Aviya Skowron

AI-hallucinated intel report almost led to the US attacking a Chinese ship, but Tristan Harris instead talks about rogue swarms when asked about it. This incident should be remedied immediately with better DoW LLM policy, but the interview never gets to that point.

CNN asks directly how to address this, and we get a tale about US-USSR comms and hopes for a diplomatic breakthrough with China. But this is something the US can address today purely through agency guidance, no Congress or diplomats required! CNN's description of the incident

Great thread by @CFGeek about switching to a career in AI policy! I'd also like to issue a belated apology: I think I answered "honestly, no" when asked if I recommend AI policy work at the Curve back in 2024 😅 I'm glad you did not listen.

@KevinBankston One risk I foresee: "whatever OAI / Anth did in their last model release" becomes "AI safety standards," which is a counterproductive approach to safety (we want a swiss cheese system of defenses), and cooperation without greater disclosure makes this worse. (2/?)

@KevinBankston This impacts you more if your business & operations differ from OAI/Anth's. Hence the concern for open models. I also have the intuition that other concerns might be more pressing, but this is what I'd highlight as a concern that affects our community specifically. (3/3)

@rajiinio It was organized by Nvidia, who solicited endorsements from industry, they were not aiming for a broader coalition with a civil society angle. So there are late joiners (including us!)

@KevinBankston Big Open Source, the rulers of the famously monolithic open source ecosystem.

Unfortunately for the "distillation = IP theft" theory, raw LLM outputs are not protected by copyright according to the US Copyright Office. Or is this saying that K3 is an infringing copy of Fable? On what basis? (Copyright protection of model weights is not guaranteed either.) x.com/mkratsios47/st…

Regarding the trade secret route — are we saying LLM outputs are trade secrets? They become trade secrets once you collect a lot of them? There's a bill in Congress trying to define this, and predictably it is overly inclusive of all kinds of uses of LLM outputs.

@Cloudflare why does it take 30 seconds to open Legiscan, and how come this "anti-bot" measure takes so long? This has gone too far

@StephenLCasper @natolambert I agree that it’s not the attitude in the EU, and that some US private actors overplayed the anti-OS attitudes for purely partisan reasons. But the impulse exists in the US, even if it’s not viable federally in 2026, and even if in general the idea falls apart under scrutiny.

@Miles_Brundage I actually talked myself out of my initial opinion by imagining what I would be doing today had I worked at Anthropic on anything that touches policy. My bad — memo is very very frustrating.

@Miles_Brundage Though to be clear it’s absolutely strategically ill-advised, yes. Would I be mortified to see it leaked if I worked at Anthropic? Yeah, my thoughts are with all government affairs & policy staff