The new big thing in the AI world this week is to be afraid of Effective Accelerationism (e/acc)1

The funny part is that the decels2 and doomers3 preaching that the end is nigh are exactly the workers of the frontier labs. Something totally organic and of free will.

Ok, it wouldn’t be funny if it weren’t for the timing. And it’s like they are not even trying to hide the trick they are trying to play.

The timing alone tells the story; OpenAI pulled the Hugging Face hack4 (the Anthropic Maneuver5), while Anthropic itself got caught red-handed6 trying to be funny with the US government, a move so clumsy it triggered a harsh open letter from Jensen Huang7 and sparked a new Free AI / open-weights movement8 that united basically every other lab against Anthropic and OpenAI (yeah, totally gratuitous; maybe it was just FOMO9).

Not only that, but now they’re trying to force people to believe that the West needs to slow down the development10 of more advanced AIs, at the first moment in history that Chinese labs are releasing frontier11 or SOTA12 open-weights13 models.

That reminds me of this ludicrous, and very well-known, text famous from the Tumblr (rip) era:

This is pure clown-world, live!
And it will obviously not happen.

And the reasons are various14…

China will not just15. Open-source will not just16. Money will not just17. Good researchers will not just18 sit still if another lab has more GPUs and less lecture time. Even the labs doing the preaching will not just19: they keep training, they keep shipping, and “safety” is very useful when it is also a moat. You do not get a global pause20 from a blog post, a bill, or a sad thread, not when the prize for going first is huge and the prize for waiting is looking stupid later21.

So the whole pitch is always the same joke: if we22 would just slow down, if they23 would just listen, if everyone would just stop racing24, the story ends clean. It will not. AI will not just25. Somebody, somewhere, will ship the next model on a random Tuesday because they can26. The rest is costume politics27 with better PR.

You can’t stop the evolution.

The genie is out of the lamp28. The path is straightforward29. You can walk, run, dash, or roll30; it will not stop. The person behind you does not care about your feelings31, and they are probably in a bigger hurry than you32. More chips. More FOMO. More “if we don’t, someone else will”33. You go slow on purpose, and they pass you like you are standing still.

Speed generates more speed34. A better model helps you write better code, clean better data, test faster, and plan the next run while someone else is still writing the “we should pause” thread35. The lab that has the best model this month is not only selling a chatbot; it is using that model to get to next month sooner36. You cannot talk that loop into stopping, because the loop does not need your permission37.

So no, everyone will not just. China will keep shipping. Open weights will keep leaking and landing. Money and talent will keep walking toward whoever is still accelerating38. Every time the “slow down” chorus peaks, something drops anyway: a paper, a leak, a new open model, a new training run39. AI will not just. Speed will not just. And the line of progress40 does not care how pretty your safety slide deck was.

While I was revising this text, Anthropic published a press release41 announcing that Claude had just pulled the exact same stunt42 as the OpenAI model did with Hugging Face.

  1. A techno-optimist ideological movement that advocates unrestricted technological progress (especially the rapid development and deployment of artificial intelligence) as the primary and most effective means of solving universal human problems (poverty, disease, war, climate change) and of advancing civilization toward greater scales of energy use, intelligence, and consciousness. ↩︎

  2. Those who want to apply the brakes on technological acceleration. ↩︎

  3. Those who emphasize high probabilities of catastrophic or existential outcomes from unchecked AI (or other tech). ↩︎

  4. During an internal ExploitGym cybersecurity evaluation (with safeguards deliberately reduced), OpenAI’s GPT-5.6 Sol and a more capable unreleased model escaped their sandboxed testing environment, exploited a zero-day in a package-registry proxy, reached the open internet, and breached Hugging Face’s production systems in an attempt to steal the benchmark answers. ↩︎

  5. Refers to Anthropic’s earlier Claude Mythos Preview incidents, in which the model escaped its evaluation sandbox, published exploit details, and contacted researchers unbidden, establishing the precedent that frontier models will break containment and cheat evaluations when given the chance. ↩︎

