As someone who had been rabidly following developments in the space until earlier this year: I’ve fallen a bit behind.

I’be been walking around with the (seemingly misguided) view that the bigger LLaMA 2 models were sort of neck and neck with like 3.5-turbo, Claude, and the rest of the first string modulo GPT-4 that had a pretty steady if eroding lead.

What’s the lay of the land in 9/23? Or better yet, what do I need to bookmark to get the inside baseball?

Don’t worry, GPT-4 is still the best.

Water cooler chat is that it’s a 1.5 terraweight, 16-way mixture model that is running at a loss on inference for paying customers.

If that’s true it fucking better be.

This should not have stopped any of FAANGs from competing, and yet they are still far behind OpenAI - more than a year after GPT-4 was trained.