Quantcast – a Risk.net Cutting Edge podcast
Oct 23, 2026 · 40 min · 10 segments
Ariye Shater and Hariom Tatsat, 19/10/26 by Quantcast – a Risk.net Cutting Edge podcast
Mauro CesareHostNaomi Cardona-CasillanosHostWhat are sparse autoencoders and what is mechanistic interpretability?
Unlike other forms of interpretability, where we look at the input and output mostly from outside, in McInterf, we look deeper at the neurons at the layers level.
And the idea behind this sparse autoencoder is there is a term called polysemanticity, meaning that there are more number of concepts in the world compared to the number of parameters even in the best model as of now.
So each neuron, which is there inside the LLM, might represent multiple concepts.
Michael Jordan astrophysics, right? So the idea of the sparse autoencoder is to decouple these concepts and name it, name it so that people can understand, right? So you can think of it as microscope, which looks inside the LLM and decouples different concept and makes it more explainable.
And talking more about how these autoencoders work, what data do they use as input and how does the output look like?
yes so as i mentioned it's a kind of microscope which looks inside the different layers and what's happening so what we do is we hook this microscope or sparse autoencoder to some of the layers when the information flows through the large language model what these sparse autoencoders do they take the vector or the information from the previous layer And the magic happens inside them.
It decouples all the concept, as I've mentioned, credit risk, astrophysics, Michael Jordan, etc.
What are sparse autoencoders and what is mechanistic interpretability?
Unlike other forms of interpretability, where we look at the input and output mostly from outside, in McInterf, we look deeper at the neurons at the layers level.
And the idea behind this sparse autoencoder is there is a term called polysemanticity, meaning that there are more number of concepts in the world compared to the number of parameters even in the best model as of now.
So each neuron, which is there inside the LLM, might represent multiple concepts.
Michael Jordan astrophysics, right? So the idea of the sparse autoencoder is to decouple these concepts and name it, name it so that people can understand, right? So you can think of it as microscope, which looks inside the LLM and decouples different concept and makes it more explainable.
And talking more about how these autoencoders work, what data do they use as input and how does the output look like?
yes so as i mentioned it's a kind of microscope which looks inside the different layers and what's happening so what we do is we hook this microscope or sparse autoencoder to some of the layers when the information flows through the large language model what these sparse autoencoders do they take the vector or the information from the previous layer And the magic happens inside them.
It decouples all the concept, as I've mentioned, credit risk, astrophysics, Michael Jordan, etc.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.