In this episode, host Stephen Bartlett interviews Daniel Kokotajlo, a prominent AI forecaster and former member of OpenAI's technical staff. Kokotajlo discusses his background in forecasting AI timelines, his role at OpenAI, and his subsequent resignation over safety concerns, which involved a highly publicized refusal to sign a $2 million non-disparagement agreement. The conversation centers on the existential risks of artificial superintelligence, the competitive dynamics driving tech CEOs, and the societal implications of rapid automation.

The discussion contrasts the alarming default trajectory outlined in Kokotajlo's previous works, including his highly accurate 2021 report What 2026 Looks Like and the dramatic timeline of AI 2027, with a newly proposed regulatory framework. This alternative path, detailed in AI 2040 Plan A, advocates for international treaties, research transparency, and economic safety nets. The episode also touches on cognitive optimization, referencing Matthew Walker's seminal work Why We Sleep Unlocking the Power of Sleep and Dreams to frame the importance of decision-making under high stakes.

The High Stakes of Sleep and Cognitive Performance

Host Stephen Bartlett opens by discussing how sleep is the single most critical investment for maintaining high performance and avoiding poor decision-making.

As detailed in Matthew Walker's Why We Sleep Unlocking the Power of Sleep and Dreams, consistently tired individuals suffer from degraded strategic thinking and cascading life failures.

Inside OpenAI: Resignation and the $2 Million Clause

Daniel Kokotajlo joined OpenAI in 2022 to work on forecasting and dangerous capability evaluations, measuring cyber abilities, persuasion, and situational awareness.

"I basically told my wife, like, let's not have any more kids. It's too uncertain." - Daniel, explaining how his collapsing AI timelines personally impacted his family planning.

Daniel resigned in 2024 after realizing OpenAI would not pause to ensure safety: "I became gradually disillusioned with how the company was going to behave." - Daniel.

Daniel refused to sign a lifetime non-disparagement clause, initially forfeiting $2 million in equity—representing 80% of his net worth—before public pressure forced OpenAI to backtrack.

The Race to Superintelligence and Dictatorial Power

Leading tech CEOs are locked in a high-stakes race, driven by mutual distrust and the fear that competitors will achieve artificial general intelligence first.

"These powerful CEOs, Dario or Sam or Elon, are racing each other... and are literally afraid that if the other guy gets there first, he might become dictator." - Daniel.

Daniel notes that Anthropic grew from $1 billion to $60 billion in annual revenue in a single year, marking what could be the fastest corporate growth in history.

Modern AI systems are artificial neural networks with up to 10 trillion parameters that learn through pruning and reinforcement, similar to how human toddlers develop.

The Default Path: AI 2027 and the Threat of Extinction

Daniel's early forecasting report, What 2026 Looks Like, proved remarkably accurate and gained attention from high-profile political figures like US Vice President JD Vance.

In the scenario paper AI 2027, Daniel outlines a default trajectory where AI labs automate their own coding and research processes, triggering an uncontrollable intelligence explosion.

"The scary open secrets in the AI industry right now is that it's possible that we'll end up essentially creating a new species that ends up ruling the world" - Daniel.

Daniel estimates a 70% chance that the default path of AI development "goes horribly wrong, like human extinction," due to the inherent difficulty of aligning superintelligent systems.

Plan A: A Regulatory Framework for a Safer Future

To counter the default race, Daniel's organization proposed AI 2040 Plan A, a policy roadmap designed to delay superintelligence until 2040 to manage existential risks.

Plan A advocates for domestic regulations and international treaties, including physical inspections of data centers in both the US and China to verify compliance.

Daniel proposes total research transparency, forcing companies to publish training data and model architectures to eliminate monopolies and foster open-science safety research.

To address massive job displacement, AI 2040 Plan A introduces a "Citizen's Dividend" starting at $25,000 annually and scaling up to $10 million as the automated economy expands.

"If we don't build powerful AI systems eventually, then we're probably going to die as a civilization eventually" - Daniel, explaining why he ultimately would not permanently ban AI development.