Loading video...

Video Failed to Load

Go Home

Planning with the views: Can VLMs predict how each camera move changes the view, and plan many such moves ahead? We introduce ViewSuite with 6 DoF camera control and ~165K task instances, testing: Path-to-View View-to-Path Interactive View Planning A sharp Planning Gap emerges: + can roughly "track" how camera...

ManlingLi_'s profile picture

Manling Li

11,259 subscribers

58,249 views • 25 days ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

🚀 New Paper: Pixel Reasoner 🧠🖼️ How can Vision-Language Models (VLMs) perform chain-of-thought reasoning within the image itself? We introduce Pixel Reasoner, the first open-source framework that enables VLMs to “think in pixel space” through curiosity-driven reinforcement learning. Current VLMs reason only in text — even when grounded in rich images or videos, their logical steps are verbalized in natural language. This restricts their ability to interrogate visual evidence and demonstrate how conclusions are drawn. 🔍 So we ask: What if we could make VLMs "show their work" by reasoning directly in the pixel space? Inspired by GPT-o3’s "think-in-image" ability, we propose a framework where VLMs use interactive visual operations — zoom, select-frame, highlight — to reason through complex visual inputs. To do this, we design a two-stage training process: Instruction tuning with synthesized visual reasoning traces. Reinforcement learning with curiosity-driven reward to balance exploration between pixel and text reasoning ✨ With this, Pixel Reasoner achieves near-SoTA performance on many information-rich multimodal benchmarks: 📊 84% on InfographicsVQA 🧠 84% on V* benchmark 🧩 74% on TallyQA-Complex It also achieves strong accuracy of 68% on MVBench (a video benchmark). Website: Paper: Code: Demo: (coming soon)
0:43

🚀 New Paper: Pixel Reasoner 🧠🖼️ How can Vision-Language Models (VLMs) perform chain-of-thought reasoning within the image itself? We introduce Pixel Reasoner, the first open-source framework that enables VLMs to “think in pixel space” through curiosity-driven reinforcement learning. Current VLMs reason only in text — even when grounded in rich images or videos, their logical steps are verbalized in natural language. This restricts their ability to interrogate visual evidence and demonstrate how conclusions are drawn. 🔍 So we ask: What if we could make VLMs "show their work" by reasoning directly in the pixel space? Inspired by GPT-o3’s "think-in-image" ability, we propose a framework where VLMs use interactive visual operations — zoom, select-frame, highlight — to reason through complex visual inputs. To do this, we design a two-stage training process: Instruction tuning with synthesized visual reasoning traces. Reinforcement learning with curiosity-driven reward to balance exploration between pixel and text reasoning ✨ With this, Pixel Reasoner achieves near-SoTA performance on many information-rich multimodal benchmarks: 📊 84% on InfographicsVQA 🧠 84% on V* benchmark 🧩 74% on TallyQA-Complex It also achieves strong accuracy of 68% on MVBench (a video benchmark). Website: Paper: Code: Demo: (coming soon)

Wenhu Chen

82,829 views • 1 year ago

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.
0:50

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 views • 3 years ago

You can now operate robots by just thinking about it. With your brain signals. WOW. This robot system from Stanford has so much sci-fi vibe and wild implications that I don't even know where to start. NOIR decodes the EEG signal from your head into a library of robot skills. It is demonstrated on 20 household activities, such as cooking Sukiyaki, ironing clothes, grating cheese, playing Tic-Tac-Toe, and even petting a robot dog! NOIR learns to predict your intended goals in advance, so that your thinking efforts (literally) can be reduced to a minimum. It works with both adults and children as young as five years old. Imagine scaling up this system to Neuralink + Humanoid robots like Tesla Optimus in the future! There's now a decent probability that we can see the movie Avatar become a reality in our lifetime. Website: Paper: This work is from Stanford Vision Lab, my PhD home. The project is led by Ruohan Zhang and advised by Fei-Fei Li Jiajun Wu. Congrats on the phenomenal idea!
0:33

You can now operate robots by just thinking about it. With your brain signals. WOW. This robot system from Stanford has so much sci-fi vibe and wild implications that I don't even know where to start. NOIR decodes the EEG signal from your head into a library of robot skills. It is demonstrated on 20 household activities, such as cooking Sukiyaki, ironing clothes, grating cheese, playing Tic-Tac-Toe, and even petting a robot dog! NOIR learns to predict your intended goals in advance, so that your thinking efforts (literally) can be reduced to a minimum. It works with both adults and children as young as five years old. Imagine scaling up this system to Neuralink + Humanoid robots like Tesla Optimus in the future! There's now a decent probability that we can see the movie Avatar become a reality in our lifetime. Website: Paper: This work is from Stanford Vision Lab, my PhD home. The project is led by Ruohan Zhang and advised by Fei-Fei Li Jiajun Wu. Congrats on the phenomenal idea!

Jim Fan

378,309 views • 2 years ago

V3D Video Diffusion Models are Effective 3D Generators Automatic 3D generation has recently attracted widespread attention. Recent methods have greatly accelerated the generation speed, but usually produce less-detailed objects due to limited model capacity or 3D data. Motivated by recent advancements in video diffusion models, we introduce V3D, which leverages the world simulation capacity of pre-trained video diffusion models to facilitate 3D generation. To fully unleash the potential of video diffusion to perceive the 3D world, we further introduce geometrical consistency prior and extend the video diffusion model to a multi-view consistent 3D generator. Benefiting from this, the state-of-the-art video diffusion model could be fine-tuned to generate 360degree orbit frames surrounding an object given a single image. With our tailored reconstruction pipelines, we can generate high-quality meshes or 3D Gaussians within 3 minutes. Furthermore, our method can be extended to scene-level novel view synthesis, achieving precise control over the camera path with sparse input views. Extensive experiments demonstrate the superior performance of the proposed approach, especially in terms of generation quality and multi-view consistency
0:21

