Thoughts in Between

by Matt Clifford

TiB 203: AI and fusion; what ARIA needs; Arbesman on science; and more...

Welcome new readers! Thoughts in Between is a newsletter (mainly) about how technology is changing politics, culture and society and how we might do it better.

It goes out every Tuesday to thousands of entrepreneurs, investors, policy makers, politicians and others. It’s free.

Forwarded this email? Subscribe here. Enjoy this email? Forward it to a friend.

Can AI accelerate practical nuclear fusion?

As we discussed in TiB 198, if you had to pick one technological breakthrough to usher in a world of abundance, you might choose vastly smarter artificial intelligence. And if you had to pick two, you might pick AI plus “nearly free” energy. One possibility is that the former ushers in the latter. Last week DeepMind published a paper in Nature, plus a blog post explainer, on using deep reinforcement learning to control plasma in a tokamak - a key component of some approaches to nuclear fusion. This thread by one of the researchers is a good summary.

If you're new to fusion and want a sense of why this is a big deal, I recommend this short piece and this video. In short - while practical fusion power remains a long way off (despite other recent advances) - this work could make new experiments in fusion much cheaper, faster and more flexible. One exciting thing about DeepMind's approach is that it demonstrates the power of simulation for training AIs in environments where getting hold of a lot of data is hard or costly. DeepMind's model was trained extensively in a simulated tokamak before being transferred to the (very expensive) real thing. Related techniques are used by a number of fusion companies; see Azeem Azhar's recent interview with Nicholas Hawker of First Light Fusion for more on the value of simulation.

I'm particularly interested in how advances like DeepMind's might unlock new innovation on the hardware side, as this tweet suggests:

@pfau I assume the current hardware design was with the expectation of a classical controller. Do you think RL controllers can unlock new tokamak design ideas, say, much more granular control of the magnetic field or asymmetric plasma chambers?

Ever since reading Sara Hooker's excellent paper, "The Hardware Lottery" (which we discussed in TiB 133) I've been interested in how hardware-software co-evolution means we can get stuck in suboptimal equilibria. That is, the hardware we have today is rarely the best possible - it's just the best for the path that software has taken. If we suddenly find a radically better solution in software, it can unlock radically better solutions in hardware (or vice versa). In the context of fusion, that might take decades... but it would be worth the wait.

Can DARPA work without the "D"?

We've talked about ARIA - the UK's new DARPA-like research agency - several times (see, e.g., TiB 153100 and 126). As I noted last week, the announcement of Peter Highnam as its founding CEO seems like very good news. Highnam is current Deputy Director of DARPA, a former DARPA programme manager and a former director of IARPA (DARPA's intelligence-focused counterpart), so it's hard to imagine a more credible figure or one more steeped in understanding how and why DARPA has been successful. Nevertheless, there are a number of structural differences between DARPA and the proposed ARIA that will require some important choices to be made to set ARIA up for success.

First, ARIA is small compared to DARPA, roughly the size of one DARPA technical office (DARPA has six, plus two for special projects). So the first big question is, should ARIA specialise? I think the answer is probably yes - in which case picking a territory where the UK has a clear competitive advantage is key. Second, the "D" in ARPA has been foundational to its success, but ARIA doesn't have a defence focus. This has two important implications. Most importantly, the DOD has been crucial to ensuring that technologies developed by DARPA are "pulled through" into useable products and services (William Bonvillian talks about an "island/bridge" model, with DARPA as the island, DOD as the bridge to real-world impact). What will be ARIA's bridge? Getting this right is critical.

Equally, DARPA's national security mission has been crucial to insulating it from political intereference. This is essential when the timelines for breakthrough research success are longer than a parliamentary term. Finding other ways to give ARIA deep bipartisan support may not be easy, but will be important if it is to have a real chance to succeed (Evidence from DARPA's history suggests that the other key ingredient of "effective political design" is ensuring that its budget is not perceived to threaten the budgets of other agencies). If ARIA finds good answers to these foundational questions, there's lots to be optimistic about.

Sam Arbesman on fixing science

This week's TiB podcast episode is a conversation with Sam Arbesman. Sam is a complexity scientist and writer, who is currently scientist-in-residence at Lux Capital, one of the world's leading "deep tech" venture capital firms. I first connected with Sam because we share an interest in institutional innovation in science. Sam is the curator of the Overedge Catalog, which collates and explores such models, so this was a natural starting point for our conversation.

I asked Sam why he thinks there's been what he calls a "Cambrian explosion" of new models for science in recent years (see, e.g., TiB 134165 and 153). Sam's starting point is that many activities (such as multidisciplinary research or writing software) are important for the progress of science, but are not valued or incentivised in traditional academia. This leaves a gap. I agree - but I'm also interested what you might call the supply side: a growing share of the world's biggest philanthropists made their wealth by innovating in science and technology and these people tend to be more interested in how to use science to improve itself than in borrowing prestige from established institutions (we discussed this in TiB 156 in the context of crypto).

Our conversation covers this and a number of other topics that will be familiar to long-time readers, including:


Quick links

  1. Thoughts and prayers. How much would you pay to (not) be prayed for? Interesting study.
  2. Incomplete. Great story about Gödel and the US constitution.
  3. Age shall not weary them. Amazing thread of modern (AI-generated) renderings of historical figures. Stuff like this can seem trivial, but I find it's an unreasonably effective way to remind yourself that the past was just as real as the present.
  4. No truck. The best broadly pro-Trudeau piece. The best definitively anti one. Where do I come out? This is surely the most important point and, indeed, possibly the most foundational political principle of all.
  5. No way in. Fun piece on locked room mysteries, for which long-time readers will remember I am a sucker (which seems as good a time as any to plug the the fact that I've written two immersive murder mystery games, available to download free of charge)

What do you think?

Thanks for reading Thoughts in Between. If you enjoy it, please do forward it to a friend or two. Or share the link on social media. I do appreciate it.

Feel free, too, to hit reply if you have comments, questions or suggestions.

Until next week,

Matt Clifford

PS: Lots of newsletters get stuck in Gmail’s Promotions tab. If you find it in there, please help train the algorithm by dragging it to Primary. It makes a big difference.