Do LLMs lack the machinery of invention? A case for new representations over scaling
No_Will_5010 · reddit · 2026-09-23
A long Reddit essay argues that LLMs excel at recall, generalization, and recombining concepts, but may lack a mechanism for recognizing when an existing representation is insufficient and autonomously constructing new conceptual frameworks — the kind of leap behind inventions like the computer.
Key points:
- For problems with missing theory (e.g. building a time machine), the challenge isn't the next step but creating the framework that makes it possible.
- More data, parameters, RL, and inference compute may improve reasoning and search without producing that invention machinery.
- Novel answers aren't invention: novel combination ≠ novel framework.
The author poses open questions: what counts as a genuinely new representation, how to distinguish invention from sophisticated recombination, and whether scaling Transformers can produce this ability or it requires a different architecture — or model plus tools, experimentation, and feedback loops.
Related event: Essay Questions Whether LLMs Can Invent Genuinely New Conceptual Frameworks(2 posts)→
More from AGI Musings
- Schmidhuber: self-driving cars since 1994 still unsolved — physical AI is a slow game — SchmidhuberAI · 2026-09-23
- Why AI progress accelerated: Claude 4.5 kicked off narrow RSI and open-weight catch-up — maksym_andr · 2026-09-23
- Hot take: narrow recursive self-improvement has already begun inside LLM pipelines — maksym_andr · 2026-09-23
- Social Work Professor Accidentally Runs an AGI Containment Exercise With His ChatGPT — shannonkish · 2026-09-23
- NeuroAI researcher: knowing which biological details to discard is key — aran_nayebi · 2026-09-23
- LLMs get misused on tasks needing real intelligence, and no one benchmarks both — generativist · 2026-09-23