Mechanistic interpretability of neural networks
Reverse-engineering the internal algorithms and representations of trained neural networks into human-understandable features and circuits.
State of the art
Sparse-autoencoder dictionary features and attribution graphs extracted from frontier LLMs (2024–2025).
- Explanations cover only a fraction of model computation
Who’s working on it
- Chris OlahResearcherAnthropic
Founded the circuits agenda; co-led Scaling Monosemanticity (2024) and the attribution-graph work (2025) that traced multi-step reasoning inside production models
- Anthropic ↗Company
Open-sourced circuit-tracing and attribution-graph tooling in May 2025 with Neuronpedia, and has made interpretability an explicit corporate priority
- Neel NandaResearcherGoogle DeepMind
Leads DeepMind's mechanistic interpretability team and shipped Gemma Scope (2024), the open sparse-autoencoder suite most external work builds on
- Goodfire ↗Startup
Raised a $50M Series A in April 2025 (Menlo Ventures, with Anthropic participating) to turn interpretability into commercial model-auditing and steering tools
- Transluce ↗Program
Nonprofit research lab building open tools for eliciting and explaining model behaviour, the main independent counterweight to in-house lab interpretability
- Coefficient Giving (formerly Open Philanthropy) technical AI safety RFPProgram
Its 2025 request for proposals funded interpretability benchmarks, white-box applications and more transparent architectures, the largest philanthropic pipeline into the field
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Automated interpretability tooling at frontier scale | $$ $10M – $100M |
Depends on
A root node: no prerequisites modeled.