bram
@bwasti.bsky.social
meta ai, focused on efficient inference (and reasoning for the next couple months)
what kinda fancy ML?
the fancy ML i see these days is usually fine-tuning rather than pre-training models - which captures a lot of priors that usually make meh data pretty workable
the fancy ML i see these days is usually fine-tuning rather than pre-training models - which captures a lot of priors that usually make meh data pretty workable
December 19, 2024 at 9:39 PM
what kinda fancy ML?
the fancy ML i see these days is usually fine-tuning rather than pre-training models - which captures a lot of priors that usually make meh data pretty workable
the fancy ML i see these days is usually fine-tuning rather than pre-training models - which captures a lot of priors that usually make meh data pretty workable
i feel like its trending the other way -- you can put 1M in the *prompt* with google's stuff
October 30, 2024 at 12:38 AM
i feel like its trending the other way -- you can put 1M in the *prompt* with google's stuff
oh interesting, whats holding you back from fine-tuning if you’ve got the tokens for it?
I might be biased because I’ve never been bit too bad by the catastrophic forgetting bug
I might be biased because I’ve never been bit too bad by the catastrophic forgetting bug
October 30, 2024 at 12:27 AM
oh interesting, whats holding you back from fine-tuning if you’ve got the tokens for it?
I might be biased because I’ve never been bit too bad by the catastrophic forgetting bug
I might be biased because I’ve never been bit too bad by the catastrophic forgetting bug