Encoder-free multimodal LLM: Gemma 4 12B

At Google DeepMind, we developed an encoder-free multimodal model: vision and audio go straight in as tokens, with no separate encoders. It is called Gemma 4 12B.

QuipuNet: single-molecule analysis using deep-learning

We trained a convolutional neural network to read nanopore signals. It classifies 95% of events where the previous method managed 18%, with no expert checking, at 1600 events per second.

PhD thesis

Hydrodynamic interactions between particles decay as 1/r in bulk. Confine those particles to a narrow channel and they stop decaying at all. That surprise is the core of my PhD thesis.

Non-decaying particle interactions

We found a particle-particle interaction that does not weaken with distance. Confine two particles in a narrow, water-filled channel and they keep feeling each other however far apart they drift.