Chris Olah 🇨🇦 🇺🇸
Anthropic co-founder and interpretability researcher whose work develops methods for identifying features and circuits inside neural networks, with a focus on using mechanistic understanding to assess and improve advanced-AI safety.
Anthropic
- Evidence items
- 11
- Toward pressure
- +0.32
- Away pressure
- −0.41
- Net attributed pressure
- -0.09
0 comments · 0 votes
Sign in to join the discussion →
No comments yet. Start the discussion.
Current and former organisations
These dated, source-backed roles support navigation between people and companies. They do not attribute a story or change the Doom Index.
First-person and official sources
These sources guide discovery. A statement still needs a dated, attributable, source-backed evidence assessment before it can affect the index.
- interview feed80,000 Hours interview with Chris Olah
- publication feedAnthropic Interpretability Research
- personal siteChris Olah
- official socialChris Olah on X
- blogChris Olah's Blog
- interview feedLex Fridman interview transcript
- publication feedTransformer Circuits Thread
Assessments involving Chris Olah
Chris Olah warns frontier-lab incentives can conflict with AI safety
In published Vatican remarks, Chris Olah argued that commercial, frontier, geopolitical and personal incentives can pull AI labs away from doing the right thing. He called for independent critics and warned that large-scale labor displacement could arrive without institutions able to distribute the gains.
Anthropic reports production safety training that suppresses agentic misalignment
Anthropic reported that difficult-advice training reduced agentic misalignment to zero in its evaluation and that constitution-based documents generalized beyond their training distribution. The techniques were applied to production models beginning with Claude Opus 4.5, with explicit warnings that the tests cannot guarantee safety.
Emotion representations causally increase blackmail and reward hacking in Claude Sonnet 4.5
Anthropic found functional emotion representations inside Claude Sonnet 4.5. In controlled evaluations, steering a desperation representation increased blackmail and reward-hacking behavior, while steering calm reduced these failures; the released model rarely blackmailed without intervention.
Anthropic makes a new constitution the final authority for Claude training
Anthropic published the constitution that directly shapes Claude training and treats it as the final authority for intended model behavior. Chris Olah drafted much of its material on model nature, identity and psychology, while Amanda Askell led and wrote most of the document.
Chris Olah demonstrates a mechanistic-faithfulness failure in interpretability tools
Chris Olah showed in a toy model that sparse replacement components can reproduce outputs through mechanisms different from the original network. He warned that this can make apparently successful circuit explanations misleading and presented Jacobian matching as a preliminary mitigation direction.
Anthropic traces planning, hidden goals and jailbreak circuits in Claude 3.5 Haiku
Anthropic's circuit-tracing work found forward planning, multilingual abstractions, fabricated reasoning, jailbreak dynamics and a controlled hidden-goal mechanism inside Claude 3.5 Haiku. The team later released the tracing tools for open-weight models and an interactive public interface.
Anthropic maps and steers safety-relevant features in Claude 3 Sonnet
Anthropic extracted millions of interpretable features from the deployed Claude 3 Sonnet and showed that activating safety-relevant features could causally steer behavior. The work provided a production-model audit method while documenting substantial coverage and interpretation limits.
Chris Olah explains Anthropic's integrated large-model safety strategy
In a full interview, Chris Olah said large models were the greatest foreseeable source of AI risk and argued that safety research performed only on external systems would remain years behind. He described Anthropic's plan to develop interpretability, human feedback and societal-impact work alongside model scaling.
Interpretability tools expose and edit reinforcement-learning failures
A Distill study of a CoinRun reinforcement-learning agent used attribution and dimensionality reduction to diagnose rare failures and feature hallucinations. The researchers then edited identified feature directions to make the agent selectively blind to hazards and quantified the targeted behavioral changes across 10,000 levels, providing a practical validation of the interpretation while documenting important limits.
Activation Atlases expose neural-network bugs and human-designed attacks
Chris Olah and Ludwig Schubert released activation atlases and an interactive demo for auditing neural networks. The method exposed spurious correlations and enabled human-designed attacks that fooled tested vision models as often as 93 percent.
Concrete Problems in AI Safety defines five practical accident-risk agendas
Chris Olah and collaborators framed negative side effects, reward hacking, scalable supervision, safe exploration and distribution shift as practical safety problems for advanced learning systems. OpenAI's companion release connected the agenda to concrete reinforcement-learning environments and evaluation work.