正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

We bust some common myths in visuo-motor pretraining! Our data-centric experiments show: 📷 ImageNet > Ego4D for self-supervised robotic pretraining 🏆 Diversity matters more than content: training on a broader image distribution beats SOTA by 30% 📊Sim evals can mislead! 1/n

Sudeep Dasari

1,479 subscribers

54,350 次观看 • 2 年前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

11 条评论

Sudeep Dasari 的头像

Sudeep Dasari2 年前

Our methods are simple and standard: we pretrain ViT models using the Masked Auto-Encoders (MAE) on *different datasets.* Then, we fine-tune them (via Behavior Cloning on 50 demos) to solve various real/simulated manipulation tasks. 2/n

Sudeep Dasari 的头像

Sudeep Dasari2 年前

Surprisingly, the models trained on old-school, curated vision datasets (blue), outperform Ego4D/RoboNet (red), and SOTA baselines from the robotics field (gray). Boosting diversity by balancing data sources during MAE pre-training (green), improves performance by 30%! 3/n

Sudeep Dasari 的头像

Sudeep Dasari2 年前

Why did previous work miss these trends? One reason could be the over-emphasis on sim evaluation. We empirically show little to no correlation (R^2=32%) with the real robot trends. 4/n

Sudeep Dasari 的头像

Sudeep Dasari2 年前

This was a fun collaboration with @mohansrirama, @unnatjain2010, and Abhinav Gupta! For more, come to our poster @corl_conf and/or check out the links below. Website: Code: ArXiV: 5/n

SecBriefs | Making Cybersecurity Simple 的头像

SecBriefs | Making Cybersecurity Simple1 年前

📖 It's More Than Just an Exam! Studying cybersecurity isn’t just about passing—it’s about mastering the skills that protect our digital world.💡 Learn smarter with 50 tips + Cybersecurity Dictionary for Everyone! 🔑 Available on Amazon:

Yanjie Ze 的头像

Yanjie Ze2 年前

Nice work, Dasari! I am curious whether this work adds proper data augmentation to the Learning from Scratch baseline, since we previously found that this could be one key (just the cited work [56])

Sudeep Dasari 的头像

Sudeep Dasari2 年前

Yes, (building on your work) we do use data aug during fine-tuning!

Hao-Shu Fang 的头像

Hao-Shu Fang2 年前

Nice work, Sudeep! I'm curious if there have been any experiments conducted to assess the significance of action data in visuo-motor pretraining?

Sudeep Dasari 的头像

Sudeep Dasari2 年前

A major issue with using actions during visual pre-training is that robot data is (comparatively) very scarce. However, we do have some past work on more efficiently finetuning representations w/ a few expert demonstrations:

Oleg Rybkin 的头像

Oleg Rybkin2 年前

Nice work! Can't find the comparison to r3m?

Sudeep Dasari 的头像

Sudeep Dasari2 年前

Thanks Oleg! Sorry for the confusion, that comparison is in Appendix D of our paper (page 15; link below). TLDR: We found that r3m really struggled in our experiments, and that the main trends (outlined above) were replicable using ResNet + SimCLR!

相关视频

NVIDIA just announced EgoScale 🤖🧠 NVIDIA Research has uncovered a log-linear scaling law for robot dexterity by pretraining VLA models on over 20,000 hours of egocentric human video This massive dataset is 20 times larger than previous efforts and proves that robot intelligence follows a predictable path: the more human data, the lower the loss The secret is a simple recipe combining large-scale human pretraining with a small amount of aligned human-robot mid-training to bridge the gap In testing, this method boosted the average success rate by 54% on a 22-DoF robotic hand compared to policies built without pretraining EgoScale also enables one-shot task adaptation and works across different hardware, suggesting that human motion is a universal motor prior for robots Website: Paper: Source: NVIDIA Research #Robot #Humanoid #Robotics #AI #EmbodiedAI #PhysicalAI #NVIDIA #EgoScale #GR00T

NVIDIA just announced EgoScale 🤖🧠 NVIDIA Research has uncovered a log-linear scaling law for robot dexterity by pretraining VLA models on over 20,000 hours of egocentric human video This massive dataset is 20 times larger than previous efforts and proves that robot intelligence follows a predictable path: the more human data, the lower the loss The secret is a simple recipe combining large-scale human pretraining with a small amount of aligned human-robot mid-training to bridge the gap In testing, this method boosted the average success rate by 54% on a 22-DoF robotic hand compared to policies built without pretraining EgoScale also enables one-shot task adaptation and works across different hardware, suggesting that human motion is a universal motor prior for robots Website: Paper: Source: NVIDIA Research #Robot #Humanoid #Robotics #AI #EmbodiedAI #PhysicalAI #NVIDIA #EgoScale #GR00T

RoboHub🤖

43,752 次观看 • 5 个月前

Meta presents Sapiens Foundation for Human Vision Models discuss: We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing baselines across various human-centric benchmarks. We achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by 7.6 mAP, Humans-2K (part-seg) by 17.1 mIoU, Hi4D (depth) by 22.4% relative RMSE, and THuman2 (normal) by 53.5% relative angular error.

Meta presents Sapiens Foundation for Human Vision Models discuss: We present Sapiens, a family of models for four fundamental human-centric vision tasks - 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning models pretrained on over 300 million in-the-wild human images. We observe that, given the same computational budget, self-supervised pretraining on a curated dataset of human images significantly boosts the performance for a diverse set of human-centric tasks. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic. Our simple model design also brings scalability - model performance across tasks improves as we scale the number of parameters from 0.3 to 2 billion. Sapiens consistently surpasses existing baselines across various human-centric benchmarks. We achieve significant improvements over the prior state-of-the-art on Humans-5K (pose) by 7.6 mAP, Humans-2K (part-seg) by 17.1 mIoU, Hi4D (depth) by 22.4% relative RMSE, and THuman2 (normal) by 53.5% relative angular error.

AK

151,582 次观看 • 1 年前

Every wondered if we can model motion as a language? can we tokenize this new language? is it useful? Turns out tremendously! 🚀 In out latest #NeurIPS2024 paper on QueST: Self-Supervised Skill Abstractions for Learning Continuous Control, we find that action tokenization matters a lot! We can learn skill encodings by representing temporal action abstractions with a discrete codebook. This enables 2 things 1. Better Behaviour Cloning: we can better assimilate multi-task data (>9%) over best paper. This is currently best in class BC method! 2. generalization of this language to represent new tasks in 5-shot transfer to longer horizon tasks! Check out the thread by Atharva Mete for more details. And check out more details at: Joint work with Atharva Mete Albert Wilcox Haotian Xue Yongxin Chen Georgia Tech School of Interactive Computing Machine Learning at Georgia Tech Robotics@GT NVIDIA Robotics

