[neilv]

True, but I think there's kitchen table opportunity in applications that don't need to do a big training, and that have tractable inferencing requirements.

The challenges I see are: (1) there's a lot of competition in the gold rush; (2) there's a lot of noise of AI slop implementations, including by anyone who sees your demo.

[Bombthecat]

You also can fine tuned LLMs. For that, you don't need big money. You also can pick up a fine tuned LLM and go from there and make it better ( for your use case)