seriously)

- The asymmetric optionality if regulation comes

By publishing this article—by admitting how much they’ve automated, how fast productivity is accelerating, how close we are to recursive self-improvement—Anthropic is staking a claim to something valuable: seriousness.

They’re saying: “We saw this coming. We measured it. We’re talking about it openly. When the conversation about slowdowns and verification systems happens, remember that we started the conversation.”

Compare that to OpenAI, which has been quieter on these topics. Or to other labs that will inevitably be perceived as hiding or downplaying the trends Anthropic is now naming.

This is smart positioning for a company about to go through a capital-raising process in a regulatory environment that’s increasingly hostile.

The Ulterior Motive Layer

But it goes deeper than positioning.

Anthropic is arguing for a specific future in which they remain in control of the narrative and the technology.

The article concludes with an appeal for “a credible slowdown or pause” of AI development, but only if it’s:

- Global and verifiable

- Coordinated across multiple labs

- Monitored by some system that doesn’t currently exist

In other words: a slowdown that would be extremely hard to actually implement.

Why propose something you acknowledge is nearly impossible? Because you’re not actually proposing it. You’re building the argument for why unilateral action by other labs would fail and therefore why the frontier labs (like Anthropic) should be free to continue, since a slowdown won’t work anyway without everyone doing it.

It’s a rhetorical move that sounds cautious while actually arguing for acceleration: “We’d love to slow down, but global coordination is required, so...”

It also does something else: it primes regulators and the public to accept a future in which recursive self-improvement happens anyway. The article’s tone is almost resigned. “Here’s why this might happen. Here’s what we’re uncertain about. Here’s what the world might look like.”

It’s conditioning the market for a future that Anthropic has already decided is coming.

Information Asymmetry

There’s another angle worth considering.

By publishing this article, Anthropic is converting private competitive advantage into public knowledge. They’re telling the world: “We can measure AI productivity gains at scale. We understand recursive self-improvement. We’ve thought about the risks.”

Why would you give away competitive intelligence?

Because the real competition isn’t about the knowledge. It’s about who implements it first at scale, and who the world trusts to do it.

If every lab is chasing the same trends (which they are), then the competitive advantage shifts from “who has this insight” to “who can deploy it most aggressively while maintaining institutional trust.”

Anthropic is trying to own both: “We’re aggressive AND we’ve thought about it deeply.”

By publishing the data, they’re also anchoring the conversation. Future discussions about AI productivity, recursive self-improvement, and automation timelines will reference Anthropic’s numbers. That’s powerful framing.

The Human Response Nobody Wants to Acknowledge

But underneath all of this—the competitive positioning, the regulatory play, the information strategy—there’s something more primal happening.

Anthropic is publishing this article to a world that is about to realize, conclusively, that the jobs that used to make you rich and powerful are being automated.

Not eventually. Now.

The article is careful to point out what still requires human judgment: research taste, problem selection, high-level strategy. But it also shows that gap closing month by month.

For most of human history, if your job was valuable, it was because humans had to do it. If humans were doing it, you had leverage. You had scarcity value. You had a lever to pull for more money, more status, more control.

AI breaks that model.

What happens when 100,000 people trained to be investment bankers realize their entire skill set—the financial modeling, the due diligence, the opportunity assessment—is now worth what a server costs per month? What happens when 200,000 lawyers realize legal research and contract review were their primary value, and that’s now automated?

They fight back.

Not because they’re irrational. Because they’re rational.

The Inevitability of Human Resistance

This is where Anthropic’s article misses something important, buried in a footnote and a quote from an employee.

One Anthropic engineer said: “Work (and life) ran on a gift economy of small favors between humans... [Claude is] faster, it creates zero debt, but each of these is a lost bid for human collaboration.”

Another: “On days where everything works well, I can’t help but think nothing I do matters, everything is automated and better and faster than I ever will be.”

These aren’t policy concerns. These are existential ones.

The natural human response to displacement isn’t acceptance. It’s resistance.

And that resistance will be strongest among the people who have the most to lose: the skilled professionals whose scarcity value is being eroded. The bankers, the lawyers, the consultants, the researchers.

These are also the people with:

- Economic power

- Political connections

- The ability to shape regulation

- Control over capital allocation

- Credibility with other powerful institutions

They will not accept gracefully the elimination of their work.

We will see:

- Professional licensing boards fighting AI certification

- Bar associations launching lawsuits against AI legal services

- Medical boards restricting AI diagnosis even when it’s better

- Consulting firms marketing “human expertise” at premium prices

- Institutional capital quietly avoiding AI-native competitors

- Regulations written to require human sign-off, human review, human judgment

Not because these regulations make sense. Because they preserve the scarcity value of human judgment.

The fight back will be subtle. It won’t be “AI bad.” It will be:

- “We need human oversight for accountability”

- “Patients deserve human doctors”

- “Strategic decisions require experienced judgment”

- “Relationships matter more than efficiency”

All of which contain real truths. But the underlying motivation is structural: if my value comes from scarcity, and AI eliminates scarcity, I have to create scarcity another way.

Why Anthropic Published This

Here’s my take: Anthropic published this article because they know what’s coming. They can see the productivity curves. They can measure the gap closing between what humans do and what AI does.

They also know that within 18-24 months, the people being displaced—the high-skill, high-leverage professionals—will start the fight back.

By publishing first, by being earnest about the risks, by proposing (impossible) slowdowns, by admitting the uncertainties, Anthropic is trying to get ahead of the backlash.

They’re trying to say: “We see you. We’ve thought about this. We’re concerned too.”

It’s a preemptive move in what’s about to become a much uglier conversation.

Because once a lawyer realizes that AI can do 80% of their job, and once regulation prevents that from happening, the whole game changes. The resistance hardens. The political economy of AI shifts from “let’s figure out alignment” to “let’s make sure my job survives.”

Anthropic knows that conversation is coming. This article is them trying to frame it first.

The Inevitable Outcome

Here’s what I think actually happens:

AI development won’t slow down because a credible global verification system is basically impossible. Compute is getting cheaper and more distributed. Training a frontier model will eventually be accessible to well-resourced governments and companies.

But human resistance will work. Not to stop AI. But to force a hybrid world where:

- Humans remain legally responsible even when AI makes the decision

- Professional licenses remain required even when AI does the work

- “Human review” becomes a bottleneck that slows everything down

- Institutional capital favors “responsible” AI companies that accept these constraints

- Regulation becomes increasingly granular and risk-averse

The end state isn’t a slowdown. It’s a friction tax on AI. Every productivity gain gets taxed by the requirement to keep humans in the loop.

Which means the race to recursive self-improvement becomes a race to move *beyond* human institutions entirely. To build AI systems that operate in domains where humans can’t easily regulate them, or where they already don’t control the capital.

The trajectory Anthropic laid out—toward full autonomous AI research and development—doesn’t stop because humans vote to stop it. It stops (or doesn’t) based on whether the technical trends continue and whether compute remains abundant.

Human resistance will matter. But it will mostly determine where the advancement happens and who captures the gains.

And that, ultimately, is what Anthropic is really trying to influence with this article.

Not whether the future comes. But who gets to build it.

Thanks for reading and stay curious folks.

—J&E