|
> Ah, the position they don't state. Come on, we're not _that_ stupid. Their alarmist tones are the height of hypocrisy when they do release powerful models after spelling doom. Their safeguards and guardrails have been broken countless times - let's see if _this_ time the jailbreak detection in J-space actually puts a stop to them. But - really - they have trillions at stake, which evaporate once their edge is commoditized and inference providers have a near-equivalent product for much less. They're firing on all PR and lobbying cylinders to delay that just enough for their IPO to succeed, and leave investors holding the bag - I'd be terrified of their inclusion in tons of pension funds which track S&P500 or whatever they'll end up with (SpaceX, should the actual floating shares increase, and OpenAI worry me just as much). On my second point: it happened, it keeps happening, it will keep happening all the time, and it's not new. Incredibly bright individual researchers, dodgy companies and state actors discovered, weaponized and traded vulns for, I don't know, 15 years before LLMs got involved? And it's not like they are out of a job now. The multiplier effect is not lost on me, but the fallout from even Mythos so far has been.... not that much, really? And I'm not completely ignorant there - I am security-adjacent and have close colleagues who specialize in that. Also, indeed I know that a powerful open weight model in malicious hands is worse than OpenAI doing a whoopsie and cutting access off once they eventually realize it. But even a closed weight frontier model served by a non-US aligned entity could serve the purpose - and while I'm thinking of China, I do wonder what Russia may do - they used to be really good at this stuff and their leaders are not exactly the friendliest bunch around. Or India, or Iran, I hear they're a bit angry lately. Unlikely but let's say they develop or acquire the capability (ultimately it's some hardware + one leak away even for closed models). Asking them politely to stop may not be effective. Western society may pay the price of applying the brake and find itself out...modeled? is that a word? Because 2 guys called Sam and Dario, and their early investors, really want those returns? Come on we're not discussing the greater good here - it's a bunch of greedy guys trying to pull the ladder up after themselves. You do realize that, right? Do we really owe them a moat? |
My overall point is, don't let their motivations, nor your cynical conclusions based on those obvious motivations, blind you to your own point of view.
The debate here shouldn't be about the signatories motivations, it should be about the risks of open-weights coming too close to the untested frontier. You've acknowledged that now: "I know that a powerful open weight model in malicious hands is worse than OpenAI doing a whoopsie".
But I'm left a little confused about your position because you follow that up with what I can only describe as a set of "what-about"-isms. Yes, closed models could still be misused, especially in the wrong hands. Still, even worst case there, it's fewer hands. China is the only set of hands that's really relevant for now, no one else is close.
Without going into deep detail, fairly sure we both don't trust the Chinese government. But we also don't have a lot of practical choice on whether they have access to a model frontier-ish. What could be different is if Russian hackers, Malaysian hackers, Iranian hackers, etc. all have access too. In the open-weight model, they sort of do (somewhat constrained by compute.. but rounding up enough compute for a bit of inference is achievable even for them).
I should make it clear at this point that I think banning US companies from using open-weight models would be a totally ineffective policy that's just self-harm. Anthropic said the same and I agreed, but just in case that stance has become blurred, I'll re-clarify. It's probably more unclear because there really isn't a policy stance on the table that backs what the "maybe there shouldn't be frontier open-weight models", because any real and effective policy there would either have to be multi-lateral, or be the result of independent actions that just happens to align. So really, all that an independent actor like Anthropic can do is say "maybe there shouldn't be frontier open-weight models", which they did and see if that leads the world toward agreeing with that position, which might then lead to some action.
If that somewhat tenuous outcome ever did occur, it wouldn't really mean a "moat", as the Moonshot, Zhipu, Alibaba, Bytedance would still have a model and offer it via API. And they could still create agreements with US/EU/etc. cloud companies to host them, in the same way as other closed-models are hosted. But there'd be more control.
Yeah, there'd be some advantages to model providers, compute providers from this. There'd be some costs to potential customers and those that would self-deploy for good purposes. As little as I support those costs, and as much as I'd prefer not to hand out those advantages, my cynicism would not prevent me from realizing that it's a bigger problem that if you don't do this, you'll fail to harm the interests of those who would self-deploy for nefarious purposes.