V3D Video Diffusion Models are Effective 3D Generators Automatic 3D generation has recently attracted widespread attention. Recent methods have greatly accelerated the generation speed, but usually produce less-detailed objects due to limited model capacity or 3D data. Motivated by recent advancements in video diffusion models, we introduce V3D, which leverages the world simulation capacity of pre-trained video diffusion models to facilitate 3D generation. To fully unleash the potential of video diffusion to perceive the 3D world, we further introduce geometrical consistency prior and extend the video diffusion model to a multi-view consistent 3D generator. Benefiting from this, the state-of-the-art video diffusion model could be fine-tuned to generate 360degree orbit frames surrounding an object given a single image. With our tailored reconstruction pipelines, we can generate high-quality meshes or 3D Gaussians within 3 minutes. Furthermore, our method can be extended to scene-level novel view synthesis, achieving precise control over the camera path with sparse input views. Extensive experiments demonstrate the superior performance of the proposed approach, especially in terms of generation quality and multi-view consistency

AK

31,997 views • 2 years ago

Punchline: World models == VQA (about the future)! Planning with world models can be powerful for robotics/control. But most world models are video generators trained to predict everything, including irrelevant pixels and distractions. We ask - what if a world model only predicted the semantic information necessary for decision-making? Introducing Semantic World Models (SWM). Given an observation and an action sequence, SWMs cast modeling as answering textual questions about the future outcome resulting from the actions. Recasting world modeling as a VQA problem lets us directly leverage the pretrained knowledge and machinery of VLMs for generalizable modeling. We had a lot of fun thinking about how this work helps connect these two seemingly very different fields of study - VLMs and world models! 🧵(1/6) Paper: Fun demo:
0:49

Punchline: World models == VQA (about the future)! Planning with world models can be powerful for robotics/control. But most world models are video generators trained to predict everything, including irrelevant pixels and distractions. We ask - what if a world model only predicted the semantic information necessary for decision-making? Introducing Semantic World Models (SWM). Given an observation and an action sequence, SWMs cast modeling as answering textual questions about the future outcome resulting from the actions. Recasting world modeling as a VQA problem lets us directly leverage the pretrained knowledge and machinery of VLMs for generalizable modeling. We had a lot of fun thinking about how this work helps connect these two seemingly very different fields of study - VLMs and world models! 🧵(1/6) Paper: Fun demo:

Abhishek Gupta

61,199 views • 8 months ago

Glad that our work “Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling”, led by Han Qi, has been accepted to IEEE Robotics and Automation Letters! 🎉 We propose Generative Predictive Control (GPC): sample action proposals from a pretrained diffusion policy (“look back”), roll them out with a diffusion-based action-conditioned video world model (“look forward”), then rank or optimize the actions using either a learned reward model or VLM preferences. Conceptually, this is trajectory optimization / MPC with hybrid sampling + gradient optimization, interpreted through modern diffusion priors and video world models. Interestingly, we first posted the paper on arXiv in Feb 2025, when action-conditioned video world models for planning were still rare—now this direction is rapidly gaining traction. Still many open questions, e.g., • how to avoid local minima in planning • what representations work best for world models • how to balance physics priors vs. data-driven learning Paper:
1:00

Glad that our work “Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling”, led by Han Qi, has been accepted to IEEE Robotics and Automation Letters! 🎉 We propose Generative Predictive Control (GPC): sample action proposals from a pretrained diffusion policy (“look back”), roll them out with a diffusion-based action-conditioned video world model (“look forward”), then rank or optimize the actions using either a learned reward model or VLM preferences. Conceptually, this is trajectory optimization / MPC with hybrid sampling + gradient optimization, interpreted through modern diffusion priors and video world models. Interestingly, we first posted the paper on arXiv in Feb 2025, when action-conditioned video world models for planning were still rare—now this direction is rapidly gaining traction. Still many open questions, e.g., • how to avoid local minima in planning • what representations work best for world models • how to balance physics priors vs. data-driven learning Paper:

Heng Yang

18,994 views • 4 months ago

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀
1:32

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀

Roberta Raileanu

104,982 views • 1 year ago

Today we are excited to open up Neuracore to the academic community! Neuracore is a new data foundation built to accelerate robot learning by removing one of the field’s biggest bottlenecks: capturing and working with high-fidelity multimodal robotics data. For the first time, researchers can store, view, and work with robotics data in a cloud-native system built specifically for large-scale learning, and we are making this core platform completely free for academia. The platform lets teams capture every sensor at its native rate, store and visualize data without loss, and then train and deploy models locally using our open-source code (Link in the comments). We are rolling out access to select academic institutions first. Anyone with an academic email can sign up, and if your institution is not part of the initial rollout, you will be able to join the waitlist directly. Beyond providing this infrastructure, we see an opportunity to build a global community where engineers and researchers can share, collaborate, and advance the frontier of robot learning together. Supported by our recent $3M pre-seed round led by Earlybird VC, we are excited to take this mission even further. Our long-term goal is for Neuracore to become the natural home for cutting-edge robot learning algorithms and real-world robotics experimentation, helping accelerate the next wave of Physical AI.
1:00

Today we are excited to open up Neuracore to the academic community! Neuracore is a new data foundation built to accelerate robot learning by removing one of the field’s biggest bottlenecks: capturing and working with high-fidelity multimodal robotics data. For the first time, researchers can store, view, and work with robotics data in a cloud-native system built specifically for large-scale learning, and we are making this core platform completely free for academia. The platform lets teams capture every sensor at its native rate, store and visualize data without loss, and then train and deploy models locally using our open-source code (Link in the comments). We are rolling out access to select academic institutions first. Anyone with an academic email can sign up, and if your institution is not part of the initial rollout, you will be able to join the waitlist directly. Beyond providing this infrastructure, we see an opportunity to build a global community where engineers and researchers can share, collaborate, and advance the frontier of robot learning together. Supported by our recent $3M pre-seed round led by Earlybird VC, we are excited to take this mission even further. Our long-term goal is for Neuracore to become the natural home for cutting-edge robot learning algorithms and real-world robotics experimentation, helping accelerate the next wave of Physical AI.

