Overview

In this solo essay-narration episode of the Dwarkesh Podcast, the host presents a tightly argued critique of the dominant narrative in AI progress, challenging both short-timeline optimists and those who attribute slow deployment to economic diffusion lags. The central thesis is that current frontier models, despite impressive benchmark performance, lack the fundamental capability for on-the-job learning and generalization that defines human intelligence—and that this gap, not corporate inertia or market friction, explains why AI has not yet transformed knowledge work. The episode has the feel of a rigorous internal debate, with Dwarkesh systematically dismantling counterarguments from both accelerationists and skeptics while laying out his own timeline expectations for a more gradual, iterative path to AGI.

---

0:00The Tension Between Short Timelines and RL Scaling

Dwarkesh opens with a pointed confusion: why do some people who predict AGI within five years simultaneously place their faith in scaling up reinforcement learning (RL) atop large language models? He argues that if we were genuinely close to a human-like learner, the entire approach of training models on verifiable outcomes through RL would be fundamentally misguided. The current paradigm, he explains, involves labs "baking in" skills through mid-training—an entire supply chain of companies building RL environments that teach models how to navigate web browsers, use Excel, or build financial models. This creates a logical dilemma: either these models will soon learn on the job in a self-directed way, making all this pre-baking pointless, or they won't, which means AGI is not imminent.

Dwarkesh draws on a blog post by Baron Milledge to sharpen the point. When we see frontier models improving at benchmarks, Milledge argues, we should think not just about increased scale and clever ML research, but about the billions of dollars paid to PhDs, MDs, and other experts to write questions and provide example answers targeting these precise capabilities. The tension is most vivid in robotics: with very little training, a human can learn to operate current hardware to do useful work. If we actually had a human-like learner, robotics would be largely solved. Instead, we must go into thousands of homes and practice millions of times on picking up dishes or folding laundry.

---

1:34The "Automated Researcher" Counterargument and Its Flaws

Dwarkesh addresses the most common counterargument from short-timeline optimists: that all this kludgy RL is in service of building a superhuman AI researcher, and then millions of copies of that automated researcher will figure out how to solve robust, efficient learning from experience. He dismisses this as reminiscent of the old joke about losing money on every sale but making it up in volume. The idea that an automated researcher will solve the algorithm for AGI—a problem humans have been working on for half a century—while lacking the basic learning capabilities that children possess strikes him as "super implausible."

Moreover, Dwarkesh notes that this belief doesn't align with how labs actually approach RL from verifiable reward. If the goal were simply to automate an AI researcher, there would be no need to pre-bake skills like crafting PowerPoint slides. The fact that labs are investing heavily in these specific skill-building pipelines suggests they hold a worldview where models will continue to fare poorly at generalization and on-the-job learning, making it necessary to build economically useful skills in advance.

He acknowledges a more reasonable counterargument: even if models could learn these skills on the job, it might be more efficient to build them in once during training rather than repeatedly for each user and company. It makes sense to bake in fluency with common tools like browsers and terminals. But Dwarkesh argues that people are underrating how much company-specific and context-specific skills are required for most jobs—and there is currently no robust, efficient way for AIs to pick up these skills.

---

3:10The Value of Human Labor and the Schleppy Training Loop Problem

Dwarkesh recounts a revealing dinner conversation with an AI researcher and a biologist who had long timelines for AGI. The biologist described a recent lab task: looking at slides and deciding whether a dot was actually a macrophage or just looked like one. The AI researcher immediately responded that image classification is a textbook deep learning problem, squarely in the sweet spot of what models can be trained to do. Dwarkesh found this exchange illuminating because it illustrated a key crux between himself and those expecting transformative economic impact within a few years.

The core insight: human workers are valuable precisely because we don't need to build custom training loops for every small part of their job. It is not net productive to build a pipeline to identify what macrophages look like given one lab's specific slide preparation method, then another pipeline for the next lab's micro-task, and so on. What is actually needed is an AI that can learn from semantic feedback or self-directed experience and generalize the way humans do every day. A typical knowledge worker must do a hundred things requiring judgment, situational awareness, and skills learned in context—tasks that differ not just across people but from one day to the next for the same person. It is not possible to automate even a single job by baking in a predefined set of skills, let alone all jobs.

Dwarkesh argues that people are underestimating how big a deal actual AGI will be because they imagine more of the current regime rather than billions of human-like intelligences on a server that can copy and merge all their learnings. He clarifies that he does expect this—actual brain-like intelligences within the next decade or two—which he acknowledges is "pretty fucking crazy."

---

5:04Economic Diffusion Lag as Cope

Dwarkesh directly challenges the argument that AIs are not yet widely deployed because technology takes a long time to diffuse across firms. He calls this "cope"—a way to gloss over the fact that these models simply lack the capabilities necessary for broad economic value. If models were actually like humans on a server, they would diffuse incredibly quickly. They would be easier to integrate and onboard than a normal human employee: they could read an entire Slack history within minutes and immediately distill all the skills that other AI employees have.

He draws a comparison to the hiring market for humans, which operates like a "lemons market"—it's hard to tell who the good people are beforehand, and hiring someone who turns out to be bad is very costly. This dynamic would not exist for AI: you would simply spin up another instance of a vetted AGI model. For these reasons, Dwarkesh expects it would be much easier to diffuse AI labor into firms than it is to hire a person—and companies hire people all the time. If capabilities were actually at AGI level, companies would be willing to spend trillions of dollars a year buying tokens from these models. Knowledge workers globally earn tens of trillions of dollars in wages annually. The fact that labs are orders of magnitude off this figure reveals that the models are nowhere near as capable as human knowledge workers.

