It adapts during inference, but since the updates are reset instead of retained (and only affect the MLP’s final projection matrix) it’s not really continual learning. It’s more like temporary in-context weight editing. We are somewhat close though. Big gains at long context.

Matrioshka Representation Learning (2024, Google Research) A common problem with embedding models is retrieval cost. Embeddings can be costly to store and retrieve, especially if you store embeddings for huge databases (e.g. 100B documents) [1/N]

The day Google stops using "experimental" is the day Google has achieved AGI

understanding deep learning by Simon Prince is an amazing book ngl. I wonder what unis use it for their courses

GPT 5.4 is too dangerous. We must keep it locked up.

You can infer the size of an AI model by how well it does Family Guy cutaways. This sounds strange, but hear me out: ranking: 1. Gemini 3.1 -- this was actually kind of funny, I could hear the voice actors voicing each founding father in my head 2. Claude 4.6 opus -- not funny

I had an old friend who was an artist and a fan of anything involving robots and AI in TV and media, yet she hated real-world AI like LLMs. She even forbade me from mentioning AI art or anything LLM-related. Truly perplexing.

And so family guy bench was born

This is horrible. Right now, I think what’s been causing it is the sort of doomer AI narrative. I think violence in the coming years might increase, and not only because of the doomer AI narrative, but due to job loss, as violence usually follows spikes in unemployment. For AI

Working on a new long horizon, abstract reasoning, AI benchmark

Like prompt engineering, harness engineering will also die out as the models get better at computer use.

Who do you think the account with the highest follower-set similarity is to yours, where similarity is measured by |A ∩ B| / |A ∪ B|. The person who shares the most amount of followers with you.