The AI Scaling Problem: Why Current LLMs Aren’t Truly Intelligent скачать в хорошем качестве

The AI Scaling Problem: Why Current LLMs Aren’t Truly Intelligent

*#AI #AGI #MachineLearning #ContinualLearning #ComputerScience #AIScaling* Is Artificial General Intelligence (AGI) actually imminent, or have we been sold a lie? While we are constantly bombarded with news of AI surpassing human benchmarks in everything from the SAT and LSAT to PhD-level physics, these metrics may be misleading. In this video, we dive into the fundamental "Scaling Problem" and why the current direction of AI research is missing the mark on what it actually means to be intelligent. *The Core Argument* Intelligence is defined as the ability to *acquire and apply* knowledge and skills. While current frontier models are excellent at applying knowledge they have already seen, they almost entirely lack the ability to *continually learn* and acquire new information once their training is complete. We discuss why the "all-knowing agent" model is unrealistic and why we need to return to the roots of AI—animal-like, continual learning systems. *The Three Pillars of the Future of AI* To reach true AGI, the field needs to shift its focus toward three critical, often overlooked areas of research: 1. *Continual Learning:* Moving away from static, frozen weights toward agents that never stop learning. We address why current workarounds like fine-tuning lead to *catastrophic forgetting* and why in-context learning is an insufficient substitute for true memory. 2. *Single Experiential Streams:* Humans learn from a continuous, lifelong sequence of observations. Current LLMs, however, are trained on billions of *disjointed segments* of text sampled randomly from the internet, which prevents them from reasoning over long time horizons or developing episodic memory. 3. *Scaling with Compute, Not Just Data:* Tech giants are running into "data problems" because current methods are grossly inefficient, requiring millions of examples to learn simple concepts. We need algorithms that can extract more intelligence from their moment-to-moment experiences by *scaling with compute* rather than just gobbling up more data. *What’s Next?* If you’re tired of generic AI hype and want to understand the actual research challenges holding the field back, this video is for you. We explore how we can move toward a general-purpose learning algorithm inspired by animal learning. *Subscribe* for more deep dives into AI research and the future of continual reinforcement learning. --- *Tags* AI Scaling, AGI, Artificial General Intelligence, Large Language Models, LLMs, Continual Learning, Machine Learning Research, Catastrophic Forgetting, Loss of Plasticity, Edan Meyer, AI Benchmarks, Reinforcement Learning, Data Efficiency, Compute Scaling, Neural Networks, Deep Learning, AI Myths, Future of AI, Single Stream Learning, In-context Learning, GPT-4, OpenAI, Meta AI, Equinox Library, Machine Learning Theory, AI Development, Human-level Intelligence.