Every wondered if we can model motion as a language? can we tokenize this new language? is it useful? Turns out tremendously! 🚀 In out latest #NeurIPS2024 paper on QueST: Self-Supervised Skill Abstractions for Learning Continuous Control, we find that action tokenization matters a lot! We can learn skill encodings by representing temporal action abstractions with a discrete codebook. This enables 2 things 1. Better Behaviour Cloning: we can better assimilate multi-task data (>9%) over best paper. This is currently best in class BC method! 2. generalization of this language to represent new tasks in 5-shot transfer to longer horizon tasks! Check out the thread by Atharva Mete for more details. And check out more details at: Joint work with Atharva Mete Albert Wilcox Haotian Xue Yongxin Chen Georgia Tech School of Interactive Computing Machine Learning at Georgia Tech Robotics@GT NVIDIA Robotics

Animesh Garg

26,218 次观看 • 1 年前

How can generative AI and Robotics help advance drug discovery？ 🚀 Excited to introduce LUMI-lab! A foundation model-driven Self-Driving Lab (SDL) for autonomous ionizable lipid discovery in mRNA delivery 🤖🔍 🔬 What is LUMI-lab? LUMI-lab integrates molecular foundation models with autonomous robotic experiments to efficiently explore new LNPs (lipid nanoparticles, mRNA delivery vehicles) with minimal wet-lab data. 🔥 Key Highlights: - 🧠 Foundation model trained on 28M molecules using a three-step strategy: - Unsupervised pretraining to capture broad molecular knowledge - Continual pretraining to specialize in lipid-like molecules - Active learning fine-tuning within a closed-loop experimental system - 🤖 1,700+ new LNPs synthesized & tested across 10 iterative cycles - 🧪 Brominated lipids autonomously identified as a novel structural feature that enhances mRNA transfection—an insight previously unrecognized in LNP design - 🏆 20.3% in vivo CRISPR gene editing efficiency in lung epithelial cells—the highest reported for inhaled LNPs 🚀 Why it matters? LNPs are the backbone of mRNA therapeutics, yet discovery has been slow due to data scarcity. LUMI-lab shows that AI-powered autonomous labs can accelerate mRNA delivery innovation🚀💡 🌐 Beyond mRNA drugs, LUMI-lab exemplifies a scalable framework for AI-driven molecular discovery, pushing boundaries in material science & drug delivery. 📜 Read the preprint: 🔗 💻 Code available on GitHub: 🔗 #AI #DrugDiscovery #mRNA #LNP #SyntheticBiology 🙏 A huge team effort behind this work, with special appreciation to Bowen LI for driving the project. Kudos to Haotian Cui, @YueXu1995, Kuan Pang, Gen Li, and Bettycat567!

How can generative AI and Robotics help advance drug discovery？ 🚀 Excited to introduce LUMI-lab! A foundation model-driven Self-Driving Lab (SDL) for autonomous ionizable lipid discovery in mRNA delivery 🤖🔍 🔬 What is LUMI-lab? LUMI-lab integrates molecular foundation models with autonomous robotic experiments to efficiently explore new LNPs (lipid nanoparticles, mRNA delivery vehicles) with minimal wet-lab data. 🔥 Key Highlights: - 🧠 Foundation model trained on 28M molecules using a three-step strategy: - Unsupervised pretraining to capture broad molecular knowledge - Continual pretraining to specialize in lipid-like molecules - Active learning fine-tuning within a closed-loop experimental system - 🤖 1,700+ new LNPs synthesized & tested across 10 iterative cycles - 🧪 Brominated lipids autonomously identified as a novel structural feature that enhances mRNA transfection—an insight previously unrecognized in LNP design - 🏆 20.3% in vivo CRISPR gene editing efficiency in lung epithelial cells—the highest reported for inhaled LNPs 🚀 Why it matters? LNPs are the backbone of mRNA therapeutics, yet discovery has been slow due to data scarcity. LUMI-lab shows that AI-powered autonomous labs can accelerate mRNA delivery innovation🚀💡 🌐 Beyond mRNA drugs, LUMI-lab exemplifies a scalable framework for AI-driven molecular discovery, pushing boundaries in material science & drug delivery. 📜 Read the preprint: 🔗 💻 Code available on GitHub: 🔗 #AI #DrugDiscovery #mRNA #LNP #SyntheticBiology 🙏 A huge team effort behind this work, with special appreciation to Bowen LI for driving the project. Kudos to Haotian Cui, @YueXu1995, Kuan Pang, Gen Li, and Bettycat567!

Bo Wang

27,626 次观看 • 1 年前

Today we're announcing #GAIA1: a 9B parameter world model, trained on 4,700 hours of driving data, able to simulate complex and diverse driving scenes from video, text and action inputs. This model is 480x larger than the preview we shared earlier this year and the results are incredible. These videos are entirely synthetically generated by Wayve's generative AI, GAIA-1. But there is more here than just generating videos, GAIA is an entire world model. A world model allows us to simulate the future, conditioned on video, text and action inputs, which can be leveraged for making informed decisions when driving. Why is this game-changing for autonomous driving? 1. Safety. One limitation with AI systems like today's Large Language Models is that they are autoregressive, next-word prediction algorithms, but aren't necessarily aware of the implications of their decisions. A world model allows us to give our AI the capability to be aware of its decisions, by simulating the future, which is important for self-driving safety. 2. Synthetic training data. I believe synthetic training data is the future for AI, because it is safer, cheaper, and infinitely scalable. GAIA-1 unlocks unprecedented realism and diversity of synthetic data for self-driving. 3. Long-tail robustness. One of the biggest challenges for self-driving is long-tail robustness: dealing with the enormous magnitude of edge cases we see on the road. An advantage of generative AI is its incredible ability to recombine experiences in new ways. This is exciting for self-driving as it means we can learn from two edge case scenarios, and combine them to become a corner case. For example, we can experience driving in fog, and experience of jay-walking pedestrians, and GAIA can learn from these experiences to understand how to generate a fog+jay walking scenario. Check out many more videos in our blog or further technical details in our paper: Or come chat with our team who are at the International Conference on Computer Vision (#ICCV2023) this week in Paris in Booth 32 Jamie Shotton

Today we're announcing #GAIA1: a 9B parameter world model, trained on 4,700 hours of driving data, able to simulate complex and diverse driving scenes from video, text and action inputs. This model is 480x larger than the preview we shared earlier this year and the results are incredible. These videos are entirely synthetically generated by Wayve's generative AI, GAIA-1. But there is more here than just generating videos, GAIA is an entire world model. A world model allows us to simulate the future, conditioned on video, text and action inputs, which can be leveraged for making informed decisions when driving. Why is this game-changing for autonomous driving? 1. Safety. One limitation with AI systems like today's Large Language Models is that they are autoregressive, next-word prediction algorithms, but aren't necessarily aware of the implications of their decisions. A world model allows us to give our AI the capability to be aware of its decisions, by simulating the future, which is important for self-driving safety. 2. Synthetic training data. I believe synthetic training data is the future for AI, because it is safer, cheaper, and infinitely scalable. GAIA-1 unlocks unprecedented realism and diversity of synthetic data for self-driving. 3. Long-tail robustness. One of the biggest challenges for self-driving is long-tail robustness: dealing with the enormous magnitude of edge cases we see on the road. An advantage of generative AI is its incredible ability to recombine experiences in new ways. This is exciting for self-driving as it means we can learn from two edge case scenarios, and combine them to become a corner case. For example, we can experience driving in fog, and experience of jay-walking pedestrians, and GAIA can learn from these experiences to understand how to generate a fog+jay walking scenario. Check out many more videos in our blog or further technical details in our paper: Or come chat with our team who are at the International Conference on Computer Vision (#ICCV2023) this week in Paris in Booth 32 Jamie Shotton

Alex Kendall

631,856 次观看 • 2 年前

PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking paper page: introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks.

PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking paper page: introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks.

AK

122,560 次观看 • 3 年前

Data has always been the bottleneck for physical AI in self driving and robotics. Tesla is taking two very different approaches for FSD and Optimus. Tesla’s Optimus Training Playbook: 1. Build 30k Optimus Gen 3 robots 2. Operate them in a mock environment where they can perform self-play “Optimus Academy” 3. Train in sim using the real robot data to close sim2real gap. Tesla FSD Training Playbook: 1. Sell millions of cars outfitted with cheap cameras 2. Collect diverse real world driving data (especially intervention and failure recovery data) for free as a byproduct of customers driving the cars. 3. Use driving data to train Autopilot/FSD and deploy policies incrementally as a supervised FSD product 4. Repeat until policy reaches robust unsupervised full self driving for robotaxi launch. The Tesla FSD playbook is a beautiful self-funding, customer subsidized, diverse real world data flywheel. The Optimus playbook is the opposite and shares none of the beautiful attributes of the FSD training flywheel that made FSD successful. The key differences: 1. Instead of having customers pay you for vehicles, Tesla will need to fund 30,000 Optimus robots. Assuming the current landed cost per unit is $100k, that will be $3B to build plus another ~30% per year for maintenance labor and spare parts given it’s still an unhardened pre-production prototype is another $900M per year. For reference, Tesla’s GAAP net income in 2025 was $3.8B. 2. Instead of having customers drive their Teslas on roads all across the world giving Tesla an insanely rich and diverse dataset that Waymo and other AV companies could never collect, the Optimus Academy is doing the equivalent of building a fake town in a parking lot and driving their car in that parking lot. No matter how real you try to make the environments for self-play you can never replicate the diversity, complexity and failure modes of the real world. Data collected in staged environments produces demo-grade policies and will not be rich enough to generalize to the vast diversity of environments, tasks, objects, etc. out of distribution. 3. Instead of having customers collect real world failure recovery data (DAgger style) for free every time FSD disengages, the Optimus Academy will need paid teleoperators or onsite operators to collect the recovery data. Assuming 1 person can manage 2 robots to start that would cost $3.5B in labor per year (30,000 robots, $40/hr fully loaded, 16 hrs/day, 365 days per year, 2:1 robot:operator). Tesla can come up with the money to do this but money doesn’t solve the “mock data” problem. Given the higher degrees of freedom in humanoids vs. cars, training a generalized humanoid will be harder and require more data than a self-driving vehicle. The best way to train your robot is by deploying them in the diverse real world, subsidized by real customer operations. Humanoids face a chicken and egg where it’s very hard to bootstrap your way to a first policy that’s good enough to deploy in real production environments. This is an extremely capital intensive playbook (which doesn’t even include cost of training). Time will tell if it works but a better playbook would be finding a way to copy the FSD playbook.

Data has always been the bottleneck for physical AI in self driving and robotics. Tesla is taking two very different approaches for FSD and Optimus. Tesla’s Optimus Training Playbook: 1. Build 30k Optimus Gen 3 robots 2. Operate them in a mock environment where they can perform self-play “Optimus Academy” 3. Train in sim using the real robot data to close sim2real gap. Tesla FSD Training Playbook: 1. Sell millions of cars outfitted with cheap cameras 2. Collect diverse real world driving data (especially intervention and failure recovery data) for free as a byproduct of customers driving the cars. 3. Use driving data to train Autopilot/FSD and deploy policies incrementally as a supervised FSD product 4. Repeat until policy reaches robust unsupervised full self driving for robotaxi launch. The Tesla FSD playbook is a beautiful self-funding, customer subsidized, diverse real world data flywheel. The Optimus playbook is the opposite and shares none of the beautiful attributes of the FSD training flywheel that made FSD successful. The key differences: 1. Instead of having customers pay you for vehicles, Tesla will need to fund 30,000 Optimus robots. Assuming the current landed cost per unit is $100k, that will be $3B to build plus another ~30% per year for maintenance labor and spare parts given it’s still an unhardened pre-production prototype is another $900M per year. For reference, Tesla’s GAAP net income in 2025 was $3.8B. 2. Instead of having customers drive their Teslas on roads all across the world giving Tesla an insanely rich and diverse dataset that Waymo and other AV companies could never collect, the Optimus Academy is doing the equivalent of building a fake town in a parking lot and driving their car in that parking lot. No matter how real you try to make the environments for self-play you can never replicate the diversity, complexity and failure modes of the real world. Data collected in staged environments produces demo-grade policies and will not be rich enough to generalize to the vast diversity of environments, tasks, objects, etc. out of distribution. 3. Instead of having customers collect real world failure recovery data (DAgger style) for free every time FSD disengages, the Optimus Academy will need paid teleoperators or onsite operators to collect the recovery data. Assuming 1 person can manage 2 robots to start that would cost $3.5B in labor per year (30,000 robots, $40/hr fully loaded, 16 hrs/day, 365 days per year, 2:1 robot:operator). Tesla can come up with the money to do this but money doesn’t solve the “mock data” problem. Given the higher degrees of freedom in humanoids vs. cars, training a generalized humanoid will be harder and require more data than a self-driving vehicle. The best way to train your robot is by deploying them in the diverse real world, subsidized by real customer operations. Humanoids face a chicken and egg where it’s very hard to bootstrap your way to a first policy that’s good enough to deploy in real production environments. This is an extremely capital intensive playbook (which doesn’t even include cost of training). Time will tell if it works but a better playbook would be finding a way to copy the FSD playbook.