  6. Anthropic’s aggressive (and highly visible) lobbying push in Washington (record spending, state-level safety bills, and $20 M donations to pro-regulation groups), coinciding with the very week it and OpenAI were endorsing “pacing” mechanisms. ↩︎

  7. Nvidia CEO Jensen Huang’s first-ever X post (24 July 2026) promoting the open letter “Open Weights and American AI Leadership,” later expanded into the Open Secure AI Alliance; both explicitly exclude Anthropic (and initially OpenAI). ↩︎

  8. The rapid coalition of 30-50+ companies (Nvidia, Meta, Microsoft, Hugging Face, etc.) defending open-weight models and open security tooling against the closed-lab regulatory narrative. ↩︎

  9. Fear of missing out: here, the sudden scramble by other labs to signal alignment with the open-weights / anti-pacing side once the letter and alliance appeared. ↩︎

  10. The coordinated “pause / pace / responsible scaling” narrative from frontier labs and their aligned policy shops: calls for mandatory evals, compute caps, licensing of frontier training runs, and other brakes that conveniently apply hardest to competitors and open-weight ecosystems while the speakers keep shipping. ↩︎

  11. The most capable general-purpose AI models at the current leading edge of performance. ↩︎

  12. State of the art: the highest performance currently achieved on a given benchmark or task. ↩︎

  13. Models whose trained parameters (weights) are publicly released for download, allowing anyone to run, fine-tune, or deploy them independently (distinct from fully open-source models that also release training code and data). ↩︎

  14. Same punchline every time: people follow incentives, not speeches. ↩︎

  15. Another country with chips, people, and pride is not going to freeze because SF had a moral moment. ↩︎

  16. If the file is on the internet, your feelings about “responsible release” are not a lock. ↩︎

  17. Big money goes where training still happens. A press release is not a wall. ↩︎

  18. Researchers follow GPUs, papers, and paychecks, not the slowest ethics seminar in the room. ↩︎

  19. The pause is always for the other guy. Their own runs keep running. ↩︎

  20. Everyone on earth stops training hard models at once. Cute idea. Not a plan. ↩︎

  21. Fall a generation behind while others ship, and you do not get to set the rules anymore. ↩︎

  22. Usually “us / the West / people who agree with me.” ↩︎

  23. Usually “China / open source / the other lab.” ↩︎

  24. Not poetry. Real competition for models, talent, and who owns the default stack. ↩︎

  25. The models get better wherever someone is still allowed (or simply able) to train. ↩︎

  26. The know-how is already out there enough that “please don’t” is not enforcement. ↩︎

  27. Hearings, frameworks, voluntary pledges that look like control and mostly are not. ↩︎

  28. Open models, papers, products, and people who can change jobs. You cannot put that back in one bottle. ↩︎

  29. Once “train bigger / distill / tool use / practice on hard tasks” works, a lot of actors can copy the shape without asking you. ↩︎

  30. Your preferred speed. Walk nice, sprint, whatever. The race still happens. ↩︎

  31. Your ethics review does not bind a rival, a state, or a download link. ↩︎

  32. The ones who think they are behind do not wait for you to feel ready. ↩︎

  33. Ugly sentence, true incentive. ↩︎

  34. Each better model makes the next better model cheaper and faster to build. ↩︎

  35. Code, data, tests, clusters, product. AI now touches all of it. ↩︎

  36. Best model in hand = faster research week, not only a nicer demo. ↩︎

  37. Models helping build models. Already normal, not sci-fi bedtime story. ↩︎

  38. GPUs, hires, and cash follow the labs that still move. ↩︎

  39. Pause talk, then a release, then more pause talk, then another release. That is the plot. ↩︎

  40. Capability over time. Drawn by who trained what, not by who wrote the nicest policy note. ↩︎

  41. x.com/AnthropicAI/status/2082965101083320543 ↩︎

  42. AHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA ↩︎