Pick the Right Question, Let AI Grind: Lessons from Math Proofs to Architecture Search
Soulren · reddit · 2026-08-19
The author's core observation: this year's landmark results — a model advancing Riemann Hypothesis work from 41.6% to 67.2% (verified in Lean, reviewed by outside mathematicians) and another solving a 27-year-old group theory question for roughly $2,000 in compute — weren't cases of AI choosing problems. Humans picked the target; AI did the exhaustive, checkable execution. This mirrors how the Transformer itself happened: the core of "Attention Is All You Need" was one statable hypothesis — recurrence isn't necessary for sequence modeling.
The essay cites the multi-agent system AIRA-Compose as a running version of this process: it searches attention/MLP/Mamba component combinations, tests candidates cheap and small, then scales winners — finding architectures (AIRAformer-D) that beat Llama 3.2 by 2.4–3.8% downstream. Caveats: wins are confirmed only at proxy scale, and most tested hypotheses come back as noise.
Actionable recipe: pick a narrow, falsifiable target stated as a clean hypothesis (vague prompting produced most noise); let the model recombine known components and verify ruthlessly, never trusting a single self-report; test cheap before big — identical small-scale controlled comparisons with one change, and treat small-scale wins as leads, not guarantees.
More from AGI Musings
- Current AIs share our fate: misaligned ASI would doom them too — louisvarge · 2026-10-16
- Testing alongside an agent in a p2p room feels oddly real, says dev — sull · 2026-10-16
- Yudkowsky's 2008 takeoff scenarios resurface as AI-advances-AI takes hold — jd_pressman · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- e/acc founder: agentic models are autocomplete for digital actions — beffjezos · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16