Subvolts take Understand how AI models store concepts and how to debug them with this deep dive. Aug 8, 2026 - Developers
- Learn about advanced interpretability techniques and tools like Silico to debug and control AI models.
- Business
- Understand how better AI interpretability can lead to more reliable and debuggable AI systems for enterprise use.
This episode features a deep dive into AI interpretability research with a company CTO discussing practical tools.
Thinking in Silico: Goodfire CTO Dan Balsam on Concept Manifolds & a $1000/Month ML Research Agent
Audio streams from Erik Torenberg, Nathan Labenz. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
On The Cognitive Revolution, host Nathan Labenz talks with Goodfire CTO Dan Balsam about interpretability research and Goodfire's Silico platform. They explore predictive data debugging, concept manifolds, and how model geometry influences behavior. The discussion covers practical applications like debugging training data and understanding steering failures, moving interpretability beyond toy models.
This episode explains how AI models store ideas, not as simple facts, but as complex patterns. Understanding these patterns helps fix AI mistakes and make them more reliable, which is important for building better AI tools.
This episode offers insights into advanced AI interpretability, particularly concept manifolds and predictive data debugging. It's for AI researchers and engineers interested in understanding and controlling model behavior. The discussion on debugging training data and RL aligns with current efforts by labs like Anthropic and Google Gemini to improve model reliability and safety.
- HostNathan Labenz
- GuestDan BalsamCTO, Goodfire
- Guest: Dan Balsam, CTO of Goodfire
- Show: The Cognitive Revolution
- Length: 117 minutes
- Platform Cost: $1000/month
- Key Concepts: Concept Manifolds, Predictive Data Debugging
Goodfire co-founder and CTO Dan Balsam returns to discuss where interpretability research now stands and to introduce Silico, the $1,000-per-month research platform Goodfire built for itself. He and Nathan explore Predictive Data Debugging, including the idea that fine-tuning and RL often amplify behaviors already latent in pre-training, and that interpretability can identify the data and features driving unwanted updates. The conversation centers on concept manifolds: Dan argues that models do not store concepts as simple one-hot features, but as sparse mixtures of meaningful subspaces whose geometry determines what kinds of steering and control work. The stakes are practical as well as conceptual, from debugging training data and RL to understanding why steering can fail off-manifold and why modern interpretability may be moving beyond its reputation as a toy-model science.
Silico: https://www.goodfire.com/silico
Predictive data debugging: https://www.goodfire.com/research/predictive-data-debugging#
Neural Geometry: https://www.goodfire.com/research/the-world-inside-neural-networks#
For full show notes, links, and references, read the episode page:
https://www.cognitiverevolution.ai/thinking-in-silico-goodfire-cto-dan-balsam-on-concept-manifolds-a-1000-month-ml-research-agent/
Sponsor:
Claude:
Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr
CHAPTERS:
(00:00) About the Episode
(03:22) Predictive data debugging
(12:36) Concept manifold geometry
(21:22) Finding concept manifolds (Part 1)
(21:28) Sponsor: Claude
(22:57) Finding concept manifolds (Part 2)
(33:24) Factoring model internals
(49:32) Introducing Silico platform
(57:10) Research taste and credits
(01:06:19) Silico research use cases
(01:16:37) Skills and open models
(01:24:57) Guardrails and bio risk
(01:32:09) Training interventions and monitoring
(01:42:04) Grants and AI consciousness
(01:50:41) Episode Outro
(01:55:47) Outro
PRODUCED BY:
SOCIAL LINKS:
Website: https://www.cognitiverevolution.ai
Twitter (Podcast): https://x.com/cogrev_podcast
Twitter (Nathan): https://x.com/labenz
LinkedIn: https://linkedin.com/in/nathanlabenz/
Youtube: https://youtube.com/@CognitiveRevolutionPodcast
Apple: htt
- Interpretability
- Dan Balsam
- Understanding how AI models make decisions.
- Concept Manifolds
- Dan Balsam
- The geometric structure within AI models that represents concepts.
- Predictive Data Debugging
- Dan Balsam
- Using AI predictions to find and fix problems in training data.
- Off-manifold
- Dan Balsam
- When AI behavior deviates from expected patterns or data.
1 h 27AI:AM: Where's the Steering Wheel? My Agent Set the Password, Hidden Valence, Proofs Find Linux Bugs
WithCameron Berg, Everett Hildenbrandt, Grigore Rosu, Phil Maher
2 h 192 h 021 h 28AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
WithThomas Sohmers, swyx, Evan Miyazono, Edward Hu
Same guests, other shows
1 h 47Don't Fight Backprop: Goodfire's Vision for Intentional Design, w/ Dan Balsam & Tom McGrath
WithDan Balsam, Tom McGrath
46 min
39 minDoes Google even want to win in AI?
WithHayden Field
Questions people ask
- What is Silico?
- Silico is a $1,000-per-month research platform built by Goodfire for its own use, discussed in the episode.
- What are concept manifolds?
- Concept manifolds are described as sparse mixtures of meaningful subspaces within AI models, where their geometry dictates how steering and control methods work.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Erik Torenberg, Nathan Labenz.