cflamant

The folklore says seven riffle shuffles are enough to randomize a deck of cards. I trained a small transformer to guess the next card after k shuffles, gave it the deck's starting order, and let it watch every card as it was dealt. One shuffle is wildly predictable; the edge shrinks fast, but a faint, statistically real signal survives all the way out to eleven shuffles.

Two frozen language models wired together through a small trainable interface on their hidden states, so they can share their thinking as it happens instead of trading text.

Sentence transformers help us identify similar sentences based on their content, topic, and implicit meaning. In this post we compare BERT and all-mpnet-base-v2, a sentence transformer that is much better suited for producing sentence embeddings that can be compared using cosine similarity.

Learning families of solutions to differential equations using deep neural networks for efficient parallel evaluation.

Created one of the first public implementations of DeepMind's neural stack architecture.