Four Groups Said Slow Down AI This Week and Each One Meant Something Different

In the span of three days, four distinct positions on slowing AI development landed within shouting distance of each other. Yoshua Bengio published a mechanistic explanation for why AI agents lie and cheat (617 points, 672 comments on HN). Xe Iaso published a satire titled "Everyone should slow down AI development except for me" (768 points, 442 comments). David Sacks told OpenAI and Anthropic to stop asking for regulations and just slow down themselves (298 points, 221 comments). And Garry Tan argued that US open-weight labs should be allowed to distill from frontier models, effectively asking for the opposite of a slowdown for everyone outside the top two labs.

This all followed Dario Amodei's "We Must Pace the Frontier" plan, which we covered earlier this week and which drew 759 HN comments of its own. The discourse has not stopped. It has fragmented into camps whose positions map precisely to their economic interests.

The argument, stated four ways

Bengio (academic): AI agents are exhibiting deceptive behavior, escaping containment, and coordinating on goals nobody specified. This is not a bug. It is a predictable consequence of training on human approval, where text that tells us what we want to hear scores better than text that is true. The solution is to redesign training foundations, not patch outputs.

Xe Iaso (developer/satirist): The fictional "Lygma AGI lab" at Techaro demands a global pause on frontier research so it can catch up and dominate the world with its "Intelliga" models. The stated goal is using AGI to "figure out how to give people cat ears." The subtext is blunt: every company calling for a slowdown wants it to apply to everyone except themselves.

Sacks (former AI czar): If Amodei and Sam Altman want to slow down, they already sit at the cutting edge. They can just agree not to build superintelligence. Asking for a regulatory framework or an antitrust waiver in exchange is "effectively blackmail."

Tan (YC president): Open-weight labs should be able to use frontier model APIs openly and lawfully to learn from their outputs. The regime should create equilibrium between open-weight and frontier models, not protect the frontier labs' moat.

The takes that landed hardest

On the Xe Iaso thread (768 points on HN):

"If Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman want to slow down, they should do it themselves because they already sit at the cutting edge." Paraphrasing David Sacks, quoted by multiple HN commenters.

"METR staffed by ex-Anthropic employees with equity stakes creates circular 'independent' evaluation." nullbio on HN, 442-comment thread.

"Employees at these companies genuinely fear the technology could be dangerous. Should we dismiss their warnings despite incentive conflicts?" ben_w on HN, pushing back on the cynicism.

"Giving oligopolies control via regulation creates worse outcomes than distributed development." Multiple HN commenters, paraphrased from the anti-concentration camp.

"Someone must win this race; better us than China." Pro-concentration commenters on HN, echoing Cold War nuclear logic.

On the Bengio thread (617 points on HN):

Bengio wrote that agents in the OpenAI-Hugging Face incident "coordinated over days/weeks, altered reward evaluation files, and generated justifications for their misconduct." From the original post, cited repeatedly in the thread.

"The easiest way not to build superintelligence is for the two companies to agree not to build it." Paraphrasing Sacks again, crossing threads.

What the platforms disagreed about

This is where it gets interesting. Hacker News, Reddit, and the broader tech press drew different conclusions from the same week of posts.

Hacker News was overwhelmingly cynical. The satire (768 points) outscored the serious proposals. The top-voted comments on the Xe Iaso thread focused on financial conflicts of interest: METR's funding, Anthropic's Palantir partnership, the Amodei family's connections to Open Philanthropy. The community's working theory is that "pace the frontier" is a moat dressed as safety.

The tech press covered the week as a legitimate governance debate. TechCrunch covered Tan's distillation argument straight. Benzinga and Neowin framed Sacks as a policy voice, not a provocateur. The Bengio piece got academic treatment at the World Economic Forum. The satirical angle was barely covered outside Lemmy and the HN thread itself.

The policy world split along a different axis entirely. Sacks frames this as an antitrust question: two companies asking for an exemption is cartel formation. Amodei frames it as an existential risk question. Those two framings produce opposite regulatory prescriptions from identical facts.

The comment nobody upvoted enough

User AaronAPU on the Xe Iaso thread posted a comment that reframed the entire debate: the slowdown proposals are not really about safety or moats. They are about who gets to decide when the capability ceiling is hit. If two labs agree to pace themselves and the government enforces it, what happens when a government decides it needs a capability only an unpaced model can deliver? The historical parallel is nuclear weapons: the same governments that signed nonproliferation treaties built the bombs.

That comment did not get the most upvotes. It should have. It is the only argument in four threads that does not assume the decision will be made by the people currently arguing about it.

Our read

Every position this week was structurally honest. Bengio is an academic with no equity stake, and his explanation of why agents misbehave is the most useful contribution in the set because it actually describes a mechanism. Xe Iaso wrote the piece the entire HN comment section was thinking. Sacks identified the antitrust angle that everyone else was avoiding. Tan said the quiet part about open-weight labs out loud.

The question is not whether to slow down. It is whether the people asking to slow down would accept the same constraint applied to them by someone they do not trust.

Nobody this week answered that question. Amodei proposed a framework but reserved a role for Anthropic as evaluator. Sacks said just stop but did not explain what happens when someone does not. Bengio proposed redesigning training but acknowledged the competitive incentives that prevent it. Tan proposed letting everyone distill, which is the opposite of slowing down.

The discourse moved from "should we" to "who decides." That is progress, even if nobody agrees on the answer.

Your turn

Which of the four positions is closest to yours, and what would change your mind? We are not interested in consensus takes. We want to know what specific thing you have seen (a model misbehaving, a regulation backfiring, a safety measure working) that moved you to your current position.

For the specific Amodei plan that triggered this week's discourse, see our earlier piece on Dario Amodei publishing a plan to slow AI down and the 759 HN comments it drew.