Thoughts in Between

by Matt Clifford

TiB 200: Imaginary books; AI alignment; crypto and modernity; 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.

A step forward in AI alignment?

This week OpenAI published a new paper and an accompanying blog post on the topic of AI alignment, which - as we discussed last week - is an important and difficult problem. The way they describe the challenge is as follows:

GPT-3 [see previous TiB coverage] is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants. In other words, these models aren’t aligned with their users

Their solution is to have human readers rank multiple potential answers written by the AI in response to user questions. The model then uses this feedback to fine tune and update. The results seem impressive: the updated model (called "InstructGPT") gives better answers that are less toxic and more truthful.

Of course, this in no way means that the alignment problem is solved, but it does seem a promising step forward (even Eliezer Yudkowsky thinks so, sort of). As the OpenAI team notes, InstructGPT still produces toxic or questionable content (an amusing example) and alignment to the average human provider of feedback may not be the right standard.

Perhaps more fundamentally, there's a risk that we just redefine what we mean by alignment. As this thread notes, given the enormous power AI models may have in future, what we want is not that a model gives the output (some) humans prefer, but the ability to formally verify that the model will or will not produce certain outcomes (This paper from Scott Niekum, the thread's author, has more). That's a much harder task, and there's a long way to go, but it's still encouraging to see top AI labs integrate safety and alignment work directly into their flagship products.

What books should exist, but don't?

Stripe Press last week announced a new editorial team led by Tamara Winter, who seems to have a lot of great ideas for new volumes. Winter kicked off her tenure with this fantastic thread, which solicits ideas for "books you'd like to exist". The replies are excellent; do take a look. I particularly liked this one on the leading AI labs and this one on financial regulation, plus this whole post by Jamie Rumbelow. I thought it was a good opportunity to list a few books I'd love to exist too (and if they do already exist, let me know!):

A history of ambition - basically a longer and much better researched version of what I outlined here. There's so much to cover here: institutions, the genesis of career paths, attitudes, social and legal regulation, immigration, and much more.

Talent, power and emerging technologies - we've talked before about the power of the tiny group of people with advanced skills in AI and how their political preferences spill into the realm of geopolitics. How did this compare with nuclear physicists in the Cold War? And people at the forefront of previous strategic technologies? Some of this is covered in Jade Leung's excellent PhD thesis (see also our podcast discussion), but it's worth a book dedicated to it.

A children's book on ranking animals! As a child, I was obsessed with questions like, "Who would win in a fight, a killer whale or a great white shark?" (relatedly, from this week…) and, now, having two young boys, it's clear that the market for such questions is not just me... My four year old (and I) would love a detailed, opinionated, well-reasoned canter through maybe 100 such questions (Maybe not one for Stripe Press, I admit)

If you have ideas of your own, email me or add them here.

Crypto, modernity and Edmund Burke

"Small-l" liberalism's disastrous 2016 on both sides of the Atlantic prompted much discussion about whether liberal democracy would survive and, if not, what would come next. The 2020 US election briefly made the topic less salient, but recently there's been a flurry of pieces on "post-liberalism" (such as this week's John Gray / Ross Douthat debate in the New Statesman). The best, though, is this excellent essay by Tanner Greer (see previous coverage), which has a lot to offer even if you're not interested in the underlying politics.

Conservative critics of modernity (of which, I probably don't need to tell you, I am not one!) tend to argue that somewhere along the way we took a wrong turn. Greer thinks so too, but says the post-liberals are wrong about when; it all started to go wrong, he says, not with the permissiveness of the 1960s or the anti-authority leanings of the Enlightenment, but with the rise of bureaucracy. He contrasts typical 19th century American workers, who would have "participated in a half dozen committees, chapters, societies, associations, councils, and congregations" that shaped their lives, with their ancestors today who are subject to a "web of routines, rules and procedures decided in faraway boardrooms".

I don't want to get all "Bitcoin fixes this", but... maybe it does? It's easy to ridicule (and I have) Web3's sometimes overblown rhetoric on "community", but if you wanted to create a 21st century version of Burke's "little platoons"* to give people a real stake in their lives and work, wouldn't decentralised, permissionless, self-governing communities be a good starting point? It’s an argument that has sympathisers on the left (see here) as well as the right. I expect many or even most DAOs (decentralised autonomous organisations) operating today will fail, but as a category they might point the way to a new kind of politics. For obvious reasons crypto tends to be labelled liberatian, but in reality it opens up possibilities across the political spectrum.

*Alas, Burke didn't actually say that, quite, but you know what I mean...

Quick links

  1. REPOST: What are NFTs actually good for? One of the most interesting arguments I've seen [Whoops, sorry to the hundreds of you who clicked on this last week and got an interesting but completely unrelated thread on protein folding...]
  2. Forgotten branches in the history of technology. Excellent thread of promising tech that never quite went mainstream.
  3. Good Samaritans? Fascinating chart of who the public in each country would be willing to help in a crisis (Doesn't look great for the UK!)
  4. What's the matter with Yorkshire? Great thread that examines why Sachsen (Germany) got much richer since 2000 and Yorkshire didn't (Spoiler: it's not the EU)
  5. Is this just fantasy? Nice video of Tolkien talking about (and writing) elvish. (Semi-related: why did Harry Potter get so big, so quickly?)

Happy 200th!

We have somehow got round to the 200th edition of Thoughts in Between. I would have given up by now were it not for you, so thank you for reading, replying to and sharing it. I do appreciate it.

This makes it a particularly good week to share it with a friend!

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.