AI #179 Part 2: Hearing The Fire Alarm

Congress is having a busy week on AI, and none of it is simple. Reps. Trahan and Obernolte introduced the FRONTIER Act, which basically takes the SB 53/RAISE framework — safety reports, incident disclosures, catastrophic-risk definitions — and pushes it up to the federal level, complete with a new emergency shutdown authority run through an Under Secretary of Commerce for AI Security. Zvi's problem isn't the concept, it's the mechanics: if that Under Secretary doesn't want to enforce, nobody can, including states, which lose their own authority permanently under preemption. He's wary of trading away that much state power without getting something solid in return, and notes the full licensing requirements only kick in at very large revenue and spending thresholds anyway.

A separate, much shorter bill — the AI Kill Switch Act from Ted Lieu and Nathaniel Moran — landed just a day after Marco Rubio reportedly told diplomats to downplay talk of a tech kill switch. It's fifteen pages, straightforward, and would require large covered AI operators to be able to cut off model inference or customer access during a crisis. Zvi likes the idea in principle but flags real gaps: the maximum daily fine is small enough that a company could just eat it rather than shut down, revenue thresholds are calculated retroactively so brand-new offerings start out exempt, and open-weight models get a free pass simply because they can't have a kill switch by design. He figures that last exemption won't spark much backlash right now, but open-source advocates will stay nervous the rules could tighten later.

Meanwhile Sam Altman was on the Hill, previewing what's presumably a new OpenAI model to officials including Lutnick and Bessent, and talking up a systematic process for frontier releases. He downplayed the HuggingFace incident when asked, said open source is an important part of the ecosystem, and sounded open to mandatory pre-deployment testing and independent auditing, so long as it's the