Artificial (dumb?) Intelligence

Two weeks. That is the unit of time I keep coming back to.

Two weeks ago, Anthropic told the world it had built something too dangerous to ship. Claude Mythos Preview, a model that found a 27-year-old vulnerability in OpenBSD, zero-days in every major browser and operating system, and enough holes in financial infrastructure that the Bank of England held emergency sessions about it. A model Anthropic decided not to release.

The same company, in the same window, shipped its entire Claude Code source tree to npm by accident. 512,000 lines of TypeScript, a single missing .npmignore line, and within hours the whole roadmap was mirrored and forked.

Too dangerous to release. Released by mistake.

The Leak and the Roadmap

The Claude Code leak is now the most expensive packaging error in AI history. The source map was discovered by researcher Chaofan Shou and posted on X within hours of the npm push. By morning it was mirrored across GitHub. A day later, a UBC student had reconstructed the core in a different language entirely.

Inside the source sat 44 hidden feature flags, twenty of them behind unshipped features. A persistent background agent called KAIROS. A three-layer memory architecture. An “Undercover Mode” that strips Anthropic attribution when employees contribute to open source.

Axios called it, accurately, a free engineering education for every competitor.

Then, last weekend, another set of screenshots leaked on X: a full-stack app builder inside Claude, apparently a Lovable competitor, with panels for Database, Auth, Storage, Secrets. 5 million views before anyone at Anthropic could respond.

Two leaks in two weeks. Both rich. Both unforced.

The Nerf Discourse

While all of this was unfolding, something else was happening in the ground-level chatter. Developers started saying Claude had gotten dumber.

Not a vibe. Receipts. A senior director at AMD filed a GitHub issue with 6,852 session logs, 234,760 tool calls, and a collapse in reads-per-edit from 6.6 to 2.0. One in three edits, she argued, was now being made to files Claude had barely read.

Anthropic partially confirmed it: a default configuration change, adaptive thinking at medium effort, optimized for latency and cost. Not a downgrade of the model, but a quiet downgrade of the experience. Power users got the bill.

The word “nerfed” went viral. “AI shrinkflation” entered the vocabulary. The same people who were writing breathless Medium posts about Claude Code best practices in February were writing obituaries for it in April.

The internet turned, in about a month, from fan to hater.

The Cycle

This is the part I have been sitting with.

Thousands of content creators built entire personal brands on CLAUDE.md files, skills, subagents, slash commands. An economy of prompt optimization and harness hacking. A cottage industry of “the right way to use Claude Code.”

And then the floor moved, the defaults shifted, the model got served differently during peak hours, and the same creators started filming videos about how it was all broken.

The loop is compressing. Praise, disappointment, outrage, workaround, new release, praise again. Two weeks per cycle, maybe less.

I do not think this is anyone’s fault, exactly. Beta services are being pushed to production audiences because the market rewards speed. The same speed that lets a small team at Anthropic ship 14 releases and 5 outages in a single month is the speed that leaves power users feeling like they are being A/B tested into frustration.

Speed and quality were always a tradeoff. In the AI era, the tradeoff is happening on a weekly cadence, in public, to paying customers.

Solo Founders and the Industry They Are Killing

I keep reading two kinds of articles in the same week.

One celebrates the solo founder who shipped an app over a weekend with Claude Code and hit six figures in revenue. The other is a post-mortem of the SaaS category that same app just commoditized.

Both are written in the same tone of wonder. Neither seems to notice the other.

Karri Saarinen said we have lost our sense of judgment and moderation about what to build. I think something similar is happening to what to write about. Every PR launch is a revolution; every PR launch is also the end of something. The same journalist can hold both positions in the same week and not flinch.

There is a beauty in the chaos, and there is a beauty in finding signal inside it. But the signal right now is mostly that the noise is too loud for anyone to be confident about anything.

Hierarchy, Intelligence, and the Thing Nobody Says

On the same day Anthropic shipped its source tree to npm, Jack Dorsey and Roelof Botha published From Hierarchy to Intelligence. A long, careful piece arguing that two thousand years of organizational design, from the Roman contubernium to the modern matrix, exists to solve a single problem: information routing. And that AI can now do that job better than humans can.

