Dev observation: models excel with a reference implementation, but struggle at designing something truly new

lucasmeijer · x · 2026-10-17

Developer lucasmeijer notes a clear capability divide: when there's a large amount of work but a reference implementation to follow, models perform impressively; but when asked to design or build something genuinely new, they fall far short. He describes the gap between the two scenarios as "an ocean."

Related event: Developer finds LLMs excel when tasks have reference implementations or perfect verification(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →