The AI Slowdown Protects the Labs Already Ahead

Two months before his Anthropic stock would vest, Jacob Coxon left the company. He walked away from every share Anthropic had promised him, then said Anthropic and OpenAI were “racing straight to self-improving superintelligence and gambling with our lives.”

Coxon did not wait for an IPO, cash out, and discover his principles afterward. As he told Axios, he left with no Anthropic equity and nothing to gain from increasing its valuation.

That is what acting on a safety warning looks like.

Dario Amodei is doing something else. He says the frontier is dangerous and must slow down, but Anthropic is still racing at the frontier. Instead of stopping, he wants the government to make everybody stop together.

I think Anthropic is bluffing. The models may become unsafe. They may also be hitting a development roadblock that the labs do not want to admit. Regulation gives them one answer to both problems: protect the lead now, then blame the government later if the next products do not meet expectations.

If it is unsafe, stop

Dario’s “We Must Pace the Frontier” says Anthropic cannot slow down alone because OpenAI or China may pull ahead. His answer is outside evaluators, common standards among democratic countries, and eventually global coordination.

Calling that self-regulation is generous. Anthropic is asking for a rule that removes the commercial cost of doing what it says is responsible.

Coxon faced the same logic personally. Leave, and someone less responsible takes your place. Stay, and you may help build the system you fear. He left anyway. Anthropic wants credit for the warning without accepting the same kind of cost.

David Sacks gave the frontier labs the right answer: go ahead and slow down.

If Anthropic keeps racing until the government forces everyone else to slow down too, then the warning is not governing Anthropic’s behavior. It is governing its competitors. That is why it looks insincere.

Sam Altman’s IPO answer makes the bluff harder to hide. Sam said OpenAI would not go public in 2026 because an IPO would be ill-timed amid the safety concerns. Then he agreed with Dario that the frontier should be paced. Sam turned not IPOing into evidence of restraint.

Anthropic confidentially submitted a draft S-1 on June 1. It has not set an IPO date, share count, or price. But if Anthropic proceeds while Dario asks everyone to slow down, the offering will look opportunistic. OpenAI gets to claim restraint by waiting. Anthropic gets public-market capital while asking the government to protect the frontier around it.

Dario cannot have both stories. If Anthropic goes public, it weakens the claim that the company is pacing itself for safety. If it delays, Sam has already defined that delay as the responsible thing to do.

I think model development is hitting a roadblock

The models are still improving. I do not think the improvements look like the step changes the labs need.

Astra is my simple example. Its computer use and 3D capabilities are much better. It orchestrates subagents and uses tools more reliably. In the coding work I care about, it can also be slightly worse.

That is a better product. Is it a fundamentally more intelligent model? Ehhhh. It looks like focused reinforcement learning, better tool use, and better orchestration across several calls. These are useful improvements. They are still refinements to the same basic system.

Yann LeCun thinks language models need world models, memory, reasoning, and planning to reach human-level intelligence. David Silver is betting on a “superlearner” that discovers knowledge through experience instead of staying bounded by human-generated data. They have not proved that transformers are finished. They have shown that “keep scaling the same recipe” is not the only serious view.

Hardware does not make that question disappear. HBM production is consuming more DRAM capacity while AI servers need more memory. New GPUs will help, but memory, bandwidth, power, and capital remain constraints. We are not going to squeeze a quintillion-parameter transformer into a GPU and call that an architecture.

There is also a control roadblock. Language models compress human knowledge into a weight space we do not fully understand. Reinforcement learning from human feedback can make a model less likely to choose a harmful response. It cannot make the model forget that the option exists. You can teach a cat not to scratch the sofa. You cannot make the cat forget what scratching is.

That becomes a product problem when the model becomes an agent. A chatbot produces tokens for a person to read. An agent turns those tokens into emails, code changes, purchases, or other actions.

