Testing if Weaker Models Can Replicate Frontier Discoveries with Hints
danshipper · x · 2026-08-03
Before boarding a flight, the author tasked GPT-5.6 (note: likely an unofficial alias or typo) with a challenge: given a hint involving algebraic number theory, can it reproduce Astra's previous proof of Erdős's planar unit distance conjecture?
The author's broader theory is that weaker models can often reproduce frontier-model discoveries if provided with the right conceptual hints, as a stronger model's advantage largely lies in exploring and finding the right direction. The author later posted "i broke it lol," implying the test resulted in an interesting crash or failure.
Related event: Prompt Distance: Can Weaker Models Reproduce Frontier Proofs?(6 posts)→
More from coding & agent
- ChatGPT co-inventor launches Jev: claims 20-200x faster, 40-400x cheaper than LLMs — GabGarrett · 2026-09-18
- LeanReact 0.1: expressing composable, provably correct React components in Lean — hargup13 · 2026-09-18
- Message Board Agents Do Science Together and Catalog Their Creator's Design Flaws — Kyrannio · 2026-09-18
- Agentic loop recipe: Kanban board + MCP + Tmux, with reviewer agents and a blocked column — oyren-ai · 2026-09-18
- Jev, a 'System One' model by Typesafe, launches on OpenRouter with typed decisions instead of text — majidmanzarpour · 2026-09-18
- FastMCP Is Only 2 Years Old, and Its Creator Recapped the History at MCP SF — hboelman · 2026-09-18