AI

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

Read full article at source ↗

Based on the episode description, not a full transcript. Source evidence retrieved 10/11/2026, 19:59:24 UTC.

Briefings

Curt Tigges from Goodfire presents research showing modern language model neurons are interpretable and functionally bounded, acting as legible conjunctions rather than opaque black boxes. This work enables dynamic sparsity without significant benchmark loss, offering practical methods for behavior monitoring and persona steering. The episode explores gated MLPs, sparsity experiments, and reverse engineering for alignment, with Tigges explaining how internal representations can be directly enumerated.

Key points

• Modern gated MLPs in language models function as legible conjunctions and control signals capable of extensive dynamic sparsity without lasting benchmark loss.
• The research demonstrates that many internal representations can be directly enumerated, providing an alternative to complex external featurizers for behavior monitoring.

Context

Curt Tigges, a technical staff member at Goodfire, discusses findings from a podcast episode published on November 10, 2026. The research challenges previous assumptions about the interpretability of individual neurons in modern language models, suggesting they are more structured and functional than previously believed. This finding offers practical avenues for behavior monitoring and persona steering in AI systems.

Limitations

The evidence is limited to a single podcast episode from November 2026. Specific experimental details, ablation results, and quantitative metrics are not fully detailed in the provided text. The episode covers multiple topics including gated MLPs, sparsity, and alignment, but the exact scope of the ablation experiments and their specific outcomes are not fully described.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 10/11/2026, 20:00:49 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihe73sgup3jqb26rzys6uwzrhuddayehr6dsh2qcjxes4z3rrrbfi