Scott Reed
@scott_e_reed
Followers
17K
Following
18K
Media
9
Statuses
1K
Research Scientist at NVIDIA working on generalist embodied agent research
Joined December 2015
True. A lot of groups gave up prematurely, or allocate ~all resources to one giant model. This leads people to spend more effort on winner-take-all gpu politics and less on just training the best models they can with moderate resources.
Hot deepseek take: before r1 blew up, a ton of western AI (and robotics!) efforts -- startups, big companies, and even academic labs -- were basically just waiting for openai to solve all their problems and it was honestly kind of sad. I hope r1 changed that.
8
10
164
@jeffclune It is a vote of no confidence in the old regime and its media organs, and an expression of hope for building a great future for America and the world.
7
1
73
Very cool idea: make the diffusion policy denoising process part of the MDP and train the whole thing with PPO.
We had a great time at the Mastering Robot Manipulation workshop at @corl_conf on Saturday! If you want a (very) short intro to DPPO, here's the 5-ish minute presentation we gave at the workshop.
1
12
73
"Um This scaling law model I made says [the world will end / company will die] if you dont give me all the GPUs and block any other team from pretraining". "No, fuck you, I will train my own model".
1
4
53
If anyone wondered what happened to Gato2, gpu game of thrones is (at least partly) what.
2
2
41
An interesting counterfactual was the Genie project, which was stubbornly cobbled together mainly out of pooled user quota. This kind of stubborn independence can lead to cool results!.
2
2
38
Really happy to collaborate with Angela Dai et. al. on 3D scan completion! Using progressively-trained 3D autoregressive models:
0
4
36
More established people tend to be less harsh in reviews. I remember a discussion in our class during phd at umich, all the first year grad students trashing every paper. Our professor Ben Kuipers advised: "Find the gold". Even flawed papers can contain great insights.
ICLR implemented a new rule this year, requiring authors who submit more than three papers to serve as reviewers. For the first time, I found many renowned professors in the reviewer pool. Interestingly, their scores tend to be higher than those of other reviewers.
1
0
26
Congrats! Cool to see that latent actions are not only useful for interactive world models (as in genie) but also as targets for self supervised learning.
Excited to share that ๐๐๐๐ has won the Best Paper Award at the CoRL 2024 Language and Robot Learning workshop, selected among 75 accepted papers!. Both @SeonghyeonYe and I come from NLP backgrounds, where everything is built around tokenization. Drawing inspiration from
0
4
24
0
1
21
Very elegant way to combine action and image diffusion for different objectives. Nice paper!.
During inference, UWM can generate samples from the policy, forward dynamics, inverse dynamics, and video prediction model by controlling the modality-specific timesteps. (5/11)
0
5
21
Telling my kids to "Think step by step" has been surprisingly effective as we go through the beast academy math books.
Good insight here. โLetโs think step by stepโ is the preamble to the closest thing we routinely say to a live transcript of thought, but it isnโt one. A Reddit reply starting that way doesnโt backtrack to fix a missing minus sign. We donโt want words; we want what picks them.
1
1
15
Amazing !
Unsupervised (!!) Image-to-Image Translation Networks #pix2pix without need for matching examplesโamazed it works!
0
4
11
Exciting work by @konradzolna, making adversarial imitation perform much better for simulated robot manipulation tasks, from pixels.
I am fortunate to have interned @DeepMindAI where we extended GAIL to make it work better for robot manipulation tasks. Task-Relevant Adversarial Imitation Learning (TRAIL) robustly learns policies and rewards from pixels.Paper Demo
0
0
10
Glad to see this; I am a fan of LeRobot.
Hugging Face and NVIDIA are teaming up to advance robotics. ๐ค. By combining @HuggingFace's LeRobot with @NVIDIAAI, @NVIDIAOmniverse, and #robotics technology, we're enabling researchers & developers to innovate. ๐ #CoRL2024
0
0
8
@zipengfu @tonyzzhao @chelseabfinn Amazing results - in retrospect very simple and elegant. Do you think such a result would have been possible 10 years ago if we had this type of hardware setup?.
1
0
6
@j_foerst I'm curious whether you can make this work from pixels directly. For example use the state based agent to generate many trajectories, then train from pixels using bc or offline RL. Would you see in context learning of new morphologies emerge?.
3
0
7
Restricting pure labor arbitrage is the technologically accelerationist position.
1
0
4
Has anyone published a thorough evaluation comparing MaskGIT versus diffusion?.
2
0
4
@_akhaliq Nice work. Looks like same strategy as our 2017 ICML paper but with modernized architecture
0
0
4
@IanOsband +100. Just used grok to tell me how to configure ssh and firewall on gpu workstation and connect from my laptop. Saved probably several hours of frustration at least!.