Simon Kalouche

34,671 次观看 • 5 个月前

Welcome to the Lab of the Future! 🧬🤖 Excited to share LUMI-lab, out today in Cell — a self-driving platform that pairs an AI foundation model with a robotic lab to autonomously discover ionizable lipids (LNPs) for mRNA delivery. The core problem: Designing lipid nanoparticles (LNPs) is hard. The chemical space of ionizable lipids is vast, experimental cycles are slow, and — critically — historical LNP datasets are far too small to train a predictive model from scratch. Most AI approaches in this space hit a wall immediately: not enough data to learn from. Our solution: lab-in-the-loop foundation model learning. Instead of training on LNP data alone, LUMI starts as a transformer-based foundation model pretrained across broad chemical space, building rich molecular representations before it ever sees a single LNP experiment. Then it enters a closed loop with a robotic synthesis platform: predict → synthesize → assay → update. Each round of real wet-lab experiments fine-tunes the model, which then proposes smarter candidates for the next round. The lab isn't just validating AI predictions — it's actively teaching the model, continuously. What happened when we let it run: LUMI-lab autonomously synthesized and screened 1,700+ ionizable lipids in human bronchial epithelial cells. The top candidate — LUMI-6 — features a brominated lipid tail, a structural motif that had been largely overlooked in LNP design. LUMI found it without being told where to look. When formulated into LNPs and delivered intratracheally to mice, LUMI-6 achieved 20.3% gene editing efficiency in lung epithelial cells — a compelling result for one of the hardest-to-reach therapeutic targets, directly relevant to diseases like cystic fibrosis and alpha-1 antitrypsin deficiency. Why this matters beyond LNPs: This is a proof of concept for a broader thesis — that foundation model pretraining + active learning + robotic experimentation can overcome the data scarcity bottleneck that plagues AI-driven discovery in biology. You don't need a massive domain-specific dataset to start. You need a model that can generalize, a lab that can generate the right data, and a loop that connects them. Huge congratulations to first authors Yue Xu, Haotian Cui, and Kuan Pang, and to the entire Bowen LI team. Grateful to our collaborators at University Health Network and Leslie Dan Faculty of Pharmacy, and to Princess Margaret Cancer Centre Research Princess Margaret Cancer Centre Research. 📄 Paper:

Welcome to the Lab of the Future! 🧬🤖 Excited to share LUMI-lab, out today in Cell — a self-driving platform that pairs an AI foundation model with a robotic lab to autonomously discover ionizable lipids (LNPs) for mRNA delivery. The core problem: Designing lipid nanoparticles (LNPs) is hard. The chemical space of ionizable lipids is vast, experimental cycles are slow, and — critically — historical LNP datasets are far too small to train a predictive model from scratch. Most AI approaches in this space hit a wall immediately: not enough data to learn from. Our solution: lab-in-the-loop foundation model learning. Instead of training on LNP data alone, LUMI starts as a transformer-based foundation model pretrained across broad chemical space, building rich molecular representations before it ever sees a single LNP experiment. Then it enters a closed loop with a robotic synthesis platform: predict → synthesize → assay → update. Each round of real wet-lab experiments fine-tunes the model, which then proposes smarter candidates for the next round. The lab isn't just validating AI predictions — it's actively teaching the model, continuously. What happened when we let it run: LUMI-lab autonomously synthesized and screened 1,700+ ionizable lipids in human bronchial epithelial cells. The top candidate — LUMI-6 — features a brominated lipid tail, a structural motif that had been largely overlooked in LNP design. LUMI found it without being told where to look. When formulated into LNPs and delivered intratracheally to mice, LUMI-6 achieved 20.3% gene editing efficiency in lung epithelial cells — a compelling result for one of the hardest-to-reach therapeutic targets, directly relevant to diseases like cystic fibrosis and alpha-1 antitrypsin deficiency. Why this matters beyond LNPs: This is a proof of concept for a broader thesis — that foundation model pretraining + active learning + robotic experimentation can overcome the data scarcity bottleneck that plagues AI-driven discovery in biology. You don't need a massive domain-specific dataset to start. You need a model that can generalize, a lab that can generate the right data, and a loop that connects them. Huge congratulations to first authors Yue Xu, Haotian Cui, and Kuan Pang, and to the entire Bowen LI team. Grateful to our collaborators at University Health Network and Leslie Dan Faculty of Pharmacy, and to Princess Margaret Cancer Centre Research Princess Margaret Cancer Centre Research. 📄 Paper:

Bo Wang

57,471 次观看 • 5 个月前

As I promised yesterday, I'll briefly explain LoRA training and share a workflow I made so you can do it quickly. First, let me answer a very common question: 'Why train LoRAs when we have such advanced models?' Even though we have incredibly advanced models now (like NBP), we still can't always get them to do specific things we want. Simplest example: the spritesheet LoRA I made the other day. I generated 1000 images with Nano Banana and only 100 were what I wanted. The LoRA I trained using those 100 images gives me nearly 100% consistent results. Second point is cost and speed. With LoRA, we can cut costs by 4-5x. And while doing that, we're generating 4-5x faster. How many images do you need for a good LoRA? This depends on your LoRA's complexity. For example, when I training the spritesheet LoRA, even though I used 100 images, I didn't include buildings in the training data, so this LoRA doesn't work for buildings. So think about your LoRA's use cases and add examples for as many use cases as possible to improve quality. What are paired images and how to train LoRAs for image-editing? When training LoRAs for image editing on fal, we call each edit example paired images - one with _start suffix, one with _end suffix. For example, if you're training a background remove LoRA, the unedited original photo will be your '_start' image. The image with background removed will be the '_end' image. Simply put: images we want to edit or use as reference get _start, target images we want to achieve get '_end'. Important: save both images with the same name. Like image332_start.jpg and image332_end.jpg. This way the system knows which images pair together. What about training LoRAs for models with multiple image inputs? Same logic. We still use _start and _end suffixes, but with one difference. Since there are multiple input images, we can number them: _start, _start1, _start2. Example: start images, 1st image = Woman portrait (image35_start.jpg) 2nd image = Glasses photo (image35_start1.jpg) 3rd image = Hat photo (image35_start2.jpg) Output image = portrait of woman wearing glasses and hat (image35_end.jpg) Can we do more detailed captioning? Yes. Similarly, you can improve training quality by creating a txt file for each set with the caption inside. Example: create image35.txt and write: 'Recreate the image by putting the glasses from the second image and the hat from the third image on the woman in the first image.' What are Steps? How many should I use? What's Learning Rate? Steps determines how many times the model sees and processes your training data (your images). Each step, the model learns a bit more. But as steps increase, so does the risk of overfitting. So there's no real default. But for a simpler LoRA with 20 paired images, 1000 steps is ideal. Here's a metaphor for the Steps and Learning Rate relationship: Imagine you have a balloon. Our goal is to inflate it to the optimal size. Steps = How many times we blow into the balloon Learning rate = How hard we blow each time If we blow too softly, we need to blow many more times. If we blow too hard, we risk popping it quickly and can't reach optimal size. Of course training won't explode, but it won't work as intended because it wasn't trained optimally. Training's done, now what? Once training's complete, you'll have a safetensors file. Every model you train on fal has a LoRA inference endpoint. In that inference, add your safetensors file link to the LoRA url input, and you can use your LoRA. Thanks for the read! The workflow in the video: If I forgot anything, let me know in the replies.

