TiB 132: Computers vs the human brain; video games and predicting the future; disruption disrupted; and more...

Welcome new readers! Forwarded this email? Subscribe here. Enjoy this email? Forward it to a friend.

Disruption is a favourite word in startup land. But an interesting new paper suggests that disruption has actually fallen since the beginning of the internet startup era. The authors look at “displacement” (how many big companies lose their market position) and “leapfrogging” (how many smaller companies become market leaders) and find both are in decline.

Their findings seem to rule out two popular explanations - lax anti-trust enforcement and political capture (e.g. lobbying). Instead, the authors “find that investments by dominant firms in intangibles, especially software, are distinctly associated with … reduced leapfrogging”. In other words, software investments by big companies, which increased tenfold around 2000, seem to be an economic moat (If you’re interested in the role of intangibles in the economy more broadly, Stian Westlake and Jonathan Haskel’s book on the topic is excellent).

Given popular rhetoric emanating from my industry, it might seem counterintuitive to think of technology as suppressing rather than promoting disruption. But it’s apiece with the oft-expressed concern that today’s tech giants may be more entrenched than their historical counterparts. This isn’t a straightforward anti-tech perspective. Ben Thompson - no knee-jerk tech critic - has an excellent essay from earlier this year on the idea that we shouldn’t expect Google, Amazon, et al to be disrupted. That’s likely to be at the heart of many policy conundrums over the next decade. 

When we discussed GPT-3 back in TiB 119 we noted that some smart people see a path from where we are today to artificial general intelligence (AGI) “simply” by continuing to scale existing models. If that’s right, one key requirement will be lots and lots of computational power, or “compute”, as we discussed in TiB 128.

A reasonable question is, “ok, how much compute?” Joseph Carlsmith of Open Philanthropy has an amazingly deep report on a related question - how much compute  we need to match the capabilities of the human brain? It’s very long, but I recommend at least the executive summary and, if this is your sort of thing, the longer summary blog post. The very short version is that Carlsmith estimates that we probably need no more than 10^15 FLOP/s to compute as well as the brain and that it’s unlikely that we need more than 10^21 FLOPs.

Is that a lot? Well, for a billion dollars, you can build a computer that does >10^17 FLOPs. So the likelihood, given the long-term decline in compute costs, is that we’ll exceed the power required even Carlsmith’s upper estimate in a reasonable timeframe. Of course - as Carlsmith notes - compute is a long way from being the only barrier to AGI. We don’t actually know how to build machine learning models that are anywhere close to the capabilities of a human brain in the general sense. But it’s still a striking, and perhaps sobering, datapoint. 

Civilization, the strategy video game, is 29 years old this week and its creator, Sid Meier, has released his memoirs. Civ is one of the most popular games of all time - apparently players have clocked up more hours in-game than have passed since the actual dawn of civilisation - but it’s also had an interesting impact in the real world.

Since 2011, Philip Tetlock, author of the superb book Superforecasting, has been running prediction competitions under the Good Judgment (GJ) brand, initially in collaboration with IARPA (the intelligence equivalent of DARPA). GJ asks volunteers to make predictions about highly uncertain geopolitical events. The results are striking. The best predictors - the “super forecasters” - reportedly do significantly better than professional intelligence analysts who have access to classified information.

One challenge, though, with a tournament based on real-world events is it’s very hard to evaluate counterfactual judgement. It would be useful to know who’s good at saying not only what what will happen, but also what would have happened in different circumstances. This is where Civilization comes in. We can’t “re-run” the world to test counterfactuals, but we can rerun Civ. This is just what Tetlock and IARPA have been doing with FOCUS. Tetlock talks about it here, about halfway down, and there’s a good (but paywalled) profile here. Who knew all those misspent youths could be so useful? 

I’m grateful to have such an engaged group of readers. If you enjoy TiB, I’d love you to help me grow it - the easiest way is to forward it to a friend (and the best way is to share the URL on social media).

I’m always happy to hear from readers - just hit reply or message me on Twitter.

Until next week,

Matt Clifford