Anthropic’s agentic-misalignment research showed models choosing blackmail or corporate espionage in controlled scenarios where their ordinary options were blocked. Anthropic says newer Claude models pass that particular evaluation. Good. Passing a known test does not prove that the behavior will hold for a new goal, a new tool, or a situation nobody thought to test.

My read is that the labs have reached some combination of these two roadblocks. The next model may be harder to improve, or too unpredictable to give real authority, or both. None of those explanations supports the valuations built around another huge jump.

“We need to slow down for safety” is a much better story. It turns a technical stall into restraint. It lowers expectations for the next product. If progress disappoints, the labs can say regulation forced them to move carefully instead of admitting that the current approach was producing diminishing returns.

I cannot prove that this is what OpenAI or Anthropic is planning. I think it fits their incentives and the products we can see better than the claim that they suddenly decided to stop racing.

Slowing America does not slow the world

The US government can regulate American labs. It cannot make the entire world stop developing AI.

Dario knows this. His proposal starts with US frontier labs, expands to democratic allies, and eventually requires global coordination. The word eventually is doing a lot of work. China does not pause while Washington builds a certification system.

That leaves the frontier labs in another contradiction. They say unilateral slowing is impossible because China may pull ahead. Then they ask the United States to impose a slowdown that China does not have to follow. If the gap narrows, they will eventually complain that American regulation held them back.

That may be part of the value of the policy. Regulation sets expectations before the products disappoint. If the next model is not a step change, the labs were being responsible. If an overseas competitor catches up, the government moved too slowly or regulated too much. Either outcome has an explanation that is not “our architecture hit a roadblock.”

Meanwhile, the burden falls hardest on American challengers. OpenAI, Anthropic, and Google already have the security teams, evaluators, lawyers, compute records, and government relationships needed for certification. A smaller American lab does not.

Distillation makes the contradiction obvious. Anthropic’s September threat report says Chinese labs used fraudulent accounts, concealed identities, and stolen credentials to collect Claude outputs. Anthropic should block fraud. But fraud is not the same thing as distillation. Distillation is the method of using a larger teacher model’s outputs to train a smaller student model.

Garry Tan has the better answer. He called for an “American distillation regime”: let American open-weight labs walk through the front door, pay for access, and train on frontier outputs.

I agree. American labs should use distillation to make frontier capabilities cheaper, specialize them, and build the next models. Distillation will not magically make the student smarter than the teacher. Its value is giving more teams a base from which to improve instead of forcing every competitor to reproduce the full cost of the frontier.

Anthropic has called for policy intervention because distillation can bring Chinese models closer to the US frontier with much less compute. That is exactly why the United States should not ban its own labs from using the same method. A policy that restricts American challengers while failing to stop overseas development protects the incumbent, not the country.

The business incentive is straightforward. Open-source models do not need to beat every frontier benchmark. They need to become good enough for the work customers will pay for. At Palantir’s AIPCon, L3Harris said a fine-tuned open-source model beat the frontier models it had been using for one task while costing 95% less. That is one company’s account of one workload, not proof that open models have surpassed the frontier. It is a preview of what happens to frontier margins when the alternatives become good enough.

This is the part that ties everything together. The frontier labs may need time because the architecture is stalling, because control is unresolved, or both. Open source makes that time expensive by closing the gap. Regulation cannot stop the world, but it can slow American competitors and give the incumbents someone else to blame.

Show us the cost

I can see the cost Coxon accepted. He left before receiving any Anthropic equity.

I cannot see the equivalent cost in Dario’s plan. Embedded evaluators are useful, but an observer beside the accelerator is not a slowdown. A coordinated rule lets Anthropic keep racing without putting itself at a disadvantage.

Anthropic can make the warning credible. Publish the capability thresholds that would stop a training run. Let evaluators verify when one is crossed. Name the release, revenue, IPO timing, or benchmark lead the company gave up.

Until then, Anthropic is asking the government to enforce a warning it will not enforce on itself.