EPISODE 2026-09-21

When Safety Tests Fail

AI safety audits, alignment funding, and AI shopping agents with Max Nadeau on how institutions adapt as capabilities accelerate.

▶ Full show on YouTube

Max Nadeau joins to discuss why AI safety audits can become box-checking, how funders evaluate uncertain alignment work, and where the real bottlenecks sit as capabilities accelerate.

The rundown

  1. 18:33Opening56 min
    OpeningAI capability progress is accelerating faster than researchers and safety teams expected. This opening explores finite-choice guardrails, multi-agent systems, emergent coordination, alignment measurements, the disruption of mathematical research, and what happens when AI development moves faster than institutions can adapt.
  2. 1:14:49Interview49 min
    Max NadeauMax NadeauAI safety audits may miss the risks that matter most if auditors lack access, independence, or incentives to investigate deeply. Max Nadeau of Coefficient Giving explains why technical safety research resists simple metrics, how funders evaluate speculative alignment work, and why talent—not money—can be the key bottleneck.
    Open segment on YouTube ↗
    • Tailwind's Grants Reach $200 Million

      0:00 / 0:00
    • The Weird Bet On Alignment Theory

      0:00 / 0:00
    • Talent, Not Money, Is The Bottleneck

      0:00 / 0:00
    • AI Could Compress A Decade

      0:00 / 0:00
  3. 2:03:20Closing27 min
    ClosingAI shopping agents are turning the customer relationship into the central battleground of online commerce. This closing discussion examines Amazon’s response to agentic shopping, the risks of banning autonomous agents, and why websites may need to rebuild their infrastructure around agents as first-class users.

In this episode

Prakash Narayanan and Nathan Labenz open with the limits of AI safety testing, including finite-choice guardrails, fast classifiers, emergent multi-agent behavior, and why alignment measurement gets harder as systems become more capable. Max Nadeau of Coefficient Giving then explains why audits can turn into box-checking, how funders assess uncertain technical safety work, and why talent may matter more than budget. The episode closes with AI shopping agents, Amazon’s response, and the question of who owns the customer relationship when agents become first-class users.

  • AI capability progress is accelerating faster than researchers and safety teams expected. This opening explores finite-choice guardrails, multi-agent systems, emergent coordination, alignment measurements, the disruption of mathematical research, and what happens when AI development moves faster than institutions can adapt.
  • AI safety audits may miss the risks that matter most if auditors lack access, independence, or incentives to investigate deeply. Max Nadeau of Coefficient Giving explains why technical safety research resists simple metrics, how funders evaluate speculative alignment work, and why talent—not money—can be the key bottleneck.
  • AI shopping agents are turning the customer relationship into the central battleground of online commerce. This closing discussion examines Amazon’s response to agentic shopping, the risks of banning autonomous agents, and why websites may need to rebuild their infrastructure around agents as first-class users.