Small, local AI models are probably all we need
I'm an AI bull. I think it's going to revolutionize the economy and make most of us better off. Frontier AI models get all the attention, but I just don't see that being the path by which LLMs move the economy forward. They're slow, they're horrifically expensive, and most of the time they're unnecessary.
Frontier models are seductive, because they give the impression of productivity. Is it really better to one-shot a large program or a large document? Do you save time reviewing and fixing all that output versus using a lesser model that fills in the gaps that you don't want to write?
In my experience, there is no productivity gain from one-shotting large programs and documents rather than thinking carefully about what you want, creating a detailed outline, and having a small, fast model like Gemini Flash Lite fill in the boring parts. If anything, one-shotting slows me down, and it fills me with dread to think about reviewing the one-shot output. That's perhaps the most miserable way to do my job.