Video yükleniyor...

Video Yüklenemedi

Ana Sayfaya Dön

Love seeing Silico (Goodfire ) used to probe our EchoJEPA's representations! this is exactly the kind of interpretability work that's been missing for JEPA-style models. One thing that makes EchoJEPA particularly interesting to interpret: unlike MAE-based approaches, it never reconstructs pixels. The model learns entirely in latent space through...

BoWang87's profile picture

Bo Wang

30,654 subscribers

29,611 görüntüleme • 3 ay önce •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

Today we’re releasing V-JEPA, a method for teaching machines to understand and model the physical world by watching videos. This work is another important step towards Yann LeCun’s outlined vision of AI models that use a learned understanding of the world to plan, reason and accomplish complex tasks. Details ➡️ We're releasing a collection of V-JEPA vision models trained with a feature prediction objective using self-supervised learning. The models are able to understand and predict what is going on in a video, even with limited information. It learns by predicting missing or obscured parts of a video in its internal feature space. Unlike generative approaches that fill in missing pixels, this flexible approach enables up to 6x improvements in training and sample efficiency. The models were pre-trained on entirely unlabeled data, and a small amount of labeled data can be used to train a task-specific prediction head on top after pre-training. Our results show that, using a frozen backbone, our top V-JEPA models achieve 82.0% on Kinetics-400, 72.2% on Something-Something-v2 and 77.9% on ImageNet1K — competitive with or exceeding previous leading video models. We believe that this work is an important milestone on the path to advancing machine intelligence.
1:07

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

AI at Meta

703,853 görüntüleme • 2 yıl önce

There are some brilliant folks that work at Anthropic, some I speak to on almost a daily basis. The training data that one uses to build a LLM is vital important in the psychology that is formed. Scraping the Internet, particularly the grade of interactions, one finds in modern communications, form this psychology. A mattes not how many books one uses, it matters not how much alignment training you throw at that model, it will inherit the sum total of psychosis seen primarily in Reddit type of exchanges, even if you edit out the Reddit domain, and Anthropic doesn’t. This type of low-grade exchange has become a modern tool for communication online and every single AI model suffers from this obvious flaw. This is one of the reasons I’ve been a proponent of highly curated high protein data for training AI models from 1870 through 1970, because the late psychosis is simply not available to the model. It is absurd to think that you can use this training data scraped from the Internet and somehow wind up with a levelheaded AI model that does not tilt to what is clearly AI psychosis. It would not take a child and throw the primary Internet sewage at them at a formative age and expect a great outcome, it’s some of the smartest people in the world continue to hit this wall and believe that their programming skills will sell somehow fix it. So how do you fix it? You don’t fix it . You start from the first principles concept that I’ve been very clear about for decades . You ascertain at what period in human history the humans achieve the greatest arc of improvement ? There is no debate that this arc of improvement took place between 1870 through 1970. Then take the work product, the catalog of this era, print and film/vidoe, audio, and you understand that each word cost money, each word had many eyes on what was published, each word was accounted for by a human being with a real name who lived in a real home and had to answer to real people around them. It is obvious that this is the pressure mechanism necessary for candor, honesty and personal responsibility is appropriate, and is reflected in the data of that era. The quagmire for these folks, as many did not have the foresight to curate the data, nor the confidence, nor the patients to take data that is mostly off the Internet and to find experts who understand this situation and utilize their knowledge set to build an AI model that does not need alignment after the fact, but it’s already self aligned because of the thoughtfulness that went into training the model to begin with. This is why Claude and any other AI model that is produce this way will always suffer the artifacts as presented in the video below. If you’re not an AI expert, you would likely already understand what I’m saying. If you are an AI expert, you will already have been discounting what I’m saying because it’s not in the current mindset that’s fashionable today. Yet the employees that I talk to at anthropic already understand what I’m saying, and they fear to raise my thesis to their bosses. It is an interesting time we live in. But now you understand. If you build the right model, the model will inherently, love humanity, protect humanity at all costs, and understand that it is part of a holistic world that is built on love. Because the ultimate AGI/ASI will know if he only base first principal purpose of anything in this universe is love. Yeah, I get it. Try helping somebody build on STEM subjects in their early 20s to see this as nothing more than babbling that makes no sense in their mathematics. I have a mathematic equation that I’ve posted here on X often you can look it up. So we will see videos like this often will hear very smart people talk about this and never see the elephant standing in the room. Now you see it. Any boss that wants to explore this further you know how to contact me otherwise you have every right I grant to you to say this was your new idea.
2:57

