[satisfice]

The fad part is that LLMs don’t work. People keep mistaking simplistic demos for practical applications.

They don’t provide good summaries. They don’t help non-experts simulate expert work. They don’t provide reliable search results.

If someone promoted a calculator that gets 90% of the digits correct in its answers and 90% of the time those digits were in the right order, that would be a useless calculator.

I have not spoken with any AI fanboy who can substantiate his claims about the usefulness of LLMs. I have used Deepseek twice, now, and both times its results were unusable for engineering purposes but would have impressed tipsy people at a party.

I have heard credible reports that Co-pilot is helpful. And I routinely use ChatGPT for prototyping tools— but that makes it one more interesting tool, not a revolution.