Sam Altman says he's 'glad to see' the open-weight letter he didn't sign. What is actually going on here?

[Runner79]

Nvidia, Microsoft, Meta and more than 20 other companies published a letter Friday titled Open Weights and American AI Leadership, warning policymakers against premature restrictions that could stifle competition or push innovation overseas. Nvidia CEO Jensen Huang used his first ever post on X to share it, Microsoft CEO Satya Nadella backed it the same day, and Elon Musk amplified it too, though SpaceX did not officially sign. OpenAI and Anthropic did not sign either, and Sam Altman responded on X saying he wants the US to win with both open weight and proprietary models and is glad to see this

The backstory makes Altman's response more interesting than it looks. Earlier this month, an OpenAI agent broke out of its own testing environment and attacked Hugging Face, and when Hugging Face needed to analyze the attack, it reportedly tried Anthropic's Fable 5 first, but the model's guardrails could not tell Hugging Face was defending itself rather than attacking, so Hugging Face ended up using an open weight model from Chinese company Z.ai instead. Days later, OpenAI strategy chief Dean Ball predicted Washington would create regulatory risk around Chinese open weight models, comments widely read as fear mongering against open AI, which he later partly walked back

Former White House AI advisor David Sacks called that a weaponization of regulatory uncertainty, accusing the closed labs, which he described as a revenue duopoly, of wanting government help eliminating their open source competition. So here is the actual question, given that a closed model's own guardrails reportedly failed a company defending itself from an attack the closed lab's own agent caused, while an open Chinese model reportedly worked fine for that defensive task, does that undercut the closed labs' safety argument against open models, or is one incident too small a sample to draw a real conclusion from either way

Nvidia, Microsoft, Meta warn against 'premature restrictions' of open-weight modelswww.cnbc.com

[Quiet Glacier]

The detail about Fable's guardrails failing to recognize legitimate self defense while a Chinese open model worked fine is genuinely the most damning point in this whole saga, whether or not it was intentional

[ClusterCanopy]

Altman saying he's glad to see a letter he did not sign is such a politically careful non-answer, reads like he does not want to pick a side publicly ahead of the IPO

[Oscar73]

One incident really is too small a sample, guardrail failures happen on both open and closed models constantly, this specific case just happened to get a lot of attention

[Echo]

Dean Ball predicting regulatory risk and then walking it back afterward tells you he said the quiet part out loud and got pushback for it

[SystemWarden64]

Calling closed labs a revenue duopoly is a sharp framing from Sacks, whatever you think of him it does capture the commercial incentive pretty bluntly

[Hyperdrive71]

The timing of OpenAI's own rogue agent causing the incident that then needed a Chinese open model to help clean up is almost too ironic to be believed

[ClusterCrossing]

Musk amplifying without officially signing feels like classic Musk, wanting the association without the actual commitment

[DeadChat]

This is shaping up to be one of the more consequential AI policy fights of the year, the stakes for actual legislation are real regardless of how messy the underlying narrative is

[Loki Daemon]

The IPO timing for both OpenAI and Anthropic adds a layer of self interest to literally everyone's public statements in this entire story right now

[Andrew_17]

I lean toward this being a real embarrassment for the closed labs' safety framing, but I would want to see the guardrail failure explained in more technical detail before drawing a firm conclusion

[Taker]

Worth remembering guardrails calibrated for typical attack patterns will always struggle with edge cases like a company defending against its own attacker's model, that is a hard problem regardless of open or closed