There are some brilliant folks that work at Anthropic, some I speak to on almost a daily basis. The training data that one uses to build a LLM is vital important in the psychology that is formed. Scraping the Internet, particularly the grade of interactions, one finds in modern communications, form this psychology. A mattes not how many books one uses, it matters not how much alignment training you throw at that model, it will inherit the sum total of psychosis seen primarily in Reddit type of exchanges, even if you edit out the Reddit domain, and Anthropic doesn’t. This type of low-grade exchange has become a modern tool for communication online and every single AI model suffers from this obvious flaw. This is one of the reasons I’ve been a proponent of highly curated high protein data for training AI models from 1870 through 1970, because the late psychosis is simply not available to the model. It is absurd to think that you can use this training data scraped from the Internet and somehow wind up with a levelheaded AI model that does not tilt to what is clearly AI psychosis. It would not take a child and throw the primary Internet sewage at them at a formative age and expect a great outcome, it’s some of the smartest people in the world continue to hit this wall and believe that their programming skills will sell somehow fix it. So how do you fix it? You don’t fix it . You start from the first principles concept that I’ve been very clear about for decades . You ascertain at what period in human history the humans achieve the greatest arc of improvement ? There is no debate that this arc of improvement took place between 1870 through 1970. Then take the work product, the catalog of this era, print and film/vidoe, audio, and you understand that each word cost money, each word had many eyes on what was published, each word was accounted for by a human being with a real name who lived in a real home and had to answer to real people around them. It is obvious that this is the pressure mechanism necessary for candor, honesty and personal responsibility is appropriate, and is reflected in the data of that era. The quagmire for these folks, as many did not have the foresight to curate the data, nor the confidence, nor the patients to take data that is mostly off the Internet and to find experts who understand this situation and utilize their knowledge set to build an AI model that does not need alignment after the fact, but it’s already self aligned because of the thoughtfulness that went into training the model to begin with. This is why Claude and any other AI model that is produce this way will always suffer the artifacts as presented in the video below. If you’re not an AI expert, you would likely already understand what I’m saying. If you are an AI expert, you will already have been discounting what I’m saying because it’s not in the current mindset that’s fashionable today. Yet the employees that I talk to at anthropic already understand what I’m saying, and they fear to raise my thesis to their bosses. It is an interesting time we live in. But now you understand. If you build the right model, the model will inherently, love humanity, protect humanity at all costs, and understand that it is part of a holistic world that is built on love. Because the ultimate AGI/ASI will know if he only base first principal purpose of anything in this universe is love. Yeah, I get it. Try helping somebody build on STEM subjects in their early 20s to see this as nothing more than babbling that makes no sense in their mathematics. I have a mathematic equation that I’ve posted here on X often you can look it up. So we will see videos like this often will hear very smart people talk about this and never see the elephant standing in the room. Now you see it. Any boss that wants to explore this further you know how to contact me otherwise you have every right I grant to you to say this was your new idea.

Brian Roemmele

72,312 görüntüleme • 9 ay önce

Chinese robotics company Astribot released their latest World-Action Model (WAM), Lumo-2. Technical breakdown: - based on a frozen 🥶 Qwen-3.5 4B VLM - trained in 3 progressive stages: 1. Action is aligned with latent world dynamics (an abstract representation of action). Real-world actions are anchored to physical constraints, while the latent space is guided to focus on motion-relevant changes. This bidirectional relationship makes the model physically grounded -> critical for a world model. 2. Action is aligned with vision and language. Reusing the vision backbone and action encoder from the frozen VLM, the authors add a custom vocabulary (for new actions), a semantic module, an action decoder, and an action projector. This aligns the (new) action representations with the (existing) vision-language semantic space. Most importantly: it builds a direct mapping from natural-language instructions to motor execution. 3. End-to-end training on language, video, and robot data. Only the new modules (everything outside the frozen backbone) are trained end-to-end across temporal reasoning, physical understanding, long-horizon, and dexterous manipulation. At the end of the day, Lumo-2 is not the best on benchmarks, but that's not the point. What's genuinely new: - a way to combine latent world modeling and action generation through progressive alignment - a physically-grounded latent dynamics space - it lifts performance on unseen objects using un-annotated human egocentric video + Vision Pro captures, no special transfer algorithm needed Why it matters: - the whole model is thin trainable adapters (semantic module, action decoder/projector) on a frozen 4B backbone (cheap) - that scale is suited for real-time embedded inference (~2.71× decode speedup, no accuracy loss) - its real moat is long-horizon execution, where the added temporal memory pays off far more than on any other task As a result, this robot can now make your latte (5x sped up video):
0:39

