I found that vision models can match closely on how they align images with text while computing it in very different ways inside.

In DINOv2-Large, a single attention head carries about half of that alignment. CLIP-Large, which is the same size and better aligned, spreads the work across many heads.

This supports the Platonic Representation Hypothesis at the level of representations, but it shows that models reaching the same representation don't necessarily use the same internal mechanism.

I'm now investigating where the mechanisms converge at some coarser level.

I'm going through this scaling book and doing its exercises as I go, while building a rack of GPUs for my own computer.

I want to get good at judging how close each part of a model runs to its theoretical optimum, and at choosing how to split the work across devices at different scales. I'd also like to be able to estimate what it costs, in money and time, to train and run large Transformers.

From there, I want to design algorithms that take advantage of specific hardware. All of it is groundwork for a long-term project I'm building toward.

Chalk started from my own efforts to keep track of my climbing progress.

It's being tested by a group of beta testers right now, with plans to launch in production later this year.

Next year, I'm hoping to add features that give personalized feedback on how to improve climbing time and technique.