Jared Kaplan explains that AI progress follows predictable scaling laws, with improvements in compute, data, and model size driving advancements in flexibility and task horizon, enabling AI to handle increasingly complex, long-term tasks across multiple modalities. He highlights the need for integrating organizational knowledge, memory, and nuanced oversight to achieve human-level AI, emphasizing ongoing human-AI collaboration and the potential for rapid future improvements in diverse fields.
Jared Kaplan, co-founder of Anthropic and former theoretical physicist, discusses the progression and future of AI through the lens of scaling laws. He explains that contemporary AI training involves two main phases: pre-training, where models learn to predict the next word in human-generated text or multimodal data, and reinforcement learning (RL), where models are fine-tuned based on feedback to perform useful tasks while being helpful, honest, and harmless. Kaplan highlights that scaling laws govern both phases, showing predictable improvements in AI performance as compute, data, and model size increase, a discovery that has given researchers confidence in the steady advancement of AI capabilities.
Kaplan emphasizes that AI capabilities can be viewed along two axes: flexibility (the range of modalities and tasks AI can handle) and task horizon (the length of time AI can effectively work on tasks). While early AI systems like AlphaGo excelled in narrow domains, modern large language models are increasingly multimodal and capable of handling longer, more complex tasks. Research shows that the length of tasks AI can perform is roughly doubling every seven months, suggesting that AI will soon be able to undertake tasks spanning days, weeks, or even years, potentially matching the output of entire human organizations or scientific communities.
Looking ahead, Kaplan identifies key ingredients needed to reach broadly human-level AI: relevant organizational knowledge, memory to track long-term progress, and improved oversight to handle nuanced, fuzzy tasks. He notes that while AI excels at clear-cut problems like coding or math, developing models that can generate nuanced reward signals for subjective tasks (e.g., humor, poetry, research quality) is crucial. He also stresses the importance of integrating AI into existing workflows and organizations, encouraging builders to experiment with AI’s current capabilities, as rapid improvements mean that products that don’t fully work today may become viable with future model releases.
During the Q&A, Kaplan discusses the incremental nature of AI progress, with each new model release improving on the last in areas like coding, memory, and supervision. He acknowledges AI’s current limitations, such as making basic mistakes, and highlights the importance of human-AI collaboration, where humans act as managers to verify AI outputs. He also points to greenfield opportunities beyond coding, such as finance, law, and business integration, where AI can add significant value. Kaplan attributes his ability to identify scaling laws to his physics background, which trained him to ask simple, precise questions about big-picture trends.
Finally, Kaplan addresses challenges like compute efficiency and the future of AI training. While current AI development focuses on unlocking frontier capabilities, he expects that over time, costs will decrease through algorithmic improvements and lower-precision computing. He also discusses the role of humans and AI in creating and supervising training tasks, suggesting a hybrid approach that leverages AI to generate tasks while still relying on human oversight for the most complex challenges. Overall, Kaplan’s insights paint a picture of steady, predictable AI progress driven by scaling laws, with exciting opportunities and challenges ahead as AI approaches human-level intelligence.