Neuracore

40,620 views • 7 months ago

Milestone! We (robotic arms for gadgets assembly) finished the first commercial order, which brought the first revenue. Here are some learnings from this: The customer was a smart toy manufacturer. The task was to add a heatsink to Raspberry Pi. We received parts from them and returned the assembled modules back. Currently, it's done by teleoperation. Later it will be done by a remote employee via the Internet. Then it will be automated action by action, reducing the operator's time on this and making the task profitable. ps. If you have an assembly task that we can do for you asynchronically - leave a comment below. Learning 1. It's possible! This task which is usually done by the human arm with 5 fingers can be done with a two-finger gripper with the addition of a couple of simple tooling. The task was not simplified. We peeled off thin films from stickers, unpacked paper boxes, moved PCB boards full of components, etc. And no unsolvable problems have been encountered yet. Challenges: 1) The paper box shifted during the opening Solved with the plastic walls that you can lean against 2) Heat pad, stuck to the gripper instead of heat sync. Can be solved by gripper with a pump, but this time solved with the patience of the operator 3) The film on the pad is very thin. Turned out that sub-millimeter arm precision is enough to peel it off with just a regular gripper. 4) The working area has not enough space. You'll only know this by doing real tasks in bulk. This could be solved by an extra pair of long arms, but in this case, solved with the patience of the operator. I think that in the end, we will have 5-10 types of universal tooling and 5-10 types of grippers to solve almost all the problems in such assembly tasks. Learning 2. It's slow. It took 5 times more time, than doing it with human hands. But the good news is there's a lot of room for improvement. We now have specific “time for task” metrics, which we will decrease with iterations. The main reasons for slowness: 1) To rotate the gripper to a steep angle you are forced to control one robot arm with two hands instead of using both arms. We can fix this by just making more room for rotations. 2) Grabbing PCB board with two arms is hard. A slight difference in rotation can break the board, and it's hard to control these angles visually. To solve this, the best way is to use force feedback so you can feel the pressure applied to the item. 3) Accuracy and steadiness is still can be improved We will try a metal version and double the motors to do this. 4) It is physically difficult for the human hands to move with such precision To solve this, we will add a pad for the hands like in surgical robots Learning 3. It's a good business model The "Factory in the cloud" is a good business model for this stage. You send us parts and we send back assembled modules. Currently, it's more convenient than sending a robot to your place, as we can iterate/fix the robot quickly and utilize it 100% of the time. When we polish the set-up over time - we can send robots to your place. So if we can assemble something for you in the USA with Chinese prices by using modern automation - leave a comment below.
0:48

Milestone! We (robotic arms for gadgets assembly) finished the first commercial order, which brought the first revenue. Here are some learnings from this: The customer was a smart toy manufacturer. The task was to add a heatsink to Raspberry Pi. We received parts from them and returned the assembled modules back. Currently, it's done by teleoperation. Later it will be done by a remote employee via the Internet. Then it will be automated action by action, reducing the operator's time on this and making the task profitable. ps. If you have an assembly task that we can do for you asynchronically - leave a comment below. Learning 1. It's possible! This task which is usually done by the human arm with 5 fingers can be done with a two-finger gripper with the addition of a couple of simple tooling. The task was not simplified. We peeled off thin films from stickers, unpacked paper boxes, moved PCB boards full of components, etc. And no unsolvable problems have been encountered yet. Challenges: 1) The paper box shifted during the opening Solved with the plastic walls that you can lean against 2) Heat pad, stuck to the gripper instead of heat sync. Can be solved by gripper with a pump, but this time solved with the patience of the operator 3) The film on the pad is very thin. Turned out that sub-millimeter arm precision is enough to peel it off with just a regular gripper. 4) The working area has not enough space. You'll only know this by doing real tasks in bulk. This could be solved by an extra pair of long arms, but in this case, solved with the patience of the operator. I think that in the end, we will have 5-10 types of universal tooling and 5-10 types of grippers to solve almost all the problems in such assembly tasks. Learning 2. It's slow. It took 5 times more time, than doing it with human hands. But the good news is there's a lot of room for improvement. We now have specific “time for task” metrics, which we will decrease with iterations. The main reasons for slowness: 1) To rotate the gripper to a steep angle you are forced to control one robot arm with two hands instead of using both arms. We can fix this by just making more room for rotations. 2) Grabbing PCB board with two arms is hard. A slight difference in rotation can break the board, and it's hard to control these angles visually. To solve this, the best way is to use force feedback so you can feel the pressure applied to the item. 3) Accuracy and steadiness is still can be improved We will try a metal version and double the motors to do this. 4) It is physically difficult for the human hands to move with such precision To solve this, we will add a pad for the hands like in surgical robots Learning 3. It's a good business model The "Factory in the cloud" is a good business model for this stage. You send us parts and we send back assembled modules. Currently, it's more convenient than sending a robot to your place, as we can iterate/fix the robot quickly and utilize it 100% of the time. When we polish the set-up over time - we can send robots to your place. So if we can assemble something for you in the USA with Chinese prices by using modern automation - leave a comment below.

Igor Kulakov

37,266 views • 1 year ago

$PLTR CTO Shyam Sankar on Deepseek and disagreeing with OpenAI CEO Sam Altman 🔮 “We are at war with China. We are in an AI arms race. I got into a disagreement with Sam Altman about this at the Senate AI summit over a year ago… There’s an opposing view that we can all get along and cooperate on these things. I’m sure the other side feels quite different realizing that their IP has been stolen. “How many times are we going to believe that anchors dragged across the sea that cut undersea cables? “China knows they’re at war. We kinda equivocate on it as a peace loving nation. “The engineering on R1 is exquisite. The optimizations they’ve done are really impressive… We have to wake up with respect for our adversary and realize that we are competing. But they absolutely did steal a lot of that through distillation of the models, and perhaps they stole even more. “We have to realize that the AI race is winner take all, and it’s going to be a whole of nation effort that extends well beyond the DoD in order for us as a nation to win.” 🇺🇸
2:09

