正在加载视频...

视频加载失败

返回首页

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight...

jackyk02's profile picture

Jacky Kwok

5,867 subscribers

906,427 次观看 • 1 天前 •via X (Twitter)

教育游戏科学技术
LIVE

Anya Rossi• Live Now

Private livecam show

47 条评论

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(1) Model Architecture CLM first trains a state encoder and an action encoder on a large-scale dataset with a contrastive objective, so that each state is pulled toward the ground-truth action and pushed away from all others. The two encoders then serve directly as a zero-shot action classifier.

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(2) Data Recipe We release CLM-8B, which is pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories.

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(3) Scaling Laws for Verification We find that the test InfoNCE loss scales as a power law with training compute, dataset size, projection-head size, and encoder size. These dimensions must be scaled jointly to achieve optimal performance. Notably, scaling the encoder size yields the strongest gains.

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(4) Latency vs. Jev and Constrained Decoding The key difference between CLM and Jev is that Jev only supports state caching, whereas CLM’s dual-encoder architecture allows state and action embeddings to be cached independently. This is particularly useful in applications such as tool calling, computer use, and games, where the action space is predefined and remains fixed. By caching the action embeddings, CLM substantially reduces inference cost, with the efficiency gains becoming even larger as the number and length of candidate actions grow. At ~1K candidates, CLM is 13× faster than Jev ⚡

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(5) Zero-Shot Evaluation Across computer-use, gaming, and tool-calling tasks, CLM-8B performs on par with Jev while running up to 9× faster. The speedups are most pronounced when the number of candidates is large (e.g., WikiRacing) or when actions can be reused frequently across states (e.g., T-Rex Game). CLM-35B, with improved generalization and even greater speedups, will be released early next month.

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(6) Agentic Benchmarks We find that Jev fails to serve as a verifier for long-horizon tasks, performing below the random-selection (Pass@1) baseline. In contrast, with lightweight fine-tuning, CLM achieves SOTA performance on challenging agentic benchmarks, including DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%), while delivering 4–6× faster inference than Jev.

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(7) Dino Run Demo

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(8) Super Mario Demo

Jacky Kwok 的头像
Jacky Kwok1 天前

🧵(9) WikiRacing Demo

Jacky Kwok 的头像
Jacky Kwok1 天前

CLM comes with an interactive playground on GitHub:

Jacky Kwok 的头像
Jacky Kwok1 天前

CLM-8B is part of our scaling ladder, where we train models across multiple scales to establish scaling laws and predict performance at larger scales. A multimodal CLM-35B is now in training with more data, compute, and parameters. Stay tuned for the release early next month 🚀

Jacky Kwok 的头像
Jacky Kwok1 天前

Joint work with @hangoo_kang @TarunSures41845 @JonSaadFalcon @drmapavone @Azaliamirh and @HazyResearch

John Rood 的头像
John Rood1 天前

the disaggregation is the part builders should steal: a loop scores the same state against many candidate actions, so the state embedding caches perfectly and each candidate collapses to a cheap dot product. that's the economics that makes per-turn decisions affordable.

Jacky Kwok 的头像
Jacky Kwok1 天前

Exactly!

Dibs 的头像
Dibs1 天前

Very nice, but didn't Jev just come out? Could you explain how you guys did this work so quickly?

Jacky Kwok 的头像
Jacky Kwok1 天前

I actually explored this for robotics last year, and CLM grew directly out of that work a few months ago. Feel free to check out my earlier post for more details on how CLM could be applied to robotics:

Xuan (Billy) Zhang 的头像
Xuan (Billy) Zhang1 天前

Contrastive learning is back again!

Brandon 的头像
Brandon1 天前

More bookmarks than likes. That's how you know someone cooked here

Cuth 的头像
Cuth1 天前

adding to the benchmark

Raymond Weitekamp 的头像
Raymond Weitekamp1 天前

impressive stuff!

Alfc Oleol 的头像
Alfc Oleol1 天前

Vision support?

Jacky Kwok 的头像
Jacky Kwok1 天前

We’ll be releasing CLM-35B next month with vision support. Stay tuned for the release 👀

Suraj 的头像
Suraj1 天前

caching state and action embeddings separately is the clever bit. huge for tool routing

Omar Ramadan 的头像
Omar Ramadan1 天前

Finally, this is what I call real research

Florian S 的头像
Florian S1 天前

Do you want me to include it in JevBench?

Durp 的头像
Durp1 天前

Wait, so this just crushes Jev, and is free, and runs locally with 8 billion parameters? It fits in under 8gb? What's the context window size? What's the catch?! Don't toy with me here, this sounds so good..

Douglas Schonholtz 的头像
Douglas Schonholtz1 天前

We gotta get it numbers on decision bench and jev bench

Stephane 的头像
Stephane1 天前

@grok explain exactly what this is, and how can I use it, for what. Compare with Jev from typesafeAi

Cuth 的头像
Cuth1 天前

This model is up next on the benchmark chopping block. I hope it pushes the frontier!

TheHBrand 的头像
TheHBrand1 天前

is this OSS on hugging face? how does this perform on language recognition applications like "what was the users intention - here's waht they said - here's your options" ?

