Dan Hendrycks

Executive Director — Center for AI Safety

Profile

Dan Hendrycks is the executive director of the Center for AI Safety (CAIS) and one of the most quietly consequential figures in AI today. If you’ve ever benchmarked a model, you’ve touched his work. He’s the lead author behind MMLU, the MATH benchmark, and the GELU activation function — the last of which is baked into BERT, GPT, and basically every transformer shipping today. He wrote the GELU paper as an undergrad.

After a PhD at UC Berkeley under Jacob Steinhardt, Hendrycks pivoted hard into AI safety. He founded CAIS and authored the one-sentence 2023 Statement on AI Risk that Sam Altman, Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, and hundreds of others signed — the document that moved existential risk from fringe to mainstream overnight. He also wrote a free open-access textbook, Introduction to AI Safety, Ethics, and Society, which is now the standard primer for the field.

He wears three hats that would normally be in tension: safety researcher, policy architect, and safety adviser at xAI (where he takes a symbolic $1 salary and holds no equity). He helped draft California’s SB 1047, the most aggressive attempt yet to regulate frontier AI. When MMLU saturated in 2024, he teamed up with Scale AI and shipped Humanity’s Last Exam in early 2025 — 3,000 expert-authored questions that state-of-the-art models initially scored under 10% on.

For developers learning AI, Hendrycks matters because his benchmarks are the measuring stick. Every model card, every leaderboard, every “GPT-X beats Claude-Y” blog post is ultimately anchored to numbers he defined. He’s also proof that you can take AI risk seriously without being a doomer — his instinct is to build the evals, write the textbook, and ship the policy, not to tweet about the apocalypse.

Books

Introduction to AI Safety, Ethics, and Society The standard textbook for AI safety — free online, covers societal risks, technical safety, and ethics across three sections.

Key Articles & Papers

Gaussian Error Linear Units (GELUs) 2016 — The activation function now used in nearly every transformer — BERT, GPT, ViT. Written while he was an undergrad. Measuring Massive Multitask Language Understanding (MMLU) 2020 — The 57-subject multiple-choice benchmark that became the default capability test for every LLM from GPT-3 onward. Measuring Mathematical Problem Solving With the MATH Dataset 2021 — 12,500 competition math problems. Exposed how badly early LLMs reasoned about STEM, and drove years of chain-of-thought work. Unsolved Problems in ML Safety 2021 — The research agenda paper — robustness, monitoring, alignment, systemic safety. A field map for anyone entering AI safety. X-Risk Analysis for AI Research 2022 — A framework for thinking about which research directions reduce or increase long-term catastrophic risk. Statement on AI Risk 2023 — One sentence, hundreds of signatures from the biggest names in AI. Moved x-risk from fringe to policy conversation. An Overview of Catastrophic AI Risks 2023 — Four risk categories: malicious use, AI race, organizational risks, rogue AIs. The canonical overview for policymakers. Introduction to AI Safety, Ethics, and Society 2024 — The full textbook as a single arXiv paper — a one-stop grounding in the field. Humanity's Last Exam 2025 — 3,000 expert-authored questions at the frontier of human knowledge, designed to resist saturation for years.

Controversies

During the SB 1047 debate, critics pointed out that Hendrycks was an investor in Gray Swan AI, a model-auditing startup that stood to benefit from the bill’s compliance requirements — a conflict of interest given his role in drafting the legislation. He responded by divesting his equity and staying on as an unpaid adviser. Reasonable people can disagree on whether that fully resolves it, but he moved quickly once the conflict was raised.