Joshua Saxe says a near-term international AI safety deal still looks hard as cyber risk rises
joshua_saxe · x · 2026-07-27
Joshua Saxe replies to a thread about an international AI safety agreement, arguing that while the medium-term policy space may be broader than it looks, a deal in the next 1-2 years is hard to imagine from experience.
He says that timeline matters because AI cyber capabilities may become dangerously strong soon, so his preferred framework is a "gradient descent" approach: focus on the next right step rather than waiting for a perfect agreement.
He still supports efforts toward safety agreements involving the US, China, and other players in the AI race.
More from AGI Musings
- A real RSI test should reach today’s frontier LLM from scratch, the author argues — kchonyc · 2026-07-27
- New co-edited book on encounters with AI arrives in print — Merzmensch · 2026-07-27
- AlphaGo is used as a case study for how superhuman AI can still improve humans — _arohan_ · 2026-07-27
- Open-weight AI will be used for harm, says quote-tweet arguing bad actors always adapt — basedjensen · 2026-07-27
- ARC-AGI’s name may overstate what the benchmark can really tell us about AGI — tedgreenwald · 2026-07-27
- Open weights may lag frontier AI by 3–12 months, but still act as a sovereign fallback — robleclerc · 2026-07-27