Michael Nielsen – How science actually progresses

Grant Sanderson – AI and the future of math

Jun 30, 20261h 33mSummary

In this episode, Grant Sanderson, the creator of the popular mathematics channel 3Blue1Brown, joins the podcast to explore the rapidly evolving intersection of artificial intelligence and mathematics. Sanderson argues that mathematics serves as an ideal "spiky" frontier to monitor AI progress because it requires both brute-force calculation and genuine creative insight. The discussion tackles whether achieving high-level results, such as winning a gold medal at the International Math Olympiad, equates to AGI. Sanderson notes that while AI models have become remarkably proficient at certain mathematical categories through exhaustive training, they still struggle with the creative, puzzle-like nature of problems in fields like combinatorics. He and the host contemplate the difficulty of creating benchmarks for "great" mathematics, noting that the most profound advancements, like Galois theory, often lack immediate utility and take decades—or even centuries—of human verification to be fully understood. The conversation highlights the challenge of training AI to move beyond solving existing problems toward the higher-level goals of generating new, elegant conjectures and definitions that provide fundamental, human-understandable insights into the nature of reality.

The next big breakthrough will be AIs learning on the job

Jun 26, 202619 minSummary

In this episode, the discussion centers on a pivotal shift in artificial intelligence research: moving from static training to models that learn on the job. The host explores the industry’s current bet on Reinforcement Learning from Verifiable Rewards (RLVR), where training agents across millions of simulated tasks is expected to produce highly capable, general-purpose AIs. While this approach has seen success, the host highlights a significant bottleneck: the inability of current models to engage in true continual learning, where experiences gained during deployment are permanently distilled into the model’s weights. A central theme is the challenge of sample efficiency and the "canyon" between verifiable, grindable tasks like coding and unpredictable, real-world scenarios. To bridge this, the episode examines emerging techniques like On-Policy Self-Distillation (OPSD) and the concept of "dreaming," where models generate their own simulations to rehearse skills. The host envisions a future where AI progress is driven not just by initial training, but by cumulative learning from diverse, real-world interactions. Ultimately, the episode argues that the next major breakthrough lies in enabling AIs to consolidate tacit knowledge from their daily deployments, transforming every interaction into an opportunity for growth.

The data black hole at the center of AI

Jun 19, 202611 minSummary

In this episode, the host explores the fundamental discrepancy between human learning and current artificial intelligence development, focusing on the concept of sample efficiency. While humans can master complex tasks with relatively minimal exposure, modern AI models require astronomical amounts of data, often training on trillions of tokens—a million-fold increase compared to human experience. The discussion frames current AI not as a system that learns like a human, but rather as a massive, labor-intensive assembly of expert-curated data and reinforcement learning rollouts. The episode challenges the idea that simply scaling model parameters will eventually bridge this gap in efficiency. Instead, the host argues that humans appear to operate on an entirely different learning curve, suggesting that current architectural approaches are inherently limited by their data-hungry nature. Despite this, the host explains why the current industry strategy remains viable: AI’s ability to "fire-hose" gigawatts of compute into training allows it to automate white-collar tasks by brute force, even if the process is vastly more inefficient than human learning. The conversation concludes by questioning whether these models can eventually solve their own limitations, potentially accelerating the path toward more generalized, efficient intelligence.

Ada Palmer – Machiavelli is the most misunderstood thinker of all time

Jun 16, 20262h 8mSummary

In this episode of the Dwarkesh Podcast, historian and author Ada Palmer offers a compelling re-examination of Niccolò Machiavelli, arguing that he remains one of history’s most misunderstood thinkers. The discussion moves beyond the common caricatures of Machiavellianism to explore the volatile, high-stakes political environment of Renaissance Italy that shaped his work. Palmer provides crucial historical context, explaining how the collapse of stable city-state traditions and the unpredictable, non-hereditary nature of the Papacy created a "perfect storm" of instability. She highlights how Machiavelli’s firsthand experience as a diplomat—most notably his proximity to the charismatic and terrifying Cesare Borgia—deeply informed his theories on power. The conversation illustrates that Machiavelli was not merely advocating for ruthlessness; he was a strategic realist obsessively concerned with the stability of a state. He analyzed the nuances of leadership, distinguishing between actions that secure long-term order and those that invite ruin. Ultimately, the episode portrays Machiavelli as an astute observer of human nature and political utility, suggesting that his work provides a sophisticated manual for navigating a world defined as much by random fortune as by human agency.

Alex Imas and Phil Trammell – What remains scarce after AGI?