---

6:34Goalpost Shifting Is Justified

Dwarkesh anticipates the objection: how can the standard suddenly become that labs must earn tens of trillions of dollars in revenue? Until recently, people were asking whether models could reason, whether they had common sense, or whether they were just doing pattern recognition. AI bulls are right to criticize bears for repeatedly moving these goalposts, and this criticism is often fair—it's easy to underestimate AI progress over the last decade. But some amount of goalpost shifting is actually justified.

He makes a striking admission: if shown Gemini Theory in 2020, he would have been certain it could automate half of knowledge work. Yet here we are, having solved what we thought were the sufficient bottlenecks to AGI—models with general understanding, few-shot learning, and reasoning—and we still don't have AGI. The rational response, Dwarkesh argues, is to recognize that there is much more to intelligence and labor than previously realized. While models have surpassed what he would have defined as AGI in the past, the fact that model companies are not making trillions of dollars in revenue reveals that his previous definition was too narrow. He expects this pattern to continue: by 2030, labs will have made significant progress on continual learning and models will earn hundreds of billions of dollars annually, but they still won't have automated all knowledge work. He will then point to additional capabilities—X, Y, and Z—that are still needed.

---

8:23RL Scaling vs. Pre-Training Scaling

Dwarkesh turns to the question of what we are actually scaling. Pre-training had an extremely clean and general trend in improvement in loss across multiple orders of magnitude in compute, albeit on a power law that is "as weak as exponential growth is strong." He argues that people are trying to "launder the prestige" of pre-training scaling—which was almost as predictable as a physical law—to justify bullish predictions about RL from verifiable reward, for which there is no well-fed, publicly known trend.

When intrepid researchers do try to piece together implications from scarce public data, they get bearish results. Dwarkesh cites Toby Board's post, which cleverly connects the dots between different O-series benchmarks. Board's analysis suggests that "we need something like a million times scale-up in total RL compute to give a boost similar to a single GPT level." This is a sobering data point for those expecting rapid gains from RL scaling alone.

---

9:18The Broadly Deployed Intelligence Explosion

Dwarkesh addresses the popular scenarios of a software-only singularity (where AI models write code for smarter successors) or a software-plus-hardware singularity (where AIs also improve computing hardware). He argues that these neglect what will likely be the main driver of further improvements: continual learning. He returns to Baron Milledge's suggestion that the future might look like continual learning agents deployed across different jobs, generating value, and bringing back their learnings to a hive-mind model that performs some kind of batch distillation.

These agents could be quite specialized, containing what Andrej Karpathy called "the cognitive core" plus knowledge and skills relevant to their deployment. Dwarkesh predicts that solving continual learning will not be a singular, one-and-done achievement. Instead, it will feel like solving in-context learning. GPT-3 already demonstrated powerful in-context learning in 2020—the paper was titled "Language Models are Few-Shot Learners"—but we didn't solve in-context learning then, and progress on comprehension and context length continues. He expects a similar progression with continual learning: labs will likely release something next year that they call continual learning, which will count as genuine progress, but human-level on-the-job learning may take another five to ten years to iron out.

This is why Dwarkesh does not expect runaway gains from the first model that cracks continual learning. If full continual learning dropped out of nowhere, it might be "game, set, match"—as Satya Nadella put it when Dwarkesh asked about this possibility on the podcast. But that's probably not what will happen. Instead, some lab will figure out initial traction, playing around with the feature will reveal how it was implemented, and other labs will soon replicate and improve upon it.

---

11:17Competition Neutralizes Runaway Advantages

Dwarkesh concludes his argument with an observation about competitive dynamics. He has a prior that competition will remain fierce between model companies, informed by the fact that all previous supposed "flywheels"—whether user engagement on chat, synthetic data, or whatever—have done very little to diminish competition. Every month or so, the big three model companies rotate around the podium, and other competitors are not far behind. There seems to be some force—potentially talent poaching, the rumor mill, or just normal reverse engineering—that has so far neutralized any runaway advantage a single lab might have had.

---

Conclusion

What stays with the listener is Dwarkesh's willingness to hold two seemingly contradictory positions simultaneously: he expects genuine AGI within a decade or two, yet he is deeply skeptical of the current trajectory and timeline predictions. His most memorable contribution is the concept of the "schleppy training loop"—the idea that the value of human labor lies precisely in not needing custom training for every micro-task. The episode matters because it provides a rigorous, internally consistent framework for understanding why AI progress on benchmarks has not translated into economic transformation, and why this gap is likely to persist longer than both optimists and pessimists expect. Dwarkesh's essay-narration format allows for unusually tight reasoning, free from the conversational digressions that can dilute podcast arguments.

---

Key takeaways

• The current RL-from-verifiable-reward paradigm is logically inconsistent with short AGI timelines: if models were close to human-like learners, pre-baking skills would be unnecessary.
• Human labor is valuable because it requires no custom training loops for each micro-task; this "schleppy training loop" problem is the fundamental bottleneck to AI-driven economic transformation.
• Claims that AI deployment is slow due to economic diffusion lag are "cope"—if models were truly AGI-level, they would diffuse faster than human hires.
• Some goalpost shifting is justified: each time we solve what we thought was the bottleneck to AGI, we discover there is more to intelligence than we realized.
• RL scaling lacks the clean, predictable trends of pre-training scaling; Toby Board's analysis suggests a million-fold RL compute increase may be needed for a single GPT-level boost.
• Continual learning will likely follow the same trajectory as in-context learning: initial breakthroughs that count as progress, but human-level on-the-job learning taking another 5-10 years.
• Competitive dynamics between labs—talent poaching, reverse engineering, and the rumor mill—have consistently neutralized any single lab's runaway advantage.