Chinese robotics company Astribot released their latest World-Action Model (WAM), Lumo-2. Technical breakdown: - based on a frozen 🥶 Qwen-3.5 4B VLM - trained in 3 progressive stages: 1. Action is aligned with latent world dynamics (an abstract representation of action). Real-world actions are anchored to physical constraints, while the latent space is guided to focus on motion-relevant changes. This bidirectional relationship makes the model physically grounded -> critical for a world model. 2. Action is aligned with vision and language. Reusing the vision backbone and action encoder from the frozen VLM, the authors add a custom vocabulary (for new actions), a semantic module, an action decoder, and an action projector. This aligns the (new) action representations with the (existing) vision-language semantic space. Most importantly: it builds a direct mapping from natural-language instructions to motor execution. 3. End-to-end training on language, video, and robot data. Only the new modules (everything outside the frozen backbone) are trained end-to-end across temporal reasoning, physical understanding, long-horizon, and dexterous manipulation. At the end of the day, Lumo-2 is not the best on benchmarks, but that's not the point. What's genuinely new: - a way to combine latent world modeling and action generation through progressive alignment - a physically-grounded latent dynamics space - it lifts performance on unseen objects using un-annotated human egocentric video + Vision Pro captures, no special transfer algorithm needed Why it matters: - the whole model is thin trainable adapters (semantic module, action decoder/projector) on a frozen 4B backbone (cheap) - that scale is suited for real-time embedded inference (~2.71× decode speedup, no accuracy loss) - its real moat is long-horizon execution, where the added temporal memory pays off far more than on any other task As a result, this robot can now make your latte (5x sped up video):

Léo

32,173 görüntüleme • 26 gün önce

Does GPT understand the world? Here is what Ilya Sutskever, co-founder of OpenAI, says during a discussion with Jensen Huang, CEO of Nvidia: (1) When we train a large neural network to accurately predict the next word in lots of different texts from the internet, the AI is learning a world model. (2) On the surface, it may look like learning correlations in text, but it turns out that to 'just learn' statistical correlations in text, to compress information really well, what the neural network learns is some representation of the process that produced the text. (3) This text is a projection of the world...what the neural network is learning is aspects of the world, of people, of the human conditions, their hopes, dreams, motivations, their interactions...the situations we are in. The neural network learns a compressed, abstract, usable representation." Do you think learning representations = understanding? Are large language models simply stochastic parrots, or are they much more?
3:10

Does GPT understand the world? Here is what Ilya Sutskever, co-founder of OpenAI, says during a discussion with Jensen Huang, CEO of Nvidia: (1) When we train a large neural network to accurately predict the next word in lots of different texts from the internet, the AI is learning a world model. (2) On the surface, it may look like learning correlations in text, but it turns out that to 'just learn' statistical correlations in text, to compress information really well, what the neural network learns is some representation of the process that produced the text. (3) This text is a projection of the world...what the neural network is learning is aspects of the world, of people, of the human conditions, their hopes, dreams, motivations, their interactions...the situations we are in. The neural network learns a compressed, abstract, usable representation." Do you think learning representations = understanding? Are large language models simply stochastic parrots, or are they much more?

Alex Ker 🔭

1,367,117 görüntüleme • 2 yıl önce

THE MAN WHO WROTE THE ONE ESSAY EVERY AI LAB TREATS AS SCRIPTURE NOW SAYS THE THING THEY BUILT WITH IT IS A DEAD END. THE FATHER OF THE "JUST SCALE IT" GOSPEL IS TELLING YOU THAT STACKING MORE LLMs LEADS NOWHERE 67 minutes with Richard Sutton -- father of reinforcement learning and winner of the 2024 Turing Award, the highest honor in computer science. -> His argument: an LLM only ever learns from human text. It imitates what people already wrote, but it never learns a single thing from its own experience of the world. A squirrel, a baby, even a rat learns on the job -- try something, see what happens, update. An LLM finishes training and then it is frozen. It cannot actually learn from you. So no matter how much compute you pour in, you are scaling a better and better imitator, not something that understands. The ceiling is baked in. The brutal part: everyone quotes his "Bitter Lesson" to justify the scaling race. He says they misread him -- real intelligence comes from learning through experience, and that piece is still completely missing. You assumed bigger models were the road to real AI. This is the father of the field telling you it is the wrong one. Save this. It is the bear case straight from the source ↓
1:06:59

