Latest Interviews

8 Predictions for the Era of Continual Learning

Dwarkesh

The discussion argues that transitioning from static to continual learning is essential for AI to perform complex tasks and maintain safety through ongoing adaptation rather than pre-deployment checks. This shift is expected to fragment the current market of identical models into diverse, experience-driven systems while creating high switching costs that allow providers to secure substantial profit margins. Consequently, major labs face intense pressure to deploy functional models immediately to leverage live data feedback, which will accelerate development cycles and reshape the economics of AI inference.

Why compute prices might 10x as AI gets smarter

Anthropic's projected tenfold revenue growth outpaces the industry's threefold compute expansion, forcing a shift toward higher inference margins, increased spot prices, and a reallocation of hardware to inference workloads. This economic divergence is exacerbated by hard supply constraints in Moore's Law, fab construction, and wafer allocation, which concentrate pricing power among frontier labs capable of charging significant premiums for compute-efficient models. Consequently, the market faces a pre-singularity regime where inelastic compute supply drives sustained price increases and intelligence concentration until automated manufacturing potentially resolves future scarcity.

What does the next training paradigm look like?

Current AI labs are betting on scaling reinforcement learning to achieve AGI, though the field faces significant stagnation in "computer use" tasks due to the lack of deterministic, replayable web simulators. While proponents argue that extended context windows can substitute for weight updates, critics point to performance degradation in long-horizon scenarios and the inefficiency of discarding inference data without feedback loops. To overcome these barriers, researchers are exploring On-Policy Self-Distillation and simulated "dreaming" to accumulate tacit knowledge from real-world deployment, aiming to shift future progress from pre-training toward continuous, weight-based learning.

The data black hole at the center of AI

The event analyzes the prevailing AI paradigm where massive data volume and compute-intensive reinforcement learning drive progress rather than sample efficiency, creating a booming market for human expert labeling. This approach contrasts sharply with human learning capabilities, as current models require millions of times more data to master tasks like driving or robotics, yet still achieve rapid open-source convergence by leveraging public data. Looking ahead, the discussion projects that while white-collar roles will expand due to AI complementing human work, the ultimate path to solving efficiency bottlenecks may lie in automating the AI research process itself.

What are we scaling?

Toby Ord, Beren Millidge

Baron Millage argues that current Reinforcement Learning strategies rely on inefficiently pre-baking skills into models due to a fundamental misunderstanding of their ability to learn like humans, which keeps AI revenue far below the potential of knowledge work automation. While the industry anticipates a 2030 surge in continual learning revenue reaching the hundreds of billions, the lack of generalizable on-the-job capabilities and the immense compute requirements for RL scaling suggest AGI remains distant despite incremental progress. This perspective challenges the "superhuman researcher" narrative by emphasizing that solving the core learning problem requires a shift from specialized training loops to systems capable of semantic, self-directed adaptation.

Some thoughts on the Sutton interview

Richard Sutton argues that current AI paradigms are inefficient because they rely on finite human data for static training rather than enabling continual, on-the-fly learning like biological systems. In response, the speaker contends that while human data acts as a necessary transitional "fossil fuel," it complements rather than opposes reinforcement learning and already facilitates world-model capabilities. Although Sutton correctly identifies current gaps in sample efficiency, the speaker predicts that while immediate successors remain LLM-based, future architectures will inevitably evolve to satisfy Sutton's vision of autonomous, continuous learning.

Why I don’t think AGI is right around the corner

Dwarkesh

A July 2025 analysis challenges industry forecasts by arguing that current large language models cannot replace white-collar workers due to a fundamental lack of continual learning and context accumulation. While dismissing the immediate arrival of autonomous computer agents, the speaker projects that end-to-end tax filing capabilities will emerge by 2028 and human-level on-the-job learning will arrive around 2032, contingent on shifting from data scaling to algorithmic breakthroughs. The presentation frames these timelines as probabilistic bets, warning that post-2030 progress will rely on overcoming physical constraints to enable a gradual intelligence explosion rather than an immediate singularity.

What will automated firms look like?

Peter Salaba

The event outlines a paradigm where artificial general intelligence transforms corporate structures into infinitely copyable digital populations that bypass human hiring bottlenecks and scale knowledge transmission through direct latent communication. By converting capital directly into compute to sustain millions of specialized entities, future AI firms achieve unprecedented evolvability and innovation rates comparable to biological leaps from prokaryotic to eukaryotic cells. Demonstrating these capabilities, the production itself was generated entirely using Google's Veo2 model to visualize complex concepts like an AGI hive mind, while the analysis posits that market feedback remains the critical anchor preventing such software-like corporations from drifting into self-referential irrelevance.

