“In short, Kurt shows that you can ablate roughly 90% of the individual neurons, keeping only those with the strongest activations, with surprisingly minimal performance degradation.”
“Showing that the original model's behavior is substantially restored relative to ablating the neuron entirely when the other LLM's guesses are patched in to the model's activations.”
Kurt Tigges · "The Cognitive Revolution" · The Neuron Basis: Goodfire's Curt Tigges on Agentic Interp, Monosemanticity & Reverse Engineering
Audio streams from the publisher’s own feed. Verbatim does not host or copy it.
What was described
validated neuron hypotheses by asking another model to predict where in text the neuron should fire and then patching those predictions into the original model’s activations
Said nearby on the same subject
“If you can get that information and you can turn that unstructured data into semi-structured data and connect it to the prediction, then you can use the storytelling power of the large language models”
“AlphaFold looks at structure prediction, the structure of proteins and how they interact with each other. Normally, that takes, you know, months or years to do x-ray crystallography and get a structure out. If you can do that with a simulation model like AlphaFold, then you don't have to go into the real world and do that. You save months”
“For the Alpha Genome family of models, we have a public API. And basically, you can install a Python package. You can basically make these predictions for free”
Moments like this, every Monday
The strongest new moments on your topics, in the speakers' own words. Your first brief is free, then USD {{ TIERS.starter.price }} a month for three topics.
No card to start. No password. Cancel any time.
Track a competitor?
Pro watches up to 25 names or phrases and emails you the morning after they are said.
See Pro, $29 a monthVerbatim quotes a short extract and links to the episode; the episode belongs to its publisher. Remove a show.