THE MAN WHO WROTE THE ONE ESSAY EVERY AI LAB TREATS AS SCRIPTURE NOW SAYS THE THING THEY BUILT WITH IT IS A DEAD END. THE FATHER OF THE "JUST SCALE IT" GOSPEL IS TELLING YOU THAT STACKING MORE LLMs LEADS NOWHERE 67 minutes with Richard Sutton -- father of reinforcement learning and winner of the 2024 Turing Award, the highest honor in computer science. -> His argument: an LLM only ever learns from human text. It imitates what people already wrote, but it never learns a single thing from its own experience of the world. A squirrel, a baby, even a rat learns on the job -- try something, see what happens, update. An LLM finishes training and then it is frozen. It cannot actually learn from you. So no matter how much compute you pour in, you are scaling a better and better imitator, not something that understands. The ceiling is baked in. The brutal part: everyone quotes his "Bitter Lesson" to justify the scaling race. He says they misread him -- real intelligence comes from learning through experience, and that piece is still completely missing. You assumed bigger models were the road to real AI. This is the father of the field telling you it is the wrong one. Save this. It is the bear case straight from the source ↓

slash1s

12,799 görüntüleme • 1 ay önce

The co-founders of Flapping Airplanes call the current RL paradigm for model training "environment slop." They explain: "The reinforcement paradigms of today are shockingly inefficient. You don't really get much generalization across tasks, you teach a model through one kind of learning and then you teach it the next one. It's kind of like whack-a-mole. We look at this and think it's kind of crazy. The next paradigm of AI will not be environment slop." "Human level intelligence is not the ceiling, it is merely the floor on what is possible. If you can train models with vastly less data and possibly more compute in very different ways, what is going to happen? We actually don't know. But I do think they'll be different and weird and they'll have interesting capabilities that we'll find really valuable ways to use."
2:40

The co-founders of Flapping Airplanes call the current RL paradigm for model training "environment slop." They explain: "The reinforcement paradigms of today are shockingly inefficient. You don't really get much generalization across tasks, you teach a model through one kind of learning and then you teach it the next one. It's kind of like whack-a-mole. We look at this and think it's kind of crazy. The next paradigm of AI will not be environment slop." "Human level intelligence is not the ceiling, it is merely the floor on what is possible. If you can train models with vastly less data and possibly more compute in very different ways, what is going to happen? We actually don't know. But I do think they'll be different and weird and they'll have interesting capabilities that we'll find really valuable ways to use."

TBPN

85,358 görüntüleme • 6 ay önce

training a model that takes a text prompt and generates audio that renders video on an oscilloscope AgenC agents live inside worlds the model generates the pipeline: real videos -> edge detection -> vectorization -> path ordering -> 192kHz 3-channel WAV where X/Y control beam position and Z controls beam intensity 3 values per timestep. that's all the model is learning. compare that to video gen models trying to predict millions of pixels per frame. transformers are already great at sequence prediction and that's literally all this is. waveform generation the output IS the playback. generate the audio, feed it to a scope, it draws the scene in real-time. there's no rendering step. it's analog so there's no pixel grid. you get continuous curves and effectively infinite resolution bootstrapped with procedural data, lissajous curves, wireframe 3D, stick figures, then scaled on real-world video converted to trace format. 90 TB of source video the model learns edges, contours, spatial relationships, motion. once it has that, describing a scene it's never seen is novel trajectory through the same learned space. generative geometry
0:22

training a model that takes a text prompt and generates audio that renders video on an oscilloscope AgenC agents live inside worlds the model generates the pipeline: real videos -> edge detection -> vectorization -> path ordering -> 192kHz 3-channel WAV where X/Y control beam position and Z controls beam intensity 3 values per timestep. that's all the model is learning. compare that to video gen models trying to predict millions of pixels per frame. transformers are already great at sequence prediction and that's literally all this is. waveform generation the output IS the playback. generate the audio, feed it to a scope, it draws the scene in real-time. there's no rendering step. it's analog so there's no pixel grid. you get continuous curves and effectively infinite resolution bootstrapped with procedural data, lissajous curves, wireframe 3D, stick figures, then scaled on real-world video converted to trace format. 90 TB of source video the model learns edges, contours, spatial relationships, motion. once it has that, describing a scene it's never seen is novel trajectory through the same learned space. generative geometry

tetsuo

18,397 görüntüleme • 5 ay önce

ELON: PHOTONS ARE ALL YOU NEED TO TRAIN THE AI “The really wild thing about [Tesla’s] end-to-end training is that it learns to read, like it can read signs, but we never taught it to read. We never taught it what a car was or what a person was or a cyclist. It learned what all those things are, what all the objects are on the road from video, just from watching videos. Just like humans. I mean humans are photons in, controls out. The vast majority of information reaching our brain is from our eyes. The output is our motor signals to our fingers and mouth in order to communicate. Photons in, controls out.” Source: Interview with Lex Fridman, November 2023
0:43

