A chart making the rounds: stop dismissing modern LLMs as next-token predictors
burny_tech · x · 2026-10-11
A developer shared that a chart posted by Fchollet has been used at least three times in two weeks to argue that modern LLMs are not mere next-token predictors regurgitating their pre-training corpora.
The chart itself isn't described in the post, but the exchange reflects the ongoing debate about LLM capability boundaries and whether the "stochastic parrot" framing still holds.
More from Models
- Rumor: GLM 5.5 lands next week, MiniMax preps Space Bunny Alpha — ChrisGPT · 2026-10-11
- Microsoft launches Fast Decision API; OpenAI's version still best on p90 latency — bindureddy · 2026-10-11
- Gemini in Gmail can't create email filters even with approval, user reports — pbaylies · 2026-10-11
- OpenAI enterprise users can now run Moonshot's Kimi K3 via Baseten — yogthos · 2026-10-11
- Prediction: $1,000/month AI subscription tier arriving by Q1 2027 — AntonObukhov1 · 2026-10-11
- Zhipu's GLM-5.4/5.5 roadmap leaks: 1T+ params and a full RSI self-training loop — airesearch12 · 2026-10-11