$PLTR CTO Shyam Sankar on Deepseek and disagreeing with OpenAI CEO Sam Altman 🔮 “We are at war with China. We are in an AI arms race. I got into a disagreement with Sam Altman about this at the Senate AI summit over a year ago… There’s an opposing view that we can all get along and cooperate on these things. I’m sure the other side feels quite different realizing that their IP has been stolen. “How many times are we going to believe that anchors dragged across the sea that cut undersea cables? “China knows they’re at war. We kinda equivocate on it as a peace loving nation. “The engineering on R1 is exquisite. The optimizations they’ve done are really impressive… We have to wake up with respect for our adversary and realize that we are competing. But they absolutely did steal a lot of that through distillation of the models, and perhaps they stole even more. “We have to realize that the AI race is winner take all, and it’s going to be a whole of nation effort that extends well beyond the DoD in order for us as a nation to win.” 🇺🇸

Jawwwn

295,441 views • 1 year ago

Wim Wenders on the "camera movement and blocking" in Wings of Desire (1987): Filmmaker Magazine: "What was your philosophy about camera movement and blocking in Wings of Desire?" Wenders: "As we very often had to “translate” the angels’ point of view, so to speak, we were extremely keen on moving the camera as much as possible. In the absence of Steadicam equipment we worked a lot on tracks, with dollies, cranes, jib-arms etc. But we also built ourselves devices so we could move through the air from one house to the next, for instance, and we shot the opening sequence on a helicopter, which was highly difficult in West Berlin at the time, as there were no private companies flying, just the Allies with their respective army pilots. We ended up shooting with a British pilot in an army helicopter without a proper camera mount. Today, you would do these things with gyroscopes and such. Blocking has always been my department. Henri [Alekan] kept out of it completely, and I did it with his operator, Agnès Godard. I have done shot lists for complicated sets, but usually I decide on location in the morning how we design the shots. I prefer to see the actors rehearse it, before I commit to any blocking. Camera moves weren’t the real challenge, though, for finding the angels’ points of view. It dawned on me early on that our camera had to do a more complex job. I told it to Henri. “Those angels have a very loving look at us humans. We have to find a way to teach our camera to look more lovingly.” Henri just stared at me as if I was out of my mind. “How do we do that?” Well, I didn’t know of course. But I figured we had to invest more care and love ourselves into every shot that represented what the angels saw. And that’s it, in the end. A camera can reflect on what you invest into its act of seeing. That sounds pretty lofty, I guess. But it does rub off, I tell you, if you try to imagine how angels would look at us. After all, they were some sort of metaphor for me for the better persons we carry inside ourselves, or for the children we somehow preserved in ourselves." — "“Imagine How Angels Would Look at Us”: Wim Wenders on Restoring Wings of Desire" by Jim Hemphill (Filmmaker Magazine, 2018)
0:52

Wim Wenders on the "camera movement and blocking" in Wings of Desire (1987): Filmmaker Magazine: "What was your philosophy about camera movement and blocking in Wings of Desire?" Wenders: "As we very often had to “translate” the angels’ point of view, so to speak, we were extremely keen on moving the camera as much as possible. In the absence of Steadicam equipment we worked a lot on tracks, with dollies, cranes, jib-arms etc. But we also built ourselves devices so we could move through the air from one house to the next, for instance, and we shot the opening sequence on a helicopter, which was highly difficult in West Berlin at the time, as there were no private companies flying, just the Allies with their respective army pilots. We ended up shooting with a British pilot in an army helicopter without a proper camera mount. Today, you would do these things with gyroscopes and such. Blocking has always been my department. Henri [Alekan] kept out of it completely, and I did it with his operator, Agnès Godard. I have done shot lists for complicated sets, but usually I decide on location in the morning how we design the shots. I prefer to see the actors rehearse it, before I commit to any blocking. Camera moves weren’t the real challenge, though, for finding the angels’ points of view. It dawned on me early on that our camera had to do a more complex job. I told it to Henri. “Those angels have a very loving look at us humans. We have to find a way to teach our camera to look more lovingly.” Henri just stared at me as if I was out of my mind. “How do we do that?” Well, I didn’t know of course. But I figured we had to invest more care and love ourselves into every shot that represented what the angels saw. And that’s it, in the end. A camera can reflect on what you invest into its act of seeing. That sounds pretty lofty, I guess. But it does rub off, I tell you, if you try to imagine how angels would look at us. After all, they were some sort of metaphor for me for the better persons we carry inside ourselves, or for the children we somehow preserved in ourselves." — "“Imagine How Angels Would Look at Us”: Wim Wenders on Restoring Wings of Desire" by Jim Hemphill (Filmmaker Magazine, 2018)

RadiantFilm

18,101 views • 5 months ago

Check out our latest work, "Actor-Critic Model Predictive Control: Differentiable Optimization meets Reinforcement Learning for Agile Flight," published in the IEEE Transactions on Robotics, where we reconcile #OptimalControl and #ReinforcementLearning, achieving the same super-human performance, but with superior generalizability, as our previous model-free deep RL! Code released! PDF: Code: Full Video: Model-free #ReinforcementLearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) provides robustness, constraint handling, and powerful online replanning capabilities. In this work, we extend our previous AC-MPC paper (Romero, ICRA'24) by taking a deeper look at how both approaches can be unified. We introduce and extend Actor-Critic Model Predictive Control (AC-MPC), a framework that embeds a differentiable MPC inside an Actor-Critic RL architecture. This integration allows the MPC-based actor to perform short-term predictive optimization, while the critic facilitates long-horizon learning and exploration. We conduct a comprehensive study that highlights AC-MPC’s key advantages: - Better out-of-distribution generalization, both against unknown disturbances and changes in the quadrotor dynamics - Improved sample efficiency - A novel empirical analysis uncovering a relationship between the critic’s value function and the MPC cost function, providing deeper insight into their interplay. We validate our method in simulation and the real world on a quadcopter flying at superhuman speeds of up to 21 m/s, matching state-of-the-art model-free RL performance, and retaining the predictive structure of MPC for more reliable out-of-distribution behavior. Reference: Actor-Critic Model Predictive Control: Differentiable Optimization meets Reinforcement Learning for Agile Flight IEEE Transactions on Robotics (T-RO), 2025 PDF: Full Video: Code: Kudos to Ángel Romero, Elie Aljalbout, Yunlong Song! University of Zurich UZH Science UZH Space Hub AUTOASSESS European Research Council (ERC) UZHai
1:46

