Joseph Suarez 🐡
@jsuarez5341
Followers
26K
Following
7K
Media
694
Statuses
8K
I build sane open-source RL tools. MIT PhD, creator of Neural MMO and founder of PufferAI. DM for business: non-LLM sim engineering, RL R&D, infra & support.
Joined March 2019
PufferLib 3.0: We trained reinforcement learning agents on 1 Petabyte / 12,000 years of data with 1 server. Now you can, too! Our latest release includes algorithmic breakthroughs, massively faster training, and 10 new environments. Live demos on our site. Volume on for trailer!
32
105
987
Hey, that's my first paper! I believe I was 19 at the time. I read a few papers, put two that I thought were complimentary together, got SOTA on a small task, and did a little bit of interesting analysis
to my surprise a whole lot of you are interested in doing research… worry not my dear guys in this 18min video featuring @jsuarez5341 I’ll show you: - the full list of prerequisites to get started - a walking fish - joseph insane dedication to RL - 2 meditation opportunities
5
21
523
$4B Lost in Memecoin Scam. Tokens like MELANIA & LIBRA exploited celebrity names to lure investors into a pump-and-dump scheme. Thousands lost billions. Read the full story.
73
110
321
Exabytes of training data or you're not really trying
Exawatts or you’re not really trying
4
2
66
Taking a vacation. Back to finish Puffer 4.0 soon! Pulled 405 raw on the last stream for now. 172 lbs 5'9
2
0
46
Biggest admin mistake I made with PufferLib 3.0: allowing PRs to current stable release. You need WAY more testing for rolling releases and stuff will break anyways. Bug fixes only starting with 4.0, new features go straight to dev.
5
0
85
Still one slot open! If your company is using raylib, consider sponsoring to support development and to be featured to thousands of low levels devs
Hey! #raylib has a new platinum sponsor: @comma_ai comma just released comma four, the new active driver assistance for your car! 🚀 comma is powered by the open source software openpilot and introduces an entirely new user interface, using raylib! 💯
3
2
95
Puffer 4.0 will be migrating to a combination of @wandb for real-time monitoring and our own new tools for aggregate visualization
7
2
116
When Medicare payments to physicians fall behind the rate of inflation, everyone feels the strain. Learn how these cuts affect you.
0
3
7
@KyleMorgenstein @jsuarez5341 many papers / codebases in robotics (and many game rl libs/envs too) don’t do this actually. Even less if not close to 0 for the codebases I looked around that were used for manipulation. the only instances of this we have found were - Original RSL RL library code ever since the
github.com
Unified framework for robot learning built on NVIDIA Isaac Sim - isaac-sim/IsaacLab
3
1
13
Dan started contributing to pufferlib with 0 RL, ML, AI... or well, programming experience at all. The Enduro environment and agent on https://t.co/T2xjeFiMO1 are both his work. This was way too big of a project for a beginner, but he did it anyways!
a small nation worth of ai beginners have asked me how they should go about getting hired in top places and a common advice is to contribute to open source project… so in this 14min video featuring @DanAdvantage I’ll show you the big secret to do that with ZERO experience
13
45
1K
1. Nice work as usual Stone 2. ... Are you shitting me roboticists. You were not doing this?
3
4
170
Playing with TinyGrad kernel gen on stream rn. Ilya podcast later
1
2
76
Puffer Drive is our biggest academic collab. 15x the perf on a 3rd gen sim. Go chat with Daphne about her lab's latest work!
I'll be at NeurIPS from tue - sat. Reach out if you'd like to chat about simulation, multi-agent RL, world models, scientific discovery, or any other fun topic
5
16
283
Been working on this for months. We built RL agents that autonomously price compute and energy in data centers, trained with @jsuarez5341 and the Puffer framework. Today, we're open-sourcing it.
Manual compute pricing is leaving millions on the table. So we built data center agents that price compute and energy for you, trained on a simulation we developed with @jsuarez5341 and Puffer, the fastest reinforcement learning framework in the world.
1
2
21
Puffer builds high-perf RL sims professionally. Most aren't open source, but this one is! Check out Arkhai's new compute marketplace sim. Trains for ~100M steps in 1 minute on a 4090. More coming soon!
Manual compute pricing is leaving millions on the table. So we built data center agents that price compute and energy for you, trained on a simulation we developed with @jsuarez5341 and Puffer, the fastest reinforcement learning framework in the world.
7
5
88
> shader god does awesome thing > hur dur see react good Okay, maybe LLMs will replace some devs
2
0
58
📣 PostgreSQL developers: Meet Azure HorizonDB the new AI-ready database service with native AI capabilities built in. For enterprise scale, security, and performance without compromise.
1
5
42
For the people saying this is expected... I've been writing torch cpp lately. The initializer syntax looks almost identical. Plus torch encourages chained ops all the time (e.g. .sum().max()). If your "antipattern" looks like your intended usage elsewhere, it's a design flaw.
The second one is way faster. Apparently torch will move gradients back to CPU or something. Thanks, great way to spend an hour of my Saturday night.
8
3
142
First and second impressions on TinyGrad are pretty good. It does a lot of things right. My initial guess of it being Jax that doesn't suck was about right. There are some things missing and somethings I don't like of course. Article early next week.
8
5
313