EPISODE 2026-08-26

Web Infrastructure and Superintelligence

AI data centers, production AI infrastructure, agent security, and recursive self-improvement with Malte Ubl, Louis Kirsch, and Damon Falck.

▶ Full show on YouTube

AI data centers, production AI infrastructure, and the path from agent reliability to recursive self-improvement.

The rundown

0:50Opening31 min
OpeningAI data centers are running into bottlenecks in power, chips, copper, and construction. Prakash Narayanan and Nathan Labenz examine China's AI capacity, YMTC and Apple, stranded gas, off-Earth compute, and the materials-science breakthroughs that could keep infrastructure scaling.
31:54Interview46 min
Malte UblMalte UblHow can AI infrastructure run itself? Vercel CTO Malte Ubl explains self-driving production systems, model-provider fallbacks, the AI Gateway, and the security work needed for agentic software. The conversation also covers AI code review, sandbox security, automated cyberattacks, and lightweight agent harnesses.
Watch
Open segment on YouTube ↗
Clips
- Everything Hackable Will Get Hacked
  0:00 / 0:00
- Offense Is Part of Defense
  0:00 / 0:00
- Rollback Is the Agent’s Superpower
  0:00 / 0:00
- Sandboxes Need More Than VMs
  0:00 / 0:00
- AI Gateway Fallbacks Buy Uptime
  0:00 / 0:00
1:18:15Interview52 min
Louis Kirsch and Damon FalckLouis Kirsch and Damon FalckHow do AI scientists improve themselves? This conversation examines Faraday, recursive self-improvement, long-horizon reinforcement learning, scientific intuition, reward hacking, and how humans can verify AI discoveries while staying in the loop. Louis Kirsch and Damon Falck explain Inherent's approach to open-ended research and human-AI teaming.
Watch
Open segment on YouTube ↗
Clips
- Recursive Self-Improvement Needs Humans
  0:00 / 0:00
- Faraday's Rare Reward-Hacking Attempts
  0:00 / 0:00
- Water-Cooler Talk Is AI's Best Data
  0:00 / 0:00
- AI Should Explain Its Proofs
  0:00 / 0:00
- The Singularity Should Not Happen Alone
  0:00 / 0:00
2:10:37Closing45 min
ClosingAI superintelligence is forcing new questions about who controls models, how agents are punished, and whether safety guardrails hold. This closing discussion covers Anthropic's privacy-preserving research access, AI persuasion, animal-welfare strategy, math essentialism, experimental science, and the future of AI agents.

In this episode

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.

AI data centers are running into bottlenecks in power, chips, copper, and construction. Prakash Narayanan and Nathan Labenz examine China's AI capacity, YMTC and Apple, stranded gas, off-Earth compute, and the materials-science breakthroughs that could keep infrastructure scaling.
How can AI infrastructure run itself? Vercel CTO Malte Ubl explains self-driving production systems, model-provider fallbacks, the AI Gateway, and the security work needed for agentic software. The conversation also covers AI code review, sandbox security, automated cyberattacks, and lightweight agent harnesses.
How do AI scientists improve themselves? This conversation examines Faraday, recursive self-improvement, long-horizon reinforcement learning, scientific intuition, reward hacking, and how humans can verify AI discoveries while staying in the loop. Louis Kirsch and Damon Falck explain Inherent's approach to open-ended research and human-AI teaming.
AI superintelligence is forcing new questions about who controls models, how agents are punished, and whether safety guardrails hold. This closing discussion covers Anthropic's privacy-preserving research access, AI persuasion, animal-welfare strategy, math essentialism, experimental science, and the future of AI agents.