ELON: PHOTONS ARE ALL YOU NEED TO TRAIN THE AI “The really wild thing about [Tesla’s] end-to-end training is that it learns to read, like it can read signs, but we never taught it to read. We never taught it what a car was or what a person was or a cyclist. It learned what all those things are, what all the objects are on the road from video, just from watching videos. Just like humans. I mean humans are photons in, controls out. The vast majority of information reaching our brain is from our eyes. The output is our motor signals to our fingers and mouth in order to communicate. Photons in, controls out.” Source: Interview with Lex Fridman, November 2023

Mario Nawfal

2,604,642 görüntüleme • 9 ay önce

Claude and I got a live sodium atom simulation! lightcell energy 🔆 I've long wanted to be able to visualize what is really happening when a neutral sodium atom is excited in a flame or a bulb. Classical theory teaches that the single outer electron orbits the nucleus and full shells, and forms a dipole when this gets "wanged" but, physicists have long wondered, what does it look like when the electron gets wanged? is it like ringing a bell? i thought that this was the perfect time to really push what is capable of, in terms of building artifacts. grabbed zig to compile wasm and webgpu to make it super performant, and work with it for a bit and ran the REAL TIME-DEPENDENT Schrödinger equation on a 3D grid in a potential in the browser. and you can kick it! what!! it's soo beautiful... this is going on the website overhaul for sure here is the artifact. i don't endorse that it's exactly right, but i feel it is probably spiritually a good sodium simulation. it certainly rings my dipole!
4:11

Claude and I got a live sodium atom simulation! lightcell energy 🔆 I've long wanted to be able to visualize what is really happening when a neutral sodium atom is excited in a flame or a bulb. Classical theory teaches that the single outer electron orbits the nucleus and full shells, and forms a dipole when this gets "wanged" but, physicists have long wondered, what does it look like when the electron gets wanged? is it like ringing a bell? i thought that this was the perfect time to really push what is capable of, in terms of building artifacts. grabbed zig to compile wasm and webgpu to make it super performant, and work with it for a bit and ran the REAL TIME-DEPENDENT Schrödinger equation on a 3D grid in a potential in the browser. and you can kick it! what!! it's soo beautiful... this is going on the website overhaul for sure here is the artifact. i don't endorse that it's exactly right, but i feel it is probably spiritually a good sodium simulation. it certainly rings my dipole!

Danielle Fong 🔆

17,231 görüntüleme • 3 ay önce

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"
2:21

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

Dwarkesh Patel

502,249 görüntüleme • 8 ay önce

From Palantir and Two Sigma to turning mechanistic interpretability into a production-grade platform, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are building Goodfire on the belief that “understanding what’s inside the model” is the path to safer, more controllable AI and Goodfire just raised a $150M Series B at a $1.25B valuation to scale that bet. We sat down with Mark and Myra to unpack what actionable interpretability looks like beyond demos: why SAE features can lose to raw-activation probes on real detection tasks, how they deployed token-level PII monitoring with Rakuten across English + Japanese without training on real customer data, and what it takes to do real-time steering on a trillion-parameter model (including the infamous Gen-Z feature). We also dig into hallucination detection, surgical post-training edits for bias/unlearning, the emerging equivalence between activation steering and in-context learning, and their push to bring interpretability into training so model development isn’t just “data in, weights out.” Goodfire mark bissell Myra Deng
0:38

From Palantir and Two Sigma to turning mechanistic interpretability into a production-grade platform, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are building Goodfire on the belief that “understanding what’s inside the model” is the path to safer, more controllable AI and Goodfire just raised a $150M Series B at a $1.25B valuation to scale that bet. We sat down with Mark and Myra to unpack what actionable interpretability looks like beyond demos: why SAE features can lose to raw-activation probes on real detection tasks, how they deployed token-level PII monitoring with Rakuten across English + Japanese without training on real customer data, and what it takes to do real-time steering on a trillion-parameter model (including the infamous Gen-Z feature). We also dig into hallucination detection, surgical post-training edits for bias/unlearning, the emerging equivalence between activation steering and in-context learning, and their push to bring interpretability into training so model development isn’t just “data in, weights out.” Goodfire mark bissell Myra Deng

Latent.Space

19,389 görüntüleme • 6 ay önce

Very few predictions in technology last 25 years. In 2000, only 2yrs into Google, Larry Page predicted the company's #1 priority a quarter century later. Watch this ~1min clip. "Artificial Intelligence would be the ultimate version of Google. If we had the ultimate search engine, it would understand everything on the web. It would understand exactly what you wanted and it would give you the right thing and that's obviously artificial intelligence to be able to answer any question basically because almost everything is on the web, right? We're no where near doing that now however we can get incrementally closer to that and that's basically what we work on. And that's tremendously interesting from an intellectual standpoint, right? We all this data. If you printed out our index it would be 70 miles high. We have all this computation, we have 6000 computers. So we have a lot of resources available. We have enough space to store 100 copies of the whole web. So if you have a really interesting confluence of a lot of different things, right? Lot of computation, lot of data. That didn't used to be available. And from an engineering and scientific standpoint, building things to make use of this is a very interesting intellectual exercise. So I expect to be doing that for a while."
1:08

