Molin Zhao is a biomedical engineering student at UC Davis who decodes brain activity, takes models apart to see how they work, and builds things.https://molinzhao.com/https://molinzhao.com/blog/contrastive-alignment/https://molinzhao.com/blog/contrastive-alignment/Contrastive alignment taught a speech encoder to pick the right sentence out of 704 a third of the time, with no phoneme labels. It did not make the encoder a better starting point for decoding.Mon, 09 Nov 2026 00:00:00 GMThttps://molinzhao.com/blog/diffusion-decoder/https://molinzhao.com/blog/diffusion-decoder/Training the language model in a speech decoder by block diffusion, not left to right, cut word error rate by about ten points. The gain disappeared once the encoder had been trained on phonemes.Mon, 09 Nov 2026 00:00:00 GMThttps://molinzhao.com/blog/inner-speech-projector/https://molinzhao.com/blog/inner-speech-projector/Inner speech looks like attempted speech, shifted and quieter. Reversing that with a linear map made the sentence easier to identify and did nothing for the synthesized voice.Mon, 09 Nov 2026 00:00:00 GMThttps://molinzhao.com/blog/place-cells-are-easy-to-find/https://molinzhao.com/blog/place-cells-are-easy-to-find/Units that pass the rodent place-cell test turn up in raw pixels and in random networks. To find out whether a vision-language model knows where it is, I had to change its activations and watch the answer move.Mon, 09 Nov 2026 00:00:00 GMT