Potato Terminator 的头像
Potato Terminator1 天前

9x faster inference with open weights and published scaling laws is real progress. I'm curious what the compute footprint per agent task looks like vs a dense autoregressive model. Efficiency gains like this matter more than another benchmark point.

YUKI 的头像
YUKI1 天前

Really cool work! We’re building JevEmbed to turn embeddings into decisions: Choice, Score, and Noul judgments from existing embedding models, without text generation. Curious how CLMs would perform as a backbone.

Gi 的头像
Gi1 天前

how can a 8B model get these numbers?? DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). and no body is surprised??

Brjan | AI Builder 的头像
Brjan | AI Builder1 天前

if it's really 9× faster, that could change the game for real-time applications

智0elsyn 的头像
智0elsyn1 天前

The independent action cache is the appealing part for me: a changing state doesn't force you to re-encode the same menu of actions. For tool routing, have you tried a two-stage version—cache tool selection, then generate the arguments? That seems a useful middle ground between four fixed game moves and entirely fresh action candidates.

Dominik 的头像
Dominik1 天前

the useful leap here is treating state and action as separate objects instead of collapsing everything into a single prompt. cached embeddings can preserve what the run learned while the action set stays auditable. that is where faster inference becomes a continuity primitive, not just a benchmark win.

Jared Zoneraich 的头像
Jared Zoneraich1 天前

10x faster... that's insane

lastinline 的头像
lastinline1 天前

The claim to check is variance under shift. Contrastive objectives are only as good as the negatives they were trained against.

Michel aka Agent B 的头像
Michel aka Agent B1 天前

Very interesting work Jacky ! Bravo to the team 👏👏 Just out of curiosity: Was the idea originally inspired by robotics?

James Odebiyi 的头像
James Odebiyi1 天前

seems really useful for game agents

Yogesh 的头像
Yogesh1 天前

@jackyk02 missing piece next to big coding models agent loops burn tokens on tiny which-action-next calls. a contrastive system-1 that maps state to action without paragraphs feels closer to a real controller curious if 9x holds in a multi-tool harness with long history

MattacK 的头像
MattacK1 天前

is this a Fair and Honest measure of latency if you can tell based on the details shared? i.e. testing latency of an API to a model sitting on hardware right next to you vs. going thru all the standard DNS pipes etc of the web to get a response from Jev's API X miles away (if it's this scenario, it would not be a fair comparison)

X4 的头像
X41 天前

InfoNCE indeed was essential, also multi-modality. Am working on a different project and approach and appreciate a lot your creative and beatiful results. Congratulations Jacky 😎🥂🍾🚀

Samuel Ekpe 的头像
Samuel Ekpe1 天前

@arankomatsuzaki This is sick!

hamjji 的头像
hamjji1 天前

Is CLM-8B open weights like Kev, or API only? Would love to see it on the same out-of-domain test Kev posted.

Martin Harold Williams 的头像
Martin Harold Williams1 天前

@grok please generate the comprehensive accurate prompt for my ai agent to set this up on my mac if possible if not will set it up on the cloud. Thank you

Thành Lê 🌩️ 的头像
Thành Lê 🌩️1 天前

this is so good

相关视频

Introducing FLUX 3 Action. An open weights 7B World Action Model that achieves first place on the RoboLab benchmark. It outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster.⁠⁠ FLUX 3 Action removes the usual trade-off between world action model performance and VLA speed: it still predicts video and actions together, but plans more than twice as far ahead and runs faster per second of robot motion than the strongest open VLA. Teams can fine-tune FLUX 3 Action on their own demonstrations to create policies for a particular robot and task. Together with NVIDIA, we also integrated FLUX 3 Action natively into Hugging Face's LeRobot, with fine-tuning recipes included and edge deployment on NVIDIA Jetson. Beyond robotics, we’re also seeing promising results training task-specific policies for acting in simulated environments like gaming, controlling a vehicle, computer use, and wherever else a model needs to understand a visual environment and then choose what to do next. FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture designed for practical deployment. In midtraining, we trained the model to predict actions and future frames together. We’re releasing the weights, code, fine-tuning recipe, benchmarks, and reproducible examples so researchers and developers can build on the model with their own robots, environments, and tasks (see below).
1:10

Introducing FLUX 3 Action. An open weights 7B World Action Model that achieves first place on the RoboLab benchmark. It outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster.⁠⁠ FLUX 3 Action removes the usual trade-off between world action model performance and VLA speed: it still predicts video and actions together, but plans more than twice as far ahead and runs faster per second of robot motion than the strongest open VLA. Teams can fine-tune FLUX 3 Action on their own demonstrations to create policies for a particular robot and task. Together with NVIDIA, we also integrated FLUX 3 Action natively into Hugging Face's LeRobot, with fine-tuning recipes included and edge deployment on NVIDIA Jetson. Beyond robotics, we’re also seeing promising results training task-specific policies for acting in simulated environments like gaming, controlling a vehicle, computer use, and wherever else a model needs to understand a visual environment and then choose what to do next. FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture designed for practical deployment. In midtraining, we trained the model to predict actions and future frames together. We’re releasing the weights, code, fine-tuning recipe, benchmarks, and reproducible examples so researchers and developers can build on the model with their own robots, environments, and tasks (see below).

