Really cool to see my and other's work on grokking covered in Quanta magazine!

I first heard Irina Rish mention Grokking #neural networks on the Paul Middlebrooks's Brain Inspired podcast! That was in early '22. Years later, here's a story on Grokking, about the follow-up detective work of Neel Nanda Ziming Liu and others for Quanta Magazine quantamagazine.org/how-do-machine…

Extremely cool work from @Saprmarks! I think this is one of my favourite SAE papers since Towards Monosemanticity. I'm particularly excited about the use of error nodes, without which SAEs are a bit too janky to do reliable circuit analysis with

Applications to my MATS stream close on Friday April 12th - in just over a week! If you're excited about the idea of getting into mechanistic interpretability research, I'd love to get your application :) The app is designed to take a while, so I'd get started soon!

Great work from my MATS scholars Callum McDougall and Joseph Bloom, in honour of today's special occasion!

Turns out SAEs contain wild features, like a Neel Nanda feature, and this perseverance feature:
lesswrong.com/posts/BK8AMsNH…

Great visualisation library for Sparse Autoencoder features from Callum McDougall! My team has already been finding it super useful, go check it out:
lesswrong.com/posts/nAhy6Zqu…

1/ Introducing Neuronpedia: an open platform for interpretability research with hosting, visualizations, and tooling for Sparse Autoencoders (SAEs).

Let's try it out! ➡️

Neuronpedia lets us instantly test activations of SAE features with custom text. Here's a Star Wars feature:

Cool challenge, though don't assume the question! Do they specialise into anything at all? That's not at all obvious to me.

I would weakly guess that any given feature may be somewhat localised to an expert, but that similar features may be distributed across experts. In…

Cool post! I was very surprised by this result. My guess is that 'adding noise with the same norm as the SAE error' isn't a fair baseline, but I don't have a great story for why. I think better understanding WTF is going on here might help me understand SAEs better!

Sparse autoencoders are currently a big deal in mech interp, but there's not a good, concise intro to what they are. I'm currently taking a stab at writing one! Here's the draft TLDR:

I'm really excited about Neuronpedia's pivot to helping with sparse autoencoder research! Johnny has made a gorgeous UI for poking around inside models and I'm excited to see what new mech interp research this can enable/accelerate!

Are you someone with management experience who would like to contribute to interpretability and AI safety? We're looking to hire another manager to help us continue growing our team, in partnership with our technical leads.

Cool work! I really like the result about adding a 'high ELO' steering vector to make it better at chess!

I've heard great things from Ethan's mentees, and they've put out cool work. If you want to get into alignment research, applying to MATS to work with Ethan is a great opportunity!

There's a new monthly update from the Anthropic mechanistic interpretability team! I'm particularly excited to see work building on my and others (David Bau Mor Geva) work reverse-engineering factual recall, this time by using the lens of SAEs

Great to see the Google DeepMind Dangerous Capability Evals team put out their first paper!

Inspired by Neel Nanda (again), we recorded a Grant Sanderson-style walkthrough for Grokking Beyond Neural Networks.

youtube.com/watch?v=--RAHz….

Arthur is great, and I expect he'll do a solid job as a mech interp mentor!