- Rapid development of OSS models. New models often require new infra, so every model starts off in Python. Predicted there would be more model-as-a-service companies today. There are a few, but they are only shipping Python libraries (maybe a mistake?)
- Organizational inertia. Seems like data teams are still owning the development. The UX isn't there for frontend devs yet.
We have seen some tokenizers make their way to JS. That's a good early signal (and necessary for the next step, running the models). For example, I compiled the sentencepiece tokenizer to WASM for https://chat.matt-rickard.com/
https://everything.intellectronica.net/p/ad-hoc-definition-of-prompt-engineering
Prompted my thinking! Great article
Matt, is it all about Python specifically these days?
Back in February, I predicted it wouldn't be https://matt-rickard.com/why-python-wont-be-the-language-of-llms
So far -- the prediction hasn't been totally correct. So taking the L on that one, at least in the short term.
I'm kind of surprised it matters so much still. Would you revise this prediction now, or does it just push the time line down a bit?
Factors I didn't really account for
- Rapid development of OSS models. New models often require new infra, so every model starts off in Python. Predicted there would be more model-as-a-service companies today. There are a few, but they are only shipping Python libraries (maybe a mistake?)
- Organizational inertia. Seems like data teams are still owning the development. The UX isn't there for frontend devs yet.
We have seen some tokenizers make their way to JS. That's a good early signal (and necessary for the next step, running the models). For example, I compiled the sentencepiece tokenizer to WASM for https://chat.matt-rickard.com/