@UnitreeRobotics Who will kill faster?
Option 1. Human Driver
Option 2. Self-driving car
Option 3. Legend Guardian (Robot)
Option 4. Idk, but I now like cute video generation
Since step-level action-video data with completion signals did not exist, we also built ๐ช๐ผ๐ฟ๐น๐ฑ๐๐๐ถ๐ฑ๐ฒ-๐๐ฒ๐ป๐ฐ๐ต: ~59K videos, 245 tasks, 27 procedural categories.
๐ช๐ผ๐ฟ๐น๐ฑ๐๐๐ถ๐ฑ๐ฒ:
- A ContextPlanner predicts the next atomic action, or a completion token, from the generated visual state
- An Executor, trained on the same step-level demonstrations, renders those actions
- Hierarchical visual memory keeps state over long rollouts
๐ฐ๐ฎ๐ป ๐ฎ ๐๐ถ๐ฑ๐ฒ๐ผ ๐บ๐ผ๐ฑ๐ฒ๐น ๐ฑ๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฎ๐ ๐๐ผ ๐ฑ๐ผ ๐ป๐ฒ๐ ๐ ๐ณ๐ฟ๐ผ๐บ ๐๐ต๐ฎ๐ ๐ถ๐ ๐ต๐ฎ๐ ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ด๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ฒ๐ฑ? More importantly, can it execute that action reliably and recognize when the task is done, instead of generating long after completion?
Much of real-world video is procedural: a clear goal, a sequence of steps, and a final outcome. Think of folding origami, assembling furniture, or cooking a meal, where one mistake affects everything that follows.
@DimitrisPapail@willcb@yacineMTB Pure RL is such a waste of resources and time. It was only successful because of pretraining and long-context pretraining.
In fact, by adding the right first tokens, base models can match RL in reasoning.
Very good blog: sophielwang.com/cues
@ChombaBupe Instability also comes from AllReduce and ring-AllReduce. When thousands of GPUs contribute to a reduction, floating-point addition is not associative.
When we train any system, we want to allow temporary parameter inconsistency in exchange for communication efficiency.
@neiltyson When we use AI, we forget that it is trained (~Overfit) on some data. However, it has become good at retrieving that data (a chunk of the data) and even generalizing for queries.
53K Followers 36 FollowingA high-throughput and memory-efficient inference and serving engine for LLMs. Join https://t.co/lxJ0SfX5pJ to discuss together with the community!
1K Followers 70 FollowingSimulated Environments for Robots.
Robots don't have enough data. So we're making virtual worlds for them to train in!
Brothers: @SaroyaShehryar @AhmadSaroya00
2K Followers 454 FollowingI head the research team at Bodhan AI (IIT Madras), and am a co-founder at AI4Bharat, a research center at IIT Madras for Indian NLP.
25K Followers 748 FollowingML Engineer @huggingface. Building https://t.co/z4nyO4pjVE. @KU_Leuven grad. General interest in machine & deep learning. Making AI more accessible for everyone!
15K Followers 328 FollowingCo-founder and CEO at @ElorianAI.
Ex (director) at Google DeepMind after 12y.
Gemini Data Area Lead.
Now building at the visual frontier.
ๆดๆๅ
6K Followers 847 FollowingMachine Learning Researcher in Tรผbingen at the ELLIS Institute & Max Planck for Intelligent Systems // Working on Safety & Efficiency of modern ML
13K Followers 6K FollowingPartner, Active Capital. Backing AI-native founders building the future of software. Weโre backing at least one new team every month with $500k to $1M checks.
254K Followers 7K FollowingOG GenAI Skeptic; spoke at US Senate. Warned about hallucinations in 2001. Advocating world models & neurosymbolic AI ever since. Author, Marcus on AI & 6 books
45K Followers 276 Following๐ Applied Maths PhD student at DAMTP, University of Cambridge
๐ช Showing others the beauty in Maths, Space, and STEM
๐ง [email protected]
49K Followers 526 FollowingDigital Geometer, Assoc. Prof. of Computer Science & Robotics @CarnegieMellon @SCSatCMU and member of the @GeomCollective. There are four lights.
311K Followers 9K FollowingFounder and CEO @tabul_ai. Creator of @trainxgb. ML ex Nvidia. Data Scientist. Physicist. Catholic. Husband. Father. Stanford Alum. Memelord. e/xgb. AMDG.
13K Followers 290 FollowingItโs Time for Peptides.
Your source of peptide news & insights ๐๏ธ
Buy 99% pure USA peptides โ https://t.co/x0tGnVpzSn
4K Followers 1K FollowingProfessor for CS at the Tuebingen AI Center @uni_tue and affiliated Professor at MIT-IBM Watson AI lab @MITIBMLab - Multimodal learning and video understanding
29K Followers 137 FollowingDirector, @PrincetonPLI and Professor @PrincetonCS. Seeks conceptual understanding of deep learning and large AI models. On leave at @DiscoLoopAI
21K Followers 436 FollowingThe #NSFfunded IceCube Neutrino Observatory is an array of optical sensors located in Antarctic ice that detects the presence of neutrinos.