Scott Reed
1,587 posts
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- Replying to @jeffcluneIt 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.
- 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.
00:00 - Replying to @scott_e_reed"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"
- Replying to @scott_e_reedIf anyone wondered what happened to Gato2, gpu game of thrones is (at least partly) what.
- Replying to @scott_e_reedAn 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!
- Really happy to collaborate with Angela Dai et. al. on 3D scan completion! Using progressively-trained 3D autoregressive models: arxiv.org/pdf/1712.10215…
- 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.
- 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 𝐋𝐀𝐏𝐀 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
00:00- Replying to @andrewgwils @gruver_nate and 2 othersNice work! Also related to arxiv.org/abs/2307.04721
- Very elegant way to combine action and image diffusion for different objectives. Nice paper!Replying to @chuning_zhuDuring inference, UWM can generate samples from the policy, forward dynamics, inverse dynamics, and video prediction model by controlling the modality-specific timesteps. (5/11)
00:00
