Latest Interviews

AI researchers debate how close we are to recursive self-improvement

John Schulman, Charlie O’Neill, Beren Millidge

Experts predict that while AI will function as full-time remote workers within three years, exponential takeoff before 2036 remains unlikely due to persistent generalization bottlenecks and sample efficiency gaps that prevent narrow benchmarks from scaling to broad real-world domains. The prevailing consensus suggests current architectures will plateau under diminishing returns on scaling, forcing a reliance on distillation and human-defined objectives rather than recursive self-improvement to drive the projected tenfold researcher productivity gains. Consequently, a dominant "ASI" model capable of mastering every cognitive and physical field is forecast only within five to ten years, contingent on overcoming fundamental barriers in non-cumulative task learning and environment creation.

Sarah Paine - Why Putin and Xi can't escape geography

Sarah Paine, Putin, Xi

The event analyzes the fundamental geopolitical divergence between continental "elephant" powers reliant on land armies and maritime "whale" powers driven by trade and naval defense, arguing that the current global instability stems from China and Russia attempting to impose a 19th-century sphere-of-influence system. Drawing on the theories of Mackinder and Spykman, the discussion highlights how maritime democracies must leverage sanctions and economic insulation rather than direct territorial conquest to counter continental aggression that seeks to hollow out post-WWII institutions. Ultimately, the presentation warns that failing to maintain this rules-based order risks a catastrophic third world war, emphasizing that maritime strategies offer the only path toward sustained positive-sum growth.

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.

Tyler Cowen — The #1 bottleneck to AI progress is humans

Tyler Cowen, Jason Crawford, Heike Larson

Economist Tyler Cowen argues that AI-driven economic growth will be constrained by non-intelligence bottlenecks such as labor shortages and regulatory inertia, resulting in a modest 0.5% annual boost rather than explosive takeoff. He posits that future success depends on scarce bundles of talent and organizational culture rather than raw intelligence, while noting a divergence where peak human capability is rising even as the middle of the distribution deteriorates. Furthermore, Cowen anticipates the fragmentation of organized social movements like Effective Altruism and advocates for writing primarily for AI training data to secure knowledge accumulation in a shifting geopolitical landscape.

What a GPT-7 Intelligence Explosion Looks Like | Carl Shulman

Carl Shulman

The discussion outlines mitigation strategies for early detection of hostile AI motivations alongside a defined productivity threshold where AI contributions match or exceed human researcher output. It details operational mechanisms such as voting algorithms, cost-effective scaling of smaller models, and self-generated curricula that enable intelligence explosions without relying solely on brute force capability. These approaches combine hard physical constraints with empirical verification to ensure safety while accelerating innovation through distributed compute and structured learning environments.

Sarah Paine – Maritime vs Continental Powers

Sarah Paine

The speaker contrasts Russia's antiquated continental model of territorial conquest with the maritime order's "win-win" system built on commerce and international law, arguing that Vladimir Putin's rejection of integration has squandered Russia's economic potential. By framing the current conflict as a strategic timeout for Russia, the analysis highlights how the Biden administration's multilateral approach has successfully mobilized former neutral nations like Finland and Sweden to enforce an impregnable border. Ultimately, the discourse advocates for preserving the post-WWII legal framework through a collaborative, non-hegemonic system that allows for Russia's eventual reintegration provided it adheres to established rules.

Are We On Path Towards Superhuman Intelligence? – Dario Amodei (Anthropic CEO)

Dario Amodei

The speaker projects that economic investment and hardware advances will drive AI capabilities to match a generally educated human within two to three years, while acknowledging that scaling laws are currently bending to yield increasing returns. Despite this rapid acceleration, the speaker cautions against precise predictions of "superhuman" universality, noting that safety regulations and the complexity of physical embodiment may introduce significant messiness and delay. This trajectory suggests models will soon lead in scientific progress and specialized domains like math, yet the exact nature of their future impact remains distinct from traditional narratives of existential threat or total autonomy.

Will an AI smart enough to win math competitions be AGI? (Grant Sanderson @3blue1brown)

Grant Sanderson

Two speakers debate the definition of Artificial General Intelligence, rejecting the idea of a discrete threshold and instead characterizing current models like GPT-4 as capable of applying a single algorithm across diverse tasks through continuous progression. While they anticipate that AI achieving a gold medal at the International Math Olympiad would represent a creative breakthrough in logical abstraction, they argue this milestone does not equate to general intelligence or significant economic disruption, as current systems still lack the context windows necessary for deep human intent analysis. Consequently, the discussion concludes that no single benchmark, including mathematical proficiency, serves as a definitive line for AGI, with current AI impact on job automation estimated to remain below one percent.