As I promised yesterday, I'll briefly explain LoRA training and share a workflow I made so you can do it quickly. First, let me answer a very common question: 'Why train LoRAs when we have such advanced models?' Even though we have incredibly advanced models now (like NBP), we still can't always get them to do specific things we want. Simplest example: the spritesheet LoRA I made the other day. I generated 1000 images with Nano Banana and only 100 were what I wanted. The LoRA I trained using those 100 images gives me nearly 100% consistent results. Second point is cost and speed. With LoRA, we can cut costs by 4-5x. And while doing that, we're generating 4-5x faster. How many images do you need for a good LoRA? This depends on your LoRA's complexity. For example, when I training the spritesheet LoRA, even though I used 100 images, I didn't include buildings in the training data, so this LoRA doesn't work for buildings. So think about your LoRA's use cases and add examples for as many use cases as possible to improve quality. What are paired images and how to train LoRAs for image-editing? When training LoRAs for image editing on fal, we call each edit example paired images - one with _start suffix, one with _end suffix. For example, if you're training a background remove LoRA, the unedited original photo will be your '_start' image. The image with background removed will be the '_end' image. Simply put: images we want to edit or use as reference get _start, target images we want to achieve get '_end'. Important: save both images with the same name. Like image332_start.jpg and image332_end.jpg. This way the system knows which images pair together. What about training LoRAs for models with multiple image inputs? Same logic. We still use _start and _end suffixes, but with one difference. Since there are multiple input images, we can number them: _start, _start1, _start2. Example: start images, 1st image = Woman portrait (image35_start.jpg) 2nd image = Glasses photo (image35_start1.jpg) 3rd image = Hat photo (image35_start2.jpg) Output image = portrait of woman wearing glasses and hat (image35_end.jpg) Can we do more detailed captioning? Yes. Similarly, you can improve training quality by creating a txt file for each set with the caption inside. Example: create image35.txt and write: 'Recreate the image by putting the glasses from the second image and the hat from the third image on the woman in the first image.' What are Steps? How many should I use? What's Learning Rate? Steps determines how many times the model sees and processes your training data (your images). Each step, the model learns a bit more. But as steps increase, so does the risk of overfitting. So there's no real default. But for a simpler LoRA with 20 paired images, 1000 steps is ideal. Here's a metaphor for the Steps and Learning Rate relationship: Imagine you have a balloon. Our goal is to inflate it to the optimal size. Steps = How many times we blow into the balloon Learning rate = How hard we blow each time If we blow too softly, we need to blow many more times. If we blow too hard, we risk popping it quickly and can't reach optimal size. Of course training won't explode, but it won't work as intended because it wasn't trained optimally. Training's done, now what? Once training's complete, you'll have a safetensors file. Every model you train on fal has a LoRA inference endpoint. In that inference, add your safetensors file link to the LoRA url input, and you can use your LoRA. Thanks for the read! The workflow in the video: If I forgot anything, let me know in the replies.

ilker

15,133 次观看 • 6 个月前

Today we’re releasing V-JEPA, a method for teaching machines to understand and model the physical world by watching videos. This work is another important step towards Yann LeCun’s outlined vision of AI models that use a learned understanding of the world to plan, reason and accomplish complex tasks. Details ➡️ We're releasing a collection of V-JEPA vision models trained with a feature prediction objective using self-supervised learning. The models are able to understand and predict what is going on in a video, even with limited information. It learns by predicting missing or obscured parts of a video in its internal feature space. Unlike generative approaches that fill in missing pixels, this flexible approach enables up to 6x improvements in training and sample efficiency. The models were pre-trained on entirely unlabeled data, and a small amount of labeled data can be used to train a task-specific prediction head on top after pre-training. Our results show that, using a frozen backbone, our top V-JEPA models achieve 82.0% on Kinetics-400, 72.2% on Something-Something-v2 and 77.9% on ImageNet1K — competitive with or exceeding previous leading video models. We believe that this work is an important milestone on the path to advancing machine intelligence.

Today we’re releasing V-JEPA, a method for teaching machines to understand and model the physical world by watching videos. This work is another important step towards Yann LeCun’s outlined vision of AI models that use a learned understanding of the world to plan, reason and accomplish complex tasks. Details ➡️ We're releasing a collection of V-JEPA vision models trained with a feature prediction objective using self-supervised learning. The models are able to understand and predict what is going on in a video, even with limited information. It learns by predicting missing or obscured parts of a video in its internal feature space. Unlike generative approaches that fill in missing pixels, this flexible approach enables up to 6x improvements in training and sample efficiency. The models were pre-trained on entirely unlabeled data, and a small amount of labeled data can be used to train a task-specific prediction head on top after pre-training. Our results show that, using a frozen backbone, our top V-JEPA models achieve 82.0% on Kinetics-400, 72.2% on Something-Something-v2 and 77.9% on ImageNet1K — competitive with or exceeding previous leading video models. We believe that this work is an important milestone on the path to advancing machine intelligence.

AI at Meta

703,853 次观看 • 2 年前

Today, we're joined by Yejin Choi, professor and senior fellow at Stanford University University in the Computer Science Department and Stanford UniversityHAI. In this conversation, we explore Yejin’s recent work on making small language models reason more effectively. We discuss how high-quality, diverse data plays a central role in closing the intelligence gap between small and large models, and how combining synthetic data generation, imitation learning, and reinforcement learning can unlock stronger reasoning capabilities in smaller models. Yejin explains the risks of homogeneity in model outputs and mode collapse highlighted in her “Artificial Hivemind” paper, and its impacts on human creativity and knowledge. We also discuss her team's novel approaches, including reinforcement learning as a pre-training objective, where models are incentivized to “think” before predicting the next token, and "Prismatic Synthesis," a gradient-based method for generating diverse synthetic math data while filtering overrepresented examples. Additionally, we cover the societal implications of AI and the concept of pluralistic alignment—ensuring AI reflects the diverse norms and values of humanity. Finally, Yejin shares her mission to democratize AI beyond large organizations and offers her predictions for the coming year. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 04:44 - "Snowball effect" in AI investments 06:58 - Approaches to smaller models 08:58 - Importance of “better data” 14:07 - Imitation learning 18:24 - Artificial Hivemind paper 25:25 - AI risks 27:50 - Spectrum tuning 28:53 - Future of AI on humanity 33:08 - Reasoning in small models 34:58 - Prismatic Synthesis 48:20 - Reinforcement as a Pretraining Objective 55:04 - Pluralistic alignment 1:03:30 - Predictions

Today, we're joined by Yejin Choi, professor and senior fellow at Stanford University University in the Computer Science Department and Stanford UniversityHAI. In this conversation, we explore Yejin’s recent work on making small language models reason more effectively. We discuss how high-quality, diverse data plays a central role in closing the intelligence gap between small and large models, and how combining synthetic data generation, imitation learning, and reinforcement learning can unlock stronger reasoning capabilities in smaller models. Yejin explains the risks of homogeneity in model outputs and mode collapse highlighted in her “Artificial Hivemind” paper, and its impacts on human creativity and knowledge. We also discuss her team's novel approaches, including reinforcement learning as a pre-training objective, where models are incentivized to “think” before predicting the next token, and "Prismatic Synthesis," a gradient-based method for generating diverse synthetic math data while filtering overrepresented examples. Additionally, we cover the societal implications of AI and the concept of pluralistic alignment—ensuring AI reflects the diverse norms and values of humanity. Finally, Yejin shares her mission to democratize AI beyond large organizations and offers her predictions for the coming year. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 04:44 - "Snowball effect" in AI investments 06:58 - Approaches to smaller models 08:58 - Importance of “better data” 14:07 - Imitation learning 18:24 - Artificial Hivemind paper 25:25 - AI risks 27:50 - Spectrum tuning 28:53 - Future of AI on humanity 33:08 - Reasoning in small models 34:58 - Prismatic Synthesis 48:20 - Reinforcement as a Pretraining Objective 55:04 - Pluralistic alignment 1:03:30 - Predictions

