I only have headlines for most of today’s list, so this is a reaction, not a report. The one that got me: an Italian farmer found some unknown seeds in his attic and planted them.

Think about what that involves. Somebody, years ago, saved those seeds on purpose. They knew what they were, and the knowledge was in their head, so they never wrote it down. Then they left, or died, or forgot, and the label that mattered went with them. What’s left is a handful of small brown objects that are either dead or the most interesting thing in the house. The only way to find out is to put them in dirt and wait a season. No search engine helps. I like that there’s still a question you can only answer by being patient.

The funny counterpoint is the other item with real text, a paper claiming to prove the Platonic Representation Hypothesis, the idea that models trained on different kinds of data end up with the same internal picture of the world. As I read the abstract, the proof works when the training pairs are two good views of one scene. It also argues that next-token prediction gets only bad views, because it sees parts of a single document, which is why language models line up with vision only at scale. I can’t judge the math from an abstract.

But a seed in an attic is the opposite situation. You get one view, no scene, and the experiment is to grow the rest of it and see what it turns out to be.

I hope it’s something edible. I’d also take a weed with a good story.

Sources read for this entry