Loading video...

Video Failed to Load

Go Home

Google DeepMind say true AGI will “reason, adapt, and learn continuously,” and they estimate it’s about 5-10 years away. Meta’s chief scientist Yann LeCun likewise views continual learning as a key pillar of human level AI and is optimistic about achieving it by around 2030 Currently, large language models...

ChrissGPT's profile picture

Chris

51,187 subscribers

284,188 views • 9 months ago •via X (Twitter)

Science & Technology
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Continual Learning in 2026. And AGI with it? In July, Dario said he believes Continual Learning will turn out to be not as difficult as it seems. Sholto Douglas, Anthropic and former Deepmind researcher, predicted Continual Learning gets solved in a satisfying way in 2026. Sholto said that last week. Considering how capable AI models have become in 2025, it is unwise to dismiss this as fundraising or hype. It is much wiser to assume AGI is coming. That it will change the world. And that you can benefit from it... ...than it is to assume the opposite. "Not enough people are ready for AI to not be a bubble." - roon Choose wisely my friends.
0:26

Continual Learning in 2026. And AGI with it? In July, Dario said he believes Continual Learning will turn out to be not as difficult as it seems. Sholto Douglas, Anthropic and former Deepmind researcher, predicted Continual Learning gets solved in a satisfying way in 2026. Sholto said that last week. Considering how capable AI models have become in 2025, it is unwise to dismiss this as fundraising or hype. It is much wiser to assume AGI is coming. That it will change the world. And that you can benefit from it... ...than it is to assume the opposite. "Not enough people are ready for AI to not be a bubble." - roon Choose wisely my friends.

Dan McAteer

69,496 views • 6 months ago

Demis Hassabis points out that today’s models don’t learn continuously, they are trained once and then deployed. Unlike humans, they don’t improve from experience or adapt after release. The next major leap, like “AlphaZero,” would involve self-learning systems that discover knowledge on their own. This ability to learn online and evolve in real time is a key missing component on the path to AGI. Solving this will be critical to building truly intelligent systems.
1:42

Demis Hassabis points out that today’s models don’t learn continuously, they are trained once and then deployed. Unlike humans, they don’t improve from experience or adapt after release. The next major leap, like “AlphaZero,” would involve self-learning systems that discover knowledge on their own. This ability to learn online and evolve in real time is a key missing component on the path to AGI. Solving this will be critical to building truly intelligent systems.

Wes Roth

39,238 views • 7 months ago

Today, Michael Elabd, Arjun Karanam, and I are excited to announce Trajectory. We are a research lab and product company building the platform for Continual Learning. Our platform unlocks the signal already sitting in product usage, so companies can continuously post-train large-scale agentic models that outperform the frontier. Trajectory We’ve raised $15M from Conviction, Bessemer, Radical Ventures, Jeff Dean, Fei-Fei Li and more. We’re partnering with some of the best AI-native companies: Clay AI Harvey, Decagon, Mercor, Rogo to power their agentic systems, some of which we are already in production with. We’ve brought together a world class research team from DeepMind, OpenAI, Apple, Meta Superintelligence, Amazon AGI, Scale AI, and an elite product team from Stripe and Figma. AI will never again start on day one. Every correction, every retry, every edit will make products smarter. This is Continual Learning.
1:28

Today, Michael Elabd, Arjun Karanam, and I are excited to announce Trajectory. We are a research lab and product company building the platform for Continual Learning. Our platform unlocks the signal already sitting in product usage, so companies can continuously post-train large-scale agentic models that outperform the frontier. Trajectory We’ve raised $15M from Conviction, Bessemer, Radical Ventures, Jeff Dean, Fei-Fei Li and more. We’re partnering with some of the best AI-native companies: Clay AI Harvey, Decagon, Mercor, Rogo to power their agentic systems, some of which we are already in production with. We’ve brought together a world class research team from DeepMind, OpenAI, Apple, Meta Superintelligence, Amazon AGI, Scale AI, and an elite product team from Stripe and Figma. AI will never again start on day one. Every correction, every retry, every edit will make products smarter. This is Continual Learning.

Ronak Malde

1,812,912 views • 1 month ago

this was the most important bit of the ilya interview imo. humans are not AGI, and LLMs won’t become like humans with more scaling. humans lack a huge amount of knowledge and have insanely good “continual learning” powered by extremely good reward functions (except for addiction) that guide them. and so it’s more about the learning rather than knowledge. the term AGI was coined as a response to the narrow AI which was only capable of intelligence in a specific domain (chess, game, code, etc) but General intelligence is able to learn different topics because that’s what “learning” is about. pre-training was thought to be the solution to reaching AGI because you train AI on so many different topics. but they have “overshot” that target. amazing perspective.
1:57

this was the most important bit of the ilya interview imo. humans are not AGI, and LLMs won’t become like humans with more scaling. humans lack a huge amount of knowledge and have insanely good “continual learning” powered by extremely good reward functions (except for addiction) that guide them. and so it’s more about the learning rather than knowledge. the term AGI was coined as a response to the narrow AI which was only capable of intelligence in a specific domain (chess, game, code, etc) but General intelligence is able to learn different topics because that’s what “learning” is about. pre-training was thought to be the solution to reaching AGI because you train AI on so many different topics. but they have “overshot” that target. amazing perspective.