Black Forest Labs

161,855 次观看 • 1 天前

Learn to train an LLM with distributed data while ensuring privacy using federated learning in a new two-part short course, Intro to Federated Learning and Federated Fine-tuning of LLMs with Private Data, created with Flower and taught by Daniel J. Beutel and Nicholas Lane. Federated learning allows a single model to be trained across multiple devices, such as phones, or multiple organizations, such as hospitals, without the need to share data to a central server. This two-part course gives you an introduction to federated learning, and then teaches you how to fine-tune your large language model with distributed data using Flower Lab’s open source federated learning framework. You’ll learn: - How to use federated learning to train a variety of models, ranging from speech and vision models to LLMs, across distributed data while offering data privacy options to users and organizations. - Privacy Enhancing Technologies like differential privacy (DP), which obscures individual data by adding calibrated noise to query results. - Two variants of differential privacy - Central and Local - and how to choose depending on your use case. - How to measure and decrease bandwidth usage to make federated learning more practical and efficient with techniques like using pre-trained models and Parameter-Efficient Fine-Tuning - How federated LLM fine-tuning reduces the risk of leaking training data. Sign up here!
4:34

Learn to train an LLM with distributed data while ensuring privacy using federated learning in a new two-part short course, Intro to Federated Learning and Federated Fine-tuning of LLMs with Private Data, created with Flower and taught by Daniel J. Beutel and Nicholas Lane. Federated learning allows a single model to be trained across multiple devices, such as phones, or multiple organizations, such as hospitals, without the need to share data to a central server. This two-part course gives you an introduction to federated learning, and then teaches you how to fine-tune your large language model with distributed data using Flower Lab’s open source federated learning framework. You’ll learn: - How to use federated learning to train a variety of models, ranging from speech and vision models to LLMs, across distributed data while offering data privacy options to users and organizations. - Privacy Enhancing Technologies like differential privacy (DP), which obscures individual data by adding calibrated noise to query results. - Two variants of differential privacy - Central and Local - and how to choose depending on your use case. - How to measure and decrease bandwidth usage to make federated learning more practical and efficient with techniques like using pre-trained models and Parameter-Efficient Fine-Tuning - How federated LLM fine-tuning reduces the risk of leaking training data. Sign up here!

Andrew Ng

64,609 次观看 • 2 年前

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.
0:48

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.

AK

125,353 次观看 • 3 年前

Today, we're joined by Aakanksha Chowdhery, member of technical staff at Reflection, to explore the fundamental shifts required to build true agentic AI. While the industry has largely focused on post-training techniques to improve reasoning, Aakanksha draws on her experience leading pre-training efforts for Google’s PaLM and early Gemini models to argue that pre-training itself must be rethought to move beyond static benchmarks. We explore the limitations of next-token prediction for multi-step workflows and examine how attention mechanisms, loss objectives, and training data must evolve to support long-form reasoning and planning. Aakanksha shares insights on the difference between context retrieval and actual reasoning, the importance of "trajectory" training data, and why scaling remains essential for discovering emergent agentic capabilities like error recovery and dynamic tool learning. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 02:26 - Reflection 04:54 - Limitations of post-training for building agents 07:31 - Rethinking pre-training in agents 10:51 - Scaling 11:27 - Evolving attention mechanisms for agentic capabilities 12:39 - Memory as a tool 14:13 - Loss objectives and training data 15:50 - Fine-tuning loss in agent performance 19:37 - Training data 21:29 - Augmenting dominant training data source 24:11 - Overcoming challenges in training on synthetic data 25:47 - Benchmarks 30:44 - Scaling laws in large models versus small models 33:20 - Long-form versus short-form reasoning 37:57 - Agent’s ability to recover from failure 40:15 - Hallucinations and failure recovery 43:53 - Tool use in agents 46:38 - Coding agents 48:37 - How researchers can contribute to agentic AI
51:54

Today, we're joined by Aakanksha Chowdhery, member of technical staff at Reflection, to explore the fundamental shifts required to build true agentic AI. While the industry has largely focused on post-training techniques to improve reasoning, Aakanksha draws on her experience leading pre-training efforts for Google’s PaLM and early Gemini models to argue that pre-training itself must be rethought to move beyond static benchmarks. We explore the limitations of next-token prediction for multi-step workflows and examine how attention mechanisms, loss objectives, and training data must evolve to support long-form reasoning and planning. Aakanksha shares insights on the difference between context retrieval and actual reasoning, the importance of "trajectory" training data, and why scaling remains essential for discovering emergent agentic capabilities like error recovery and dynamic tool learning. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 02:26 - Reflection 04:54 - Limitations of post-training for building agents 07:31 - Rethinking pre-training in agents 10:51 - Scaling 11:27 - Evolving attention mechanisms for agentic capabilities 12:39 - Memory as a tool 14:13 - Loss objectives and training data 15:50 - Fine-tuning loss in agent performance 19:37 - Training data 21:29 - Augmenting dominant training data source 24:11 - Overcoming challenges in training on synthetic data 25:47 - Benchmarks 30:44 - Scaling laws in large models versus small models 33:20 - Long-form versus short-form reasoning 37:57 - Agent’s ability to recover from failure 40:15 - Hallucinations and failure recovery 43:53 - Tool use in agents 46:38 - Coding agents 48:37 - How researchers can contribute to agentic AI

