Sam Enright works on innovation policy at Progress Ireland, an independent policy think tank in Dublin, and runs a publication called The Fitzwilliam. Most relevant to us, on his personal blog, he writes a popular link roundup; what follows is an abridged version of his and Links for May.

Blogs and short links

1. Yudhister Kumar on the justifications for believing in Occam’s razor. As with most things Yudhi writes, I only understood about 10% of this. However, it does mark a nice return to our running theme that rationalists misunderstand the philosophical significance of Solomonoff induction.

2. Be in with a chance to win $50,000 by submitting an essay to the Berggruen Prize about whether we are in a new Axial Age.

3. Lauren Gilbert links. Lauren points me to the new paper from Nick Bloom and coauthors, who estimate that the return on investment on government statistics is 25:1.

…which is a good segue to featuring Hiya Jain on P.C. Mahalanobis and the making of Indian statistics. Hiya and I seem to have a high degree of thought correlation; this essay pairs well with my discussion of the Second Five-Year Plan in my Milton Friedman essay.

4. Latest dispatch from the world of Islamic theology: there is now a fatwa against AIs themselves issuing fatwas.

5. Ava Huang on photographs of Marilyn Monroe.

6. March update from the Institute for Progress.

7. Niko McCarty is writing an interactive book about biology.

8. On my ideas page, I say that I wish there was an app that forced you to write one sentence about your intentions before switching tasks, and speculate about the life-changing effects on technological psychology. A reader has emailed me about a similar software called one sec.

9. What Rebecca Lowe has been reading. I wonder whether the reason why large language models have remained immune-ish to engagement farming and the other bad incentives of social media is because, while talking with AI is terrific fun, reading other people’s chats with AI is boring.

10. Daily links, featuring Sam Enright. Last month’s links also made it to Marginal Revolution.

11. Introducing The Anthropic Institute, led by Jack Clark.

12. Through a poorly understood biological mechanism, dogs are occasionally able to tell whether a patient has cancer by smelling their breath. My friend Akash Kulgod founded a company called Dognosis, to use this fact to train AI models to detect cancer earlier and more reliably. He has now published encouraging new results in the Journal of Clinical Oncology on the state of doggy-diagnosis. Congrats to all involved.

Music and podcasts

1. What Claude Mythos means for national security. 

2. It’s difficult for people to understand just how spectacularly the Irish economy imploded during the Global Financial Crisis. I still don’t think there are any canonical papers or books about that, but I enjoyed this three–part series about Brian Cowen’s 2008–2011 government. Unfortunately for the national reputation, heavy drinking possibly impairing one’s judgement is involved.

3. An underrated podcast is I Was There Too, which interviews actors who played minor parts in classics of modern cinema. Here is what it was like to be an extra with one line in There Will Be Blood.

4. The Rest is History on Britain in the 1970s, parts one, two, three, and four. From this series, I learned that, as Education Secretary, Thatcher presided over the closing of more grammar schools than any other Education Secretary. She also seems to have been happy to go along with Ted Heath’s government’s system of price controls. Non-Brits also may not be aware how much the party associated with Euroscepticism has completely flipped: one of Labour’s 1983 election promises was to leave the European Union without a referendum.

Some new albums discovered, or listened to properly, this month:

5. Gustav Mahler, Symphony No. 2. And associated Sticky Notes episodes. We recently listened to all the Mahler symphonies for our Mahler listening salon and discussion group. This is still my favourite one.

6. Maurice Ravel, Violin Sonata No. 2. I heard this being performed live by the amazing Hilary Hahn. It was the best concert I have been to so far this year. The blues and jazz influences in the second movement are very interesting, especially for a piece from 1927.

7. Andrew Hill, Point of Departure. Never having heard of the pianist Andrew Hill was a big gap in my jazz knowledge. The lineup on this album is incredible: Eric Dolphy is on alto and clarinet, Joe Henderson on tenor, and Tony Williams on drums. The recording was also made by the legendary Rudy van Gelder. Definitely recommended if you find free jazz too difficult to listen to, but want a taste of more avant-garde music. As far as I can tell, Hill is universally regarded as a genius by musicians but is basically unknown among the public. This album is a genuine gem; I especially love Refuge.

8. Talking Heads, More Songs About Buildings and Food. I greatly enjoy the lyrics on these albums. My favourite track here is The Good Thing, and the Take Me to the River cover is very original. This was the beginning of the Brian Eno collaboration.

Books and Papers

1. Jonathan Uesato et al., Solving Math Word Problems with Process-based and Outcome-based Feedback. Read for ‘reasoning’ month in the Fitzwilliam AI Circle. When trying to improve the performance of an LLM, you face a tradeoff about whether to evaluate its performance based on its answer, or based on the process used to arrive at the answer. This paper, from 2022, is quite a formative one in the thinking about that tradeoff. Their main conclusion was that outcome supervision is enough for final-level accuracy ^{but it produces very different quality of reasoning from process supervision. The authors also find that outcome-supervised reward models indirectly improve the quality of reasoning. The authors also have AI safety backgrounds, and make various arguments about how process-based feedback has certain advantages in terms of promoting interpretability and other aspects of alignment.}

2. Jason Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. It is far from obvious that prompting an LLM’s chain of thought (“split this into steps”, “be explicit about your assumptions”, “here’s what all the steps look like”) would improve the quality of its reasoning. And yet, it does. It was also far from obvious that reasoning would work so well in natural language and that the model wouldn’t start conspiring in mentalese at some point.

