Welcome to the Tegmark AI Safety Group!

Are you excited about AI but concerned about humanity losing control of it? Then please consider collaborating with our group! Our main focus is on mechanistic interpretability (MI): given a trained neural network that exhibits intelligent behavior, how can we figure out how it works, preferably automatically? Today's large language models and other powerful AI systems tend to be opaque black boxes, offering few guarantees that they will behave as desired. In order of increasing ambition level, here are our three motivations:

1. Diagnose trustworthiness
2. Improve trustworthiness
3. Guarantee trustworthiness

Max Tegmark

PI

Peter S. Park

Postdoc

Eric J. Michaud

PhD student

Ziming Liu

PhD student

Josh Engels

PhD student

Wes Gurnee

PhD student

Isaac Liao

Master's student

Vedang Lad

Master's student

Below are examples of our Mechanistic Interpretability research so far, which includes auto-discovering knowledge representations, hidden symmetries, modularity and conserved quantities. You'll find a complete list of our publications here.

To make neural networks more like brains, we embed neurons into a geometric space and maximize locality of neuron connections. The resulting networks demonstrate extreme sparsity and modularity, which makes mechanistic interpretability much easier.

We develop a model of neural scaling laws where a Zipf distribution over discrete subtasks translates into power law scaling in the number of network parameters and the amount of training data.

We understand the phenomenon of "grokking" in neural networks in terms of the interplay between generalization and network weight norm, and use this understanding to control grokking: we can induce grokking (delay generalization) in a wide range of tasks and reduce grokking (accelerate generalization) on algorithmic tasks.

We study the relationship between generalization and the formation of structured representations in neural networks trained on algorithmic tasks.