How do scaling laws in AI work and what do they predict?

Scaling laws show that as AI models increase in size (compute, data, parameters), their performance predictably improves. They suggest a continued, smooth curve towards more capable AI, including human-level AI.

What are the key ingredients needed for human-level AI beyond current scaling capabilities?

Beyond scaling, AI needs relevant organizational knowledge to work within structures, memory to track progress on long tasks, and improved oversight for nuanced tasks beyond simple right/wrong answers.

What advice is given for preparing for the future of AI?

It's recommended to build things that don't quite work yet, experiment on the boundaries of AI capabilities, leverage AI for AI integration, and identify areas ripe for rapid AI adoption.

How does Claude 4 improve upon previous versions for tasks like coding?

Claude 4 is designed to act more effectively as an agent for coding and other applications, with improved supervision to follow directions better and enhanced memory capabilities to handle complex tasks across multiple context windows.

Is human-AI collaboration or full automation the future?

While AI is becoming capable of end-to-end task automation, human-AI collaboration is crucial for the most advanced tasks. In the long term, more tasks are expected to be fully automated, but human oversight might still be needed.

What are some potential greenfield areas for AI development beyond coding?

Areas requiring significant computer-based data interaction like finance and potentially law are seen as greenfield. Integrating AI into existing businesses by rethinking workflows, similar to early electricity adoption, also holds great leverage.

How can individuals stay relevant in a future dominated by highly capable AI?

Focus on understanding how AI models work, efficiently leveraging and integrating them into workflows, and building at the frontier of AI capabilities are key to staying relevant.

Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan

Y Combinator

Science & Technology6 min read41 min video

Jul 29, 2025|57,375 views|922|37

YC Y Combinator

Save to Pod

Key Moments

On this page

TL;DR

AI scaling laws predict predictable progress toward human-level AI, enabled by pre-training and reinforcement learning.

Key Insights

Intelligence, particularly in AI, scales predictably with increased compute and data, similar to physical laws.

Modern AI training involves two phases: pre-training for pattern recognition and reinforcement learning for task optimization.

Scaling laws apply to both pre-training and reinforcement learning, leading to systematic improvements in AI performance.

AI capabilities are expanding beyond task flexibility to longer task horizons, suggesting a path toward AGI.

Future AI development requires advancements in organizational knowledge, memory, and nuanced oversight.

Building at the frontier of AI capabilities and leveraging AI for AI integration are key strategies for innovation.

FROM PHYSICS TO ARTIFICIAL INTELLIGENCE

Jared Kaplan transitioned from theoretical physics, driven by fundamental questions about the universe, to AI research. Initially skeptical of AI's potential, he was convinced by the progress and the people involved, eventually co-founding Anthropic. His physics background, particularly the pursuit of understanding grand trends and asking fundamental questions, proved instrumental in his approach to AI research, especially in identifying and quantifying scaling laws.

THE TWO PILLARS OF MODERN AI TRAINING

Contemporary AI models like Claude and ChatGPT are trained in two primary phases. The first is pre-training, where models learn to imitate human-written text and understand underlying correlations by predicting the next word in vast datasets, or even multimodal data. The second phase is reinforcement learning, which refines the model's behavior using human feedback to encourage helpful, honest, and harmless responses, optimizing for useful task completion by reinforcing preferred actions and discouraging undesirable ones.

THE POWER OF SCALING LAWS IN AI

A foundational insight in AI development has been the discovery of scaling laws, which demonstrate a predictable relationship between the resources used for training (compute, data, model size) and the resulting performance. These laws, as precise as those found in physics, apply to both pre-training and reinforcement learning. This predictability provides strong conviction that AI capabilities will continue to improve systematically as computational resources increase, driving progress exponentially.

EXPANDING AI HORIZONS AND CAPABILITIES

AI capabilities are advancing along two axes: flexibility (handling diverse modalities) and task horizon. While AI is becoming more versatile, the ability to perform tasks of increasing duration is particularly significant. Research suggests that the length of tasks AI can handle is doubling regularly, potentially leading to AI systems capable of undertaking complex, multi-day, or even multi-year projects, mirroring the work of human organizations or scientific communities.

KEY INGREDIENTS FOR HUMAN-LEVEL AI

Achieving human-level AI requires more than just scaling. Key ingredients include relevant organizational knowledge to enable AI to work within established contexts, robust memory systems for tracking progress on long tasks, and enhanced oversight capabilities for handling nuanced and fuzzy problems. Expanding AI to more complex tasks and modalities, such as moving from text to multimodal systems and robotics, remains a critical area of development.

