Working with AI on novel problems: engineer it to stumble into 100 ideas and try them all
ivan_bezdomny · x · 2026-09-27
- The author describes a deliberate workflow when tackling new problems with Astra: instead of expecting the model to be right first try, he engineers a setup where it can stumble into 100 candidate ideas, then tries them all.
- He has applied this to some language-related work and sees Jevons-type models as a way to scale this idea-generation approach further.
- The approach echoes the thread's thesis: LLMs may solve problems less through intelligence than through brute-force enumeration, sidestepping human memory limits — like Deep Blue in chess.
Related event: Practitioner's method: getting AI to generate 100 candidate ideas at once(2 posts)→
More from AGI Musings
- DARPA funding for quantum computing is far bigger than most people realize — jwt0625 · 2026-09-27
- Why AI agents may not go mainstream this decade: most lives aren't complex enough — menhguin · 2026-09-27
- A practical AGI test: replace one human validating 150,000 document conversions — burkov · 2026-09-27
- Are frontier models' inner optimizers obsolete? CoT as goal-directed search sparks alignment debate — xuanalogue · 2026-09-27
- DHH roasts frontier-lab doomer class at Rails World 2026: more P-Bloom, less P-Doom — rohanpaul_ai · 2026-09-27
- Grinding LeetCode today is like cramming for the 1910 imperial exam, says ex-Tesla AI lead — henloitsjoyce · 2026-09-27