Researcher Decries Open Source AI Shift From Safety To Sovereignty
GPT-OSS development was delayed to ensure robust safety guardrails.
Entities: Aidan Clark, GPT-OSS
In a post on X, researcher Aidan Clark said he is "pretty dismayed" that open-source AI talk has shifted from safety to sovereignty, arguing that with GPT-OSS, "the hard part was ensuring the built-in safeguards were strong enough we could sleep at night." His point lands in a debate where "digital sovereignty" is framed, in a Hugging Face blog post, as a way for countries and institutions to govern and control their own AI infrastructure.
Combined views
16.6K
7 posts, first seen 6h ago
Reactions from ranked influencers
7 posts
Pretty dismayed the open source narrative has moved from safety to sovereignty. GPT-OSS was obviously good for everyone, the hard part was ensuring the built-in safeguards were strong enough we could sleep at night (and we ended up delaying to ensure this).
@_aidan_clark_ I don’t think the people making those arguments have ever been concerned with safety.
Btw this is basically my safety take
@aiamblichus I hope you stay far away from frontier model development! We believe in iterative deployment, because we think eventually these models will pose a risk of significant harm, that that point is hard to identify ahead of time, and therefore that you should move carefully and update
@xeophon it’s a good model sir
I'm not sure "has moved from safety to sovereignty" is the most accurate framing. There are four things happening in parallel: - max accuracy is being displaced by accuracy-per-dollar as "good enough AI" emerges and tokenmaxxing is viewed as uneconomical - closed source is being threatened by the accuracy of open source (both max accuracy and accuracy-per-dollar) - company-level AI narratives are being displaced by nation-level AI narratives as nations take the reigns - the safety conversation is on hold until we know who is going to be in power... at which point that party will set the tone on safety. The location of the greatest capability is more foundational than safety... which shows up in the national conversation as "sovereignty". Labs have been making a company-level version of the same argument for years. Gotta produce the most accurate model within your own 4 walls in order to be able to hold the pen on advancing safety. When nobody really knows who will hold the pen on advanced AI capability in 6-12 months (US? China? OpenAI? Moonshot? Kimi? Open? Closed?), talking pragmatically about AI safety over longer time horizons becomes a bit fuzzy.
💯
@aiamblichus I hope you stay far away from frontier model development! We believe in iterative deployment, because we think eventually these models will pose a risk of significant harm, that that point is hard to identify ahead of time, and therefore that you should move carefully and update