Noticing gpu models

A few weeks ago a new type of small model popped up on my feed: gpu-* models.

Instead of writing a grammar for every programming language, gpu-lexer notices the shape of the language's syntax and highlights it. It's 41k parameters, ships as a 30KB npm package, and runs on WebGPU in the browser. A few days later gpu-time followed in the same vein, turning phrases like "every other Friday at noon" into real dates with 25k parameters.

Compared to the small models that I train, these are utility-based classifiers. They take something in and hand back a label. This token is a keyword, this word is a weekday, this phrase a duration or a time of day. There's a menu of answers and the model deciphers your text to pick one. Mine go the other way. Give them a line and they write the next one, with no menu at all.

It's interesting to imagine a world of flexible software packages like these. The name suggests use within the browser, but I could imagine them as small single-use tools that a model could use to augment itself. Since the first post a small family has grown around them, each one noticing its own thing:

• gpu-code-spacer — reads your code and puts the blank lines back where the thoughts break
• gpu-color — describe a color and get back one you can copy
• gpu-cron — "every weekday at 9:30am" becomes a cron line
• gpu-lexer — the one that started it, 75 languages and no grammars
• gpu-query — turns a search into filters without ever seeing the schema
• gpu-sankhya — reads "sava lakh" as 125,000, in whatever script you write it
• gpu-time — the follow-up, phrases in and dates out