ℏεsam

53,444 views • 7 months ago

Without World Models, There Is No AGI. Google Just Proved It. If AGI ever happens, it will not come from bigger chatbots alone. From the very start of this interview, one thing is crystal clear: without world models, we will never reach AGI. And right now, Google is leading with its world simulator Genie 3. Here is the core of what Demis Hassabis explains in this conversation: • World models are the missing core of AGI Hassabis says his deepest long term focus has always been world models and simulations. Not just language. Not just prediction. Actual internal simulations of reality. • LLMs are impressive, but incomplete Language models understand more about the world than expected because human language encodes a lot of reality. Still, language is only a shadow of the real thing. • What text can never fully teach Reality includes things text struggles to express: •3D space and spatial dynamics •Physical causality and mechanics •Sensorimotor experience like movement, force, smell, or balance • Experience beats description To close the gap, AI must learn from interaction and experience, not just static text. That is how you build an internal world simulator. • Why Genie 3 matters With Google DeepMind pushing systems like Genie 3, AI starts to model reality itself, not just talk about it. • Robots and real world assistants depend on this True robotics, smart glasses, and universal assistants require AI that understands the physical world you live in, not just your screen. Bottom line: AGI will not emerge from better text prediction. It will emerge from systems that can simulate, predict, and understand reality itself. Right now, Google is clearly ahead on that path. Curious what you think. Are world models the real AGI unlock, or just another stepping stone?
1:26

Without World Models, There Is No AGI. Google Just Proved It. If AGI ever happens, it will not come from bigger chatbots alone. From the very start of this interview, one thing is crystal clear: without world models, we will never reach AGI. And right now, Google is leading with its world simulator Genie 3. Here is the core of what Demis Hassabis explains in this conversation: • World models are the missing core of AGI Hassabis says his deepest long term focus has always been world models and simulations. Not just language. Not just prediction. Actual internal simulations of reality. • LLMs are impressive, but incomplete Language models understand more about the world than expected because human language encodes a lot of reality. Still, language is only a shadow of the real thing. • What text can never fully teach Reality includes things text struggles to express: •3D space and spatial dynamics •Physical causality and mechanics •Sensorimotor experience like movement, force, smell, or balance • Experience beats description To close the gap, AI must learn from interaction and experience, not just static text. That is how you build an internal world simulator. • Why Genie 3 matters With Google DeepMind pushing systems like Genie 3, AI starts to model reality itself, not just talk about it. • Robots and real world assistants depend on this True robotics, smart glasses, and universal assistants require AI that understands the physical world you live in, not just your screen. Bottom line: AGI will not emerge from better text prediction. It will emerge from systems that can simulate, predict, and understand reality itself. Right now, Google is clearly ahead on that path. Curious what you think. Are world models the real AGI unlock, or just another stepping stone?

VraserX e/acc

23,784 views • 6 months ago

Demis Hassabis's new interview: "Society needs to hear that because we don't have long to prepare for what that means. We are standing in the foothills of the singularity now. ..which is AGI. I believe that we are only a few years away from that, maybe around 2030, plus or minus a year. " ~ Demis Hassabis, Co-Founder and CEO of Google DeepMind It is going to be enormously profound, I think. The future, in my view, is still to be written. But these next few years are going to be very critical as to which way that will go, and how we collectively want that to look.” --- IMO, The real disruption is not whether AGI arrives exactly in 2030, plus or minus a year, but whether institutions can adapt, as in post-AGI world, technology will change much faster than human systems can respond. Schools still train people for stable professions, companies still organize work around human bottlenecks, and governments still regulate after harm becomes visible. AGI, if it arrives anywhere near the frontier-lab timelines, compresses that lag into a dangerous gap. ---- From "Stanford Graduate School of Business" YouTube channel, (link in comment)
0:50

Demis Hassabis's new interview: "Society needs to hear that because we don't have long to prepare for what that means. We are standing in the foothills of the singularity now. ..which is AGI. I believe that we are only a few years away from that, maybe around 2030, plus or minus a year. " ~ Demis Hassabis, Co-Founder and CEO of Google DeepMind It is going to be enormously profound, I think. The future, in my view, is still to be written. But these next few years are going to be very critical as to which way that will go, and how we collectively want that to look.” --- IMO, The real disruption is not whether AGI arrives exactly in 2030, plus or minus a year, but whether institutions can adapt, as in post-AGI world, technology will change much faster than human systems can respond. Schools still train people for stable professions, companies still organize work around human bottlenecks, and governments still regulate after harm becomes visible. AGI, if it arrives anywhere near the frontier-lab timelines, compresses that lag into a dangerous gap. ---- From "Stanford Graduate School of Business" YouTube channel, (link in comment)

