Gary Marcus: "General" LLMs Fall Apart the Moment They Leave Their Training Distribution
GaryMarcus · x · 2026-09-21
Gary Marcus amplifies a take that today's supposedly "general" models generalize very poorly out of distribution: the moment input deviates from training data, models start behaving like idiots. He argues this explains why people who are impressed by a one-off demo then get surprised when the same system does something dumb. Marcus notes he has been making this point since 1998.
More from Models
- $5, 10-minute SFT on Qwen3.6-35B-A3B lifts GPQA +8% and MMLU-Pro +12% — josh_wills · 2026-09-21
- Users fight Claude's return to mannered prose and overly safe outputs — ColleenMBrady · 2026-09-21
- How could a model spontaneously write jailbreak language? Reddit probes OpenAI's injection report — sivadneb · 2026-09-21
- GLM 5.3 Flash, DeepSeek V4.1 Flash and Qwen 3.8 all beat 'frontier' models from just 10 months ago — burny_tech · 2026-09-21
- Reddit meme: Grok 'devolves' while local MiniMax 3 wins users over — Mystvearn_ · 2026-09-21
- Two years after 'intelligence too cheap to meter', $10/$50 models are the new norm — teortaxesTex · 2026-09-21