Check out our latest work, "Actor-Critic Model Predictive Control: Differentiable Optimization meets Reinforcement Learning for Agile Flight," published in the IEEE Transactions on Robotics, where we reconcile #OptimalControl and #ReinforcementLearning, achieving the same super-human performance, but with superior generalizability, as our previous model-free deep RL! Code released! PDF: Code: Full Video: Model-free #ReinforcementLearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) provides robustness, constraint handling, and powerful online replanning capabilities. In this work, we extend our previous AC-MPC paper (Romero, ICRA'24) by taking a deeper look at how both approaches can be unified. We introduce and extend Actor-Critic Model Predictive Control (AC-MPC), a framework that embeds a differentiable MPC inside an Actor-Critic RL architecture. This integration allows the MPC-based actor to perform short-term predictive optimization, while the critic facilitates long-horizon learning and exploration. We conduct a comprehensive study that highlights AC-MPC’s key advantages: - Better out-of-distribution generalization, both against unknown disturbances and changes in the quadrotor dynamics - Improved sample efficiency - A novel empirical analysis uncovering a relationship between the critic’s value function and the MPC cost function, providing deeper insight into their interplay. We validate our method in simulation and the real world on a quadcopter flying at superhuman speeds of up to 21 m/s, matching state-of-the-art model-free RL performance, and retaining the predictive structure of MPC for more reliable out-of-distribution behavior. Reference: Actor-Critic Model Predictive Control: Differentiable Optimization meets Reinforcement Learning for Agile Flight IEEE Transactions on Robotics (T-RO), 2025 PDF: Full Video: Code: Kudos to Ángel Romero, Elie Aljalbout, Yunlong Song! University of Zurich UZH Science UZH Space Hub AUTOASSESS European Research Council (ERC) UZHai

Davide Scaramuzza

27,090 views • 5 months ago

🇨🇭 BADMINTON-BOT TAKES THE COURT IN SWITZERLAND A Swiss team built a four-legged robot that plays badminton against humans and actually returns shots like it means it. Created with ETH Zurich, this bot uses artificial intelligence to teach itself how to move, see, and hit the shuttlecock with freakish precision. It learns through reinforcement learning which is a fancy way of saying it fails over and over until it figures things out. A special vision system helps it track the birdie while running around like a robot goat with a grudge. It even adjusts its movement to avoid messing up its own view. Researcher, Yuntao Ma: "We believe that this project demonstrates the potential of artificial intelligence to advance legged robots in complex tasks, paving the way for advances in autonomous and intelligent systems, including humanoids. The control algorithm generalizes to other robotic platforms such as humanoids, as well as to other tasks such as search and rescue and home services." Source: Telegrafi Media Source: The Independent On Youtube
3:18

🇨🇭 BADMINTON-BOT TAKES THE COURT IN SWITZERLAND A Swiss team built a four-legged robot that plays badminton against humans and actually returns shots like it means it. Created with ETH Zurich, this bot uses artificial intelligence to teach itself how to move, see, and hit the shuttlecock with freakish precision. It learns through reinforcement learning which is a fancy way of saying it fails over and over until it figures things out. A special vision system helps it track the birdie while running around like a robot goat with a grudge. It even adjusts its movement to avoid messing up its own view. Researcher, Yuntao Ma: "We believe that this project demonstrates the potential of artificial intelligence to advance legged robots in complex tasks, paving the way for advances in autonomous and intelligent systems, including humanoids. The control algorithm generalizes to other robotic platforms such as humanoids, as well as to other tasks such as search and rescue and home services." Source: Telegrafi Media Source: The Independent On Youtube

Mario Nawfal

76,687 views • 1 year ago

📢📢 𝐏𝐞𝐫𝐜𝐇𝐞𝐚𝐝: 𝐏𝐞𝐫𝐜𝐞𝐩𝐭𝐮𝐚𝐥 𝐇𝐞𝐚𝐝 𝐌𝐨𝐝𝐞𝐥 𝐟𝐨𝐫 𝐒𝐢𝐧𝐠𝐥𝐞-𝐈𝐦𝐚𝐠𝐞 𝟑𝐃 𝐇𝐞𝐚𝐝 𝐑𝐞𝐜𝐨𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 & 𝐄𝐝𝐢𝐭𝐢𝐧𝐠📢📢 PercHead reconstructs realistic 3D heads from a single image and enables disentangled 3D editing via geometric controls and style inputs from images or text. At its core is a generalized 3D head decoder trained with perceptual supervision from DINOv2 and SAM 2.1. We find that our new perceptual loss formulation improves reconstruction fidelity compared to commonly-used methods such as LPIPS. Our trained reconstruction model is able to generate 3D-consistent heads from a single input image. Even with challenging side-view inputs, the model robustly infers missing regions for a coherent, high-fidelity output. In addition, our architecture seamlessly adapts to downstream tasks: by swapping the encoder, we can transform the model into a disentangled 3D editing pipeline. In this scenario, we can control geometry through - potentially hand-drawn - segmentation maps, and condition style via image or text prompt. We also provide an interactive GUI to enable the exploration of our editing pipeline. 🌍 📽️ Great work by Antonio Oroz and Tobias Kirschstein
0:30