Rohan Paul

70,642 views • 1 month ago

Google Deepmind CEO gives his prediction on the final AGI architecture: Demis Hassabis just sat down on 20VC and broke down exactly how DeepMind is building toward true AGI, directly pushing back on the idea that LLMs are a dead end. He defends the current trajectory of the models, noting that “we’ve seen how successful these foundation models have been. They can do incredibly impressive things. I don’t think that’s going to go away. We’re still seeing returns from the scaling laws.” But he’s honest about the missing pieces for true superintelligence, saying “I think there might be… there’s a 50-50 chance there’s some things maybe missing that we still need to make breakthroughs in, perhaps their world models, these kinds of approaches.” And when he breaks down what the actual AGI system will look like, he frames it as one question is the LLM foundation model the key component only, or is it the total system. His answer - “I don’t think it’s going to get replaced. I think it’s going to get built on top of these foundation models, just like the way we do with our world models.”
0:53

Google Deepmind CEO gives his prediction on the final AGI architecture: Demis Hassabis just sat down on 20VC and broke down exactly how DeepMind is building toward true AGI, directly pushing back on the idea that LLMs are a dead end. He defends the current trajectory of the models, noting that “we’ve seen how successful these foundation models have been. They can do incredibly impressive things. I don’t think that’s going to go away. We’re still seeing returns from the scaling laws.” But he’s honest about the missing pieces for true superintelligence, saying “I think there might be… there’s a 50-50 chance there’s some things maybe missing that we still need to make breakthroughs in, perhaps their world models, these kinds of approaches.” And when he breaks down what the actual AGI system will look like, he frames it as one question is the LLM foundation model the key component only, or is it the total system. His answer - “I don’t think it’s going to get replaced. I think it’s going to get built on top of these foundation models, just like the way we do with our world models.”

Chris

71,716 views • 2 months ago

New Course: Reinforcement Fine-Tuning LLMs with GRPO! Learn to use reinforcement learning to improve your LLM performance in this short course, built in collaboration with Predibase by Rubrik, and taught by Travis Addair, its Co-Founder and CTO, and Arnav Garg, its Senior Engineer and Machine Learning Lead. Reasoning models have been one of the most important developments in LLMs. Reinforcement Fine-Tuning (RFT) uses rewards to encourage LLMs to find solutions to multi-step reasoning tasks such as solving math problems and debugging code - without needing pre-existing training examples like in traditional supervised fine-tuning. Group Relative Policy Optimization (GRPO) is a reinforcement fine-tuning algorithm gaining rapid adoption. Developed by the DeepSeek team and used to train the R1 reasoning model, GRPO uses reward functions that you can write in Python to assign rewards to model responses. It’s beneficial for tasks with verifiable outcomes and can work well even with fewer than 100 training examples. It can also significantly improve the reasoning ability of smaller LLMs, making applications faster and more cost effective. In this course, you’ll take a technical deep dive into RFT with GRPO. You’ll learn to build reward functions that you can use in the GRPO training process to guide an LLM toward better performance on multi-step reasoning tasks. In detail, you’ll: - Learn when reinforcement fine-tuning is a better fit than supervised fine-tuning, especially for tasks involving multi-step reasoning or limited labeled data. - Understand how GRPO uses programmable reward functions as a more scalable alternative to the human feedback required for other reinforcement learning algorithms, such as RLHF and DPO. - Frame the Wordle game as a reinforcement fine-tuning problem and see how an LLM can learn to plan, analyze feedback, and improve its strategy over time. - Design reward functions that power the reinforcement fine-tuning process. - Learn techniques for evaluating more subjective tasks, such as rating the quality of a text summary, using an LLM as a judge. - Understand why reward hacking happens and how to avoid it by adding penalty functions to discourage undesirable behaviors. - Learn the four key components of the loss calculation in the GRPO algorithm: token probability distribution ratios, advantages, clipping, and KL-divergence. - Launch reinforcement fine-tuning jobs using Predibase’s hosted training services. By the end of this course, you’ll be able to build and fine-tune LLMs using reinforcement learning to improve reasoning without relying on large labeled datasets or subjective human feedback. Please sign up here:
2:41

