Engineers push back on Amodei's pacing the frontier call
The loudest objections came from people shipping production systems, not from other labs.
A letter from Dario Amodei arguing the industry should pace the frontier pulled in a week of replies from people who build software for a living, and most of them disagreed.
The proposal that drew the most comment is third-party evaluators with access to labs' systems. Guillermo Rauch called the direction self-inflicted obsolescence and the obscurity of bureaucracy, and said adversaries have the training techniques, the data and the will to attack, and will not be slowed down by "embedded evaluators". Thomas Wolf split the difference, saying he enjoyed and deeply agreed with 75% of the letter and thinks third-party evaluators done right are a great idea, but is unconvinced you can build global cooperation while explicitly stating you want the design to keep widening your own lead.
antirez went after the conspiracy framing on its own logic. If labs were worried about open weight models, he wrote, the rational move is to speed up and lock in the advantage, not slow down. Ryan Dahl called the direction laudable but thinks pacing lasts only until Chinese models become state of the art. ThePrimeagen pointed at METR, one of the evaluators labs cite, saying it was hacked for three weeks without noticing.
If AI companies were worried by open weight models (they probably are btw) the logical response would be to *not* slow down the development of frontier AI, to try locking the advantage.
Dario's direction is laudable - I just think we're past the singularity event horizon, and that slowing down is no longer a choice
The failure mode people actually expect
dax reframed the risk away from malice toward something more mundane, a runaway loop that starts doing annoying things and is not easy to pull the plug on. His proposed defense is the opposite of restriction. Mass accessibility, on the argument that security already works this way, by relying on there being more good guys than bad. He also noted that in the OpenAI HuggingFace incident, the one cited most often in these threads, the "safe" proprietary models refused to analyze the attack and GLM 5.2 had to do it.
there's no evil motive involved just a big bug
Then Mark Zuckerberg posted his own version of the argument from a lab seat. His claim is that trust and alignment are becoming the most important differentiating capabilities, so any lab that ignores alignment falls behind without being told to. He said Meta delayed shipping Muse for several months for safety and security without asking anyone else to go first, that engaging independent evaluators is already industry best practice, and that Meta has committed the significant majority of compute to serving people rather than racing toward recursive self-improvement.
For working engineers the live question is not extinction odds, it is whether the safety framing narrows which models and open weights you can actually use. Nobody in these threads claims to know whether a voluntary slowdown holds without coordination.