NAVIGATING THE FUTURE OF AI DEVELOPMENT

Preparing for the rapid evolution of AI involves building and experimenting at the boundaries of current capabilities, as future models will likely overcome current limitations. Leveraging AI for its own integration into products and processes is also crucial for accelerating adoption. Identifying areas ripe for rapid AI integration, beyond coding, is key to unlocking widespread value and innovation as AI continues its relentless progress.

IMPROVEMENTS IN CLAUDE 4 AND BEYOND

The release of Claude 4 brings notable advancements, particularly in its ability to act as an agent for complex tasks like coding and search. It features improved supervision for better adherence to instructions and code quality, alongside enhanced memory capabilities for managing long-horizon tasks across multiple context windows. These incremental improvements, driven by scaling, are pushing AI towards more capable and collaborative roles.

THE ROLE OF HUMANS IN AN AI-DRIVEN WORLD

While AI is becoming increasingly capable of performing tasks end-to-end, humans will continue to play vital roles, especially in managing and sanity-checking AI outputs, particularly for complex tasks. Human oversight is crucial for AI to develop nuanced judgment, learn to generate creative content like jokes or poems, and possess good taste in research. This human-AI collaboration is expected to be the most fruitful area for advanced AI applications.

APPLYING PHYSICS PRINCIPLES TO AI CHALLENGES

Kaplan's physics background influenced his approach to AI by focusing on macro trends and precision. He emphasizes asking fundamental, sometimes naive, questions like the exact nature of convergence (power law versus exponential) to sharpen understanding. This methodical approach to quantifying trends is vital for identifying areas for improvement, such as finding better 'slopes' in scaling laws to gain a competitive advantage in AI development.

INTERPRETABILITY AND THE FRONTIER OF AI

Interpretability in AI is seen as more akin to biology or neuroscience, requiring an understanding of complex systems. Unlike the human brain, AI systems offer complete measurability, allowing for detailed reverse-engineering of their workings. While advanced mathematical techniques are involved, the core challenge lies in understanding AI's intricate features, similar to deciphering how the brain functions, but with the advantage of full data access.

THE RELIABILITY OF SCALING LAWS

Scaling laws have consistently held true over vast ranges, suggesting their fundamental nature in AI development. When these laws appear to fail, it's often an indication of errors in the AI training process, architectural flaws, or algorithmic bottlenecks, rather than a breakdown of the scaling principle itself. This perspective reinforces the belief that continued adherence to scaling principles, with improved methodologies, will drive future AI progress.

COMPUTE EFFICIENCY AND THE FUTURE OF AI

The drive for AI progress is currently focused on unlocking frontier capabilities, which requires significant computational resources. While AI training and inference are becoming more efficient through algorithmic improvements and hardware scaling, the immense value in frontier models means efficiency might not immediately lead to lower costs. As AI becomes more widespread, cost reduction will likely become a greater focus, potentially involving lower-precision computations.

ADVICE FOR STAYING RELEVANT IN AI

To remain relevant in a rapidly advancing AI landscape, individuals should focus on understanding how AI models work, efficiently leveraging and integrating them, and building at the frontier of AI capabilities. Continuous learning and adaptation are essential, with an emphasis on developing skills that complement AI's rapidly expanding functionalities and exploring novel applications across various domains.

THE MYSTERY OF EXPONENTIAL GROWTH IN TASK HORIZONS

The observed exponential growth in the time horizon of tasks AI can accomplish, despite linear progress in scaling laws related to compute, is a compelling empirical finding. This suggests that improvements in AI's ability to self-correct and adapt when encountering reality, even modest ones, can significantly extend the duration and complexity of tasks it can manage, leading to disproportionately large gains in overall capability.

GENERATING TASKS FOR AI TRAINING

The process of creating tasks for reinforcement learning in AI training involves a hybrid approach. While AI models, particularly large language models, are increasingly used to generate tasks, especially in domains like coding, human input remains essential. As AI capabilities advance, the complexity of tasks also increases, necessitating continued human involvement in task design to ensure robust and comprehensive AI development.

Mentioned in This Episode

●Software & Apps

●Organizations

●Concepts

●People Referenced

Common Questions

The two main phases are pre-training, where models imitate human-written data to understand correlations, and reinforcement learning, where models are optimized based on feedback to perform useful tasks like being helpful, honest, and harmless.