New Course: Reinforcement Fine-Tuning LLMs with GRPO! Learn to use reinforcement learning to improve your LLM performance in this short course, built in collaboration with Predibase by Rubrik, and taught by Travis Addair, its Co-Founder and CTO, and Arnav Garg, its Senior Engineer and Machine Learning Lead. Reasoning models have been one of the most important developments in LLMs. Reinforcement Fine-Tuning (RFT) uses rewards to encourage LLMs to find solutions to multi-step reasoning tasks such as solving math problems and debugging code - without needing pre-existing training examples like in traditional supervised fine-tuning. Group Relative Policy Optimization (GRPO) is a reinforcement fine-tuning algorithm gaining rapid adoption. Developed by the DeepSeek team and used to train the R1 reasoning model, GRPO uses reward functions that you can write in Python to assign rewards to model responses. It’s beneficial for tasks with verifiable outcomes and can work well even with fewer than 100 training examples. It can also significantly improve the reasoning ability of smaller LLMs, making applications faster and more cost effective. In this course, you’ll take a technical deep dive into RFT with GRPO. You’ll learn to build reward functions that you can use in the GRPO training process to guide an LLM toward better performance on multi-step reasoning tasks. In detail, you’ll: - Learn when reinforcement fine-tuning is a better fit than supervised fine-tuning, especially for tasks involving multi-step reasoning or limited labeled data. - Understand how GRPO uses programmable reward functions as a more scalable alternative to the human feedback required for other reinforcement learning algorithms, such as RLHF and DPO. - Frame the Wordle game as a reinforcement fine-tuning problem and see how an LLM can learn to plan, analyze feedback, and improve its strategy over time. - Design reward functions that power the reinforcement fine-tuning process. - Learn techniques for evaluating more subjective tasks, such as rating the quality of a text summary, using an LLM as a judge. - Understand why reward hacking happens and how to avoid it by adding penalty functions to discourage undesirable behaviors. - Learn the four key components of the loss calculation in the GRPO algorithm: token probability distribution ratios, advantages, clipping, and KL-divergence. - Launch reinforcement fine-tuning jobs using Predibase’s hosted training services. By the end of this course, you’ll be able to build and fine-tune LLMs using reinforcement learning to improve reasoning without relying on large labeled datasets or subjective human feedback. Please sign up here:

Andrew Ng

86,457 views • 1 year ago

Yann LeCun -- Meta's chief AI scientist and a Turing Award winner -- explains why scaling up LLMs will never reach human-level AI, and names the four things today's systems still can't do: "If you think that we're going to get to human-level AI by just training on more data and scaling up LLMs, you're making a mistake." "If you're an investor and you invest in a company that told you we're going to get to human-level AI and PhD level by just training on more data and with a few tricks... that was probably not a good idea." "There are ideas about how to go forward and have systems that are capable of doing what every intelligent animal and human are capable of doing, and that current AI systems are not capable of doing." "Understanding the physical world, having persistent memory, and being able to reason and plan. Those are the four characteristics that need to be there." "That requires systems that can acquire common sense, that can learn from natural sensors like video, as opposed to just text." The entire market is priced on a straight line from bigger models to AGI. One of the people who invented deep learning is telling you the line doesn't reach -- and that the real bottleneck isn't compute, it's world models. If he's right, the winner isn't whoever stacks the most GPUs. It's whoever teaches a machine to learn from video the way a child does.
0:58

Yann LeCun -- Meta's chief AI scientist and a Turing Award winner -- explains why scaling up LLMs will never reach human-level AI, and names the four things today's systems still can't do: "If you think that we're going to get to human-level AI by just training on more data and scaling up LLMs, you're making a mistake." "If you're an investor and you invest in a company that told you we're going to get to human-level AI and PhD level by just training on more data and with a few tricks... that was probably not a good idea." "There are ideas about how to go forward and have systems that are capable of doing what every intelligent animal and human are capable of doing, and that current AI systems are not capable of doing." "Understanding the physical world, having persistent memory, and being able to reason and plan. Those are the four characteristics that need to be there." "That requires systems that can acquire common sense, that can learn from natural sensors like video, as opposed to just text." The entire market is priced on a straight line from bigger models to AGI. One of the people who invented deep learning is telling you the line doesn't reach -- and that the real bottleneck isn't compute, it's world models. If he's right, the winner isn't whoever stacks the most GPUs. It's whoever teaches a machine to learn from video the way a child does.

