Eugene Vinitsky 🍒🦋
@EugeneVinitsky
Followers
19K
Following
12K
Media
403
Statuses
8K
All politics, science at 🦋. Anti-cynic. Asst. Prof at @nyutandon. Lab website: https://t.co/vrtbcqR11P
Joined October 2018
summer student project presentations are incredible
21
518
10K
We'Re GOiNG tO STAPLe a GrEEn cARD to EVeRY sTem VIsA
84
277
6K
Stackoverflow was such a wonderful resource, got scraped into chatgpt, and now gets no use while OpenAI makes money. Not sure what the answer here is but it feels pretty bad.
214
197
4K
Teaching is so great for sneakily learning all the stuff you didn't quite understand the first go-round.
30
352
4K
Fairly high score NeurIPS submission rejected because “simulator is written in C++ and people don’t know C++” is a fun new low.
88
109
3K
We have 100GB of data that we need to make publicly accessible. What do folks use? S3 seems wildly expensive per download.
324
41
2K
What if we just didn't bold the overlapping distributions
24
166
2K
During my PhD I struggled with the sense that I wasn't doing enough every day. In the hopes of sparing my students this, I wrote up a little expectation setting doc that I've been sharing with potential students; I figured it might be useful to others:
25
161
1K
We’re open-sourcing and arxiving GPUDrive, a GPU-accelerated 2.5D multi-agent driving simulator that runs at over a million FPS. Hundreds of scenes on one GPU means scalable multi-agent planning
19
144
934
Becoming a better researcher is easy. Just spend your time like so:.40% reading papers.40% writing code.40% coming up with new ideas.40% studying textbooks .40% talking to other researchers.
36
37
930
Deeply pleased with the quotes going into the lectures
17
33
850
Dropped out of a physics PhD only to discover I’ve been doing physics all along. Thanks Nobel committee.
10
28
846
We've built a simulated driving agent that we trained on 1.6 billion km of driving with no human data. It is SOTA on every planning benchmark we tried. In self-play, it goes 20 years between collisions.
32
96
885
This page of common pytorch mistakes is pretty invaluable
4
107
826
A PhD is really great for folks that don't yet have the skills they want. For me I exited undergrad knowing:.- basically 0 programming.- 0 optimization.- 0 ML.(curse you physics).but a PhD let me take time and develop those skills.
25
34
800
AI skeptics logging onto twitter and seeing the news each morning
17
54
762
So much of research effectiveness is just good emotional regulation:.- ignoring the deluge of papers and continuing to read despite analysis paralysis.- finding motivation to work even though the amount of work is technically infinite.
4
67
736
The real difference between PyTorch, TensorFlow, and Jax is one has ptrblck and the others don't.
13
44
724
A weird thing I've learned this year: most doctors don't reason about tricky medical issues. They have a tree chart in which the leaf nodes say "lol, guess you have a chronic condition now.".
20
39
663
There is a giant stretch of highway in Tennessee where all the variable speed limits are controlled by multi-agent RL
11
58
621
Since we are doing “are PhD students miserable” discourse again, a reminder that you can avoid this by:.- picking a supportive supervisor.- not entering uncertain funding situations.- only joining if it’s a subject you’re obsessed with.Don’t take a bad job, don’t join a bad PhD.
15
40
512
@dr_alphalyrae It’s partly special packages for engineering that don’t exist in other languages. For example, there’s a robust control package that’s pretty popular.
8
4
498
Hiring researchers and engineers for a stealth, applied research company with a focus on RL x foundation models. Folks on the team already are leading RL / learning researchers. If you think you'd be good at the research needed to get things working in practice, email me.
10
35
525
🚩🚩🚩🚩someone says neural networks work because of the universal approximation theorem 🚩🚩🚩🚩.
25
11
496
You can't make up some dumb story about how this is about Harvard, this is country-wide.
12
8
502
So. I'm going to be an incoming professor @nyutandon in civil engineering in 2023! Likely be working on multi-agent learning and RL for mixed human-autonomous systems (think autonomous vehicles, zero-shot coordination, transportation systems, etc.).
63
10
465
As PhD rejections start, just to normalize your expectations I went:.0 / 13 my first time applying.3 / 11 the second time.1 / 2 after dropping out of a physics PhD and switching to controls.and honestly my life is so much better for those rejections.
17
29
460
I learned a ton of subtle GPU things that I did not understand before this year from this very clear talk by Stephen Jones:
3
59
471
I genuinely don't care if this current era of deep learning leads to AGI or not; the tools we're building are astonishingly useful.
16
37
448
If you are concerned about leadership at OpenAI and inexplicably want to do a PhD, we’re hiring.