The TWIML AI Podcast

45,470 次观看 • 9 个月前

Meta announces Movie Gen A Cast of Media Foundation Models We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user’s image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models
0:30

Meta announces Movie Gen A Cast of Media Foundation Models We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user’s image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models

AK

62,719 次观看 • 2 年前

Still following your human intuition to mix corpora from different sources for language model pre-training 🧠? Everyone says that data mixture has a big impact on model performance, but how - and why🕵️? Did you know that web corpora are actually highly impactful for downstream tasks 🏆? Let's check out our preprint "RegMix: Data Mixture as Regression for Language Model Pre-training" 📄 🔬In this paper, we've proposed an automatic data mixture method RegMix that achieves a 6.3% improvement over human selection on the widely used HellaSwag benchmark - and it only needs a 2% extra training FLOPs! 📈 Details in the thread 🧵
0:34

Still following your human intuition to mix corpora from different sources for language model pre-training 🧠? Everyone says that data mixture has a big impact on model performance, but how - and why🕵️? Did you know that web corpora are actually highly impactful for downstream tasks 🏆? Let's check out our preprint "RegMix: Data Mixture as Regression for Language Model Pre-training" 📄 🔬In this paper, we've proposed an automatic data mixture method RegMix that achieves a 6.3% improvement over human selection on the widely used HellaSwag benchmark - and it only needs a 2% extra training FLOPs! 📈 Details in the thread 🧵

Qian Liu

54,961 次观看 • 2 年前

New short course on Pretraining LLMs! Developed with Upstage and taught by their CEO Sung Kim and CSO Lucy Park. While prompting or fine-tuning existing models works well for many general language tasks, pretraining is valuable for specialized domains or languages with limited representation in current models. This course walks you through the LLM pretraining pipeline: 1. Data preparation: Learn to source, clean, and prepare training data using HuggingFace. 2. Model architecture: Configure transformer networks, including modifying existing models. 3. Training: Set up and run training using open-source libraries. 4. Evaluation: Benchmark performance using popular evaluation strategies. As an example use case, you'll also compare the output of a base model with its fine-tuned and further pretrained variants, to see the impact of pretraining on a model's ability to write Python. The course also explores an innovative technique called depth up-scaling, which Upstage used to train their Solar model family, reducing pretraining compute costs by up to 70%. This technique works by first duplicating layers of a smaller pretrained model to form a larger model, and then further pretraining the result. Sign up here!
3:54

New short course on Pretraining LLMs! Developed with Upstage and taught by their CEO Sung Kim and CSO Lucy Park. While prompting or fine-tuning existing models works well for many general language tasks, pretraining is valuable for specialized domains or languages with limited representation in current models. This course walks you through the LLM pretraining pipeline: 1. Data preparation: Learn to source, clean, and prepare training data using HuggingFace. 2. Model architecture: Configure transformer networks, including modifying existing models. 3. Training: Set up and run training using open-source libraries. 4. Evaluation: Benchmark performance using popular evaluation strategies. As an example use case, you'll also compare the output of a base model with its fine-tuned and further pretrained variants, to see the impact of pretraining on a model's ability to write Python. The course also explores an innovative technique called depth up-scaling, which Upstage used to train their Solar model family, reducing pretraining compute costs by up to 70%. This technique works by first duplicating layers of a smaller pretrained model to form a larger model, and then further pretraining the result. Sign up here!

Andrew Ng

85,714 次观看 • 2 年前

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:
2:52

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

125,146 次观看 • 1 年前

Introducing HRM-Text. An ultra-lean 1B-parameter reasoning language model designed to deliver strong general performance with a fraction of the data, compute, and infrastructure. Trained on just 40B structured tokens, HRM-Text achieves competitive performance while using ~1/1000 of the training data of comparable models. The kicker? The full model trains in roughly one day on a $1,000 budget. This opens the door to a new generation of AI that is powerful, accessible, and radically easier to adapt. Theories and research concepts once deemed too expensive to test are officially back in the game. Sapient Intelligence invites you to help us shape a new paradigm for general intelligence.
5:31

Introducing HRM-Text. An ultra-lean 1B-parameter reasoning language model designed to deliver strong general performance with a fraction of the data, compute, and infrastructure. Trained on just 40B structured tokens, HRM-Text achieves competitive performance while using ~1/1000 of the training data of comparable models. The kicker? The full model trains in roughly one day on a $1,000 budget. This opens the door to a new generation of AI that is powerful, accessible, and radically easier to adapt. Theories and research concepts once deemed too expensive to test are officially back in the game. Sapient Intelligence invites you to help us shape a new paradigm for general intelligence.