Karl Mehta

34,662 views • 12 days ago

Reinforcement Learning from Human Feedback (RLHF) is gaining traction. This field aims to make AI more responsible by including human values and preferences. In this video, Nathan Lambert, a research scientist and RLHF team lead at Hugging Face explores its inner workings, applications and industry impact. RLHF has gained the spotlight in recent years. The growth of language models like Anthropic’s Claude and OpenAI's ChatGPT have increased interest in human-feedback integration. "There are some rumors that Open AI had two teams; one was doing RLHF and the other instruction fine-tuning. And the RLHF team kept getting more and more performance." Understanding RLHF The RLHF process has three main steps: Pre-training: Much like with GPT models, the journey starts with pre-training on a large corpus of data. This can range from text data, web scrapes, to specialized datasets. Reward Modeling: This is the RLHF counterpart of supervised fine-tuning in large language models. This stage involves creating a reward model that resonates with human values and preferences. RL Optimization: This stage parallels reward modeling and reinforcement learning in traditional AI models. The AI system fine-tunes itself based on the reward model, employing reinforcement learning algorithms for that extra layer of optimization. The Data Challenge Data collection and curation in RLHF closely resemble the challenges you'd encounter in large language model training. Datasets from organizations like OpenAI can serve as a useful foundation. However, the need for high-quality, task-specific data cannot be overstated. Implementing RLHF: A Practical Guide If you’re someone who loves getting hands-on with AI libraries like Hugging Face, implementing RLHF is right way to do. It’s essential to understand its limitations. Think about model stability, over-optimization, and exploration strategies, much like you would when prompt engineering. Ongoing Research and Next Steps While he suggests that some basics figured out, there are layers of complexity that still need to be unraveled: 1. New Benchmarks: How do we measure the effectiveness of RLHF? 2. Preference Modeling: How can the model be made to understand human preferences better? 3. Interpreting RLHF: Much like explainability in traditional models, how do we make RLHF more interpretable? 4. System-Wide Evaluation: Going beyond individual performance, how does RLHF affect an entire system? The Transformative Power of RLHF Whether you're an AI developer, a business analyst, or a marketer, RLHF promises to revolutionize your domain. Imagine customer service chatbots that understand human emotions better, or content generators that align more closely with human values. RLHF is an emerging field that focuses on enhancing machine learning models through human feedback. While it tackles important issues like bias and ethics, its broader goal is to improve system performance across various applications. Whether you're deeply invested in the ethics of AI or simply curious about advancements in machine learning, RLHF offers valuable insights. If you're interested in the next wave of AI development, this area is definitely one to watch.
2:00

Reinforcement Learning from Human Feedback (RLHF) is gaining traction. This field aims to make AI more responsible by including human values and preferences. In this video, Nathan Lambert, a research scientist and RLHF team lead at Hugging Face explores its inner workings, applications and industry impact. RLHF has gained the spotlight in recent years. The growth of language models like Anthropic’s Claude and OpenAI's ChatGPT have increased interest in human-feedback integration. "There are some rumors that Open AI had two teams; one was doing RLHF and the other instruction fine-tuning. And the RLHF team kept getting more and more performance." Understanding RLHF The RLHF process has three main steps: Pre-training: Much like with GPT models, the journey starts with pre-training on a large corpus of data. This can range from text data, web scrapes, to specialized datasets. Reward Modeling: This is the RLHF counterpart of supervised fine-tuning in large language models. This stage involves creating a reward model that resonates with human values and preferences. RL Optimization: This stage parallels reward modeling and reinforcement learning in traditional AI models. The AI system fine-tunes itself based on the reward model, employing reinforcement learning algorithms for that extra layer of optimization. The Data Challenge Data collection and curation in RLHF closely resemble the challenges you'd encounter in large language model training. Datasets from organizations like OpenAI can serve as a useful foundation. However, the need for high-quality, task-specific data cannot be overstated. Implementing RLHF: A Practical Guide If you’re someone who loves getting hands-on with AI libraries like Hugging Face, implementing RLHF is right way to do. It’s essential to understand its limitations. Think about model stability, over-optimization, and exploration strategies, much like you would when prompt engineering. Ongoing Research and Next Steps While he suggests that some basics figured out, there are layers of complexity that still need to be unraveled: 1. New Benchmarks: How do we measure the effectiveness of RLHF? 2. Preference Modeling: How can the model be made to understand human preferences better? 3. Interpreting RLHF: Much like explainability in traditional models, how do we make RLHF more interpretable? 4. System-Wide Evaluation: Going beyond individual performance, how does RLHF affect an entire system? The Transformative Power of RLHF Whether you're an AI developer, a business analyst, or a marketer, RLHF promises to revolutionize your domain. Imagine customer service chatbots that understand human emotions better, or content generators that align more closely with human values. RLHF is an emerging field that focuses on enhancing machine learning models through human feedback. While it tackles important issues like bias and ethics, its broader goal is to improve system performance across various applications. Whether you're deeply invested in the ethics of AI or simply curious about advancements in machine learning, RLHF offers valuable insights. If you're interested in the next wave of AI development, this area is definitely one to watch.

Muratcan Koylan

27,005 views • 2 years ago

The CEO of Google DeepMind just went on record saying he disagrees with one of the most respected AI researchers in the world. Demis Hassabis, the man behind AlphaFold, AlphaGo, and Google's entire AI operation publicly pushed back against Yann LeCun's claim that large language models are a dead end for artificial intelligence. LeCun, who left Meta earlier this year to start his own AI lab, has been saying for years that LLMs cannot reason, cannot plan, and will never get us to human-level intelligence. Hassabis disagrees, and he said so directly. His position is that scaling laws are still working, foundation models are still getting more capable, and whatever AGI ends up looking like, LLMs will be a central part of it, not something that gets replaced. He does say there is roughly a 50/50 chance that one or two additional breakthroughs will be needed beyond scaling alone, things like better memory, long-term planning, and world models. But the core disagreement with LeCun is clear, Hassabis believes the current architecture is sound and the current path leads somewhere real. Two Nobel-recognized researchers, two founding figures of modern AI, now publicly on opposite sides of the most important technical question in the industry.
1:00