📢📢 𝐏𝐞𝐫𝐜𝐇𝐞𝐚𝐝: 𝐏𝐞𝐫𝐜𝐞𝐩𝐭𝐮𝐚𝐥 𝐇𝐞𝐚𝐝 𝐌𝐨𝐝𝐞𝐥 𝐟𝐨𝐫 𝐒𝐢𝐧𝐠𝐥𝐞-𝐈𝐦𝐚𝐠𝐞 𝟑𝐃 𝐇𝐞𝐚𝐝 𝐑𝐞𝐜𝐨𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 & 𝐄𝐝𝐢𝐭𝐢𝐧𝐠📢📢 PercHead reconstructs realistic 3D heads from a single image and enables disentangled 3D editing via geometric controls and style inputs from images or text. At its core is a generalized 3D head decoder trained with perceptual supervision from DINOv2 and SAM 2.1. We find that our new perceptual loss formulation improves reconstruction fidelity compared to commonly-used methods such as LPIPS. Our trained reconstruction model is able to generate 3D-consistent heads from a single input image. Even with challenging side-view inputs, the model robustly infers missing regions for a coherent, high-fidelity output. In addition, our architecture seamlessly adapts to downstream tasks: by swapping the encoder, we can transform the model into a disentangled 3D editing pipeline. In this scenario, we can control geometry through - potentially hand-drawn - segmentation maps, and condition style via image or text prompt. We also provide an interactive GUI to enable the exploration of our editing pipeline. 🌍 📽️ Great work by Antonio Oroz and Tobias Kirschstein

Matthias Niessner

18,855 views • 8 months ago

Today, we are launching the first publicly available AI Scientist, via the FutureHouse Platform. Our AI Scientist agents can perform a wide variety of scientific tasks better than humans. By chaining them together, we've already started to discover new biology really fast. With the platform, we are bringing these capabilities to the wider community. Watch our long-form video, in the comments below, to learn more about how the platform works and how you can use it to make new discoveries, and go to our website or see the comments below to access the platform. We are releasing three superhuman AI Scientist agents today, each with their own specialization: A general-purpose agent (Crow); An agent to automate literature reviews (Falcon); and An agent to answer the question “Has anyone done X before” (Owl). We are also releasing an experimental agent, Phoenix, that has access to a wide variety of tools for planning experiments in chemistry. More on that below. The three literature search agents (Crow, Falcon, and Owl) have benchmarked superhuman performance. They also have access to a large corpus of full scientific texts, which means that you can ask them more detailed questions about experimental protocols and study limitations that general-purpose web search agents, which usually only have access to abstracts, might miss. Our agents also use a variety of factors to distinguish source quality, so that they don’t end up relying on low-quality papers or pop-science sources. Finally, and critically, we have an API, which is intended to allow researchers to integrate our agents into their workflows. Phoenix is an experimental project we put together recently just to demonstrate what can happen if you give the agents access to lots of scientific tools. It is not better than humans at planning experiments yet, and it makes a lot more mistakes than Crow, Falcon, or Owl. We want to see all the ways you can break it! The agents we are releasing today cannot yet do all (or even most!) aspects of scientific research autonomously. However, as we show in the video, you can already use them to generate and evaluate new hypotheses and plan new experiments way faster than before. Internally, we also have dedicated agents for data analysis, hypothesis generation, protein engineering, and more, and we plan to launch these on the platform in the coming months as well. Within a year or two, it is easy to imagine that the vast majority of desk work that scientists do today will be accelerated with the help of AI agents like the ones we are releasing today. The platform is currently free-to-use. Over time, depending on how people use it, we may implement pricing plans. If you want higher rate limits, especially for research projects, get in touch. Michael Skarlinski, Andrew White 🐦‍⬛, Tyler Nadolski, Remo Storni, James Braza, Ludovico Mitchener, Michaela Hinks, as well as Jason Carman and his team for making such fantastic videos of us!
1:21

Today, we are launching the first publicly available AI Scientist, via the FutureHouse Platform. Our AI Scientist agents can perform a wide variety of scientific tasks better than humans. By chaining them together, we've already started to discover new biology really fast. With the platform, we are bringing these capabilities to the wider community. Watch our long-form video, in the comments below, to learn more about how the platform works and how you can use it to make new discoveries, and go to our website or see the comments below to access the platform. We are releasing three superhuman AI Scientist agents today, each with their own specialization: A general-purpose agent (Crow); An agent to automate literature reviews (Falcon); and An agent to answer the question “Has anyone done X before” (Owl). We are also releasing an experimental agent, Phoenix, that has access to a wide variety of tools for planning experiments in chemistry. More on that below. The three literature search agents (Crow, Falcon, and Owl) have benchmarked superhuman performance. They also have access to a large corpus of full scientific texts, which means that you can ask them more detailed questions about experimental protocols and study limitations that general-purpose web search agents, which usually only have access to abstracts, might miss. Our agents also use a variety of factors to distinguish source quality, so that they don’t end up relying on low-quality papers or pop-science sources. Finally, and critically, we have an API, which is intended to allow researchers to integrate our agents into their workflows. Phoenix is an experimental project we put together recently just to demonstrate what can happen if you give the agents access to lots of scientific tools. It is not better than humans at planning experiments yet, and it makes a lot more mistakes than Crow, Falcon, or Owl. We want to see all the ways you can break it! The agents we are releasing today cannot yet do all (or even most!) aspects of scientific research autonomously. However, as we show in the video, you can already use them to generate and evaluate new hypotheses and plan new experiments way faster than before. Internally, we also have dedicated agents for data analysis, hypothesis generation, protein engineering, and more, and we plan to launch these on the platform in the coming months as well. Within a year or two, it is easy to imagine that the vast majority of desk work that scientists do today will be accelerated with the help of AI agents like the ones we are releasing today. The platform is currently free-to-use. Over time, depending on how people use it, we may implement pricing plans. If you want higher rate limits, especially for research projects, get in touch. Michael Skarlinski, Andrew White 🐦‍⬛, Tyler Nadolski, Remo Storni, James Braza, Ludovico Mitchener, Michaela Hinks, as well as Jason Carman and his team for making such fantastic videos of us!