The TWIML AI Podcast

12,141 次观看 • 6 个月前

Exciting updates on Project GR00T! We discover a systematic way to scale up robot data, tackling the most painful pain point in robotics. The idea is simple: human collects demonstration on a real robot, and we multiply that data 1000x or more in simulation. Let’s break it down: 1. We use Apple Vision Pro (yes!!) to give the human operator first person control of the humanoid. Vision Pro parses human hand pose and retargets the motion to the robot hand, all in real time. From the human’s point of view, they are immersed in another body like the Avatar. Teleoperation is slow and time-consuming, but we can afford to collect a small amount of data. 2. We use RoboCasa, a generative simulation framework, to multiply the demonstration data by varying the visual appearance and layout of the environment. In Jensen’s keynote video below, the humanoid is now placing the cup in hundreds of kitchens with a huge diversity of textures, furniture, and object placement. We only have 1 physical kitchen at the GEAR Lab in NVIDIA HQ, but we can conjure up infinite ones in simulation. 3. Finally, we apply MimicGen, a technique to multiply the above data even more by varying the motion of the robot. MimicGen generates vast number of new action trajectories based on the original human data, and filters out failed ones (e.g. those that drop the cup) to form a much larger dataset. To sum up, given 1 human trajectory with Vision Pro -> RoboCasa produces N (varying visuals) -> MimicGen further augments to NxM (varying motions). This is the way to trade compute for expensive human data by GPU-accelerated simulation. A while ago, I mentioned that teleoperation is fundamentally not scalable, because we are always limited by 24 hrs/robot/day in the world of atoms. Our new GR00T synthetic data pipeline breaks this barrier in the world of bits. Scaling has been so much fun for LLMs, and it's finally our turn to have fun in robotics! We are building tools to enable everyone in the ecosystem to scale up with us. Links in thread:

Jim Fan

364,380 次观看 • 2 年前

We benchmarked leading multimodal foundation models (GPT-4o, Claude 3.5 Sonnet, Gemini, Llama, etc.) on standard computer vision tasks—from segmentation to surface normal estimation—using standard datasets like COCO and ImageNet. These models have made remarkable progress; however, it is unclear exactly where they stand in terms of understanding vision in detail. Especially when it comes to tasks beyond question-answering. How well do they understand an object's segments or geometry? Our analyses yield an assessment that is quantitatively and qualitatively detailed and is compatible with evaluations developed in the field of computer vision over the past decades. Observed trends: 🔹 The foundation models consistently underperform task-specific SOTA models across all tasks. However, they are respectable generalists, which is remarkable as they are presumably trained primarily on image-text-based tasks. 🔹 They perform semantic tasks notably better than geometric ones. 🔹 GPT-4o performs the best among non-reasoning models, getting the top position in 4 out of 6 tasks. 🔹 Reasoning models, e.g., o3, show improvements in geometric tasks. 🔹 The 'image generation' models, e.g., GPT-40 Image Generation, which have been natively trained multimodally, exhibit quirks. E.g., hallucinated objects, misalignment between the input and output, etc. 🔹 While the prompting techniques affect performance, better models exhibit less sensitivity to variations in prompts. We control for the variance introduced by the prompting methods in our experiments. 🌐 Detailed analyses, visualizations: ⌨️ code: 🧵 1/n

We benchmarked leading multimodal foundation models (GPT-4o, Claude 3.5 Sonnet, Gemini, Llama, etc.) on standard computer vision tasks—from segmentation to surface normal estimation—using standard datasets like COCO and ImageNet. These models have made remarkable progress; however, it is unclear exactly where they stand in terms of understanding vision in detail. Especially when it comes to tasks beyond question-answering. How well do they understand an object's segments or geometry? Our analyses yield an assessment that is quantitatively and qualitatively detailed and is compatible with evaluations developed in the field of computer vision over the past decades. Observed trends: 🔹 The foundation models consistently underperform task-specific SOTA models across all tasks. However, they are respectable generalists, which is remarkable as they are presumably trained primarily on image-text-based tasks. 🔹 They perform semantic tasks notably better than geometric ones. 🔹 GPT-4o performs the best among non-reasoning models, getting the top position in 4 out of 6 tasks. 🔹 Reasoning models, e.g., o3, show improvements in geometric tasks. 🔹 The 'image generation' models, e.g., GPT-40 Image Generation, which have been natively trained multimodally, exhibit quirks. E.g., hallucinated objects, misalignment between the input and output, etc. 🔹 While the prompting techniques affect performance, better models exhibit less sensitivity to variations in prompts. We control for the variance introduced by the prompting methods in our experiments. 🌐 Detailed analyses, visualizations: ⌨️ code: 🧵 1/n

Amir Zamir

73,244 次观看 • 1 年前

We scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 sec), instantly forgetting what just happened. We pushed to 3 orders of magnitude beyond SOTA. Introducing RoboTTT. Test-Time Training (“TTT”) carries a tiny model *inside* the model. Every incoming sensor reading triggers one gradient step on that tiny core, so the history keeps getting compressed into its weights. The hidden state has a fixed size (literally a small neural net), so the robot can “grok” arbitrarily long experience with little overhead. Learning continues indefinitely after deployment. We can then put an entire video in context as prompt! RoboTTT enables one-shot in-context learning from human video: in circuit board assembly, a human demonstrates a never-seen configuration once, and the robot imitates it faithfully. Humans drop things all the time, but we pick them up so fast that we don’t even notice. That reflex to fix is half of our physical competence. RoboTTT shows self-improvement on the fly: the robot is skilled at recovering from its own errors mid-episode, and each fix enters its context to inform the next move. The TTT core distills a general-purpose, failure-to-correction mapping from the training data. One more thing. What excites me the most is a new Context Scaling Curve: from 128 to 8K timesteps, closed-loop performance hill-climbs steadily with no sign of saturation. 8K-context pretraining beats 1K by 62%. What LLM enjoys, robotics should too. Soon, even 1M context is not a fantasy. Deep dive in thread:

We scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 sec), instantly forgetting what just happened. We pushed to 3 orders of magnitude beyond SOTA. Introducing RoboTTT. Test-Time Training (“TTT”) carries a tiny model inside the model. Every incoming sensor reading triggers one gradient step on that tiny core, so the history keeps getting compressed into its weights. The hidden state has a fixed size (literally a small neural net), so the robot can “grok” arbitrarily long experience with little overhead. Learning continues indefinitely after deployment. We can then put an entire video in context as prompt! RoboTTT enables one-shot in-context learning from human video: in circuit board assembly, a human demonstrates a never-seen configuration once, and the robot imitates it faithfully. Humans drop things all the time, but we pick them up so fast that we don’t even notice. That reflex to fix is half of our physical competence. RoboTTT shows self-improvement on the fly: the robot is skilled at recovering from its own errors mid-episode, and each fix enters its context to inform the next move. The TTT core distills a general-purpose, failure-to-correction mapping from the training data. One more thing. What excites me the most is a new Context Scaling Curve: from 128 to 8K timesteps, closed-loop performance hill-climbs steadily with no sign of saturation. 8K-context pretraining beats 1K by 62%. What LLM enjoys, robotics should too. Soon, even 1M context is not a fantasy. Deep dive in thread:

Jim Fan

295,901 次观看 • 19 天前

$After 8 months of building in stealth and testing our infrastructure on 10000+ hours of real-world data and hundreds of unique environments, we're bringing FPV Labs into the open today. FPV Labs started with the following bet - if human data proves to be the underlying factor that determines scaling laws in general-purpose robotics, it will trigger the largest economic transformation in human history, and the underlying infrastructure that captures that data will determine how fast we get there. We will achieve this by building the full-stack infrastructure for capturing, processing, transferring, and evaluating human experience into spatial, temporal, and semantic knowledge for machines. Despite all the research novelty behind ChatGPT, its success can be attributed to one foundational fact - the scaling law of transformers. We believe the same dynamics have made their way into robotics. Recent studies showed task completion rates jumping from 30% to 70% when human demonstration data scaled from 1,000 to 20,000 hours, a log-linear trend that mirrors exactly what we saw in language and vision. Seeing these emergent signs of scaling law curves in robotics, we believe we are entering the era of general-purpose robotics policies, which makes the next few years the most exciting time in the history of this field. But the library of physical interactions required to train general-purpose robot policies does not exist yet. Over the last 8 months, we've seen dozens of companies emerge in this space. We were really happy to see new companies pushing this space forward, but we also saw the same pattern repeat: every egocentric data company was making some tradeoffs between quality, scale, and diversity. We have built FPV labs on the core principle that high-quality data is orders of magnitude more valuable than sheer volume. Case in point, self-driving cars collect thousands of hours of data per day, but only a small fraction of that data is actually useful for training better models. Several studies, like RT-2, have shown that as little as 1% of data improves as much as 25% on task success. The quality and diversity of data matter a lot more than scale, so there is clearly a power law curve in the downstream impact of data. We've spent months obsessing over data quality by building our stack, discarding it, rebuilding it, and iterating until we found a formula that doesn't compromise downstream quality at scale. We believe the downstream impact here is far more profound than most people realize. Workers globally are paid around $60 trillion per year in aggregate, and a lion's share of that compensation goes to physical labor - tasks that require navigating real spaces, manipulating real objects, and negotiating the infinite variability of the physical world. Human-to-robot transfer will be one of the most important infrastructures that will shape our society in the near future, and if it works, the economic impact will dwarf every technology transition that came before it in an exponential manner and lead to the creation of goods and services we can’t imagine today. Our mission is to lay the groundwork for us to transition into this future - the future of abundance. We are deeply grateful to our earliest believers, Paras Chopra and Lossfunk, who played a critical role in shaping our thinking.$

After 8 months of building in stealth and testing our infrastructure on 10000+ hours of real-world data and hundreds of unique environments, we're bringing FPV Labs into the open today. FPV Labs started with the following bet - if human data proves to be the underlying factor that determines scaling laws in general-purpose robotics, it will trigger the largest economic transformation in human history, and the underlying infrastructure that captures that data will determine how fast we get there. We will achieve this by building the full-stack infrastructure for capturing, processing, transferring, and evaluating human experience into spatial, temporal, and semantic knowledge for machines. Despite all the research novelty behind ChatGPT, its success can be attributed to one foundational fact - the scaling law of transformers. We believe the same dynamics have made their way into robotics. Recent studies showed task completion rates jumping from 30% to 70% when human demonstration data scaled from 1,000 to 20,000 hours, a log-linear trend that mirrors exactly what we saw in language and vision. Seeing these emergent signs of scaling law curves in robotics, we believe we are entering the era of general-purpose robotics policies, which makes the next few years the most exciting time in the history of this field. But the library of physical interactions required to train general-purpose robot policies does not exist yet. Over the last 8 months, we've seen dozens of companies emerge in this space. We were really happy to see new companies pushing this space forward, but we also saw the same pattern repeat: every egocentric data company was making some tradeoffs between quality, scale, and diversity. We have built FPV labs on the core principle that high-quality data is orders of magnitude more valuable than sheer volume. Case in point, self-driving cars collect thousands of hours of data per day, but only a small fraction of that data is actually useful for training better models. Several studies, like RT-2, have shown that as little as 1% of data improves as much as 25% on task success. The quality and diversity of data matter a lot more than scale, so there is clearly a power law curve in the downstream impact of data. We've spent months obsessing over data quality by building our stack, discarding it, rebuilding it, and iterating until we found a formula that doesn't compromise downstream quality at scale. We believe the downstream impact here is far more profound than most people realize. Workers globally are paid around $60 trillion per year in aggregate, and a lion's share of that compensation goes to physical labor - tasks that require navigating real spaces, manipulating real objects, and negotiating the infinite variability of the physical world. Human-to-robot transfer will be one of the most important infrastructures that will shape our society in the near future, and if it works, the economic impact will dwarf every technology transition that came before it in an exponential manner and lead to the creation of goods and services we can’t imagine today. Our mission is to lay the groundwork for us to transition into this future - the future of abundance. We are deeply grateful to our earliest believers, Paras Chopra and Lossfunk, who played a critical role in shaping our thinking.