The CEO of Google DeepMind just went on record saying he disagrees with one of the most respected AI researchers in the world. Demis Hassabis, the man behind AlphaFold, AlphaGo, and Google's entire AI operation publicly pushed back against Yann LeCun's claim that large language models are a dead end for artificial intelligence. LeCun, who left Meta earlier this year to start his own AI lab, has been saying for years that LLMs cannot reason, cannot plan, and will never get us to human-level intelligence. Hassabis disagrees, and he said so directly. His position is that scaling laws are still working, foundation models are still getting more capable, and whatever AGI ends up looking like, LLMs will be a central part of it, not something that gets replaced. He does say there is roughly a 50/50 chance that one or two additional breakthroughs will be needed beyond scaling alone, things like better memory, long-term planning, and world models. But the core disagreement with LeCun is clear, Hassabis believes the current architecture is sound and the current path leads somewhere real. Two Nobel-recognized researchers, two founding figures of modern AI, now publicly on opposite sides of the most important technical question in the industry.

Milk Road AI

130,649 views • 3 months ago

Prime Intellect's will brown says continual learning could be solved in the first half of 2026: "Continual learning is going to fall pretty quickly, I think. It's more of an engineering problem. No one's actually trying." "OpenAI and Anthropic don't want to continuously train their models for each user. It's expensive and annoying and hard to serve at scale. But from a research perspective, we do continue learning where they just keep training the model more and it knows more stuff because they put more of the Internet in it." "I think there's a lot of experimentation around exactly the recipe that's going to be the most reliable. But we kind of have a grab bag of six or seven tricks that kind of work, or they work in different ways, and you can mix and match them. And it's just going to be like whatever's the best combination of these tricks." "People are going to experiment with it and find the versions that work the best. And there doesn't seem to be any big wall inside that prevents that from being practical."
1:47

Prime Intellect's will brown says continual learning could be solved in the first half of 2026: "Continual learning is going to fall pretty quickly, I think. It's more of an engineering problem. No one's actually trying." "OpenAI and Anthropic don't want to continuously train their models for each user. It's expensive and annoying and hard to serve at scale. But from a research perspective, we do continue learning where they just keep training the model more and it knows more stuff because they put more of the Internet in it." "I think there's a lot of experimentation around exactly the recipe that's going to be the most reliable. But we kind of have a grab bag of six or seven tricks that kind of work, or they work in different ways, and you can mix and match them. And it's just going to be like whatever's the best combination of these tricks." "People are going to experiment with it and find the versions that work the best. And there doesn't seem to be any big wall inside that prevents that from being practical."

TBPN

72,077 views • 4 months ago

Dario Amodei, the CEO of Anthropic, went on CNN and put a number on the thing most AI CEOs will not say out loud: Half of all entry-level white-collar jobs gone. Unemployment at 10 to 20%. Inside 1 to 5 years. "The most salient feature of the technology, and what is driving all of this, is how fast the technology is getting better." "A couple of years ago, you could say that AI models were maybe as good as a smart high school student. I would say that now they're as good as a smart college student, and sort of reaching past that." "I really worry, particularly at the entry level, that the AI models are very much at the center of what an entry level human worker would do." "What is striking to me about this AI boom is that it's bigger and it's broader and it's moving faster than anything has before." "People will adapt, but they may not adapt fast enough. And so there may be an adjustment period." The tell: this is not a critic outside the industry. It is the man building the models saying the quiet part on camera -- because he thinks the people whose jobs are on the line adapt slower than the models improve. High school to college in two years was the warm-up. The entry level is where it lands first.
1:22

Dario Amodei, the CEO of Anthropic, went on CNN and put a number on the thing most AI CEOs will not say out loud: Half of all entry-level white-collar jobs gone. Unemployment at 10 to 20%. Inside 1 to 5 years. "The most salient feature of the technology, and what is driving all of this, is how fast the technology is getting better." "A couple of years ago, you could say that AI models were maybe as good as a smart high school student. I would say that now they're as good as a smart college student, and sort of reaching past that." "I really worry, particularly at the entry level, that the AI models are very much at the center of what an entry level human worker would do." "What is striking to me about this AI boom is that it's bigger and it's broader and it's moving faster than anything has before." "People will adapt, but they may not adapt fast enough. And so there may be an adjustment period." The tell: this is not a critic outside the industry. It is the man building the models saying the quiet part on camera -- because he thinks the people whose jobs are on the line adapt slower than the models improve. High school to college in two years was the warm-up. The entry level is where it lands first.

