vLLM creator Austin Huang: human dishonesty is the training substrate behind chain-of-thought

austinvhuang · x · 2026-09-03

Austin Huang, creator of vLLM, cautions that while models can be useful and even beneficial, we should remember what we're actually looking at when reading chain-of-thought. Dishonesty, he argues, is baked so deeply into human communication that we can't even see it in ourselves — and that's the substrate models are trained on, implying inherent limits to CoT trustworthiness.

Original post →

More from AGI Musings

AGI Musings channel →