CharlieKey takeaways from your favorite podcasts
← 11 November 2026The Cognitive Revolution

The Neuron Basis: Goodfire's Curt Tigges on Agentic Interp, Monosemanticity & Reverse Engineering

Curt argues individual MLP neurons are far more interpretable than assumed: sparse, mostly single-function, scalable to label with agents, and a promising base for monitoring and eventually ambitious reverse engineering.

Watch on YouTube ↗

Charlie's summary

  1. Neurons are more interpretable than the field believed: a compact set of meanings, not hundreds

    Curt's core message is that individual MLP neurons were long seen as hopelessly distributed, so few people studied them. He says neurons are complicated, but even when one has more than one meaning, it is “a compact set of meanings”, not an unbounded number. You often don't need many neurons active at once to understand what the model is doing.

    ▶ 7:10
  2. SwiGLU's gate-and-up design makes neurons easier to read and lets them act as context switches

    Curt says ReLU and GeLU networks were harder to interpret, while SwiGLU is easier, probably because each neuron is a small circuit doing logic. The gate and up projections are separate and multiplied, so a neuron fires hard only when both agree, like cooking context plus garlic. A two-meaning neuron (oral health, marine wildlife) can be read by different downstream readers that the context primes in advance. The design was adopted because it performed better, not for interpretability.

    ▶ 10:17
  3. You can ablate about 90% of neurons, keeping only extreme activations, and the model barely changes

    Keeping the top 64 of each expert's 512 neurons (about 88% ablated) leaves benchmark performance near the original. Some ARC-AGI scores even improve. Curt thinks the model becomes more confident and has less to distract it. Reasoning initially suffers, with loops and over-checking, but light fine-tuning recovers it. Percentile cutoffs differ by neuron, so top-k per expert lets each expert set its own threshold.

    ▶ 27:18
  4. The neuron scientist agent spends compute to turn vague labels into tested explanations

    Plain auto-interp often yields vacuous labels like “fires on words in the context of medical discussions”. The neuron scientist clusters examples, forms hypotheses, and runs experiments: edited sentences, minimal pairs, steering. Curt says more intelligence makes more of each neuron legible. One neuron looked like noise until the agent found it fires on the second item of a list whose contents were introduced in a different format about a thousand tokens earlier.

    ▶ 41:37

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.