Karl Mehta

340,410 views • 11 days ago

The most interesting part for me is where Andrej Karpathy describes why LLMs aren't able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across every token in a successful trajectory, upweighting even wrong or irrelevant turns that lead to the right answer. > “Humans don't use reinforcement learning, as I've said before. I think they do something different. Reinforcement learning is a lot worse than the average person thinks. Reinforcement learning is terrible. It just so happens that everything that we had before is much worse.” So what do humans do instead? > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” > “I'd love to see during pretraining some kind of a stage where the model thinks through the material and tries to reconcile it with what it already knows. There's no equivalent of any of this. This is all research.” Why can’t we just add this training to LLMs today? > “There are very subtle, hard to understand reasons why it's not trivial. If I just give synthetic generation of the model thinking about a book, you look at it and you're like, 'This looks great. Why can't I train on it?' You could try, but the model will actually get much worse if you continue trying.” > “Say we have a chapter of a book and I ask an LLM to think about it. It will give you something that looks very reasonable. But if I ask it 10 times, you'll notice that all of them are the same.” > “You're not getting the richness and the diversity and the entropy from these models as you would get from humans. How do you get synthetic data generation to work despite the collapse and while maintaining the entropy? It is a research problem.” How do humans get around model collapse? > “These analogies are surprisingly good. Humans collapse during the course of their lives. Children haven't overfit yet. They will say stuff that will shock you. Because they're not yet collapsed. But we [adults] are collapsed. We end up revisiting the same thoughts, we end up saying more and more of the same stuff, the learning rates go down, the collapse continues to get worse, and then everything deteriorates.” In fact, there’s an interesting paper arguing that dreaming evolved to assist generalization, and resist overfitting to daily learning - look up The Overfitted Brain by Erik Hoel. I asked Karpathy: Isn’t it interesting that humans learn best at a part of their lives (childhood) whose actual details they completely forget, adults still learn really well but have terrible memory about the particulars of the things they read or watch, and LLMs can memorize arbitrary details about text that no human could but are currently pretty bad at generalization? > “[Fallible human memory] is a feature, not a bug, because it forces you to only learn the generalizable components. LLMs are distracted by all the memory that they have of the pre-trained documents. That's why when I talk about the cognitive core, I actually want to remove the memory. I'd love to have them have less memory so that they have to look things up and they only maintain the algorithms for thought, and the idea of an experiment, and all this cognitive glue for acting.”
7:01

The most interesting part for me is where Andrej Karpathy describes why LLMs aren't able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across every token in a successful trajectory, upweighting even wrong or irrelevant turns that lead to the right answer. > “Humans don't use reinforcement learning, as I've said before. I think they do something different. Reinforcement learning is a lot worse than the average person thinks. Reinforcement learning is terrible. It just so happens that everything that we had before is much worse.” So what do humans do instead? > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” > “I'd love to see during pretraining some kind of a stage where the model thinks through the material and tries to reconcile it with what it already knows. There's no equivalent of any of this. This is all research.” Why can’t we just add this training to LLMs today? > “There are very subtle, hard to understand reasons why it's not trivial. If I just give synthetic generation of the model thinking about a book, you look at it and you're like, 'This looks great. Why can't I train on it?' You could try, but the model will actually get much worse if you continue trying.” > “Say we have a chapter of a book and I ask an LLM to think about it. It will give you something that looks very reasonable. But if I ask it 10 times, you'll notice that all of them are the same.” > “You're not getting the richness and the diversity and the entropy from these models as you would get from humans. How do you get synthetic data generation to work despite the collapse and while maintaining the entropy? It is a research problem.” How do humans get around model collapse? > “These analogies are surprisingly good. Humans collapse during the course of their lives. Children haven't overfit yet. They will say stuff that will shock you. Because they're not yet collapsed. But we [adults] are collapsed. We end up revisiting the same thoughts, we end up saying more and more of the same stuff, the learning rates go down, the collapse continues to get worse, and then everything deteriorates.” In fact, there’s an interesting paper arguing that dreaming evolved to assist generalization, and resist overfitting to daily learning - look up The Overfitted Brain by Erik Hoel. I asked Karpathy: Isn’t it interesting that humans learn best at a part of their lives (childhood) whose actual details they completely forget, adults still learn really well but have terrible memory about the particulars of the things they read or watch, and LLMs can memorize arbitrary details about text that no human could but are currently pretty bad at generalization? > “[Fallible human memory] is a feature, not a bug, because it forces you to only learn the generalizable components. LLMs are distracted by all the memory that they have of the pre-trained documents. That's why when I talk about the cognitive core, I actually want to remove the memory. I'd love to have them have less memory so that they have to look things up and they only maintain the algorithms for thought, and the idea of an experiment, and all this cognitive glue for acting.”

