Skip to content

Dan Klein

Co-Founder and CTO of Scaled Cognition

  • 1 episode on Chain of Thought
  • 3 topics covered
  • 1 full transcript
  • Apr 2026 first on the show
Dan Klein

Dan Klein is the co-founder and CTO of Scaled Cognition and a professor of computer science at UC Berkeley, where he leads the NLP group and won the ACM Grace Murray Hopper Award. On Chain of Thought he explained why LLMs are plausibility engines rather than truth engines, and how his team built APT1 for under $11 million.

Cover art for Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition

Latest episode with Dan Klein · EP 54

Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition

Dan Klein, Scaled Cognition ·

Show notes & transcript → All listening options →

Episodes

What the conversations cover

These notes summarize the episodes featuring Dan Klein, including other speakers in panel conversations. Follow a source link to read the discussion in context.

EP 54 Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition

  • Dan Klein describes prompting as a flexible but imprecise control surface. Rewording an instruction can change behavior, but natural-language instructions alone do not provide a guarantee that the model will follow them. Transcript at 5:40
  • Klein argues that teams need a more precise definition of hallucination. A retrieved answer, a paraphrase, and a changed meaning require different checks; treating all mistakes as one phenomenon obscures their causes. Transcript at 29:24
  • Klein cautions that adding models to check other models brings cost, latency, and additional failure modes. His concern is whether a complex checking system can provide dependable behavior across a whole conversation. Transcript at 57:46

In their words

And that's why even if you could make this work, it's still better to have a model that in its natural operation doesn't make the mistake in the first place. And that's absolutely one of the advantages of APT1 is it gets it right in its natural operation rather than after seven other models have signed off.

Dan Klein Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition · 1:00:15

Transcript excerpts; ellipses mark omitted words. Some automatically selected quotes have filler words removed.