OK, this is exciting: we now have four alternatives with benchmarks that put them in the same class as GPT-4 - up from zero contenders less than a month ago
@simon it seems like none of these are self-hostable. I wonder how long it'll take to get that type of LLM running locally, and just how much compute will be needed.
@nogweii yeah I've been wondering about that a lot - I don't have a mental model at all for what kind of computer these things need
I wonder what the most powerful model you could run on an M3 MacBook Pro with 196GB of RAM would theoretically look like, if it could use the RAM with the GPU and cores on that thing?
@simon If I understand correctly, one problem with understanding GPT-4 is that it's more opaque than previous models, a lot less has been revealed about how it was trained or tuned or how it is structured. Is that right? Is the story better for these newer models?
@not2b that's a great question - I've not spent the time with the technical papers for these new models to see if they provide significantly more detail than the GPT-4 paper did
@simon Any examples of significant superiority over GPT-4? The benchmarks in their official announcement were only a marginal improvement, and the difference in cost is not marginal at all.
Add comment