Henry Prior

carpal tunnel @GoogleDeepMind

[Arthur Douillard @Ar_Douillard]

The DiLoCo team at Google DeepMind and Google Research is proud to release Decoupled DiLoCo, the next frontier for resilient AI pre-training.

Decoupled DiLoCo enables training with datacenters across the world, using heterogeneous hardware, and never halting the system despite hardware failures.

[Lucio Dery Jnr Mwinm @derylucio]

New paper alert: "Latent Space Communication via K-V Cache Alignment" https://t.co/MyYYTvQs7z

We propose a method for Large Language Models to communicate directly via their internal states, bypassing the need for discrete text generation 🧵

[Henry Prior @priorupdates]

@jxnlco bro’s finally moving to manhattan?? 👀

[Henry Prior @priorupdates]

@jxnlco insightful tweet, jason

[Henry Prior @priorupdates]

@jxnlco

[Henry Prior @priorupdates]

@willdepue if you think this is one of the greatest albums of all time, guess what? it is

[Henry Prior @priorupdates]

@_arohan_ Are those initials?

[Henry Prior @priorupdates]

@giffmana @ZeyuanAllenZhu +1!

[Henry Prior @priorupdates]

Scalable complexity is the future of LM evals!

[Kenneth Marino @Kenneth_Marino]

Excited to announce our latest #ICLR work on long-context/relational reasoning evaluation for LLMs ReCogLab!

https://t.co/jRk6kJv4uK
https://t.co/TLOle1HJYr

Work with Andrew Liu, @priorupdates @gargi_balasu @neuro_kim and others at @GoogleDeepMind

[Henry Prior @priorupdates]

@amydeng_ Pretty consistently I’ve found that the best “ML” engineers are the ones who push for the least ML — being knowledgeable enough to persuade people against unnecessary complexity makes the times you do use ML stand out in impact!

[Henry Prior @priorupdates]

@linaeons we already cooking https://t.co/Fk7JhEyswA

[Henry Prior @priorupdates]

I’m at NeurIPS! Reach out if you’re interested in topics like multi-turn dialogue, LLM planning, or synthetic data.

[Henry Prior @priorupdates]

@finbarrtimbers @EugeneVinitsky DeepMind Playroom, now with real babies!

[Henry Prior @priorupdates]

personal update: i’ve joined @GoogleDeepMind as a research engineer this week! excited to help improve large models’ ability to solve real-world problems

[Henry Prior @priorupdates]

@_Dave__White_ Numerical Linear Algebra by Trefethen is my favorite. Great examples, intuitive explanations, and explicit connections called out between a new idea and those previously covered.

[Henry Prior @priorupdates]

@EugeneVinitsky @araffin2 @markus_with_k I really should benchmark it on other tasks, I'll try to get around to that soon and ping here if/when I do

[Henry Prior @priorupdates]

@EugeneVinitsky @araffin2 @markus_with_k The graph in the README is episode return on cartpole:swingup! It's close/slightly worse than SAC. I ran it on a couple of other envs one-off but didn't want to add graphs for 1 random seed and didn't have compute access to run a bunch

[Henry Prior @priorupdates]

@araffin2 @markus_with_k @EugeneVinitsky I have a very minimal implementation in JAX/Flax! https://t.co/2jVGTwShFq

[Henry Prior @priorupdates]

@_Dave__White_ interesting.. I think for the system to resolve to something that looks like a typical distribution of skill you need the likelihood of non-transitivity to decay with rating. will think about this more