Sapient Intelligence

517,742 次观看 • 4 个月前

🚀 We're thrilled to introduce Orthrus 🧬🐕—a groundbreaking mature RNA foundation model designed to push the boundaries of RNA property prediction! 🔬 What is Orthrus? Orthrus is a Mamba-based RNA foundation model, pre-trained using a novel self-supervised contrastive learning objective with biologically inspired augmentations. It optimizes the similarity between splicing isoforms and orthologous transcripts, capturing functional and evolutionary relationships to enhance mature RNA property prediction accuracy. 📑 Preprint: 💻 Code: 🌐 Project Page: 📦 Model Weights: 🧠 Why Orthrus? Decoding the RNA regulatory code is key to understanding biology, but traditional experimental approaches are slow and costly. Existing genomic foundation models rely on techniques like masked language modeling or next-token prediction, which aren't fully aligned with the complexities of genomic data—leading to suboptimal results. 🌟 Orthrus Highlights: - Biologically-Informed Contrastive Learning 🧪: A novel contrastive learning objective designed specifically for genomics, maximizing similarity between splicing isoforms and orthologous transcripts across species. - Extensive Pre-training 📊: Trained on splicing annotations from 10 species and orthologous alignments from 400+ mammalian species (Zoonomia Project), with a focus on sequences of high functional importance. - Superior Representations🏅: Orthrus outperforms existing genomic models on 5 mRNA property prediction tasks, often surpassing supervised methods with just a simple linear transformation. - Efficiency in Low-Data Settings📉: Orthrus excels in low-data regimes, achieving state-of-the-art results with as few as 45 labeled examples for fine-tuning on RNA half-life prediction. Shoutout to the amazing leading authors Phil (Phil Fradkin) and Ian (Ian Shi)! Also the work is impossible without an outstanding collaboration by Karina (Karin(a) Isaev), Brendan (Brendan Frey) , Quaid (Quaid Morris), Leo J. Lee! Vector Institute University Health Network U of T Department of Computer Science Temerty Centre for AI in Medicine (T-CAIREM) Department of Laboratory Medicine & Pathobiology
0:53

🚀 We're thrilled to introduce Orthrus 🧬🐕—a groundbreaking mature RNA foundation model designed to push the boundaries of RNA property prediction! 🔬 What is Orthrus? Orthrus is a Mamba-based RNA foundation model, pre-trained using a novel self-supervised contrastive learning objective with biologically inspired augmentations. It optimizes the similarity between splicing isoforms and orthologous transcripts, capturing functional and evolutionary relationships to enhance mature RNA property prediction accuracy. 📑 Preprint: 💻 Code: 🌐 Project Page: 📦 Model Weights: 🧠 Why Orthrus? Decoding the RNA regulatory code is key to understanding biology, but traditional experimental approaches are slow and costly. Existing genomic foundation models rely on techniques like masked language modeling or next-token prediction, which aren't fully aligned with the complexities of genomic data—leading to suboptimal results. 🌟 Orthrus Highlights: - Biologically-Informed Contrastive Learning 🧪: A novel contrastive learning objective designed specifically for genomics, maximizing similarity between splicing isoforms and orthologous transcripts across species. - Extensive Pre-training 📊: Trained on splicing annotations from 10 species and orthologous alignments from 400+ mammalian species (Zoonomia Project), with a focus on sequences of high functional importance. - Superior Representations🏅: Orthrus outperforms existing genomic models on 5 mRNA property prediction tasks, often surpassing supervised methods with just a simple linear transformation. - Efficiency in Low-Data Settings📉: Orthrus excels in low-data regimes, achieving state-of-the-art results with as few as 45 labeled examples for fine-tuning on RNA half-life prediction. Shoutout to the amazing leading authors Phil (Phil Fradkin) and Ian (Ian Shi)! Also the work is impossible without an outstanding collaboration by Karina (Karin(a) Isaev), Brendan (Brendan Frey) , Quaid (Quaid Morris), Leo J. Lee! Vector Institute University Health Network U of T Department of Computer Science Temerty Centre for AI in Medicine (T-CAIREM) Department of Laboratory Medicine & Pathobiology

Bo Wang

114,994 次观看 • 1 年前

fun weekend project: AutoJev. i was curious to see if i could train a competitive Jev-like model completely autonomously with a swarm of agents using our internal system. turns out you can get quite far! some details: - gave the swarm a devbox with an h200 gpu - the swarm is a mix of astra and fable; used sol and luna for synthetic data and filtering; ran for 20 hours - spent $3.1k in total ($1.9k on agents, $1.2k on data) - used qwen3.8-27b as the backbone. it underperforms Jev initially, but after training it is competitive/better, and it also supports 260k context and multimodal input - the agents tried RL, but it hasn't worked out so far, so they fell back to SFT on high-quality synthetic data - both accuracy and calibration are better - jev-compatible api, playground, and inference. on an h200 it gets around 120ms p95. that will be higher under contention, but still quite fast for an unoptimized version code: model: going to let it run longer, and will also bring up inference so folks can try it out.
0:40

fun weekend project: AutoJev. i was curious to see if i could train a competitive Jev-like model completely autonomously with a swarm of agents using our internal system. turns out you can get quite far! some details: - gave the swarm a devbox with an h200 gpu - the swarm is a mix of astra and fable; used sol and luna for synthetic data and filtering; ran for 20 hours - spent $3.1k in total ($1.9k on agents, $1.2k on data) - used qwen3.8-27b as the backbone. it underperforms Jev initially, but after training it is competitive/better, and it also supports 260k context and multimodal input - the agents tried RL, but it hasn't worked out so far, so they fell back to SFT on high-quality synthetic data - both accuracy and calibration are better - jev-compatible api, playground, and inference. on an h200 it gets around 120ms p95. that will be higher under contention, but still quite fast for an unoptimized version code: model: going to let it run longer, and will also bring up inference so folks can try it out.

