Gonzo in the Valley of Heart’s Delight

Trip 8 to 9 September 2025 · Written up 9 September 2025 · Last expanded 29 September 2026

My first Imagination in Action, at Google’s Moffett Place in Sunnyvale. Another gonzo adventure in the Valley of Heart’s Delight.

One piece of advice before anything else: do not try to attend two conferences at the same time. It is a lot of notes to take and considerably more to sieve through afterwards (I am probably going to do it again anyway, so do as I say, not as I do).

Teller on moonshots, and where I part company

Astro Teller gave a fun talk on how Google X separates moonshots from other Alphabet projects, the claim being that the more imperative a project is to succeed, the worse it performs when given arbitrary resources.

I am not sure I buy it. The Manhattan Project is an obvious counterexample, though you could argue its real limiting factor was time and not money.

Maybe what Teller actually means is that important projects should not automatically be staffed with the most expensive senior talent, since labour is usually the largest cost. Or perhaps time, framed as a resource, makes his dictum trivially true.

Either way it raises a question I keep coming back to: do constraints sharpen innovation, or are there contexts, nation-state initiatives among them, where throwing resources at a problem is the only path? Is achieving 10x harder or easier as you move up in scale? My answer is that it depends on the time horizon, and that the horizon itself has to be scaled appropriately when we are judging things a whole society undertakes.

One wonders what a Teller versus Teller debate (Edward against Astro) would look like.

Preterm birth from routine ultrasounds

Weeks later, heads-down on my MVPs, one lightning talk kept coming back to me while I edited my notes: Robert Bunn of Ultrasound AI, titled Do You Have the Audacity to Prove the Experts Wrong?

Complications from preterm birth are the leading cause of death among children under five worldwide, so the stakes are about as high as they get.

In the PAIR study, run by the University of Kentucky with Ultrasound AI, a model trained on routine obstetric ultrasounds predicted days-to-delivery with strong fit; early versions reported R² near 0.90 for term births, improved on retraining, and held across trimesters. These are routine scans, with nothing exotic in the acquisition, which is the impressive part.

It resonated with a session I had missed a year earlier, thanks to Memorial Drive traffic, and watched afterwards: François Charton and Geordie Williamson at the CMSA on how mathematics and machine-learning collaborations work in practice, including the ones that fail. It is worth your time if you have not seen it.

Hostage exchange as AI policy

One panel dug into whether open-weight and open-source releases, and perhaps open datasets, can serve as confidence-building measures for AI safety and policy against the sovereign-AI tensions between the US and China, and third or non-state actors.

A metaphor I have been chewing on since is the Æsir and Vanir truce from Norse myth, where two rival tribes of gods end a ruinous war and exchange hostages to stabilise relations. It is an imperfect analogy, and in the myth the exchange nearly collapsed (the Vanir beheaded Mímir on suspicion of bad faith), but they did iterate back toward truce in the game-theoretic way. As a provocation it asks whether shared standards, shared evaluations and selective openness could move us off a purely adversarial footing.

Who I met

• Astro Teller, Google X. The moonshot framework, and the resource claim I still do not fully accept.
• Robert Bunn, Ultrasound AI. The preterm-birth prediction work, and the best-argued lightning talk of the event.
• The open-weights panel: Dinesh Maheshwari (Groq), Karl Zhao (DeepSeek), Charles Fan (MemVerge) and Jose M. Plehn (BrightQuery), with MacKenzie Sigalos on stage moderating.

Photographs

Links

The Charton and Williamson CMSA talk, the PAIR study, and my longer Day 1 writeup. To be added.

Open entry
This page is not finished, and is not meant to be. I add photographs, people and links as I dig them out, so it keeps growing after publication. Why the whole site works this way is on the colophon.