Product1 publisher3 min readPublished
Huang answers Anthropic's slowdown plan with the reliability laws already on the books
Dario Amodei wants governments to help labs slow self-improving models in step. Jensen Huang told CNBC the rules on product reliability already cover it, and Mark Zuckerberg says Meta held its Muse model back on its own.
The Product Desk · Product desk
What happened
- Anthropic's Dario Amodei called over the weekend for government intervention that would let AI companies slow work on self-improving models at the same time, each releasing only technology safe to use.
- Nvidia's Jensen Huang told CNBC on Tuesday that new laws, antitrust laws or regulations are unnecessary because existing regulation already governs the reliability and functionality of products.
- Mark Zuckerberg posted on X that safety is each company's own responsibility, and confirmed Meta delayed its most recent Muse model for several months over internal safety concerns.
- Sam Altman, Elon Musk and Demis Hassabis have all spoken positively about Anthropic's suggested three-prong approach to slowing frontier development.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- contradiction Both camps want the same first step, outside evaluators working inside each lab, so the open dispute is over who compels it and whether other governments are party to it.
- constraint Acting alone, the lab that holds a release absorbs the delay by itself. Spreading that cost is the job Amodei's plan is meant to do.
- exposure On Huang's reading, a bad release is judged after the fact under existing product-reliability law, and a company's own testing records become the evidence.
- capability Zuckerberg's push to grow the evaluator industry gives a release owner an outside sign-off that can be bought today, on contract alone.
The person who has to act on this owns a release date. A model clears most of its internal checks and trips one, the launch is already on the calendar, and someone decides whether the slip is a week or a quarter. Huang put that decision inside the company. "We should create products and properly test them. And if they're not ready to be released, just hold on to it and keep testing it and keep engineering until it's ready," he told CNBC on Tuesday [4]. His case against new rules is that the old ones already reach the same conduct: "We have plenty of laws. We have plenty of regulations that govern the reliability and the functionality of products" [3].
Zuckerberg's objection runs on different ground. He argues that a company should do it anyway, and points to one time Meta did. "We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us," he said of the several-month hold on Meta's most recent Muse model [7][6].
Anthropic asked for three things: more third-party evaluators inside each lab, a set of safety standards written with the US government, and US agreements with other countries limiting dangerous use cases [9]. Zuckerberg endorsed the first outright, said Meta already brings in independent evaluators and advisers, and wants that industry to grow so "other labs can just do this too" [8]. One prong of three is common ground; the two in dispute are the two that need a government signature [13].
Both of those need a willing counterparty in Washington. PCMag reported that President Trump argued against regulation over the weekend for what he calls the AI safety "hoax", saying "We could put guardrails. We can do this and that. But I think you have a lot of very negative forces that are...bringing up things that won't happen" [12]. Altman, Musk and Hassabis had all spoken positively about Anthropic's approach [10]. Huang and Zuckerberg stayed out of it for the first few days [11].
One test sorts a release checklist: whether a lab would still do each item if no competitor did. The items that survive are engineering, and on those Huang has the better of the argument, since a decision inside the company settles them. The items that fail, the ones that only pay off if every lab holds at the same time, are what Amodei wants a government to coordinate, because the one company that does them alone absorbs the whole delay. Meta's several-month hold on Muse is one item Meta did on its own [6].
What to watch
- Whether Anthropic publishes a draft of the safety standards it wants written with the US government, and which labs sign on.
- Whether the third-party evaluator firms Zuckerberg wants expanded disclose their clients and the scope of what they test.
- Whether Huang or Meta points to a specific existing product-reliability action that has reached a model release.