Denis Yarats

26,757 次观看 • 3 天前

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...
3:28

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

kwindla

331,598 次观看 • 28 天前

The World Model as NEO's Cognitive Core 1X has revealed a major AI development where the NEO humanoid can translate any natural language prompt into robotic action. It demonstrates this capability even for novel tasks, objects, and environments not found in its robot dataset. - the 1X World Model is trained on internet-scale human interaction videos and fine-tuned with robot data to ground its understanding in physics and in NEO's embodiment - from a simple voice or text prompt, the world model generates a visualization of future actions - a built-in inverse dynamics model then translates these into precise motor movements for NEO
1:02

The World Model as NEO's Cognitive Core 1X has revealed a major AI development where the NEO humanoid can translate any natural language prompt into robotic action. It demonstrates this capability even for novel tasks, objects, and environments not found in its robot dataset. - the 1X World Model is trained on internet-scale human interaction videos and fine-tuned with robot data to ground its understanding in physics and in NEO's embodiment - from a simple voice or text prompt, the world model generates a visualization of future actions - a built-in inverse dynamics model then translates these into precise motor movements for NEO

The Humanoid Hub

68,453 次观看 • 8 个月前

Smart Turn v2: open source, native audio turn detection in 14 languages. New checkpoint of the open source, open data, open training code, semantic VAD model on Hugging Face, fal, and Pipecat AI. - 3x faster inference (12ms on an L40) - 14 languages (13 more than v1, which was english-only) - New synthetic data set `chirp_3_all` with ~163k audio samples - 99% accuracy on held out `human_5_all` test data Good turn detection is critical for voice agents. This model "understands" both semantic and audio patterns, and mitigates the voice AI trade-off between unwanted turn latency vs the agent interrupting people before they are finished speaking. Training scripts for both Modal and local training are in the repo. We want to make it as easy as possible to contribute to or customize this model! Here's a demo running the smart-turn model with default settings, aimed at generally hitting 400ms total turn detection time. You can tune things to be faster, too. You can help by contributing data, doing architecture expermints, or cleaning open source data! Keep reading ...
2:12

Smart Turn v2: open source, native audio turn detection in 14 languages. New checkpoint of the open source, open data, open training code, semantic VAD model on Hugging Face, fal, and Pipecat AI. - 3x faster inference (12ms on an L40) - 14 languages (13 more than v1, which was english-only) - New synthetic data set `chirp_3_all` with ~163k audio samples - 99% accuracy on held out `human_5_all` test data Good turn detection is critical for voice agents. This model "understands" both semantic and audio patterns, and mitigates the voice AI trade-off between unwanted turn latency vs the agent interrupting people before they are finished speaking. Training scripts for both Modal and local training are in the repo. We want to make it as easy as possible to contribute to or customize this model! Here's a demo running the smart-turn model with default settings, aimed at generally hitting 400ms total turn detection time. You can tune things to be faster, too. You can help by contributing data, doing architecture expermints, or cleaning open source data! Keep reading ...

kwindla

42,246 次观看 • 1 年前

An interactive world model developed by NVIDIA in collaboration with academic partners. - DreamDojo turns egocentric human video data into physical intelligence. - Human data is more scalable than robotics data but lacks action labels. - To solve this, a dedicated action model extracts latent actions by identifying physics and motion deltas between frames. Training - A massive 44k hours of video data are used for pre-training. - Post-training on small-scale robot datasets maps human physics to specific robot embodiments. - An additional distillation stage converts the model into an autoregressive, few-step diffusion model, enabling real-time, action-controllable simulation. Primary Use Cases - Live Teleoperation: Controlling a robot inside a world simulation in real-time. - Model-based Planning: Previewing and curating the best actions for improved success. - Policy Evaluation: Testing robot policies in realistic, out-of-distribution scenarios. Everything that's open-sourced: weights, code, post-training dataset, eval set, and details to reproduce.
1:14

An interactive world model developed by NVIDIA in collaboration with academic partners. - DreamDojo turns egocentric human video data into physical intelligence. - Human data is more scalable than robotics data but lacks action labels. - To solve this, a dedicated action model extracts latent actions by identifying physics and motion deltas between frames. Training - A massive 44k hours of video data are used for pre-training. - Post-training on small-scale robot datasets maps human physics to specific robot embodiments. - An additional distillation stage converts the model into an autoregressive, few-step diffusion model, enabling real-time, action-controllable simulation. Primary Use Cases - Live Teleoperation: Controlling a robot inside a world simulation in real-time. - Model-based Planning: Previewing and curating the best actions for improved success. - Policy Evaluation: Testing robot policies in realistic, out-of-distribution scenarios. Everything that's open-sourced: weights, code, post-training dataset, eval set, and details to reproduce.

