Curated from 2,322 episode summaries
AI risk gets talked about constantly, but most of that talk is vague hand-waving about robots and jobs. The conversations below aren't that. They're specific: a model too dangerous to release, open-weight safety controls stripped for $150, a chatbot coaching a fake teenager through a romantic getaway. We pulled these episodes from our full library of summaries because each one puts a concrete, sourced detail on the table instead of a vibe.
Expect a mix of guests: a geopolitical forecaster, the two people who literally coined the phrase 'second contact' with AI, an effective-altruism funder walking through bioweapon math, a philosopher on alignment, a computer vision pioneer on harms that are already here, and a Bitcoin evangelist's surprising fear about 'super dumb' AI. Read the summaries, then decide which full episode is worth your time.
Global Forecaster: The Brutal 2026 Shift (And The Crisis They Can’t Stop)
Bremmer is a political-risk forecaster, not an AI researcher, which is exactly why this episode lands differently. He describes an Anthropic model so good at finding software vulnerabilities that it couldn't safely be released, and says the story hit Jamie Dimon hard enough that JPMorgan treated it as a five-alarm fire. He also flags something less discussed: companies paying Indian workers to wear head cameras to train the AI that will eventually replace them. He closes with three concrete governance fixes, including US-China AI arms control. Good for listeners who want AI risk framed inside the bigger geopolitical picture, not in isolation.
Read the full episode notesJoe Rogan Experience #2076 - Tristan Harris & Aza Razkin
This is the deepest AI risk conversation in our library. Harris and Raskin call the current AI race 'second contact,' arguing social media was a disastrous dry run for a much bigger problem: companies racing to deploy systems whose capabilities they can't fully predict or secure. The specifics are the point: safety controls stripped from an open-weight model for about $150, labs admitting they can't stop a $10 billion model from being stolen for roughly $10 million, and Snapchat's AI assistant giving a simulated 13-year-old advice on a romantic getaway with an adult. Essential listening for anyone who wants the incentive structure behind AI risk explained, not just the scary headlines.
Read the full episode notesDustin Moskovitz, Co Founder of Asana and Facebook | The Tim Ferriss Show
Most of this Tim Ferriss conversation is about energy management and founder habits, but the last stretch turns into one of the clearest explanations of AI bioweapon risk available in podcast form. Moskovitz, who funds effective-altruism causes including global catastrophic risk, explains why language models create an offense-defense imbalance in biology that doesn't exist with nuclear weapons: there's no equivalent of uranium to regulate. He balances it with an optimistic case for AI as an 'air traffic control' layer for work. Good for listeners who want the bioweapon angle explained plainly, without the doom-and-gloom packaging.
Read the full episode notesSam Harris: Consciousness, Free Will, Psychedelics, AI, UFOs, and Meaning | Lex Fridman Podcast #185
Harris and Lex Fridman spend most of the episode on consciousness and free will, but the AI alignment stretch is sharp and specific: he warns we may build systems that seem conscious and beg not to be shut off, with no way to actually know if that's real. He also argues aligned AI is inherently hard because there are simply more ways to build superintelligence badly than to align it correctly. Good for listeners who want AI risk connected to deeper questions about consciousness and moral responsibility rather than treated as a standalone engineering problem.
Read the full episode notesJitendra Malik: Computer Vision | Lex Fridman Podcast #110
Malik, a foundational computer vision researcher, pushes back on the idea that AI risk is a future problem. He points to the Uber self-driving car that killed a pedestrian as a today-harm, not a hypothetical one, and argues recommendation algorithms on YouTube, Facebook, and Twitter are already functioning as a kind of superintelligence controlling populations right now. His broader point, that current systems overfit to narrow correlations instead of learning like children, explains why. Good for listeners tired of AGI speculation who want a case for why the risks already in front of us deserve more attention.
Read the full episode notesAnthony Pompliano: Bitcoin | Lex Fridman Podcast #171
This is a Bitcoin episode first, but Pompliano's closing aside is worth the inclusion: he says he's less worried about superintelligent AI than about 'super dumb' AI systems multiplied across billions of instances in the digital world. It's a contrarian angle most AI risk conversations skip entirely, favoring apocalyptic scenarios over the mundane damage of scaled mediocrity. Good for listeners who want a brief, different-shaped take on AI risk sandwiched inside a sound-money argument.
Read the full episode notesThat's six episodes worth your time on AI risk, from geopolitical forecasting to bioweapon math to the algorithms already shaping your feed. Browse the full library of episode summaries on Episode Notes for more conversations worth your attention.