sigmoid.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
A social space for people researching, working with, or just interested in AI!

Server stats:

593
active users

The thing I'm most interested in, this morning, are the leaks suggesting that GPT-4 is a "faster" version of ChatGPT. Could be nonsense. But if not ... making these things lighter weight (or able to handle longer contexts at the same weight) would really change the business case and (also, selfishly) matter a lot for academics.

E.g., imagine we had a model that could do reliable summarization of 10,000-word texts and also run on one GPU. A year later the text-analysis part of @dh would look completely different. Everyone would be using that model.

Ted Underwood

This also explains why there has been no rush to publish the GPT-4 paper. An abstractly interesting architecture? Great, publish. But improvements that make LLMs more *deployable* ... you probably wouldn't publish those until you were ready to deploy.