KAN Author Ziming Liu: LLM Training Intuition Has Four Limits, 'Science of AI' Language Doesn't Exist Yet
ZimingLiu11 · x · 2026-10-24
Ziming Liu, first author of the KAN paper and MIT researcher, laid out four structural limitations of LLMs' training intuition:
- The intuition is imperfect, sometimes sub-human;
- It's empirical, not structured;
- It's not maximally condensed;
- It's insensitive to dataset properties.
His conclusion: the language of "Science of AI" hasn't been created by humans yet, so LLMs have never seen it. The logic: if training intuition behaved like math or coding, more compute would yield better performance — but it doesn't. An accompanying chart contrasts the rising scaling curve of math/coding with the flat curve of training intuition versus compute.
More from AGI Musings
- Yacine on open weights: 'If your AI is open weights, your country is the product' — yacineMTB · 2026-10-24
- Researcher blasts Hinton: anthropomorphism, hype disconnect and ideological fear-mongering — gerardsans · 2026-10-24
- Grok lays out a path to billions of humanoid robots by the 2030s — Dr_Singularity · 2026-10-24
- Researchers argue a human-controlled AI future is worth a cost to long-term efficiency — mobav0 · 2026-10-24
- Greg Isenberg lists 9 obvious-but-good startup ideas for the agent economy — AI_Andrew · 2026-10-24
- Swedish startup IntuiCell's digital nervous system lets a robot dog learn to stand without pre-training — CurieuxExplorer · 2026-10-24