Thoughts in Between

by Matt Clifford

TiB 221: Increasing the supply of great companies; AI-created ideas; AI dystopias; 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.

$158m to increase the supply of great founders

I try to avoid excessive “day job self-promotion” here, but as it's just over a year since the last Entrepreneur First-related post and as it touches on many TiB themes... I'm excited to announce that EF has raised $158m from a large coalition of individual and institutional investors, including Patrick and John Collison (loyal TiB readers!), Reid Hoffman, Taavet Hinrikus, Matt Mullenweg, and many more. More here. For those of you who don't know, Entrepreneur First is a new way of building technology companies. Our mission is to make great companies happen that otherwise wouldn't exist, by backing great individuals before they have a company and being the best place in the world for them to find a co-founder and get started.

There are a few TiB-aligned reasons why this is exciting, beyond being good news for EF and the founders we back. First, I've been a big advocate here of institutional and structural innovation in funding - and one of the important things about this raise is that we didn't structure it as a fund, but as a first step towards giving EF the company permanent capital (see also TiB 189). This is important and valuable because it allows us to take a very long-term perspective and, crucially, to think like a product company and innovate on the core ways we fund entrepreneurs. As I said back in TiB 176:

A more open-ended iteration of EF might be tricky to package as a traditional LP-backed fund (our current business model), but I think for less than $100m you could build a fairly scalable (and very profitable) evergreen capital version of it

Turns out more than $100m is better!

Second, as I've argued here several times, the world still allocates way too little of its best talent to entrepreneurship (No, there are not "too many entrepreneurs"). As we discussed in TiB 165 and TiB 167 there's good evidence that when extremely talented people who otherwise wouldn't start companies are compelled to do so, they actually do better than the average entrepreneurs in their ecosystem. I find this a remarkable and still underrated fact! And it's the whole EF thesis in a sentence: the world is missing out on some of its best founders - and with this new capital we're one step closer to changing that.

I should say that if you want to start a company, you can apply to Entrepreneur First here. And if you want to fund EF companies, let me know…

Can AIs generate their own ideas?

Julio Gonzalez-Diaz and Ignacio Palacios-Huerta have a fascinating new paper on whether and how AI can discover new ideas. The reason these questions are important will be familiar to long-time readers: if ideas are getting harder to find, as we've discussed, several times, then enlisting AI to help us find them may be a path out of (actual or potential) scientific and economic stagnation. The idea that AI might be able to discover new ideas is both somewhat controversial and difficult to test, so the authors focus on an usefully concrete and measurable domain: chess. They look at the impact of AlphaZero (see TiB 136) on elite human play. Did AlphaZero discover novel and useful chess ideas that the top human players were able to adopt? If so, the authors argue, that's an existence proof for AI-manufactured ideas.

The actual results are a puzzling to me. The paper finds that there is quite compelling evidence of AI-generated ideas that provide "alpha" (e.g. the early advance of the h-pawn and "innovative sacrifices in exchange for 'harmony, initiative, dynamism, aggresivity'") being adopted by a human... but for Magnus Carlsen only. Clearly Carlsen is an all-time great, but surely adopting AlphaZero's insights in this way is low-hanging fruit? Carlsen himself believes AlphaZero improved his play, but I'm struggling to understand the model whereby only he benefits. Tyler Cowen has further (mildly skeptical) commentary here.

That said, AI ideas become interesting way below this hurdle. The authors' point - which I agree is fascinating - is that AlphaZero is important because it's not trained on human data (it learned chess by playing against itself); its strategies are the Platonic ideal of AI-generated ideas. But AI could make a huge contribution to scientific progress simply by making better use of existing human knowledge. We've discussed this paper by Eric Schmidt and Maithra Raghu on the uses of machine learning in science before (TiB 114). And I'm bullish on the benefits of having AI models "read" and make connections across whole corpuses of academic literature (see TiB 72). Fortunately we don't need to be Magnus Carlsen to make use of such ideas!

Is AI inevitably communist? (No)

Peter Thiel once said, "crypto is libertarian, AI is communist". His less ideological framing, less often quoted, is that crypto is decentralising, AI is centralising. I was thinking about this again this week because of this interesting Twitter conversation between Glen Weyl, Vitalik Buterin and Henry Farrell, among others. One of the strands of the discussion is whether with sufficient amounts of computational power, we could see a much more centralised future with whole economies run by AI central planners. A related, dystopian future is sometimes posited: that a technologically advanced authoritarian regime could employ AI to exceed their 20th century totalitarian counterparts and dictate all aspects of human activity.

I think this fear is overblown. First, as we discussed back in TiB 103, when looking at this excellent paper, AI cannot rescue central planning from its flaws, because the essential problem is behavioural, not computational; that is, it's a problem of incentives and capabilities, not intractability. If you asked me to name my favourite ever blog post, it might well be "In Soviet Union, Optimization Solves You" by Cosma Shalizi, which makes a similar point from a more left-wing perspective. It's a long and brilliant argument, but suffice to say: the USSR's economic challenges can't be reduced to the need for better linear programming! (Weyl also notes the problem of communications constraints)

Second, I think the value of human freedom in generating high quality decision making - whether by AI or humans - remains underrated (and as a result the advantage of authoritarian states in AI is overrated). We discussed this in TiB 93 in the context of Farrell's essay, "Seeing Like a Finite State Machine" and the problem that autocratic regimes have in gathering reliable data. But this is a much broader topic that Farrell has been exploring for years. His paper with Shalizi on "Cognitive Democracy" goes beyond Hayek's classic argument to explore and contrast the different information properties of democracy, markets and hierarchies. (He summarises his ideas in this excellent podcast from last year: "Democracy as a Problem Solving Mechanism"). As you know, I'm not remotely complacent about the risks of AI, but I don't think inevitable authoritarian is high on the list.

Quick links

  1. The stubborn fragility of the physical? Striking video of a steel mill being destroyed by a cyber attack.
  2. A very big/small house in the country. Home size is an underrated dimension of variation across countries (Imagine the psychological effect of moving from a typical Australian home to a typical Hong Kong one).
  3. How to know if you've clicked. Interesting study looking at the relationship between conversational response time and emotional connection.
  4. Family planning! Birth month seems to matter quite a lot.
  5. "A cheap sensationalist, clumsy and vulgar". Hilarious / fascinating write-up of Nabokov's take on other writers.

We've got to stop meeting like this

As always, thanks for reading and making it all the way to the end. Shares and forwards are always appreciated!

If you have comments, feedback or suggestions, do feel free to hit reply.

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.