13
13
449
A lot of blog posts on research vision, less so on the day to day of managing your research, experiments, and productivity. I’ve repeated this to students enough times so I’ve written it up
13
71
458
An illustrated guide to learning absolutely nothing
8
48
453
Reading a paper from a more "principled" subfield of ML and couldn't help but think of this
5
19
452
Introducing Nocturne, a simulator for partially observed multiagent coordination. The challenge: to build human-level decentralized agents. Paper: Code: w/ (equal coauthors*) @nathanlichtle*, @BIT_silence*, @brandondamos, @j_foerst
8
77
449
PhD application advice: emails of the form “I admire your excellent research in <NAME OF FIELD> and enjoyed your paper <NAME OF PAPER>” are ineffective and a waste of your own time! If you’re going to cold email, at least do it properly:
5
56
415
Lyft deployed tabular, online RL for driver matching and made millions off it:.
14
54
384
This (from is a really succinct summary of why I mostly don't read new deep RL papers
6
38
375
@mikewawszczak I don’t think these people deserve good faith assumptions at this point. It is more reasonable to assume they will try to extend this.
8
2
388
Today's my last day at Apple and have to say, if you're looking for a place that lets you make wildly ambitious research bets, Apple is hard to beat.
8
3
354
Don’t let anyone convince you that you *should* expect to be depressed during your PhD. Intellectual struggle isn’t the same as depression. depression is a symptom that something is wrong.
7
21
318
A short list of tips for keeping a clean, organized ML codebase for new researchers:
7
42
362
Thrilled to announce the first annual Reinforcement Learning Conference @RL_Conference, which will be held at UMass Amherst August 9-12!.RLC is the first strongly peer-reviewed RL venue with proceedings, and our call for papers is now available:
4
55
337
Interesting to revisit “deep reinforcement learning doesn’t work” six years later It really was like this! Everything was hard and bad and now we can train amazing agents in minutes or hours.
8
32
327
I can't stop watching this, NeRFs are getting out of this world.
Finally accessing Mirror Teleportation and Dimensional Storage superpowers. Look at how the reflection in the mirror has been reconstructed and synthesized - I'm able to get in the reflection! NeRF / Instant-ngp getting out of control...#AI #NeRF #synthesis #volumetric
6
30
313
Human-level performance without search at Stratego, an imperfect information game, using regularized model-free RL plus some clever dynamics? This is completely wild. Haven't finished reading it yet but this is a HUGE accomplishment
5
49
298
Some personal news:.1. First week at Deepmind! It is genuinely a magical place. 2. Moving to NYC, the greatest city in the US, in March. If you're in the area and want to chitchat, always psyched to meet people (w/ 6 feet of space).
10
4
298
@NathanpmYoung Say I give some money to scientists I employ, occasionally check in on how they're doing, and yell at them when they're not going fast enough. The scientists produce great research. Am I doing science?.
15
0
260
Recruiting PhD students this year to the EMERGE lab! The unifying theme is seriously scaling our capability to solve real, hard decision making problems in multi-agent autonomy and transportation w/ RL, LLMs, whatever works
4
44
283
Honestly love having undergrads in the lab (particularly the ones we currently have!) so much.
1
0
284
Alternative grad success metrics besides paper writing:.1. graduating.2. not graduating and leaving because you found a cool job.3. getting really good at programming.4. Not writing a single paper for 4 years because you were thinking a lot and now you have such deep thoughts.
10
13
277
Interested in multi-agent reinforcement learning and its applications to autonomous vehicles, transportation, and robotics? I'm looking for PhD students @nyutandon for Fall 2023 so reach out!.
8
56
272
95% of the good opportunities in my life are direct outcomes of cold-emailing someone rather than fortuitous accident. Research projects, mentors, grad school admissions, etc. If you're not contacting people you respect and want to work with, you should try it.
2
25
267
Fun fact, the expression for the policy gradient is wrong in almost every algorithm, this isn't widely acknowledged, and the expression that we use isn't apparently the gradient of anything?.
6
24
265
RL struggles as a field because it does not yet have one common, simulated, easily reproduced benchmark where progress would immediately improve an exciting application.
34
6
253
It's 2017, I'm a new PhD student, debugging coordinate frame errors.It's 2024, I'm a new professor, debugging coordinate frame errors.You can't escape.
6
8
246
Today's my last day as a visitor at FAIR. In the fall I'm starting as a research scientist at Apple in Vladlen Koltun's group, then starting as an assistant prof at NYU in Fall 2023.
16
3
253
This guy is out here making some of the only really hand-holding game-solving videos on the internet: It's really helpful as an intro!.
5
27
250
Are you an RL PhD at Harvard who has had your funding wrecked by the government and working on topics related to multi-agent? Reach out, I am happy to try to find a way to support you.
5
18
264
It's not sensible to ask empirical papers to provide fake theoretical explanations of why they worked. The empirical finding is inherently interesting.
9
12
249
Trying to reproduce some deep RL results from scratch
7
10
252
There’s definitely a quiet revolution in RL brewing up. Folks will be surprised at all the things that are suddenly possible.
17
4
239
Think it'd be fun to run a "your ML method doesn't work" conference where only papers that propose simplifications, find mistakes in prior work, or show that a method isn't robust can be submitted.