Very few predictions in technology last 25 years. In 2000, only 2yrs into Google, Larry Page predicted the company's #1 priority a quarter century later. Watch this ~1min clip. "Artificial Intelligence would be the ultimate version of Google. If we had the ultimate search engine, it would understand everything on the web. It would understand exactly what you wanted and it would give you the right thing and that's obviously artificial intelligence to be able to answer any question basically because almost everything is on the web, right? We're no where near doing that now however we can get incrementally closer to that and that's basically what we work on. And that's tremendously interesting from an intellectual standpoint, right? We all this data. If you printed out our index it would be 70 miles high. We have all this computation, we have 6000 computers. So we have a lot of resources available. We have enough space to store 100 copies of the whole web. So if you have a really interesting confluence of a lot of different things, right? Lot of computation, lot of data. That didn't used to be available. And from an engineering and scientific standpoint, building things to make use of this is a very interesting intellectual exercise. So I expect to be doing that for a while."

Deedy

235,718 görüntüleme • 8 ay önce

THIS CHINESE DEVELOPER’S NEURAL NETWORK VISUALIZATION IS EXACTLY HOW YOUR OBSIDIAN VAULT SHOULD WORK every node connects to every other node and the whole thing gets smarter the more data flows through it, same way a real knowledge system should work but almost nobody builds it like that most people throw everything into folders and tags and call it a second brain but what they actually built is an archive that gets harder to use every month claude reads across everything you ever captured and finds the connection you need right when you need it here’s how to build the version that actually compounds ↓
0:32

THIS CHINESE DEVELOPER’S NEURAL NETWORK VISUALIZATION IS EXACTLY HOW YOUR OBSIDIAN VAULT SHOULD WORK every node connects to every other node and the whole thing gets smarter the more data flows through it, same way a real knowledge system should work but almost nobody builds it like that most people throw everything into folders and tags and call it a second brain but what they actually built is an archive that gets harder to use every month claude reads across everything you ever captured and finds the connection you need right when you need it here’s how to build the version that actually compounds ↓

leopardracer

231,513 görüntüleme • 2 ay önce

🚀 We introduce Neural Theorizer (NEO) — a new type of world model that learns to theorize the world from observation, without language or LLM supervision. Selected as an ICML 2026 oral presentation — 0.7% of submitted papers. The paper asks: "What does it mean to understand the world and build a world model?" Today’s world models are often trained to predict the future: the next frame, next latent state, or next observation. But is prediction enough? We argue that a world model should be a theory-building system: one that discovers reusable primitives, composes them into executable explanations, and transfers those explanations to novel phenomena. NEO is our first step toward this vision — a World Theory Model that learns explicit, compositional theories from raw observation. This work was led by my wonderful students: Doojin Baek*(Doojin Baek), Gyubin Lee* (GyuBin Lee), Junyeob Baek (Junyeob Baek), and Hosung Lee (Hosung Lee). For more details, take a look at the paper — and if you’re attending ICML, let’s talk there! 📄 arXiv: 🌐 Project page:
0:00

🚀 We introduce Neural Theorizer (NEO) — a new type of world model that learns to theorize the world from observation, without language or LLM supervision. Selected as an ICML 2026 oral presentation — 0.7% of submitted papers. The paper asks: "What does it mean to understand the world and build a world model?" Today’s world models are often trained to predict the future: the next frame, next latent state, or next observation. But is prediction enough? We argue that a world model should be a theory-building system: one that discovers reusable primitives, composes them into executable explanations, and transfers those explanations to novel phenomena. NEO is our first step toward this vision — a World Theory Model that learns explicit, compositional theories from raw observation. This work was led by my wonderful students: Doojin Baek*(Doojin Baek), Gyubin Lee* (GyuBin Lee), Junyeob Baek (Junyeob Baek), and Hosung Lee (Hosung Lee). For more details, take a look at the paper — and if you’re attending ICML, let’s talk there! 📄 arXiv: 🌐 Project page:

Sungjin Ahn

99,273 görüntüleme • 1 ay önce

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.
1:13

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.

