This submission is specifically about ARC-AGI-PUB, so that's what I was discussing.
I'm aware that LLMs can solve problems other than coloring grids, and I'd tend to agree those are likely to be more near-term useful. Those applications (coding, medicine, law, education, etc.) have been endlessly discussed, and I don't think I have much to add.
In my own work I've found some benefits, but nothing commensurate to the public mania. I understand that founders of AI-themed startups (a group that I see includes you) tend to feel much greater optimism. I've never seen any business founded without that optimism and I hope you succeed, not least because the entire global economy might now be depending on that. I do think others might feel differently for reasons other than simple ignorance, though.
In general, performance on benchmarks similar to tests administered to humans may be surprisingly unpredictive of performance on economically useful work. It's not intuitive at all to me that IBM could solve Jeopardy and then find no profitable applications of the technology; but that seems to be what happened.