Notes on China

Dwarkesh

A two-week investigative trip across major Chinese cities examined the nation's unique state-subsidized economic model, stark urban infrastructure, and the fragmented sentiments of its youth and public. While observers noted a complex societal landscape marked by high youth stress, limited Western-style free speech, and a capital-constrained AI sector, they also discovered a disconnect between official narratives and local realities regarding political criticism and minority relations. The findings suggest that direct travel yields new strategic questions but that assessing the true state of the AI race and war risks requires direct access to elite decision-makers rather than surface-level observation.

Satya Nadella shows me the first Majorana 1 Quantum Computing chip

Satya Nadella

Microsoft has achieved a pivotal physics breakthrough by fabricating Majorana zero modes, creating the stable topological qubits necessary to build utility-scale quantum computers. This foundation supports the upcoming "Majorana 1" chip, which aims to host a million physical qubits capable of scaling to thousands of logical, error-corrected units by 2029. The initiative integrates quantum simulation with AI-driven emulation to accelerate discovery in chemistry and materials science while targeting a strategic shift from classical high-performance computing for specific data-light tasks.

The Limits of American Power – Sarah Paine

Sarah Paine

The speaker argues that U.S. policy in 1940s China failed due to a catastrophic misallocation of resources compared to the Marshall Plan, exacerbated by the absence of indigenous institutions required to sustain foreign aid or enforce decommunization. Historical warnings from competent Foreign Service officers regarding the infeasibility of a Nationalist-Communist coalition were suppressed during McCarthy-era purges, preventing the U.S. from recognizing that Chiang Kai-shek lacked the necessary peasant support to defeat the Communists. Ultimately, the intervention is presented as a strategic error driven by the inability to compel primary adversaries to share power and the domestic political constraints of an isolationist American public unwilling to fund an indefinite civil war.

"The Brilliance of Communism" – Sarah Paine

Sarah Paine

This analysis examines the mechanisms of Communist power consolidation through Mao Zedong's purges of rivals like Peng Dehuai and Liu Shaoqi, which triggered the Cultural Revolution to preempt posthumous criticism. The speaker details how Xi Jinping's reverence for Mao and economic re-centralization stem from his traumatic upbringing during that period and a psychological drive to emulate Stalin and Hitler. While these regimes are credited with seizing power but failing to deliver prosperity, the narrative argues that modern public adherence to such leaders persists due to a collective national tendency to avoid confronting historical crimes similar to the U.S. legacy of slavery.

Was The War Against Japan Avoidable? - Sarah Paine

Sarah Paine

Historian discussions analyze how the Smoot-Hawley Tariff fostered global economic instability that enabled militarist factions in Japan, led by War Minister Tojo, to assassinate pro-peace figures and drive nations into World War II. The dialogue examines the United States' strategic dilemma in responding to the oil embargo, weighing the moral necessity of denying resources to an aggressor against the risk of triggering a broader conflict or inadvertently aiding a potential Nazi victory in Eurasia. Ultimately, the analysis concludes that Japan's modern post-war strength was a contingent outcome rather than an inevitable result, highlighting the extreme difficulty of predicting how specific policy choices might have altered the war's trajectory.

Why Space Elevators Can't Mine Black Holes – Adam Brown

Adam Brown

Physicists have determined that while black holes offer a theoretically perfect mass-to-energy conversion, mechanical mining of solar-mass black holes is physically impossible because even carbon nanotubes lack the tensile strength required to extract Hawking radiation at a useful rate. Instead, efficient energy extraction requires a civilization to ingest baryonic matter into a small, controlled black hole, a process that bypasses the conservation laws limiting chemical and nuclear reactions to yield near 100% efficiency. This approach transforms the black hole into an ideal power plant by converting protons and neutrons entirely into radiation, provided the civilization can prevent the structure from growing and capture all emitted particles including elusive gravitons.

How Far Are We From An AI Einstein? - Adam Brown

Adam Brown

In a discussion regarding the future of artificial intelligence, a speaker predicts that Large Language Models will likely achieve the ability to derive General Relativity from Newtonian physics within a decade, marking a potential terminal milestone for human intellectual discovery. This projection is supported by private evaluations from a Stanford professor who observed LLMs rapidly advancing from scoring zero on graduate-level General Relativity exams to essentially acing them in just three years, rendering the assessment obsolete. Despite this exponential growth in solving existing problems, the speaker notes a persistent gap between a model's capacity to translate knowledge and its ability to make the intuitive conceptual leaps characteristic of historic figures like Einstein.