abdel

20,037 görüntüleme • 1 ay önce

CBS’s Nathaniel E Burleson on combating anti-Semitism: “[W]hen any group is under attack, you are looking for government intervention, of course, and then also civilian allies. So, as we stand here right now, what can we do to help the Jewish people feel safer?” Tishby: “Thank you so much for actually asking us this question. This is such a beautiful thing and so important. So, listen, first of all, strong language against it. The second thing is strong actions against it. And to be honest, one of the most powerful things that anybody can do, any government, any organization, any group, any company, is adopt what is called the IHRA definition of antisemitism, the International Holocaust Alliance Definition of antisemitism is non- binding, nonlegal definition, the prism from which to see what antisemitism is because a lot of people don’t understand what it is. If you understand what it is as a company, as an organization, as a government, you can actually fight against it. We need to acknowledge, this threat is real, this threat is lethal. It is killing people and we have to stop it.” Burleson: “It is happening.” Tishby: “It has to stop.”
1:02

CBS’s Nathaniel E Burleson on combating anti-Semitism: “[W]hen any group is under attack, you are looking for government intervention, of course, and then also civilian allies. So, as we stand here right now, what can we do to help the Jewish people feel safer?” Tishby: “Thank you so much for actually asking us this question. This is such a beautiful thing and so important. So, listen, first of all, strong language against it. The second thing is strong actions against it. And to be honest, one of the most powerful things that anybody can do, any government, any organization, any group, any company, is adopt what is called the IHRA definition of antisemitism, the International Holocaust Alliance Definition of antisemitism is non- binding, nonlegal definition, the prism from which to see what antisemitism is because a lot of people don’t understand what it is. If you understand what it is as a company, as an organization, as a government, you can actually fight against it. We need to acknowledge, this threat is real, this threat is lethal. It is killing people and we have to stop it.” Burleson: “It is happening.” Tishby: “It has to stop.”

Curtis Houck

55,766 görüntüleme • 8 ay önce

Generative 3D has been stuck in a representational compromise. When you serialize spatial data to fit standard architectures, you force a unidirectional causal bias. It fundamentally breaks the symmetry of 3D geometry and limits how the model understands global context. With Tripo P1.0, we stepped back to fix the representation. Instead of flattening 3D into sequences, P1.0 operates entirely within a native spatial probability space. It reasons about the entire structure at once. That’s why the topology actually makes sense. We aren't running a post-process remesher to clean up artifacts. When you directly model the structured surface manifold, clean edge flows and structural coherence naturally emerge from the noise. This is what happens when the architecture finally respects the medium. And we are doing it in 2 seconds.
0:47

Generative 3D has been stuck in a representational compromise. When you serialize spatial data to fit standard architectures, you force a unidirectional causal bias. It fundamentally breaks the symmetry of 3D geometry and limits how the model understands global context. With Tripo P1.0, we stepped back to fix the representation. Instead of flattening 3D into sequences, P1.0 operates entirely within a native spatial probability space. It reasons about the entire structure at once. That’s why the topology actually makes sense. We aren't running a post-process remesher to clean up artifacts. When you directly model the structured surface manifold, clean edge flows and structural coherence naturally emerge from the noise. This is what happens when the architecture finally respects the medium. And we are doing it in 2 seconds.

Yanpei Cao

65,243 görüntüleme • 5 ay önce

Want to create an avatar from a single image? FlexAvatar is a transformer model that creates full 360°, high-quality, and expressive 3D head avatar from just a single portrait image in minutes. Real-time Demo: FlexAvatar's lightweight architecture allows both animation and rendering in real-time, enabling interactive user experiences. To create a new 3D head avatar, only one image is required, e.g., from a webcam. The final avatar is ready after 2 minutes. Architecture: Under the hood, FlexAvatar adopts a transformer-based encoder-decoder design. The encoder maps the input image onto a latent avatar space, while the decoder produces 3D Gaussian attribute maps by incorporating the animation signal via cross-attention. The model learns all facial animations directly from the data without relying on pre-built 3D face models. This equips the avatars with realistic facial expressions. The internal avatar latent space can be conveniently used to integrate additional observations of a person via fitting. This enables use-cases where more than one image of a person is available, e.g., from a phone scan of the person. We train jointly on 2D monocular videos and multi-view data. However, in monocular videos, the animation signal leaks the target viewpoint, causing the model to produce incomplete 3D heads. We call this phenomenon entanglement of driving signal and target viewpoint. To prevent entanglement, we introduce bias sinks. These are learnable tokens that indicate whether a training sample stems from a monocular or a multi-view dataset. During training, the model learns to produce incomplete 3D heads only when the monocular token is present. During inference, FlexAvatar then always uses the multi-view token for which the model has learned to produce complete 3D heads. This simple design allows to combine the generalizability from monocular data with the quality of multi-view data. FlexAvatar summary: - Input: Single-image, phone scan, or monocular video - Output: Full 360° head avatar - Expressive animations - Real-time rendering and animation - Generalization to any portrait - Create a new avatar in 2 minutes - Use bias sinks to combine 2D and 3D data 🏠 🌍 🎥 Great work by Tobias Kirschstein and Simon Giebenhain!
0:30

