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

Jared Kaplan, co-founder of Anthropic, explains how the discovery of predictable, physics-like scaling laws in AI training provides a clear roadmap for progress. He details the two main phases of model training (pre-training and RL), discusses how scaling compute predictably unlocks longer-horizon task capabilities, and outlines the remaining challenges—memory, nuanced oversight, and organizational knowledge—on the path to human-level AI.

July 29, 2025 4 min read

Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan — source: https://www.youtube.com/watch?v=p8Jx4qvDoSo

Jared Kaplan’s journey from theoretical physics to co-founding Anthropic was driven by a search for fundamental principles. In AI, he found one: intelligence scales in a predictable, almost physical way. This insight into “scaling laws” has become a foundational driver of progress in large language models, suggesting a systematic road toward human-level AI.

The Two-Phase Engine of Modern AI

Contemporary AI models like Claude are built using a two-phase process:

Pre-training: This is the foundational phase where a model learns from a massive corpus of human-generated data (text, and now, multimodal data). The core task is simple: predicting the next word. This process ingrains an understanding of patterns, correlations, and knowledge embedded in the data, effectively allowing the model to “imbibe all of human civilization’s knowledge.”
Reinforcement Learning (RL): After pre-training, the model is refined to be useful and aligned with human intentions. Through interfaces where humans or crowdworkers rank different model responses, the AI is optimized to produce outputs that are helpful, honest, and harmless. This RL phase steers the model’s vast knowledge toward desirable behaviors.

The Surprising Predictability of Scaling Laws

The most profound discovery in modern AI has been the existence of precise scaling laws. By asking simple, “dumb” questions inspired by his physics background—like “How much does more data help?"—Kaplan and his colleagues uncovered a startlingly predictable trend: as you increase compute, dataset size, and model parameters, the model’s performance improves along a smooth, predictable curve.

“We found that there’s actually something very very very precise and surprising underlying AI training. This really blew us away that there are these nice trends that are as precise as anything that you see in physics or or astronomy.”

These trends, observable over many orders of magnitude, apply to both the pre-training and RL phases. This predictability gives researchers immense conviction that continued investment in scale will lead to steadily more intelligent systems. Progress isn’t just the result of clever new architectures; it’s about systematically “turning the crank” on a known, reliable process.

The Expanding Frontiers of AI Capability

Scaling unlocks capabilities along two primary axes:

Flexibility: The ability of AI to operate across various modalities (text, images, etc.), essentially meeting humans where they are.
Task Time Horizon: A more critical axis, this measures the length of a task an AI can successfully complete. An empirical finding by METER showed this horizon is doubling roughly every seven months.

This exponential growth in task complexity suggests that AI is on a trajectory to handle tasks that take not just minutes or hours, but days, weeks, and eventually, the kind of work performed by entire organizations or scientific communities.

The Remaining Challenges for Human-Level AI

While scaling provides a clear path forward, several key ingredients are still needed to achieve broadly capable, human-level AI:

Relevant Organizational Knowledge: Models must move beyond being a “blank slate” and integrate the specific context and knowledge of the company or organization they are working within.
Memory: For long-horizon tasks, models need a persistent memory to track progress, store intermediate results, and build upon previous work. This is a feature beginning to be implemented in models like Claude 4.
Oversight for Nuanced Tasks: A major challenge is applying RL to “fuzzy” domains where correctness is subjective, such as telling a good joke, writing a compelling poem, or having good taste in research. Developing more sophisticated, AI-assisted reward signals is crucial for improving performance on these nuanced tasks.

Implications for Builders and the Future

For those building with AI, Kaplan offers several recommendations:

Build on the Frontier: Create products that don’t quite work with today’s models. As the capability line advances, your product may become viable with the next generation, like a “Claude 5.”
Human-AI Collaboration: The current paradigm is one of partnership. Humans excel at providing oversight and sanity-checking AI outputs, acting as managers for AI agents. This is especially true for advanced tasks where AI can make brilliant leaps but also simple errors.
The Value of the Frontier: While AI will become more efficient and cheaper, the Jevons paradox suggests that as intelligence improves, demand will grow. Kaplan expects that the most significant value will continue to be at the frontier, with the most capable models that can handle complex, end-to-end tasks without requiring humans to break them down into smaller pieces.