I was pleased to be invited back to talk on Rick Torseth’s ‘10,000 Swamp Leaders‘ podcast to discuss AI and some of its implications, especially from a leadership perspective.

Rick’s intro notes: Mick Yates makes a return visit to the swamp. This time he brings his deep knowledge and experience working inside the Artificial Intelligence movement. Mick has been a long-time student and teacher of AI. In our conversation, we discuss some of the many opportunities AI is bringing to our lives and work. But he doesn’t stop there. He shares his concerns about the power of AI to dominate our lives. This portion of the discussion also addresses ethics and the challenges that are bound to arise when so much AI power and control reside in a small group of wealthy individuals.

In the conversation I try to keep things practical. We start with what today’s AI actually is – why a large language model is essentially a very large prediction machine, and what changes once you give it agency and it can go off and do things on your behalf. That distinction matters more than most of the headlines: the Internet never woke up on a Monday morning and did something on its own, and an AI agent can. Rick asked me to explain a little more what agents are and what they can do.

From there we look at where AI genuinely helps on bounded, measurable problems (like drug discovery), and where it is much harder going, which is the unbounded, conceptual, wicked problems most leaders actually face. I make the case that leaders cannot outsource everything to the IT department, that the regulatory questions are closer to nuclear energy than to the issues with the web, and that far too much of the value of the technology is currently accruing to far too few people.

The title points to a thread running under the whole conversation: today’s AI is disembodied. LLMs have read about the world without ever having been in it. I make the point with the recent eclipse, where a pub full of people explored what was happening, happily chatted about it and compared notes. It was a physical event with social implications. This is a different kind of knowing from anything in LLM training data, rather as my grandchildren learned about wasps by being stung and not by reading about them. We then briefly discussed ‘World Models’ as next generation AIs, hence the title.

I also got to talk a little about what I have been building – The Studio, The Gallery and The Library – and about one rule that I keep coming back to: the apparatus cannot mark its own homework.

Finally, Rick asks the awkward questions, including a couple he got Claude to write for him, which was a fair ‘gotcha’ about whether I am being in some way ‘assimilated’ into my own system. I conceded the point, and said why I think ‘Rule Δ₁’ – a quality protocol that my system uses against marking itself – is one defence.

I end where I would want others to start: please go and find out for yourselves, and tell me where I have got it wrong.

You can find the podcast here, or on any of your favourite channels (e.g. Apple Podcasts).