LeCun's "Autoregressive LLMs are doomed" slide resurfaces: errors diverge exponentially

john__allard · x · 2026-11-22

johnallard digs up Yann LeCun's old slide deck titled "Auto-Regressive Generative Models Suck!" — now a classic time-capsule meme given how events played out.

Core claims on the slide: autoregressive LLMs "are doomed" and can't be made factual, non-toxic, or controllable; if each token has error probability e, the probability a length-n answer is correct scales as (1-e)^n, diverging exponentially — "not fixable without a major redesign," citing Dziri et al., ArXiv:2305.18654.

Original post →

More from Fun

Fun channel →