1
0
4
@DrJimFan We were able to run 1.2B gato model on rtx 3090 at about 15hz to control sawyer arm. So it may be just a temporary hardware limitation.
2
0
4
Seeing those wavenet results for the first time at the research team meeting was quite mind blowing. At that time the feeling was that even policy "imitation learning" was a bit uncool for not learning purely from reward.
I was working on autoregressive models around that time, but instead of RNNs and language modelling, we were trying to make convolutional nets generate music by producing 16000+ audio waveform amplitude values per second. No regrets๐ (Never got tempted by RL๐คญ).
0
0
3
I agree with that framing. @wrathofgnon (I think) has a related saying: your life is a gift from your ancestors that you pass on to your descendants.
Modern culture cares deeply about the social contract with others around us, and almost not at all about the generational contract with our ancestors. They made sacrifices far beyond what parenting requires today, to continue lineages weโre now giving up on with barely a fight.
0
0
3
LeRobot code and documentation is also very helpful. I highly recommend!.
1
0
3
0
0
3
@danielblawson9 @marktenenholtz That is very cool. I think not having image-specific pos embeddings is better and more general. I would guess that scaling up to 8B vs 1B in Gato helps.
0
0
3
@AndrewYNg @derrickharris @architechtpub Thanks for posting this. I have been disappointed by NYTimes repeated clickbait-style reporting on this topic.
0
0
3
@kaush_trip @alexkoch_ai Just the arms so far. Maybe next weekend add two more arms to do bimanual tasks.
1
0
1
@gdb Could he join full-time as a researcher after graduating high school? I think that would be awesome actually.
1
0
2
@guanzhi_wang @XueFz @YangYou1991 @ZangweiZheng @NiJinjie @frankkklee @Francis_YAO_ @xiangyue96 @YiTayML @m__dehghani @drjwrae @DrJimFan @yukez @AixinSG @hzhang26 Congratulations!!.
1
0
1
Need a way to clean this type of thing from LLM pretraining data at scale.
0
0
2
Seems clearly true in the long term. Also worth studying how and in which cases industrial policy has gotten great results in China and elsewhere, and apply those lessons.
Nothing is more expensive than โcheapโ goods. The price of cheap goods from China are factories never built in the US, industrial technologies never discovered, younger generations that look down on factories, deep process knowledge that dies out, and dependence on an adversary.
1
0
2
Progress!.
Last year, $9B of the $35B that the National Institutes of Health (NIH) granted for research was used for administrative overhead, what is known as โindirect costs.โ Today, NIH lowered the maximum indirect cost rate research institutions can charge the government to 15%, above
0
0
0
@yoavgo Now more of your grant can be wasted on administrative bloat (more consultants). Maybe funding agencies could disincentivise this somehow.
0
0
2
@wrathofgnon @Cindy_c05 Seems important: โPeople went through a big mental change . ,โ Oku said, โbecause we had to survive as a town.โ Beyond financial incentives, people need a shared purpose, so that raising a family is viewed as valuable and good.
0
0
1
@egrefen @rsalakhu @awmcmu @CRAtweets also maybe bloated university bureaucracy and confiscatory administrative overhead?.
1
0
2
We make it look easy ๐.
the weight that Software Engineers carry on their back. no one knows the sacrifice that we make. silently working, the world relying on us, for nothing more than life changing income in an air conditioned office . .
0
0
1
@amycatalinac @Patrick_J_Egan I would be surprised to see that in academia, but maybe some industry labs would have that benefit.
0
0
1
@ZhangLunjun One advantage of maskgit may be better representation learning, and maskgit models may be more readily suited to be used as encoders of clean (non-noised) data.
0
0
1
You can turn negative energy from twitter arguments into reps; I found this very effective. Is this what our gym bros are doing on their phone between every set??.
2
0
1
Directionally correct, but not within 3 orders of magnitude of what needs to happen.
Speaking to a number of constituents today, the message is overwhelming - deport foreign child rapists. We must say what needs to be said, even if some in Parliament are offended. I want these savages out of our country, and rotting behind bars in Pakistan. That is the aim.
0
0
1
๐.
America has just been through a (Mao-style) Cultural Revolution. On precedent, truly great things happen after they're shut down.
0
0
1
Gwern's website is great! Worth supporting.
After the interview, I convinced @gwern to make a donation page so that people can help sustain what heโs up to.
0
0
1
๐.
No more โclimate changeโ bs, we have the highest density of universities and engineers per capita in Bay Area.
0
1
0
This encourages automation and industrial capacity, and raises the standard of living for everybody.
1
0
1
Also a major factor contributing to the North winning the civil war.
0
0
1
Text rendering of image generation models has gotten pretty good in the last year!.
1
0
1
@NandoDF In the long term, better job market for UK nurses, leading to better wages and conditions, strengthening the domestic training pipeline.
2
0
1