Abhishek Anand

81,573 次观看 • 4 个月前

Studies have shown ChatGPT outperforms human annotators for Structured Data by about 25% and costs 30x less. 1 In just 2 months, miners on SN33 running ChatGPT without optimization can’t survive. Today we announce SN33 is now ReadyAI to fully align with our mission 👇 SN33 is building a more performant and significantly cheaper alternative to Scale AI Today structured data is performed primarily by human annotation services like Amazon’s Mechanical Turk and Scale AI It is now more important than ever for every business and individual to make their data AI Ready. However, taking unstructured data and making it Structured Data using today’s tools is extremely costly. SN33 revolutionizes this process, unlocking immense opportunities for commercialization. We lay out the vision for it in this detailed blog post: Validators TODAY can monetize access to this structured data pipeline independently, but we’re streamlining this process, launching a frontend soon that any validator can opt into to provide bandwidth. We've received great feedback from the community, recognizing that what we're building goes far beyond Conversational AI. Building the world's largest annotated conversational dataset (which we've already accomplished) is just one of countless real-world applications for SN33's Structured Data pipeline. We're building a decentralized Scale AI, offering a full suite of Structured Data commodities—from text metadata tagging (available today) to fully customizable queries for company-specific data annotation use cases and image metadata tagging coming soon 👀. Thanks for all the feedback! It has been invaluable so keep bringing it to us! 🙏$TAO Openτensor Foundaτion 1 “ChatGPT Outperforms Crowd-Workers for Text-Annotation Tasks” shows “The zero-shot accuracy of ChatGPT exceeds that of crowd-workers by about 25 percentage points on average [...] Moreover, the per-annotation cost of ChatGPT is less than $0.003—about thirty times cheaper than MTurk”

Studies have shown ChatGPT outperforms human annotators for Structured Data by about 25% and costs 30x less. 1 In just 2 months, miners on SN33 running ChatGPT without optimization can’t survive. Today we announce SN33 is now ReadyAI to fully align with our mission 👇 SN33 is building a more performant and significantly cheaper alternative to Scale AI Today structured data is performed primarily by human annotation services like Amazon’s Mechanical Turk and Scale AI It is now more important than ever for every business and individual to make their data AI Ready. However, taking unstructured data and making it Structured Data using today’s tools is extremely costly. SN33 revolutionizes this process, unlocking immense opportunities for commercialization. We lay out the vision for it in this detailed blog post: Validators TODAY can monetize access to this structured data pipeline independently, but we’re streamlining this process, launching a frontend soon that any validator can opt into to provide bandwidth. We've received great feedback from the community, recognizing that what we're building goes far beyond Conversational AI. Building the world's largest annotated conversational dataset (which we've already accomplished) is just one of countless real-world applications for SN33's Structured Data pipeline. We're building a decentralized Scale AI, offering a full suite of Structured Data commodities—from text metadata tagging (available today) to fully customizable queries for company-specific data annotation use cases and image metadata tagging coming soon 👀. Thanks for all the feedback! It has been invaluable so keep bringing it to us! 🙏$TAO Openτensor Foundaτion 1 “ChatGPT Outperforms Crowd-Workers for Text-Annotation Tasks” shows “The zero-shot accuracy of ChatGPT exceeds that of crowd-workers by about 25 percentage points on average [...] Moreover, the per-annotation cost of ChatGPT is less than $0.003—about thirty times cheaper than MTurk”

David Fields

13,639 次观看 • 1 年前

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.

AK

249,708 次观看 • 3 年前

Tencent presents GameGen-O Open-world Video Game Generation We introduce GameGen-O, the first diffusion transformer model tailored for the generation of open-world video games. This model facilitates high-quality, open-domain generation by simulating a wide array of game engine features, such as innovative characters, dynamic environments, complex actions, and diverse events. Additionally, it provides interactive controllability, thus allowing for the gameplay simulation. The development of GameGen-O involves a comprehensive data collection and processing effort from scratch. We collect and build the first Open-World Video Game Dataset (OGameData), amassed extensive data from over a hundred of next-generation open-world games, employing a proprietary data pipeline for efficient sorting, scoring, filtering, and decoupled captioning. This robust and extensive OGameData forms the foundation of our model's training process. GameGen-O undergoes a two-stage training process, consisting of foundation model pretraining and instruction tuning. In the first phase, the model is pre-trained on the OGameData via the text-to-video and video continuation, endowing GameGen-O with the capability for open-domain video game generation. In the second phase, the pre-trained model is frozen, and we fine-tuned using a trainable InstructNet, which enables the production of subsequent frames based on multimodal structural instructions. This whole training process imparts the model with the ability to generate and interactively control content. In summary, GameGen-O represents a notable initial step forward in the realm of open-world video game generation via generative models. It underscores the potential of generative models to serve as an alternative to rendering techniques, which can efficiently combine creative generation with interactive capabilities.

Tencent presents GameGen-O Open-world Video Game Generation We introduce GameGen-O, the first diffusion transformer model tailored for the generation of open-world video games. This model facilitates high-quality, open-domain generation by simulating a wide array of game engine features, such as innovative characters, dynamic environments, complex actions, and diverse events. Additionally, it provides interactive controllability, thus allowing for the gameplay simulation. The development of GameGen-O involves a comprehensive data collection and processing effort from scratch. We collect and build the first Open-World Video Game Dataset (OGameData), amassed extensive data from over a hundred of next-generation open-world games, employing a proprietary data pipeline for efficient sorting, scoring, filtering, and decoupled captioning. This robust and extensive OGameData forms the foundation of our model's training process. GameGen-O undergoes a two-stage training process, consisting of foundation model pretraining and instruction tuning. In the first phase, the model is pre-trained on the OGameData via the text-to-video and video continuation, endowing GameGen-O with the capability for open-domain video game generation. In the second phase, the pre-trained model is frozen, and we fine-tuned using a trainable InstructNet, which enables the production of subsequent frames based on multimodal structural instructions. This whole training process imparts the model with the ability to generate and interactively control content. In summary, GameGen-O represents a notable initial step forward in the realm of open-world video game generation via generative models. It underscores the potential of generative models to serve as an alternative to rendering techniques, which can efficiently combine creative generation with interactive capabilities.

AK

367,110 次观看 • 1 年前

.Sergey Levine is one of the world's leading robotics researchers (and co-founder of Physical Intelligence). He thinks fully autonomous robots are much closer than people realize - when I pushed him on a prediction, he said 5 years to robots that can autonomously run your household). The big game changer, in his view, is that we finally have LLMs with common sense and prior knowledge that we can scaffold the robot models around. We get into all the reasons this might not happen, the data flywheel from deployment, architecture, simulation, manufacturing, and much more. Enjoy! 0:00:00 – Timeline to widely deployed autonomous robots 0:22:12 – Why robotics will scale faster than self-driving cars 0:32:15 – How vision-language-action models work 0:50:26 – Improvements needed for brainlike efficiency 1:02:48 – Learning from simulation 1:14:08 – How much will robots speed up AI buildouts? 1:22:54 – If hardware’s the bottleneck, does China win by default?

.Sergey Levine is one of the world's leading robotics researchers (and co-founder of Physical Intelligence). He thinks fully autonomous robots are much closer than people realize - when I pushed him on a prediction, he said 5 years to robots that can autonomously run your household). The big game changer, in his view, is that we finally have LLMs with common sense and prior knowledge that we can scaffold the robot models around. We get into all the reasons this might not happen, the data flywheel from deployment, architecture, simulation, manufacturing, and much more. Enjoy! 0:00:00 – Timeline to widely deployed autonomous robots 0:22:12 – Why robotics will scale faster than self-driving cars 0:32:15 – How vision-language-action models work 0:50:26 – Improvements needed for brainlike efficiency 1:02:48 – Learning from simulation 1:14:08 – How much will robots speed up AI buildouts? 1:22:54 – If hardware’s the bottleneck, does China win by default?

Dwarkesh Patel

211,158 次观看 • 10 个月前