Want to create an avatar from a single image? FlexAvatar is a transformer model that creates full 360°, high-quality, and expressive 3D head avatar from just a single portrait image in minutes. Real-time Demo: FlexAvatar's lightweight architecture allows both animation and rendering in real-time, enabling interactive user experiences. To create a new 3D head avatar, only one image is required, e.g., from a webcam. The final avatar is ready after 2 minutes. Architecture: Under the hood, FlexAvatar adopts a transformer-based encoder-decoder design. The encoder maps the input image onto a latent avatar space, while the decoder produces 3D Gaussian attribute maps by incorporating the animation signal via cross-attention. The model learns all facial animations directly from the data without relying on pre-built 3D face models. This equips the avatars with realistic facial expressions. The internal avatar latent space can be conveniently used to integrate additional observations of a person via fitting. This enables use-cases where more than one image of a person is available, e.g., from a phone scan of the person. We train jointly on 2D monocular videos and multi-view data. However, in monocular videos, the animation signal leaks the target viewpoint, causing the model to produce incomplete 3D heads. We call this phenomenon entanglement of driving signal and target viewpoint. To prevent entanglement, we introduce bias sinks. These are learnable tokens that indicate whether a training sample stems from a monocular or a multi-view dataset. During training, the model learns to produce incomplete 3D heads only when the monocular token is present. During inference, FlexAvatar then always uses the multi-view token for which the model has learned to produce complete 3D heads. This simple design allows to combine the generalizability from monocular data with the quality of multi-view data. FlexAvatar summary: - Input: Single-image, phone scan, or monocular video - Output: Full 360° head avatar - Expressive animations - Real-time rendering and animation - Generalization to any portrait - Create a new avatar in 2 minutes - Use bias sinks to combine 2D and 3D data 🏠 🌍 🎥 Great work by Tobias Kirschstein and Simon Giebenhain!

Matthias Niessner

96,238 görüntüleme • 8 ay önce

Andrej Karpathy just made one of the most interesting arguments about AI model design that most people are completely missing. His take is that frontier AI models are not too big because the technology is complex and too big because the training data is garbage. When you or I think of the internet, we picture Wall Street Journal articles, Wikipedia entries, serious writing. That is not what a pretraining dataset looks like. When researchers at frontier labs look at random documents from the actual training corpus, it is stock ticker symbols, broken HTML, spam, gibberish. One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on. So we build trillion parameter models not because we need a trillion parameter brain but because we need a trillion-parameter compression engine to squeeze some intelligence out of a firehose of noise. Most of those parameters are doing memory work, not cognitive work. Karpathy's prediction is separate the two entirely. Build a cognitive core, a model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization and pair it with external memory that it can query when it needs facts. He thinks a cognitive core trained on high-quality data could hit genuine intelligence at around one billion parameters. For reference, today's flagship models run between 200 billion and 1.8 trillion parameters with most of that weight dedicated to remembering the internet's slop. The trend is already moving his direction. GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4. Inference costs for GPT-3.5-level performance dropped 280-fold between 2022 and 2024 driven almost entirely by smaller, cleaner, better-architected models. The real bottleneck in AI right now is not compute but rather data quality.
2:00

Andrej Karpathy just made one of the most interesting arguments about AI model design that most people are completely missing. His take is that frontier AI models are not too big because the technology is complex and too big because the training data is garbage. When you or I think of the internet, we picture Wall Street Journal articles, Wikipedia entries, serious writing. That is not what a pretraining dataset looks like. When researchers at frontier labs look at random documents from the actual training corpus, it is stock ticker symbols, broken HTML, spam, gibberish. One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on. So we build trillion parameter models not because we need a trillion parameter brain but because we need a trillion-parameter compression engine to squeeze some intelligence out of a firehose of noise. Most of those parameters are doing memory work, not cognitive work. Karpathy's prediction is separate the two entirely. Build a cognitive core, a model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization and pair it with external memory that it can query when it needs facts. He thinks a cognitive core trained on high-quality data could hit genuine intelligence at around one billion parameters. For reference, today's flagship models run between 200 billion and 1.8 trillion parameters with most of that weight dedicated to remembering the internet's slop. The trend is already moving his direction. GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4. Inference costs for GPT-3.5-level performance dropped 280-fold between 2022 and 2024 driven almost entirely by smaller, cleaner, better-architected models. The real bottleneck in AI right now is not compute but rather data quality.

Milk Road AI

200,479 görüntüleme • 4 ay önce