super excited to announce our collaboration to the world! here’s the backstory: we started with a shared goal of using AI to generate novel discoveries that would be non-trivial for human scientists to make
my team is hiring for a research associate! we're looking for ambitious people right out of their undergrad who want to get 1-2 years of experience in a fast moving environment before starting a PhD (at or outside of Retro)
https://t.co/PLqXgVLxPW
@Meaningness EY basically told you to predict the next token and take it as a bug report if your epistemology is not helping you predict the next token. If an agent wants to make its Brier score go down eventually it *has* to learn to balance various incomplete systems or it gets stuck.
This is unfortunate because the ostensible benefit of aggregators is matching people with diverse preferences to the diverse menu of options available. But instead we get "cultural mode collapse".
the fact that Apple sets this as default-on and only allows you to turn it off manually from each app’s individual settings implies that they are collecting a ton of data from this
you don’t need dark patterns for settings that aren’t valuable to you
discusses tricks & techniques for model training and inference at scale:
- model compilation
- kernel fusion
- KV-caching
- gradient accumulation
- low-rank finetuning
- sharding & data parallelisation
+ more, go check it out!
https://t.co/3YoZQUEzBU