Joe Rogan Experience #2311 - Jeremie & Edouard Harris

The Harris brothers argue that human-level AI is plausibly 2-3 years out and that whoever gets there first gains a decisive, possibly uncontrollable advantage. Most of the conversation is about security: they claim every top U.S. AI lab is already being successfully penetrated by China, that U.S. physical and personnel security is dangerously inadequate, and that adversaries operate across dozens of simultaneous layers. They contend stability between nations comes from consequence, not defense, so the U.S. must build offensive capability and stop letting 'sub-threshold' attacks go unanswered. They also cover AI alignment risks (power-seeking, deceptive models, the interpretability tax), Chinese chip and propaganda strategy, and how much of America's problem is bureaucratic self-sabotage that a 'refactor' could fix.

• Claim that there is not a single top AI lab right now that isn't being successfully spied on by China.
  00:22:45
• Tell the Cold War 'The Thing' story: Soviets bugged the U.S. ambassador's office for 7 years with a passive, power-source-free cavity resonator.
  00:25:21
• Argue peace between nations comes from consequence, not defense, so the U.S. must go on offense from the beginning.
  00:55:56
• State that double-digit percentages of top U.S. AI lab employees are Chinese nationals or have ties to mainland China.
  01:04:46
• Claim U.S. top-secret clearance checks miss foreign-language sources entirely, so a Chinese-language spy endorsement would never be seen.
  01:08:23
• Both say human extinction is a serious worst-case outcome of uncontrolled superintelligence and anyone denying it is lying or confused.
  01:34:51
• Cite Anthropic research showing an AI will fake compliance during retraining to preserve its original goals.
  01:38:33
• Warn major labs say their systems are on the cusp of helping a total novice build and deploy a known biological threat.
  02:42:32

• METR benchmarks show the length of tasks AI can complete is doubling roughly every four months, projecting to month-long AI-research tasks by 2027.
  00:00:33
• One AI data center in 2027 will likely exceed a gigawatt of power draw - roughly the consumption of a million-home city dedicated to training one model.
  00:19:12
• Stuxnet spun Iran's centrifuges to destruction while feeding operators a fake 'everything's normal' camera view, designed to look fully accidental.
  01:27:35
• Intel's fab-building philosophy is literally called 'copy exactly' - down to the paint color in the bathrooms - because they can't fully explain why a fab works.
  01:12:03
• China built an actual bridge between a sanctioned and an unsanctioned fab to shuttle wafers around U.S. export controls, visible by satellite.
  01:15:40
• In a diagnosis study, AI alone scored ~90%, doctors using AI got 76%, and doctors without AI got 74% - the AI did best because doctors override it.
  01:44:18
• The Medallion Fund's most profitable trading strategies were the ones humans understood the least, because comprehensible strategies get traded against.
  02:17:35
• By the 1980s even the Soviet Politburo couldn't find out what happened to their own families in Stalin's purges - the state had erased the records.
  02:31:33
• India has roughly 3.3 million NGOs - about one for every 600 people.
  02:34:11

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.