AI Research. Previously @mila_quebec, @mcgillu, ElementAI, @msftresearch and @ias_tudarmstadt.
Joined February 2014
- Tweets65
- Following620
- Followers526
- Likes667
I made this transformer variant explorer. It goes down to the most basic math and covers variants like MLA, MoE, RoPE, MTP, hybrid attention. Link and more below ↓
1
2
324
simonramstedt.com/tools/tran…
It's an interactive reference. Everything is broken down into simple scalar operations with explicit indices.
It took some time to get the ui right but seeing multiple levels of abstraction at once like this is valuable, imo.
1
179
Today @kshitijkgupta, @DanielZKaplan, @Hessian_AI and I are releasing new vision language models based on the excellent @MistralAI's Mistral-7B and @teknium's OpenHermes-2.5:
github.com/AGI-Collective/Ro…
We reproduced LLaVA, then added vis. enc. finetuning, SigLIP & Mistral support
Full team: @kshitijkgupta, @DanielZKaplan, @simonramstedt, Alexis Roger, @CluelessAndrew, Edwin Fennell, @GeorgeAdams101, Prateek Humane, @QuentinAnthon15, @rishika2110, Sun Qi, Yuchen Lu, @irinarish (PI)
Compute was generously provided by @Hessian_AI and @OLCFGOV
A German-Canadian delegation stopped by
@Mila_Quebec yesterday. We're only a couple of Germans at Mila and they wanted to say hi and chat with us for a bit!!
📷: Paul Chiasson/The Canadian Press
.@Bundeskanzler Scholz and I often talk about how our two countries can work together. Today, @MILA_Quebec showed us what that can look like in AI – with real-world applications – and how that partnership can create good jobs and new opportunities for both Canadians and Germans.
ALT Prime Minister Justin Trudeau and German Chancellor Olaf Scholz are standing beside each other and smiling. They are speaking with a group of young people, who are standing in front of them and around them.
Having trouble staying focused? Try my new Focus Chrome extension 🤓 It hides feeds and recommendations from Twitter, Youtube, etc. while still keeping the sites usable for research.
Install: chrome.google.com/webstore/d…
Code: github.com/rmst/focus (single file, easy to mod)
9
Happy to release Uniton github.com/uniton-dev/uniton! It allows you to control and capture almost everything in any Unity project from Python without having to write C# or recompile 🧵
1
8
Or you can easily built your own environments as in this demo video youtu.be/7BHYa1Ycb-A
1
Does @OpenAI loose money on GPT-3 inference?
I suspected they were but @ethancaballero disagreed. We ended up doing some back-of-the-envelope calculations. Looks like OpenAI is actually doing pretty well.
I was surprised how cheap inference is even with huge models 🧵
1
6
The hardware cost is more significant. The DGX 320GB costs 200k$ [5]. Let's assume the 640GB version costs 400k$.
However, even then, it would only take 6.6B tokens or 10 days of running GPT-3 to recover the hardware cost.
[5] anandtech.com/show/15801/nvi…
1
1
Was glad to see @ID_AA_Carmack still being so involved with VR. Hopefully soon we won't be glued to our screens but the screens will be glued to us. And not just for games but for productive work.
Big question: Will we still use keyboards then?
youtube.com/watch?v=sXmY26pO…
2
Our paper on Reinforcement Learning with Random Delays has been accepted to ICLR 2021!
- paper + reviews: openreview.net/forum?id=QFYn…
- code: github.com/rmst/rlrd
Joint work with @yannbouteiller (equal contribution) and our advisors @jumpjoe78, @chrisjpal and @jjbinas!
Simon Ramstedt retweeted
Everything is a POMDP (and you can solve POMDPs with memory)
Our new paper, MELD, shows that meta-RL can work with real-world robots from images. Meta-RL corresponds to a POMDP problem, and such POMDP problems can be solved with latent state-space models. We use the SLAC model directly to tackle meta-RL.
sites.google.com/view/meld-l…
more ->
Reinforcement Learning with Random Delays arxiv.org/abs/2010.02966 - Check out our new paper with Yann Bouteiller, @jumpjoe78, @chrisjpal and @jjbinas!
