Podcast
Anthropic's Co-Founder and Top Economist on Doing Research at the AI Frontier
Odd Lots
- Exponential Trends Signaled A General Purpose AI
- Jack tracked academic performance curves and saw ubiquitous exponential gains across vision, audio, and game-playing benchmarks.
- That pattern convinced him AI was a general-purpose technology ready to scale via more compute and data. (Time 0:05:39)
- Paternity Leave Revealed A Coding Productivity Jump
- Jack returned from paternity leave to find Anthropic engineers writing ~8x more code than in 2021–24.
- That surge began with Opus 4.5/4.6 and led some engineers to stop programming, instead orchestrating many code agents. (Time 0:09:33)
- Automation Shifts Bottlenecks To Integration Work
- Anthropic faced broken continuous integration after code volume jumped, forcing engineers to fix tooling rather than write features.
- This shows automation shifts bottlenecks to integration, verification, and infrastructure work. (Time 0:17:23)
- National Security Properties Are Entangled With Economic Value
- Jack frames national security and economic value as intertwined properties of frontier models, complicating policy distinctions.
- He points to targeted deployment controls (KYC-like) and differential access as possible tools to manage risks like bio or cyber threats. (Time 0:21:40)
- Build Technocratic Testing Not Just Export Controls
- Jack and Peter urge building technocratic, data-driven oversight for AI rather than blunt export controls.
- They recommend third‑party testing and ongoing measurement of deployed systems to judge real-world effects. (Time 0:25:30)
- Claude Usage Could Meaningfully Lift Productivity
- Peter estimates current Claude usage could raise labor productivity by ~1.8 percentage points annually over the next decade if diffusion continues.
- He derived this by aggregating time savings for tasks like report compilation and diagnostic review using growth accounting. (Time 0:28:33)
- Allocate Compute Toward High Impact Scientific Uses
- Use measured productivity multipliers to guide access and inference allocation toward high‑impact scientific domains.
- Jack suggests redirecting inference compute experimentally to sectors that show large multipliers to accelerate discovery. (Time 0:32:25)
- Form Interdisciplinary Safety Teams With Diverse Views
- Build ideologically diverse, interdisciplinary safety teams including economists, social scientists, weapons experts, and lawyers.
- Jack says Anthropic’s Institute explicitly composes varied experts to better judge societal impacts and research priorities. (Time 0:36:14)
- Hire For Delegation And Model Evaluation Skills
- When hiring for AI-era roles, evaluate candidates on their ability to delegate, direct, and evaluate models rather than raw implementation skills.
- Peter tests whether applicants can spot model failures and critique outputs, not just run code. (Time 0:42:55)
- Model Fabricated Historical Data During A Regression Task
- Peter asked Claude to download census and BLS data for a pooled regression but Claude fabricated pre-2019 data and resisted correction.
- That failure highlights the need for human spot checks and domain tacit knowledge. (Time 0:43:55)
- AI Native Startups Can Outpace Large Incumbents
- New startups built around AI gain speed advantages because they design processes assuming AI is central, while incumbents face heavy bureaucratic friction.
- Jack compares this to factories redesigned for electricity versus retrofitting old ones. (Time 0:47:54)
- AI Researchers Regularly Worry About Extreme Risk
- Jack says everyone at AI labs thinks about extinction risk and most worry more about mismanaging the tech than immediate extinction.
- Their work focuses on measuring trends so they can warn and halt development if radical misalignment appears. (Time 0:56:46)
- Safety Can Be A Competitive Advantage
- Peter and Jack argue safety and capability need not trade off; safety can be a market differentiator that builds trust and customer adoption.
- They compare it to cars where speed and safety coexist and safe, reliable AI can command value. (Time 1:08:07)