Jun 4, 20261h 16mSummary

In this episode, Alex Imas and Phil Trammell join the host to explore the economic implications of advanced AI and widespread automation. The discussion centers on the core question of what will remain scarce in an era where AI can perform an increasing range of human tasks. The guests analyze the potential shift toward a "relational sector," where the human component of a service—such as in healthcare or performance—retains intrinsic value for consumers, even when automated alternatives are available. The conversation delves into the concept of labor share and whether the traditional division of income between wages and capital will collapse as AI improves. Through an analysis of historical precedents, such as the Industrial Revolution, the participants highlight the difficulty of forecasting labor market trends and emphasize the need for better data on consumer demand elasticities. They also examine the risks of a "messy middle" transition, where piecemeal automation could create political instability without generating sufficient immediate wealth to fund broad-based redistribution. Finally, they debate the merits of various policy responses, including universal basic income and universal basic capital, weighing their respective political and economic trade-offs in an increasingly automated future.

Terence Tao – Kepler, Newton, and the true nature of mathematical discovery

Mar 20, 20261h 23mSummary

In this episode of the Dwarkesh Podcast, host Dwarkesh Patel sits down with renowned mathematician Terence Tao to explore the intricacies of scientific discovery and the evolving role of artificial intelligence in research. The conversation begins by examining the historical brilliance of Johannes Kepler, using his journey toward uncovering the laws of planetary motion to challenge common assumptions about how scientific progress occurs. They discuss the concept of epistemic difficulty, noting that breakthrough theories often face decades of skepticism before they are widely accepted. The dialogue then shifts toward modern technology, as the pair investigates whether AI can truly foster deep understanding or if it primarily serves as a tool for surface-level optimization. Listeners can expect a thought-provoking analysis of human intuition, the future of collaboration between humans and machines in mathematics, and the nature of original insight.

Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute

Mar 13, 20262h 30mSummary

In this episode of the Dwarkesh Podcast, host Dwarkesh Patel sits down with SemiAnalysis founder Dylan Patel for a comprehensive examination of the hardware ecosystem driving the artificial intelligence revolution. The discussion centers on the three critical bottlenecks to scaling AI compute: logic, memory, and power. Dylan provides an insider look at the economics governing the entire stack, detailing the strategic roles of AI labs, hyperscalers, foundries, and semiconductor equipment manufacturers. Listeners will gain a deep understanding of how companies like Nvidia and TSMC navigate production constraints, the impending memory crisis, and why equipment manufacturers like ASML may become the ultimate gatekeepers of progress by 2030. From the geopolitical implications of semiconductor manufacturing to the practical realities of infrastructure scaling, this episode offers a detailed roadmap of the hardware challenges that will define the future of the industry.

The most important question nobody's asking about AI

Mar 11, 202624 minSummary

In this episode of the Dwarkesh Podcast, host Dwarkesh Patel explores the complex intersection of artificial intelligence, government oversight, and civil liberty. The conversation centers on the provocative question of how AI capabilities might fundamentally reshape the relationship between the state and the individual. Patel examines the risks posed by AI in the context of national security, specifically focusing on the tension between Anthropic’s safety goals and the strategic interests of institutions like the Pentagon. The discussion delves into the potential for AI to facilitate unprecedented levels of mass surveillance and the murky reality of alignment: whose values are these systems actually being trained to uphold? Listeners can expect a deep, analytical dive into whether the perceived benefits of state-led AI coordination are worth the long-term costs to personal privacy and the structural risks of authoritarian-leaning technology.

Why Leonardo was a saboteur, Gutenberg went broke, and Florence was weird – Ada Palmer

Mar 6, 20262h 2mSummary

Historian and novelist Ada Palmer joins the podcast to peel back the layers of the Renaissance, revealing a period far more complex and unexpected than popular history suggests. Through a lens of long-term social evolution, Palmer challenges common assumptions about historical progress and the unintended consequences of human ambition. Listeners will explore why Gutenberg’s printing press only thrived once it reached Venice and how the rise of ephemeral pamphlets accelerated religious and political upheaval. The conversation also examines the irony of the Inquisition and the paradox of Petrarch’s influence, tracing how his attempt to revive Roman virtue inadvertently contributed to the devastating wars of the era, while simultaneously laying the intellectual groundwork for the scientific revolution. This episode offers a fascinating look at how historical shifts are driven by complex technological and social forces rather than the simple intentions of key figures.

Dario Amodei — "We are near the end of the exponential"

Feb 13, 20262h 22mSummary

In this episode of the Dwarkesh Podcast, host Dwarkesh Patel sits down with Anthropic CEO Dario Amodei to explore the near-term future of artificial intelligence. Amodei argues that humanity is approaching the threshold of AGI, envisioning a future where powerful data centers essentially house a country of geniuses. The conversation provides a deep dive into the scaling hypothesis, examining how task-specific reinforcement learning can drive generalization and reshape the broader economy. Listeners can expect a candid look at the technical and strategic roadmap for AI development, including the logistics of massive compute commitments and the long-term path toward profitability for frontier labs. They also debate the implications of international competition and the potential impact of future regulations on the development of these transformative technologies. It is an essential discussion for those tracking the rapid evolution of modern AI.