Dwarkesh Podcast Notify me• Dec 23, 2025• 12:31Interview

What are we scaling?

From The Dwarkesh Podcast

Dwarkesh Patel(host)

Get the full transcript next time Dwarkesh Podcast releases an episode

Summary, key quotes, top claims, and the searchable transcript - emailed automatically. No card needed.

Executive Summary

Continue your research

Keep pulling the thread on Microsoft Excel.

The Continual Learning Bottleneck Pre-trained Skills vs. On-the-Job Learning

12 quotes

Concerns Raised

Current AI models lack the crucial capability of continual, on-the-job learning, limiting their economic impact.
The industry's reliance on pre-baking skills via reinforcement learning is an inefficient and unscalable approach to AGI.
Scaling laws for pre-training do not apply to RL, which may require orders of magnitude more compute for meaningful gains.
The hype around AI capabilities is outpacing their actual economic utility and deployment in real-world jobs.

Opportunities Identified

Solving continual learning represents the most significant opportunity to unlock true AGI and trillions of dollars in economic value.
A future AI architecture of specialized, continually learning agents contributing to a central 'hive mind' could be a powerful paradigm.
The massive market for knowledge work (tens of trillions in wages) provides a clear benchmark and incentive for developing more capable AI.

Key Themes

Research Findings12

According to Baron Milledge, improvements in frontier AI models are driven not only by increased scale and ML research but also by billions of dollars paid to human experts like PhDs and MDs to create high-quality training data.

Dwarkesh Patel predicts the development of "actual brain-like intelligences" within the next 10 to 20 years.

Dwarkesh Patel predicts that by 2030, AI models will be generating hundreds of billions of dollars in annual revenue.

Dwarkesh Patel predicts that by 2030, major AI labs will have made significant progress on the problem of continual learning.

The predictable scaling laws observed in AI pre-training do not apply to reinforcement learning from verifiable reward, for which there are no well-established scaling trends.

According to an analysis by Toby Ord, a one-million-fold increase in total reinforcement learning compute is needed to achieve a performance boost equivalent to a single GPT-level improvement.

Dwarkesh Patel predicts that achieving human-level on-the-job learning capabilities in AI systems will take an additional 5 to 10 years to fully develop.

A supply chain of companies exists to build reinforcement learning (RL) environments designed to teach AI models specific skills, such as web browser navigation or using Microsoft Excel for financial modeling.

A common argument for a near-term AI takeoff, within five years, is that current reinforcement learning methods are a temporary step to build a superhuman AI researcher, likened to an "automated Ilya," which will then solve more fundamental learning problems.

The slow diffusion of AI in businesses is not due to normal technology adoption lags, but because current models lack the necessary capabilities to provide broad economic value.

The cumulative annual wages for knowledge workers globally amount to tens of trillions of dollars.

The revenue of AI labs is orders of magnitude lower than the total wages of knowledge workers because current AI models are not nearly as capable as their human counterparts.

Topics

AGI Timelines Continual Learning Reinforcement Learning (RL)On-the-Job Learning AI Economics Knowledge Work Automation Scaling Laws AI Hype Cycle AI Competition Model Generalization AI Capabilities vs. Utility AI Moats Future AI Architectures Robotics Pre-training

Processed Apr 9, 2026Daily intelligence brief → yt-dlp + mlx-whisper + Gemini

What are we scaling?

Continue your research

Concerns Raised

Opportunities Identified

Key Themes

The Continual Learning Bottleneck

Pre-trained Skills vs. On-the-Job Learning

AI Capabilities vs. Economic Utility

Competitive Dynamics and Incremental Progress

Research Findings12

Topics