What happened?
At The New York Times' DealBook Summit, two opposing views clashed on an AI safety panel moderated by Michael Barbaro, host of The Daily. Dan Hendrycks, director of the Center for AI Safety, called for an indefinite pause on training runs for frontier AI models. Max Hodak, a former co-founder of Neuralink who also spoke on the panel, said that actually enforcing such a ban would be impossible without building a surveillance state that continuously monitors computing infrastructure.
Why it matters
This isn't the industry's first pause call. In 2023, the Future of Life Institute's open letter called for a six-month halt on training systems more powerful than GPT-4, signed by figures like Elon Musk and Steve Wozniak. The letter didn't stop a single company; OpenAI, Google, and Anthropic kept their training timelines moving at the same pace. Hendrycks himself authored the brief statement from that era declaring that mitigating extinction risk from AI should be a global priority — a statement also signed by senior executives at OpenAI and DeepMind.
The time since then has shown that the pause idea gained no real traction in practice. Model training runs didn't shrink — they grew. Hendrycks today also serves as a safety adviser to xAI, meaning his call comes from someone inside the industry who sees the pace of scaling up close. That sets him apart from the 2023 signature campaign — this is no longer an outside warning, but an objection raised from within the lab.
Hodak's objection represents a different camp altogether: the accelerationist circle that views regulation itself as a bigger risk than AI misuse. This view aligns with the "techno-optimist" rhetoric championed by a16z and with the framing of US-China competition. In their view, the only thing that would make a training moratorium meaningful is a mechanism that continuously inspects every data center and GPU cluster on the planet — a regime resembling nuclear disarmament verification, one that takes decades to build and still leaves gaps.
Why doesn't an AI training moratorium work in practice?
Existing oversight tools fall far short of a full stop. US export restrictions on Nvidia's most advanced AI chips already struggle to monitor a narrow target — chip flows — as chips get rerouted through third countries and gray markets emerge. The oversight needed to halt a training run is far more extensive than that: it would require real-time monitoring of every company's and every cloud provider's computing capacity.
| Side | Main argument | Risk priority |
|---|---|---|
| Pause camp (Hendrycks) | Scaling speed has outrun risk assessment | Existential / loss-of-control risk |
| Accelerationist camp (Hodak) | Oversight mechanism would require totalitarian scope | Regulatory overreach / competitive loss risk |
Both sides propose different solutions to the same problem, but neither describes a workable enforcement mechanism. Hendrycks says "stop," without saying who would enforce it. Hodak says "you can't stop it," without offering an alternative risk-management approach.
What's next?
No binding step should be expected in the short term. The EU's AI Act classifies risk by use case rather than by training process, while the regulatory trend in the US has recently shifted toward looser rules and fewer restrictions on AI companies. China, meanwhile, keeps investing despite every outside call for a pause. The real policy tool is likely to remain not a training moratorium, but narrower measures — chip export controls and limited reporting requirements for cloud service providers.
Hendrycks' call is a morally consistent warning. But Hodak's objection isn't empty either: a ban that can't be enforced stops being a ban and becomes nothing more than a statement of intent on paper. The real gap in this debate is that both sides hit the same wall — neither offers a concrete answer to the question of who would actually do the enforcing.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.