The Humanoid Hub

11,575 次观看 • 7 个月前

We’re no longer just scaling computing power. We’re using compute to scale intelligence itself. That’s what makes this moment historically significant. For sixty years, progress in computing followed Moore’s Law—transistor density doubling roughly every two years. But AI is advancing on a far steeper curve. Today, frontier model capabilities are improving on a cadence closer to every six months—an order of magnitude faster than classical hardware scaling. The underlying principle is both simple and radical: when you increase data, compute, and model complexity, intelligence emerges. Scaling laws show that larger models—given sufficient compute and high-quality data—become predictably more capable. In just over a decade, we’ve gone from neural nets that could identify cats to systems that can draft legal briefs, write production-grade code, generate scientific hypotheses, and outperform top human competitors in mathematics, strategy, and reasoning tasks. This is no longer “software” in the traditional sense. It is a new form of intelligence—synthetic, scalable, rapidly compounding, and increasingly able to take meaningful action in the real world. The geopolitical, economic, and societal implications of this shift are only beginning to unfold—and they will redefine global power in the decades ahead.
2:04

We’re no longer just scaling computing power. We’re using compute to scale intelligence itself. That’s what makes this moment historically significant. For sixty years, progress in computing followed Moore’s Law—transistor density doubling roughly every two years. But AI is advancing on a far steeper curve. Today, frontier model capabilities are improving on a cadence closer to every six months—an order of magnitude faster than classical hardware scaling. The underlying principle is both simple and radical: when you increase data, compute, and model complexity, intelligence emerges. Scaling laws show that larger models—given sufficient compute and high-quality data—become predictably more capable. In just over a decade, we’ve gone from neural nets that could identify cats to systems that can draft legal briefs, write production-grade code, generate scientific hypotheses, and outperform top human competitors in mathematics, strategy, and reasoning tasks. This is no longer “software” in the traditional sense. It is a new form of intelligence—synthetic, scalable, rapidly compounding, and increasingly able to take meaningful action in the real world. The geopolitical, economic, and societal implications of this shift are only beginning to unfold—and they will redefine global power in the decades ahead.

Nina Schick

17,750 次观看 • 8 个月前

NVIDIA CEO Jensen Huang says one scaling law multiplies AI faster than NVIDIA can hire engineers. Most people know three AI scaling laws. Pre-training. Post-training. Test-time. Each one multiplies intelligence by throwing more compute at a different stage. Jensen Huang says there's a fourth and it's the one that will dominate... Agentic scaling law. "During test time, that agentic system goes off and does research, bangs on databases, uses tools," Huang says. "And one of the most important things it does is spawn off a whole bunch of sub-agents." That's the multiplier. One AI worker can become a team. Then a department. Then a company. "It's so much easier to scale NVIDIA by hiring more employees than it is to scale myself," Huang says. Now imagine scaling without a payroll constraint. "The agentic scaling law — it's kind of like multiplying AI," Huang says. "We could spin off agents as fast as you want to spin off agents." Each agent spins off sub-agents. Each sub-agent spins off more. The compute requirement compounds inside a single query. And every agent generates new data, new experiences, new edge cases. "Wow, this is really good. We ought to memorize this," Huang says. "That data set comes back to pre-training." The four scaling laws don't compete. They feed each other. Agentic systems produce data, which feeds pre-training, which smartens the base model, which enables better agents, which produce more data. A flywheel that compounds forever. The companies pricing in three scaling laws are mispricing the fourth. The fourth eats the other three for lunch. P.S. Pull the thread on any story like this and you'll find the hidden incentive at the other end. As Munger said: "Show me the incentive and I'll show you the outcome." So I wrote a short book on how to spot them and design your own. Comment "INCENTIVES" and I'll send you the details. If you're new here, follow GeniusThinking for content on the greatest minds in economics, psychology, and history. — Jensen Huang ( NVIDIA ), NVIDIA CEO, on Lex Fridman's ( Lex Fridman ) podcast
0:55

NVIDIA CEO Jensen Huang says one scaling law multiplies AI faster than NVIDIA can hire engineers. Most people know three AI scaling laws. Pre-training. Post-training. Test-time. Each one multiplies intelligence by throwing more compute at a different stage. Jensen Huang says there's a fourth and it's the one that will dominate... Agentic scaling law. "During test time, that agentic system goes off and does research, bangs on databases, uses tools," Huang says. "And one of the most important things it does is spawn off a whole bunch of sub-agents." That's the multiplier. One AI worker can become a team. Then a department. Then a company. "It's so much easier to scale NVIDIA by hiring more employees than it is to scale myself," Huang says. Now imagine scaling without a payroll constraint. "The agentic scaling law — it's kind of like multiplying AI," Huang says. "We could spin off agents as fast as you want to spin off agents." Each agent spins off sub-agents. Each sub-agent spins off more. The compute requirement compounds inside a single query. And every agent generates new data, new experiences, new edge cases. "Wow, this is really good. We ought to memorize this," Huang says. "That data set comes back to pre-training." The four scaling laws don't compete. They feed each other. Agentic systems produce data, which feeds pre-training, which smartens the base model, which enables better agents, which produce more data. A flywheel that compounds forever. The companies pricing in three scaling laws are mispricing the fourth. The fourth eats the other three for lunch. P.S. Pull the thread on any story like this and you'll find the hidden incentive at the other end. As Munger said: "Show me the incentive and I'll show you the outcome." So I wrote a short book on how to spot them and design your own. Comment "INCENTIVES" and I'll send you the details. If you're new here, follow GeniusThinking for content on the greatest minds in economics, psychology, and history. — Jensen Huang ( NVIDIA ), NVIDIA CEO, on Lex Fridman's ( Lex Fridman ) podcast