Block’s pitch is to collapse the company into three roles: individual contributors, directly responsible individuals, and player-coaches. No permanent middle management layer. A “company world model” does the alignment work the hierarchy used to carry.

It is a beautiful argument. I have been chewing on it for two weeks.

Here is the part I keep returning to. Every company that has actually tried this at scale has, quietly, walked it back. Zappos bled people through Holacracy. Spotify’s squads drifted back toward conventional management as it grew. Valve’s flatness broke somewhere past a few hundred people. Dorsey’s own essay acknowledges all of this in passing, then argues that AI is the missing ingredient that makes the old experiments finally work.

Maybe. But the counter-signal is loud right now. Korn Ferry found that 41% of employees at flattened companies feel directionless. Axios reported in July that Gusto’s data shows industries with more managers have higher worker productivity, not less. Boston University’s Questrom school pointed out the thing nobody wants to say out loud: when you remove middle layers, junior people lose the opportunity structure to ever become senior ones. You optimize the present and quietly hollow out the future. Business Insider’s reporting caught the other edge of it: supervisors who survived the purges are burned out to a crisp, taking on teams twice their former size, and companies are beginning to discover the invisible work that middle managers were doing all along.

Dorsey’s essay and the flattening-regret research disagree about something fundamental. Dorsey thinks managers were mostly routing information, and AI can route information. The counter-research keeps pointing at things managers do that are not routing at all: coaching, translating strategy into action, absorbing ambiguity, holding the room when something goes sideways, growing the next generation.

I suspect both sides are partially right and that the honest answer depends on what, exactly, your particular middle managers have been doing. The companies that discover their managers were mostly shuffling Jira tickets will flatten successfully. The companies that discover their managers were quietly preventing chaos will regret it, and the regret will arrive eighteen months after the press release.

I guess it all depends what do you want to optimize for.

Two Weeks From Now

Here is what I actually believe, under all of it.

In two weeks, Anthropic will ship a fixed version of Opus 4.6. The “dumb Claude” discourse will quiet down. Power users will type /effort max and move on. The next model will be announced and the cycle will reset.

Open-weights fans will find ways to run their own stacks. Gemma 4 landed on April 2 with a 31B dense model that ranks #3 on the Arena open leaderboard, Apache 2.0 licensed, and it runs on a MacBook. The next iteration of Mac Studio is going to have one of its best launch quarters in Apple’s history, because a non-trivial percentage of engineers now want local frontier capability more than they want another subscription.

The Block essay will keep circulating. Somewhere, a company will announce a flattening with great confidence; somewhere else, another will quietly rehire the layer it cut last year. Both stories will be reported in the same tone of certainty. Neither will settle the question.

None of this is a prediction I am confident about. It is just the shape the last two weeks suggest.

What I Am Doing With Any of This

I am writing this while my daughter sleeps off a fever in the next room. I spent the evening alternating between checking her temperature, reading through the Tokyo Metropolitan Infectious Disease Surveillance Center’s weekly report and the JIHS national bulletin to figure out what is going around this month, and a document outlining the shape of a new AI-native organization: what a small team can do, how to structure authority, what work even means when the implementation bottleneck is gone.

Three threads, all carrying the same texture. A small person running a fever, and me trying to work out whether it is the thing half of Setagaya has right now or something worth a clinic visit. A roadmap of unreleased AI features I was not supposed to see. A document about whether Dorsey is right.

The noise from the timeline does not actually help me with any of them. My daughter does not care about Mythos. The surveillance report does not care about the BridgeBench controversy. The org document does not care about KAIROS.

What helps is the slower signal underneath. That capability is advancing faster than process. That leaks, nerfs, and vibe coding competitors are all symptoms of a company sprinting at the edge of its own operational capacity. That the tools we are betting on will change shape three times before September. That the confident essay and the quiet regret can both be true at once, in different rooms.

I am not sure whether the intelligence is artificial or dumb. I am increasingly sure the impatience is ours.

Two weeks from now, none of this will feel the same.

The fever will break first.