Sure, but does slowing down the development of new models count as "adopting safety standards"? I very much doubt it.
> Anthropic believes their work has a 10% chance of killing all humans
This ignores the other part of what they believe, which is that they are the people most likely to make a model that doesn't do that. So, in their view, letting other companies win would increase the probability of human extinction.
> with one half of the frontier duopoly gone, the other half would have no incentive to race forward
I don't see how this could possibly be true, with at least half a dozen companies being just months behind what the frontier labs are releasing.
Of course! What slows down the development is the adoption of specific safety conditions the companies draw up. They can just do that, and it will hold up in court.
> in their view, letting other companies win would increase the probability of human extinction
Yes, I've heard: "We must be in charge even if we end up killing everyone in the process." I personally think that proposition is invalid, but we're all entitled to our opinions.
People used to append IANAL to such statements :-)
> I personally think that proposition is invalid
I agree, but that's meaningless in this context. I was responding to the claim that "if they wanted to pace the frontier, would simply do it", which is false, given what they actually believe.
Do you think their beliefs deserve some sort of special treatment?
No.
I mean, the point (if the claim is to be believed) isn't just to "slow down the development", it's to take more time during development to properly assess the risks the models pose, develop methodologies to reduce that risk, and standardize that across companies. I doubt those wouldn't count, especially in the eyes of regulators of an administration calling for that same slow down.
I hope this is a joke. But most of the LLM research and inventions come from China? The best papers are from DeepSeek? You either get the data by stealing from humans or distilling from bigger models?
Their fear-mongering about GLM 5.3 got me to try it out. Its very good, I'll only go back to Claude if GLM isn't available (it forgot how to do tool calls yesterday).
Interestingly enough, it seems to compact at about 10% of the 1mn context, which makes sense if they're trying to run profitably.
I recommend trying Coralbricks with GLM due to their cheap input cache prices. GLM can get quite expensive elsewhere.