Goodfire CTO on Concept Manifold Geometry & a $1000/Month ML Research Agent
The Cognitive Revolution · rss · 2026-08-09
Goodfire co-founder and CTO Dan Balsam returns to discuss the current state of AI interpretability research and introduces Silico, an internal AI research agent platform costing around $1,000 per month.
Interpretability & Data Debugging
- Predictive Data Debugging: Fine-tuning and RL often amplify behaviors already latent in pre-training data. Interpretability tools can identify the exact data and features driving unwanted model updates.
- Concept Manifold Geometry: Dan argues that models do not store concepts as simple one-hot features, but as sparse mixtures of meaningful subspaces. This internal geometry determines what kinds of steering and control interventions work, explaining why steering can fail off-manifold.
The Silico Platform
- Silico is an agent platform designed specifically for ML research, capable of autonomously running experiments, writing code, and utilizing tools.
- Dan shares concrete research use cases for Silico, discussing research taste, compute credits, and the role of open models. The conversation also covers how these interpretability techniques can be applied to training interventions, model monitoring, and mitigating bio-risk.
More from coding & agent
- “The agent is working in the sandbox”: Usually implies a spectacular fail — haydendevs · 2026-08-09
- Vibe Coding Too Pricey: Dev Burns 1.5M Tokens in Minutes via Codex CLI — afl_ext · 2026-08-09
- AI Code Review Pitfall: Vague Prompts Disguise Core QA Pipeline Issues — ClickOk5811 · 2026-08-09
- Liquid AI's New Model Outperforms 4x Larger Peers in Private Benchmarks — JosephJacks_ · 2026-08-09
- Swarms Launches Auto Agent Builder to Generate Agent Teams from One Prompt — KyeGomezB · 2026-08-09
- Experiment Shows Naive LLM Code-Auditing Loops Destroy Correct Code — sebpaquet · 2026-08-09