lateinteraction on AI's jagged frontier: solves millennium problems, still dumb at 1000 things
lateinteraction · x · 2026-10-27
lateinteraction pushes back on recent AI hype in a multi-post thread:
- Jaggedness cuts both ways: a tool that can solve millennium prize problems yet fails at 1000 smaller things reshapes his understanding of the territory more than his judgment of the vehicle — compute vs. humans are mutually jagged.
- Brute force vs. generalization badly missed his (already low) expectations: an AI that solved Navier-Stokes still needs help writing English or explaining itself, after seeing the whole Web 10x over.
- Next-gen won't be fundamentally different: "public frontier AI is simply not that smart or reliable just yet," and the common argument "it just solved <major problem>, so…" doesn't hold.
He also links a quick-and-dirty writeup expanding on the same theme.
More from AGI Musings
- AutoTavern: Self-Improving LLM Agents Build a Hearthstone AI Rivaling Top Players — Changlian Ma · 2026-10-27
- Researcher argues model-interaction continual learning is 'the next 10 years of work' — arthurcolle · 2026-10-27
- Reddit take: open weights and reverse engineering are dissolving Big Tech moats — RealisticDonkey9302 · 2026-10-27
- 'Humans will never again lead the frontier of understanding' — pivoting to managing the transition — AndyMasley · 2026-10-27
- Brundage: Anthropic's 'pace the frontier' just means 'we won't slow down unilaterally' — Miles_Brundage · 2026-10-27
- If even Microsoft and Amazon workers aren't safe, what career path is? — Equivalent_Can_4029 · 2026-10-27