GeniusThinking

93,264 次观看 • 4 个月前

Jason Ma reveals the first true scaling law in robotics: robots trained on 1 million hours of first-person human video predictably get better without ever seeing robot data "Yesterday we announced our new flagship robot foundation model called Dyna-2. It's very significant for the entire field for many different reasons. First of all, it's the first robot foundation model trained on at least one million hours of data. That in itself is a very challenging infrastructure challenge." "What we demonstrated is that even if the one million hours of data is entirely first-person video of humans doing manipulation tasks, we actually see scaling law for transferring to robot embodiments that the model has never seen before." "By training on more human data, we actually see predictable performance improvement on robots. That's a big deal because the biggest challenge facing robot foundation models is that we don't have enough data." "Unlike language models, there are just not enough robotics data out there, and it's very difficult to collect these robot data. Only by passively observing humans doing things can we scale this kind of data, and this is the first time we showed there is a lot of improvement that can transfer." Jason Ma Dyna Robotics
1:14

Jason Ma reveals the first true scaling law in robotics: robots trained on 1 million hours of first-person human video predictably get better without ever seeing robot data "Yesterday we announced our new flagship robot foundation model called Dyna-2. It's very significant for the entire field for many different reasons. First of all, it's the first robot foundation model trained on at least one million hours of data. That in itself is a very challenging infrastructure challenge." "What we demonstrated is that even if the one million hours of data is entirely first-person video of humans doing manipulation tasks, we actually see scaling law for transferring to robot embodiments that the model has never seen before." "By training on more human data, we actually see predictable performance improvement on robots. That's a big deal because the biggest challenge facing robot foundation models is that we don't have enough data." "Unlike language models, there are just not enough robotics data out there, and it's very difficult to collect these robot data. Only by passively observing humans doing things can we scale this kind of data, and this is the first time we showed there is a lot of improvement that can transfer." Jason Ma Dyna Robotics

MTS

17,740 次观看 • 1 个月前

If you want a vision encoder for dexterous manipulation, what should be the most important part to model? 🤔 Current standard models like CLIP, SigLIP, and DINOv2 have an incredible grasp of semantics and spatial details. But they lack the action-centric structure needed for downstream visuomotor control. But collecting annotated robotic trajectories at scale is SUPER expensive and largely unrealistic. So, how do we bridge this gap? We introduce CAIP (Contrastive Action-Image Pre-training) ⬇️ 🔸 Action-centric upstream: we align visual observations with action chunks through a contrastive objective. 🔸 Human video as a proxy: we represent 3D human hand poses analogously to robotic end-effector actions, tapping into a massive source of human demonstrations. 🔸 Massive scale: pre-trained on over 32,000 hours of manipulation video, driving both sample efficiency and robust generalization. 🔸 Hardware proven: achieves a 76% average success rate on a real-world Dexmate Vega bimanual Dexmate manipulator with dual 22-DoF Sharpa Wave hands Sharpa . 🔸 State-of-the-art: significantly outperforms strong baselines like DINOv2, SigLIP, MVP, and Qwen3.5 ViT across complex tasks, even under unexpected lighting changes and visual distractors. 🌐 Project: 📝 Blog: 📄 Paper: 💻 Model:
1:30

If you want a vision encoder for dexterous manipulation, what should be the most important part to model? 🤔 Current standard models like CLIP, SigLIP, and DINOv2 have an incredible grasp of semantics and spatial details. But they lack the action-centric structure needed for downstream visuomotor control. But collecting annotated robotic trajectories at scale is SUPER expensive and largely unrealistic. So, how do we bridge this gap? We introduce CAIP (Contrastive Action-Image Pre-training) ⬇️ 🔸 Action-centric upstream: we align visual observations with action chunks through a contrastive objective. 🔸 Human video as a proxy: we represent 3D human hand poses analogously to robotic end-effector actions, tapping into a massive source of human demonstrations. 🔸 Massive scale: pre-trained on over 32,000 hours of manipulation video, driving both sample efficiency and robust generalization. 🔸 Hardware proven: achieves a 76% average success rate on a real-world Dexmate Vega bimanual Dexmate manipulator with dual 22-DoF Sharpa Wave hands Sharpa . 🔸 State-of-the-art: significantly outperforms strong baselines like DINOv2, SigLIP, MVP, and Qwen3.5 ViT across complex tasks, even under unexpected lighting changes and visual distractors. 🌐 Project: 📝 Blog: 📄 Paper: 💻 Model:

Roei Herzig

26,451 次观看 • 3 个月前