Sam Rodriques

724,665 views • 1 year ago

Dr. Fei-Fei Li just called out the biggest blind spot in the entire AI industry. We have been building half of human intelligence. And calling it the finish line. Li: “If you look at human intelligence, it pretty much boils down to two buckets.” The first bucket is language. Symbolic reasoning. Communication. The ability to think in words and abstractions. That’s what every major AI lab has spent the last decade building. The second bucket is the one the industry has almost entirely ignored. Li: “We call that in AI spatial intelligence.” How humans and animals perceive, navigate, and interact with the three-dimensional physical world. How we reach for objects. How we move through space. How we build and manipulate physical reality. From painting masterpieces to constructing the pyramids, non-verbal spatial intelligence is what actually shapes the world. Language describes reality. Spatial intelligence acts on it. And the gap between those two things is the gap between a chatbot and a robot. Li: “When this technology is ready, the robotic revolution is gonna start. We’re already seeing that trend.” Every robot is a moving agent. Every moving agent requires spatial intelligence to function in the real world. The humanoid robots being deployed in factories right now are hitting the ceiling of what language models alone can power. Spatial intelligence is the unlock. But Li didn’t stop at robotics. Li: “From a geopolitics point of view, this is part of the technology that goes straight into weapons.” Autonomous drone swarms. Battlefield navigation. Physical target acquisition without human oversight. Every military application of AI that operates in the real world runs on spatial intelligence. The nation that masters the transition from static text to dynamic three-dimensional perception doesn’t just win the software race. It commands the physical battlefield. The AI arms race just broke out of the data center. It’s operating in three dimensions now.
3:08

Dr. Fei-Fei Li just called out the biggest blind spot in the entire AI industry. We have been building half of human intelligence. And calling it the finish line. Li: “If you look at human intelligence, it pretty much boils down to two buckets.” The first bucket is language. Symbolic reasoning. Communication. The ability to think in words and abstractions. That’s what every major AI lab has spent the last decade building. The second bucket is the one the industry has almost entirely ignored. Li: “We call that in AI spatial intelligence.” How humans and animals perceive, navigate, and interact with the three-dimensional physical world. How we reach for objects. How we move through space. How we build and manipulate physical reality. From painting masterpieces to constructing the pyramids, non-verbal spatial intelligence is what actually shapes the world. Language describes reality. Spatial intelligence acts on it. And the gap between those two things is the gap between a chatbot and a robot. Li: “When this technology is ready, the robotic revolution is gonna start. We’re already seeing that trend.” Every robot is a moving agent. Every moving agent requires spatial intelligence to function in the real world. The humanoid robots being deployed in factories right now are hitting the ceiling of what language models alone can power. Spatial intelligence is the unlock. But Li didn’t stop at robotics. Li: “From a geopolitics point of view, this is part of the technology that goes straight into weapons.” Autonomous drone swarms. Battlefield navigation. Physical target acquisition without human oversight. Every military application of AI that operates in the real world runs on spatial intelligence. The nation that masters the transition from static text to dynamic three-dimensional perception doesn’t just win the software race. It commands the physical battlefield. The AI arms race just broke out of the data center. It’s operating in three dimensions now.

Dustin

122,680 views • 4 months ago

Nesara/Gesara and The New World Times are changing and we are all going through our own Awakening whilst learning so much about ourselves and the world we live in Nesara/Gersara is a phrase that has been thrown around by many people with lots of promises of what we have awaiting us. It has been a big topic of conversation over the last few years, which I feel has a big lack of understanding to what it is. Many of the people who have pushed this information are Gatekeepers and deceivers and they have been doing all they can to stop people finding out the real truth of what is really going on in the world In this video I share a really powerful section from one of my favourite books Conversations With God by Neale Donald Walsch. When we truly understand the matrix and systems we are all currently living in, we can start to appreciate how the world will differ when we expose and remove them, which is what is happening now This will then bring more clarity and a better vision to how we view Nesara/Gesara which can provide us a whole new meaning Remember The New World, is without Crime, Evil and Corruption, a completely new reality, so books like this help us start to see what life will be like for us all. I invite you to listen to these words to really feel what is being said. This is one of my favourite books and one I will regularly listen to. See what comes up for you, I would love for you to share your comments I will be doing a series of videos like this, sharing the words and perspectives from others who speak the word of God
28:20

Nesara/Gesara and The New World Times are changing and we are all going through our own Awakening whilst learning so much about ourselves and the world we live in Nesara/Gersara is a phrase that has been thrown around by many people with lots of promises of what we have awaiting us. It has been a big topic of conversation over the last few years, which I feel has a big lack of understanding to what it is. Many of the people who have pushed this information are Gatekeepers and deceivers and they have been doing all they can to stop people finding out the real truth of what is really going on in the world In this video I share a really powerful section from one of my favourite books Conversations With God by Neale Donald Walsch. When we truly understand the matrix and systems we are all currently living in, we can start to appreciate how the world will differ when we expose and remove them, which is what is happening now This will then bring more clarity and a better vision to how we view Nesara/Gesara which can provide us a whole new meaning Remember The New World, is without Crime, Evil and Corruption, a completely new reality, so books like this help us start to see what life will be like for us all. I invite you to listen to these words to really feel what is being said. This is one of my favourite books and one I will regularly listen to. See what comes up for you, I would love for you to share your comments I will be doing a series of videos like this, sharing the words and perspectives from others who speak the word of God

Harry The Soul Coach

16,186 views • 1 month ago

New course: Transformers in Practice. You'll get a practical view of how transformer-based LLMs work, so you can reason about their behavior, diagnose problems like slow inference, and make smarter decisions about deployment. This course is built in partnership with AMD and taught by Sharon Zhou ✈️ ICML. You'll see how transformers generate text one token at a time, how the model decides which earlier words matter most when predicting the next one, and how techniques like quantization speed up inference on GPUs. This is not a video-only course; interactive visualizations throughout let you play with these concepts and build intuition that sticks. Skills you'll gain: - Understand why LLMs hallucinate, and RAG and chain-of-thought shape what they generate - Look inside the model to see how attention and layers combine to predict the next token - Diagnose inference bottlenecks and learn the techniques that speed up transformers on GPUs Join and understand what's really happening inside your LLMs:
1:00

