AI Explained Ep 3: How Will AGI Arrive? 3 Bets, 3 Catches #AGI #WorldModels #AIExplained

Publicado em

Vídeo em Inglês, publicado pelo canal Pay Grade AI.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video discusses three competing bets in the pursuit of artificial general intelligence (AGI): making models bigger through scaling laws, giving AI a model of the physical world, and letting AI learn by doing through experience and trial. It highlights the catches and signals of success for each approach, noting how they are beginning to merge in current research.

Key points

  • Scaling laws show that model prediction errors fall predictably as computing power, data, and model size increase, but public human text data is projected to run out between 2026 and 2032.
  • Allowing models to think step-by-step before answering significantly improves performance on complex tasks, such as OpenAI's o1 model outperforming GPT-4o on math tests.
  • Bet two involves building world models, supported by companies like AMI Labs and Google DeepMind, to help AI understand physical environments and interaction, though maintaining these worlds over long periods remains a challenge.
  • Bet three focuses on letting AI learn by doing through reinforcement learning, creating its own data and learning from experience, similar to how AlphaZero mastered chess without human games.
  • Current developments show these three distinct bets are merging together, combining scale, world models, and experiential learning into advanced agents like SIMA 2.
  • The video suggests a practical skill for users is shifting from asking AI simple questions to delegating complete tasks and reviewing the results like a manager.

Topics

Artificial General IntelligenceScaling LawsWorld ModelsReinforcement LearningData ExhaustionAI Agents
Video transcript

Transcript from available captions; it may contain errors.

400,000 years. That's how long it would take you to read all the text online, says Yann LeCun. By his math, a four-year-old has already taken in about that much data, just by looking. So is more text really the road to AGI, human-level AI? Your next skill depends on that answer. So do billions of dollars. No hype, no filler. Short answer: three big bets, and nobody knows which wins. Make models bigger, give them a model of the world, or let them learn by doing. Each has a catch, and each makes predictions you can check. One bet may even fix another's biggest problem: running out of data. And all three point to one skill you can start building this week. Here's our ladder. Today's AI is climbing toward the middle rung. Three roads lead up to it, and each one has a weak spot. Bet one: make it bigger. More chips, more data, bigger models. Why bet on size? In 2020, researchers found something striking about language models, the engines inside chatbots. Their prediction error, how wrong their next-word guesses are, falls on a predictable curve. That happens as you add computing power, data and size. They called these scaling laws. So the industry scaled. Look at the computing power used to train the biggest models. From 2010 to 2024, it grew four to five times a year. GPT-6 Astra was trained on more than 100,000 GPUs, the chips that run AI. There's a second way to scale: let the model think longer before it answers. In September 2024, OpenAI tested this on a hard math contest. GPT-4o solved 12% of the problems. Then o1, trained to reason step by step first, took the same test. Guess its score. 74%. That's about six times the score. And the longer o1 was allowed to think, the better it did. The catch for bet one is data. In 2024, the research group Epoch AI made an estimate. Their best guess: usable public human text adds up to about 300 trillion tokens. A token is a chunk of a word. If trends continue, models will have used all of it sometime between 2026 and 2032. That window is open now. So where does the next fuel come from? Hold that question. Bet three has an answer. Bet two: give AI a model of the world. When you toss your keys to a friend, you don't do the physics. You predict where they'll land. That's a world model. LeCun argues chatbots are good with words but don't understand the physical world. In March, his new company, AMI Labs, raised over a billion dollars to build world models. AMI stands for advanced machine intelligence, his alternative to the word AGI. Systems that learn by watching, and predict the main idea of what happens next, not every pixel. Google DeepMind bets on world models too, but a different kind. In August 2025, it showed Genie 3, which draws every frame. It builds worlds you can walk through, live, at 24 frames a second. DeepMind calls world models a key stepping stone on the path to AGI. The catch for bet two: those worlds don't last. DeepMind says Genie 3 supports a few minutes of continuous interaction, not hours. And an agent, an AI that takes actions, can only do a limited set of things inside them. A world model that holds together for minutes is a start. Real life runs for years. Quick ask: if this helps, subscribe. The first three episodes are a free course. Bet three: let AI learn by doing. It answers the data problem: an agent that acts makes its own data. In 2025, two pioneers of reinforcement learning, or learning by trial and reward, made this case. David Silver and Richard Sutton argued that high-quality human data is running out. The next era, they wrote, belongs to agents that learn mostly from their own experience. In their words: agents will inhabit streams of experience, rather than short snippets of interaction. And remember Episode one. Agents keep finishing longer software tasks, at a 50% success rate. There's a precedent. In 2017, AlphaZero started from random moves, knowing only the rules of chess. Within 24 hours, it played at a superhuman level. No human games at all. Here's the catch. AlphaZero had a perfect rulebook, and it could play millions of practice games against itself in hours. The real world has no rulebook, and experience there is slow and expensive. So which bet wins? Our read: they're merging. In November 2025, Google DeepMind showed an agent called SIMA 2. Gemini, Google's big AI model, is its core. After learning from human players, it can pick up new games through its own play. DeepMind even had it teach itself inside new worlds made by Genie 3. Scale, a world model and learning by doing, side by side. Scale gives knowledge. A world model gives physics. Experience gives skill. Which mix gets there first, and when? Nobody knows. Here's how to tell which bet is paying off. If scale is winning, each bigger model keeps beating the last, even after the human text runs out. If world models are winning, robots suddenly get good at objects they've never seen. If experience is winning, agents get better the longer they work, without being retrained. Watch those three signals, and you'll read AI news better than most headlines do. Pay Grade Check. Every bet aims at AI that finishes longer tasks on its own. So the skill to build is handing AI whole tasks, not just questions. Anthropic, which makes Claude, tracks how people use it. In late 2024, 27% of conversations handed over a complete task. By August 2025, it was 39%. That's one company's data, about its own users. So this week, give an AI one complete task, not a question. Write the brief. Define what done looks like. Then review the result like a manager. Quick test before you go. Can you name the three bets? Bigger, a model of the world, and learning by doing. Each has a catch, and Google DeepMind is already combining them. If you remember one thing: knowledge, physics and skill. AGI probably needs all three. Next Thursday: will AGI take your job? We'll look at real payroll data, and who's feeling it first. Subscribe and tap the bell so you don't miss it. Start the course from Episode 1 and watch it next.

Assistir ao vídeo original no YouTube

Assuntos: AGI, ANI, ASI. Publicado no YouTube em 2026-11-05.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.

Comentários

Postagens mais visitadas deste blog

How to build a JARVIS AI Voice Agent for FREE | Full Livekit Tutorial (2026)

How Does AI Actually Work? 🤖 | AI, Machine Learning & Generative AI Explained