8
14
234
In our new paper, we find that LLMs can efficiently do RLHF in-context! .Our method, in-context preference learning (ICPL), iterates LLMs writing reward functions, training agents, and putting preferences into context. We see a 30x boost in query efficiency over baseline RLHF!
2
25
225
Andrew Barto coming out of retirement to give an amazing history of RL
11
5
233
Hoping that the inevitable collapse of the LLM market doesn't take the rest of AI down with it.
29
11
225
This snippet of text, from when Diplomacy bot Cicero (France) went down and had to explain its absence, will live rent-free in my head forever.
6
28
230
Kinda wild that the only thing that has *really* made RL work in robotics problems is to turn it into dynamic programming.
10
10
221
As PhD applications season draws closer, I have an alternative suggestion for people starting their careers in artificial intelligence/machine learning: . Don't Go to Industry for Machine Learning ❌. (or, at least, not right now). 1/4 🧵.
7
21
220
Just discovered Sanjeev Arora's lecture notes for "A Theorists Toolkit": It's such a wonderful concept for a class, a concise summary of useful mathematical tools.
3
44
226
Neuroscience-washing is when you vaguely gesture at a neuroscience concept to justify your heuristic ML algorithm.
11
15
219
Can't understand how this paper, showing that given a reasonable amount of data and a "simple" self-play objective yields superhuman diplomacy play does not have 1000 citations.
11
14
226
We deployed 100 deep RL cruise controllers into rush-hour highway traffic to smooth traffic flow and reduce everyone’s energy consumption. Our AVs were decentralized and used standard radar, making our controllers deployable on most cars.1/n
9
26
221
I’ll be voting for Kamala. Trump will make things worse on fundamental issues I care about:.1) he will make it hard for international students to come here and to visit their families.2) he will make life worse for trans students.3) he will attempt to deport children of.
7
6
217
My wife is using LLMs to understand the tax code so we're either about to save a bunch of money or go to jail.
8
4
218
I don’t really care if papers are rejected, game is the game, but this is just really shocking.
1
1
212
Some personal news: I've upgraded from pretty lucky person to luckiest person
13
0
216
Turns out research is more fun when you’re not stressed? What a concept.
6
7
208
@moultano @mattyglesias This is an obvious social good; I genuinely don't understand the criticism. Why would we want it to be easier for people to run red lights or block bike lanes?.
3
1
196
The whole system is simply not set up for any type of investigative work. If I hadn't found a good physical therapist who was willing to think things through, I'd be in dire straits now.
1
2
200
You: RL is too sample inefficient to ever be useful.NVIDIA researchers: what do you mean.
We are excited to share Isaac Gym tech report Physics simulation data is directly passed to pytorch without ever going through any CPU bottlenecks in the process allowing blazing fast training on many challenging environments. @viktor_m81
3
15
201
Happy ICML rejection day to all who celebrate the holiday.
5
6
206
Research rabbit is incredible, an almost instant boost over using google scholar for paper finding.
10
8
210
Life update!.Excited to spend the next year or so as a visting researcher at @facebookai in NYC working with @brandondamos on reinforcement learning!. It has been a wild few months 😀.
10
2
201
I'm happy to announce that I went 0/2 on ICLR papers, that this is totally fine, and that most of life is failures spotted with occasional successes and not the reverse.
3
2
200
"This technology is so dangerous we must slow down open source reproductions" and "we are actively developing and commercializing new versions of this" feel like incompatible positions. What makes folks at OpenAI uniquely wise and capable enough that they should be trusted?
19
21
194
This excellent tutorial on implicit differentiation and implicit layers (neural ODES, differentiable optimization, etc.) really brings together the techniques, upsides, and downsides in a clear way:
4
35
189
As an undergrad I asked a faculty "why are exams so time constrained, why not just give double the time needed?" and he said "what do you mean, that's what I do".
5
3
192
Want to highlight this wonderful example of how to do experimentation in RL appropriately: I (and I suspect many others) definitely knew very little in the way of good stats practices when getting started in RL. This paper is a huge help.
4
37
195
Huge fraction of all written text as its training data and GPT-4 still makes basic reasoning mistakes; kinda makes you suspicious about imitation learning as a robust approach in other settings e.g. driving, robotics.
18
8
185
Last year NeurIPS gave an extension for folks affected by the US policing protests, a great call IMO. Might be good to have such a plan in hand for countries w/ an unchecked COVID epidemic.
3
17
186
The best interpretation of the bitter lesson is "don't pursue things that can't scale" not its more common interpretation of "don't think, just scale".
4
13
180
Closing Deepminds Edmonton office is such terrible news both for the people who work there and for research. That place has consistently generated some of the best multi-agent research.
4
6
176
Folks were asking for all the @RL_Conference keynote talks. Well, they're all up now! Talks from.- David Silver.- Andrew Barto.- Emma Brunskill.- Finale Doshi-Velez.- Sergey Levine.- Peter Stone.
2
22
181