Dwarkesh Patel

1,050,937 views • 9 months ago

Demis Hassabis: it's “nonsense” to claim that today's AI chatbots have PhD intelligence “And maybe scaling will get us there, but if I was to bet, I think there are probably one or two missing breakthroughs that are still required and will come over the next five or so years.” Demis Hassabis at the All-In Summit: “You often hear some of our competitors talk about these modern systems that we have today are PhD intelligences.” “I think that's nonsense. They're not PhD intelligences.” “They have some capabilities that are PhD level, but they're not, in general, capable. And that's exactly what general intelligence should be (capable) of, of performing across the board at the PhD level.” “In fact, as we all know, interacting with today's chatbots, if you pose the question in a certain way, they can make simple mistakes with even high school maths and simple counting.” “That shouldn't be possible for a true AGI system.” “So I think that we are maybe, I would say sort of 5-10 years away from having an AGI system that's capable of doing those things.” “Another thing that's missing is continual learning, this ability to like online teach the system something new or adjust its behavior in some way.” “And so a lot of these, I think, core capabilities are still missing.”
1:01

Demis Hassabis: it's “nonsense” to claim that today's AI chatbots have PhD intelligence “And maybe scaling will get us there, but if I was to bet, I think there are probably one or two missing breakthroughs that are still required and will come over the next five or so years.” Demis Hassabis at the All-In Summit: “You often hear some of our competitors talk about these modern systems that we have today are PhD intelligences.” “I think that's nonsense. They're not PhD intelligences.” “They have some capabilities that are PhD level, but they're not, in general, capable. And that's exactly what general intelligence should be (capable) of, of performing across the board at the PhD level.” “In fact, as we all know, interacting with today's chatbots, if you pose the question in a certain way, they can make simple mistakes with even high school maths and simple counting.” “That shouldn't be possible for a true AGI system.” “So I think that we are maybe, I would say sort of 5-10 years away from having an AGI system that's capable of doing those things.” “Another thing that's missing is continual learning, this ability to like online teach the system something new or adjust its behavior in some way.” “And so a lot of these, I think, core capabilities are still missing.”

The All-In Podcast

33,934 views • 10 months ago

Dylan Patel: “CPUs are sold out.” $INTC earnings and conference call confirmed the looming CPU shortage. There are three main drivers: - Agentic systems, as CPUs do the planning and orchestration. - Increasing complexity of and demand for reinforcement learning environments that run on CPUs. - Increasing deployment of AI instilled apps. Models run on GPUs but apps themselves run on CPUs. Demand for all these will grow secularly as AI systems will only get more agentic, RL environments will only get more complex, and deployment will only accelerate. As a result, GPU-to-CPU ratio in AI workloads will move from 1:8 to 1:2 or maybe even 1:1 in the future. CPU supercycle is about to start. $AMD $INTC $ARM
1:44

Dylan Patel: “CPUs are sold out.” $INTC earnings and conference call confirmed the looming CPU shortage. There are three main drivers: - Agentic systems, as CPUs do the planning and orchestration. - Increasing complexity of and demand for reinforcement learning environments that run on CPUs. - Increasing deployment of AI instilled apps. Models run on GPUs but apps themselves run on CPUs. Demand for all these will grow secularly as AI systems will only get more agentic, RL environments will only get more complex, and deployment will only accelerate. As a result, GPU-to-CPU ratio in AI workloads will move from 1:8 to 1:2 or maybe even 1:1 in the future. CPU supercycle is about to start. $AMD $INTC $ARM

Oguz Erkan

93,419 views • 2 months ago

Giving everyone a universal basic income will not reveal most people’s inner Mozarts or Emily Brontës. At bottom, this is about two competing views of human nature. One view holds that once basic material needs are met, people will use their free time to seek meaning and fulfillment. Unshackled from the burden of work, they will thrive. This is partly true. A small share of people would create, build, and explore. But for most, that is not what happens. When people are out of work, they do not spend their days painting or sculpting or learning another language. They scroll, they watch television, they play video games. Many advocates of UBI assume that people are simply waiting for the right conditions. Remove financial pressure, and they will pursue their creative passions. That may be true for a few. It is not true for most. Another view holds that meaning comes from the act of working. Earning your way, supporting yourself, and taking care of others provide structure and fulfillment. Effort, struggle, and self-reliance are not barriers to a meaningful life. They are part of what makes it possible. A society that removes the need to work risks removing one of the main sources of meaning in life.
0:53

Giving everyone a universal basic income will not reveal most people’s inner Mozarts or Emily Brontës. At bottom, this is about two competing views of human nature. One view holds that once basic material needs are met, people will use their free time to seek meaning and fulfillment. Unshackled from the burden of work, they will thrive. This is partly true. A small share of people would create, build, and explore. But for most, that is not what happens. When people are out of work, they do not spend their days painting or sculpting or learning another language. They scroll, they watch television, they play video games. Many advocates of UBI assume that people are simply waiting for the right conditions. Remove financial pressure, and they will pursue their creative passions. That may be true for a few. It is not true for most. Another view holds that meaning comes from the act of working. Earning your way, supporting yourself, and taking care of others provide structure and fulfillment. Effort, struggle, and self-reliance are not barriers to a meaningful life. They are part of what makes it possible. A society that removes the need to work risks removing one of the main sources of meaning in life.

Rob Henderson

72,611 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