With millennium-prize-tier model swarms, should AI safety conjectures be written down?
inductionheads · x · 2026-09-29
bayeslord argues that if reliable AI is partly solvable as a math problem, it seems worth writing down a list of conjectures now that we have model swarms with millennium-prize-solver-tier capability.
The retort: isn't this what MIRI was supposed to be doing for 15 years? The exchange touches on whether formal/mathematical AI safety approaches are being revived by frontier model capabilities.
More from AGI Musings
- Enterprises use GenAI for email and calendars, not core systems — small self-hosted models may fill the gap — DavidLinthicum · 2026-09-29
- AI adoption dip is just a lull, argues teortaxesTex: diffusion and compute are the bottlenecks — teortaxesTex · 2026-09-29
- Jensen Huang is our Andrew Carnegie — but who'll be AI's J.P. Morgan? — kevinnbass · 2026-09-29
- Thought experiment: how do we cope when AI reveals our forgotten secrets at will? — PierceLilholt · 2026-09-29
- "Human language is a compressed mathematical isomorphism of cognition" — one dad's GPT-2 epiphany — repligate · 2026-09-29
- Dario's entry-level job loss prediction starts now, not in 2030, Ben Todd notes — ben_j_todd · 2026-09-29