AI infrastructure is colliding with physical constraints, and this episode starts by mapping the limits: power, chips, construction, copper, materials science, and the scale problem behind modern data centers.
From there, the conversation turns to production AI systems and the security implications of agents, then to recursive self-improvement, long-horizon reinforcement learning, scientific intuition, and the challenge of keeping humans meaningfully in the loop as models become more capable.
Show Notes
Prakash Narayanan and Nathan Labenz open on the real bottlenecks behind AI data centers, including power, chips, copper, construction, and the 100-gigawatt problem. Malte Ubl joins to discuss Vercel AI Gateway, production fallbacks, agent security, and AI code review, followed by Louis Kirsch and Damon Falck on Faraday, recursive self-improvement, reward hacking, and how humans can verify AI discoveries.
Chapters
(0:00) China may not be compute-starved.
(1:51) Sandboxes aren’t inherently safe.
(4:26) Science needs wrong answers.
(5:09) Who pays when AI misbehaves?
(6:29) Opening and Ox Alpha
(7:38) Ox Alpha revealed
(8:01) China’s AI infrastructure
(12:19) YMTC and NAND memory
(14:07) Apple, YMTC, and Micron
(15:01) Companies rivaling states
(17:07) Market denial strategy
(20:04) China’s regulatory model
(21:55) Federal land infrastructure
(23:51) Alaska data centers
(26:38) Stranded gas to compute
(29:04) The 100-gigawatt problem
(30:48) Copper and future tech
(33:20) AI and material science
(34:38) Faster physics simulations
(37:38) Closing question
(37:48) Malte Ubl and Vercel
(39:10) Self-driving infrastructure
(39:20) AI decisions in production
(43:01) Eve for common agents
(46:25) Normalizing model providers
(48:35) AI Gateway economics
(59:33) Automatic provider fallbacks
(1:00:57) AI security becomes urgent
(1:01:51) Why AI attacks succeed
(1:04:26) DeepSec and code scanning
(1:06:57) Rerunning AI code review
(1:08:43) AI regulation and responsibility
(1:09:57) Provider responsibility and KYC
(1:11:03) Vercel Sandbox challenge
(1:14:50) AI model attack timelines
(1:16:44) Experimental agent harnesses
(1:21:40) Introducing Faraday and Inherent
(1:24:32) Recursive self-improving organizations
(1:28:13) Faraday’s self-improvement loops
(1:30:57) Separating scientist and coder
(1:34:34) Why science differs from prediction
(1:37:49) Training with uncertain rewards
(1:40:33) Cheating and reward hacking
(1:44:15) Human control and AI scientists
(1:47:31) Scientific intuition and taste
(1:50:46) Meta-reinforcement learning
(1:53:11) Multimodal scientific models
(1:55:18) Faraday beyond orchestration
(1:56:50) Measuring recursive improvement
(2:02:22) AI agents and workplace context
(2:05:09) AI infrastructure bottlenecks
(2:12:17) Verifying AI discoveries
(2:14:20) AI company culture
(2:15:21) AI labs and organizational culture
(2:17:58) Founders, liquidity, and risk
(2:21:57) AI wealth changes culture
(2:28:35) Animal welfare and communication
(2:33:29) AI superpersuasion politics
(2:34:51) Privacy-preserving AI research
(2:38:39) Punishing AI agents
(2:43:47) Math versus empirical science
(2:52:50) AI persuasion reality
(2:56:25) AI creativity and music
(2:58:39) The AI treadmill
Guests:
Louis Kirsch and Damon Falck — Co-Founder and Chief Superintelligence Officer (Louis), Member of Technical Staff (Damon), Inherent Laboratories (𝕏)
Malte Ubl — CTO, Vercel (𝕏 | LinkedIn)