New course: Transformers in Practice. You'll get a practical view of how transformer-based LLMs work, so you can reason about their behavior, diagnose problems like slow inference, and make smarter decisions about deployment. This course is built in partnership with AMD and taught by Sharon Zhou ✈️ ICML. You'll see how transformers generate text one token at a time, how the model decides which earlier words matter most when predicting the next one, and how techniques like quantization speed up inference on GPUs. This is not a video-only course; interactive visualizations throughout let you play with these concepts and build intuition that sticks. Skills you'll gain: - Understand why LLMs hallucinate, and RAG and chain-of-thought shape what they generate - Look inside the model to see how attention and layers combine to predict the next token - Diagnose inference bottlenecks and learn the techniques that speed up transformers on GPUs Join and understand what's really happening inside your LLMs:

Andrew Ng

117,443 views • 2 months ago

We have HOT3D! I've started using Claude to port more datasets into Rerun and exoego-forge. I'd been meaning to bring in the HOT3D dataset from Meta for a while, but with Claude, it's way easier. My goal is to take any egocentric, exocentric, or both datasets and ingest them into a standardized schema. Getting everything into Rerun means we can easily query and transform data via the in-memory OSS server. This lets us generate SQL-like queries such as: "Find me all frames that only contain left hands in the leftmost camera view." Most people think of Rerun as a viewer, but this is the actual superpower. So far we have: 1. HOT3D 2. Hocap 3. UmeTrack 4. Assembly101 5. EgoDex Planning to add more, and with every addition, it gets easier as we build up agent skills and better code examples. Hoping to make it almost fully automatic for adding new datasets. The next few I'm looking at are Harmony4D and Aria Pilot Gen2 After we have enough samples, I'll work on bringing in all the different algorithms I've worked on to transform the data 🙂
0:29

We have HOT3D! I've started using Claude to port more datasets into Rerun and exoego-forge. I'd been meaning to bring in the HOT3D dataset from Meta for a while, but with Claude, it's way easier. My goal is to take any egocentric, exocentric, or both datasets and ingest them into a standardized schema. Getting everything into Rerun means we can easily query and transform data via the in-memory OSS server. This lets us generate SQL-like queries such as: "Find me all frames that only contain left hands in the leftmost camera view." Most people think of Rerun as a viewer, but this is the actual superpower. So far we have: 1. HOT3D 2. Hocap 3. UmeTrack 4. Assembly101 5. EgoDex Planning to add more, and with every addition, it gets easier as we build up agent skills and better code examples. Hoping to make it almost fully automatic for adding new datasets. The next few I'm looking at are Harmony4D and Aria Pilot Gen2 After we have enough samples, I'll work on bringing in all the different algorithms I've worked on to transform the data 🙂

Pablo Vela

35,662 views • 3 months ago

Why AI Can Now Make Discoveries - my conversation with Dan Roberts, Lead of the Foundations of Reinforcement Learning team at OpenAI 00:00 Intro: AI's wild week in mathematics 01:21 What OpenAI's Foundations of RL team does 03:08 Dan's journey: from black holes and quantum gravity to frontier AI 07:04 Are AI systems becoming useful for real science 08:21 The AI math moment: Erdős, OpenAI, DeepMind, and Anthropic 08:52 Why the OpenAI result was an act of exploration 10:25 OpenAI vs. DeepMind: informal reasoning vs. formal proof 12:13 RL 101: learning by doing, not just watching 15:10 Why reinforcement learning works 15:58 How RL breaks: sparse feedback and long-horizon tasks 17:03 RLHF: how human feedback shaped early language models 18:48 Move 37, self-play, and the search for novel strategies 22:16 Explore vs. exploit in scientific discovery 24:49 Why RL may now be "the cake," not the cherry on top 25:46 Why RL started working with large language models 27:29 Is RL "sucking supervision through a straw"? 28:47 Why language may be the grounding layer for intelligence 31:46 A contrarian take on the Bitter Lesson 32:41 What test-time compute actually is 34:50 How RL gives models the ability to think 35:40 Verifiable rewards, math, coding, and the messy real world 38:00 What physics can teach us about AI 42:08 Is there a thermodynamics of AI? 43:08 From Erdős problems to Einstein-level AI 45:16 Is AI already doing original science? 45:51 How far are we from AI automating AI research 47:41 Why Dan is excited about the future of science
49:06

Why AI Can Now Make Discoveries - my conversation with Dan Roberts, Lead of the Foundations of Reinforcement Learning team at OpenAI 00:00 Intro: AI's wild week in mathematics 01:21 What OpenAI's Foundations of RL team does 03:08 Dan's journey: from black holes and quantum gravity to frontier AI 07:04 Are AI systems becoming useful for real science 08:21 The AI math moment: Erdős, OpenAI, DeepMind, and Anthropic 08:52 Why the OpenAI result was an act of exploration 10:25 OpenAI vs. DeepMind: informal reasoning vs. formal proof 12:13 RL 101: learning by doing, not just watching 15:10 Why reinforcement learning works 15:58 How RL breaks: sparse feedback and long-horizon tasks 17:03 RLHF: how human feedback shaped early language models 18:48 Move 37, self-play, and the search for novel strategies 22:16 Explore vs. exploit in scientific discovery 24:49 Why RL may now be "the cake," not the cherry on top 25:46 Why RL started working with large language models 27:29 Is RL "sucking supervision through a straw"? 28:47 Why language may be the grounding layer for intelligence 31:46 A contrarian take on the Bitter Lesson 32:41 What test-time compute actually is 34:50 How RL gives models the ability to think 35:40 Verifiable rewards, math, coding, and the messy real world 38:00 What physics can teach us about AI 42:08 Is there a thermodynamics of AI? 43:08 From Erdős problems to Einstein-level AI 45:16 Is AI already doing original science? 45:51 How far are we from AI automating AI research 47:41 Why Dan is excited about the future of science

Matt Turck

64,952 views • 1 month ago