Skip to content

Chip Huyen

  • 1 episode on Chain of Thought
  • 6 topics covered
  • 1 full transcript
  • Dec 2024 first on the show
Chip Huyen

Chip Huyen is the author of Designing Machine Learning Systems and, at the time of the conversation, was working as a storyteller at Tép Studio after a stint as a VP at Voltron Data. On Chain of Thought she dug into the practical side of evaluating generative AI, including how to assess AI coding assistants and agentic systems.

Cover art for Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang

Latest episode with Chip Huyen · EP 5

Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang

Chip Huyen & Vivienne Zhang ·

Show notes & transcript → All listening options →

Episodes

What the conversations cover

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

EP 5 Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang

  • The harder the task, the harder the eval. Chip Huyen’s core point: “almost everyone can tell whether the solution to a first grade math problem is wrong, but it’s really hard to tell whether the solution to a PhD math level question is wrong.” As AI handles expert work, the pool of people qualified to judge its output shrinks — Vivienne Zhang notes NVIDIA’s ChipNeMo co-pilot answers chip-design questions she, without a hardware background, can’t even parse, let alone grade. Episode takeaways
  • Most teams ship on a vibe check — and that’s the gap. Vivienne Zhang says curating a benchmark dataset and measuring against it “sets you probably ahead of the big crowd already, because a lot of application developers also just do a vibe check and then send it to production.” The fix is treating eval as a loop: capture user feedback and usage data post-launch, then systematically fix what tripped up the LLM. Episode takeaways
  • “Coherence” and “relevancy” are not exact metrics. Chip Huyen’s #1 observed mistake is teams not defining clear evaluation guidelines, then treating LLM-judge scores as fixed numbers when they’re “highly dependent on the models you use, on the prompt you use for the judge.” She’s found typos in nearly every open-source eval prompt she’s inspected; the same metric can swing 90% from one day to the next as prompts change. Episode takeaways

In their words

So anyone, like almost everyone can tell whether the solution to a first grade math problem is wrong, but it's really hard to tell whether the solution to a PhD math level question is wrong.

Chip Huyen Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang · 22:00

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