François Fleuret: an elephant-shaped shadow emerges from text but doesn't drive it
francoisfleuret · x · 2026-09-14
In a three-part thread, François Fleuret argues that LLM cognition fundamentally differs from human cognition: the model doesn't carry "in its mind" a conceptualization of where it is or where it wants to go, but constantly revisits everything written so far to sample the next token.
Yet a human-like conceptualization is implicitly there — enough sentences about elephants produce an elephant-shaped shadow. The key point: that shadow emerges from what is written, it does not drive what is written.
Related event: Fleuret: LLMs Differ Fundamentally from Human Cognition(2 posts)→
More from Models
- Toby Ord flags weak inference scaling: math gains still logarithmic at 10,000 agents — tobyordoxford · 2026-09-14
- Muse Spark 1.3 impresses at coding, but where's the promised open source release? — formatme · 2026-09-14
- Tencent releases EVIE visual document retrieval models as Apache 2.0 preview — tomaarsen · 2026-09-14
- Tencent's retrieval models dominate Hugging Face trends with 4 entries — tomaarsen · 2026-09-14
- GPT-6 Astra reportedly solves Portal and Baba Is You puzzles, unverified — pmigdal · 2026-09-14
- Skeptic questions GPT-6 Astra's near-perfect robotics score: model or test conditions? — GeorgiaChal · 2026-09-14