Approximately Correct: An AI Podcast from Amii

The episode contains several genuinely non-obvious ideas - why the IID assumption breaks in economic settings, the 'AI is not rocket science' tech-transfer critique, and the limits of in silico LLM experimentation - but the live format and conversational meandering dilute the density. Long stretches of setup and Amii-plugging dilute the payload.

“the core reason is that you don't see IID samples from one distribution for a couple of different ways... if you're building an economic system, people's behavior in that system absolutely does change in response to the system”

“AI is not rocket science because instead of making rockets, we make absolutely everything”

The 'twins separated at birth' framing for AI and economics is fresh and well-argued from first principles, and the 'AI is not rocket science' policy argument is a genuinely clever structural critique rather than a recycled talking point. The critique that LLMs may pattern-match economics questions without actually reasoning is sharp. Some segments (Canada adoption challenges, LLM evaluations) cover ground that is increasingly standard discourse.

“economics and AI are sort of two twins, separated at birth with completely different upbringings, but exactly the same genetic code”

“AI is not rocket science because instead of making rockets, we make absolutely everything”

Leyton-Brown is a legitimate senior researcher - Royal Society fellow, CIFAR Chair, textbook co-author on multi-agent systems - who has deployed a real system transacting tens of millions of dollars in agricultural commodities, not merely theorized. His weakness for a B2B audience is that he is a pure academic practitioner rather than an operator who has scaled a company, limiting direct commercial transferability.

“it transacted, um, tens of millions of dollars US in agricultural commodities in Uganda”

“I was part of a group of people at UBC who wrote a policy piece for CD Howe last year”

There are some concrete anchors - Canada's productivity growth being 15x less than the US, 80% subsistence agriculture in Uganda, tens of millions transacted, the SMS-based market design - but many claims remain at a moderate level of abstraction without citations, timelines, or deeper quantification. The policy proposal section is largely conceptual.

“our rate of productivity growth...has been 15 times less than the U.S.”

“80% of, uh, the economy in Uganda, at least at the time that I was working there, like a decade ago, um, was subsistence agriculture”

The hosts ask functional follow-up questions and land one genuine pushback ('But to me that sounds like reinforcement learning'), but most questions are open invitations rather than sharp probes. Claims about productivity statistics, policy proposals, and the Uganda project's ultimate failure go largely unchallenged, and a tangent about Amii self-promotion eats several minutes of a short episode.

“But to me that sounds like reinforcement learning”

“But it's sort of inescapable. I mean, that's where we're going, right?”