Michael Matthews

@mitrma

SR @GoogleDeepMind | PhD student @FLAIR_Ox

Oxford, United Kingdom mtmatthews.com Joined December 2011

424 Following 980 Followers 215 Posts

Latest posts 5

Pinned post

Michael Matthews @mitrma May 26

Hindsight Experience Replay has become the ubiquitous method for goal-conditioned reinforcement learning, but leaves open the question of which goal to relabel with.

In this work, accepted at ICML, we propose instead simply Learning Everything All at Once (LEO).
1/

0:29

Michael Matthews @mitrma Nov 2

Cool stuff @rldoeswork

fred Verified account @rldoeswork Nov 2

i trained a 14M-parameter world model on 2D Minecraft in under 10 minutes for $0.1 and turned it into an autoresearch benchmark.
no game engine. the network draws every pixel. play it locally on your phone.
maharishiva.github.io/tinywmcraft/

View quoted post on X

Michael Matthews @mitrma Sep 30

Excellent and informative writeup by Roger - and congratulations to him (and OpenAI) for the first artificial agent to beat Craftax!

Ironic that the original selling point of high SPS and JAX integration was not even relevant 😅

Lots more interesting work to be done!

Roger Creus Castanyer Verified account @creus_roger Sep 29

x.com/i/article/2104…
View quoted post on X

Michael Matthews @mitrma Sep 14

Great progress and writeup on Craftax!

Daphne Cornelisse Verified account @daphnesolves Sep 14

x.com/i/article/2098…
View quoted post on X

Michael Matthews @mitrma Sep 9

Wow... not how I expected Craftax to fall 😵
Next stop @NetHack_LE ?

Roger Creus Castanyer Verified account @creus_roger Sep 9

uuuuh guys... Astra is fighting the final boss in the graveyard 👀

i'm putting together a blog post on how we got here (hopefully by then it will have beaten the game 🚀)

View quoted post on X