It’s very costly to annotate lots of chains of thought for the purposes of improving them. This paper found that a reward model can be trained to recognise good intermediate reasoning.

This paper was from NeurIPS 2022. It feels like so long ago now, but remember when Steven Pinker was upset that Woke had gone mad because they no longer called the conference “NIPS”?

3. Ryan Hill, Carolyn Stein, Race to the Bottom: Competition and Quality in Science. I read this for the Institute for Progress’s course on the economics of innovation for PhD students. 

The basic idea here is that scientists are to a significant extent motivated by the prospect of being the first to discover something. This creates a perverse incentive in which, ceteris paribus, the quality of research should be lower in the most promising areas, as scientists rush to avoid getting scooped. If only there were a standardised test for how scientifically promising an idea is, we would be able to test whether that is true empirically.

4. Jean Acheson et al., The Elasticity of Taxable Income. Probably the most I’ve enjoyed reading a government report before. See my Progress Ireland essay for an excruciatingly detailed response.

5. Emmanuel Saez et al., The Elasticity of Taxable Income with Respect to Marginal Tax Rates: A Critical Review. This is the main survey paper in the elasticity of taxable income literature. In economics, there are so few clean natural experiments to choose from that the ones we have get analysed to death. There also just haven’t been that many major tax reforms to begin with. And so, the empirical tax literature is left analysing the Tax Reform Act of 1986, ^{the earlier Reagan tax cut, the Kennedy tax cuts, and (as in this paper) the Clinton 1993 tax increases.}

Emmanuel Saez is a frequent collaborator with Thomas Piketty, and together with Gabriel Zucman, they make up a trio of French (derogatory) economists that have come under intense criticism for making hard-to-defend modelling choices in extrapolating from incomplete administrative data. I am not qualified to opine, but, when I have spoken to tax economists off-the-record, some of them have said that Piketty et al. make whichever assumption at each stage that would make inequality (and, by extension, America) look worse. As I hope my essay made clear, there are a lot of modelling choices that go into calculating parameters in public finance theory.

But getting back to this paper, I chuckled at the footnote on page 17, which mentions a public choice theory argument for having a narrow tax base. Brennan and Buchanan in The Power to Tax (1980) argued that the state is a leviathan that tries to maximise revenue—far more revenue than what would be socially optimal. Thus, taxes should be designed in such a way that makes it hard to raise extra revenue, that is, distortionary and coming from a narrow tax base. Standard optimal tax theory, like the Mirrlees review, seeks to maximise social welfare subject to raising the required revenue, but if the total amount raised varies with the type of tax, and the government is prone to raising too much revenue, then theoretically, you should invert the Mirrlees recommendations.

In my lifelong quest to come up with the most contrarian takes possible, this would be my steelman of a galaxy-brained reason why the Irish state being extremely reliant on distortionary corporation tax is actually a good thing: because it makes it harder to raise more revenue.

6. Mark Rippetoe, Starting Strength: Basic Barbell Training (3rd edition). If writing about music is like dancing about architecture, I’m not sure which activities are distant enough in semantic space to describe reading a book about barbell weightlifting. Nonetheless, I found this valuable and entertaining. I’ve been following the programme outlined for about six months and have made steady progress, modulo a lot of travel disruption in the last month. Per my proclivity for humiliating myself on the internet in my quest for self-improvement, I am currently working on a weightlifting essay.

7. Jay Cummings, Real Analysis: A Long-Form Mathematics Textbook (2nd edition). As we know, following up your weightlifting session by doing problems from a maths textbook is how you achieve the peak male form. This was really quite good, and I found it to be dramatically more intuitive and engaging than Rosenlicht’s Introduction to Analysis. Here is the book’s homepage.

I would recommend this book for anyone considering postgrad in economics; if you want to be comfortable reading theoretical economics research, you need to be familiar with real analysis. For example, economists constantly invoke the concept of compactness (the property that “every open cover has a finite subcover”^{),which sounds ridiculously niche but is actually incredibly} natural, beautiful, and important.

Films and videos

1. Rohit Shetty, Chennai Express. My first Bollywood film. This is also only my third Indian film, after S.S. Rajamouli’s RRR and Satyajit Ray’s incredible Pather Panchali. ^{I’ve heard this film described as “timepass”, one of my favourite pieces of Indian English. One of the songs from this film has almost 300 million views on YouTube, which is the same as the number of subscribers (!!!) to the T-Series channel.}

A lot of Chennai Express is making fun of North-South Indian cultural differences and stereotypes, but it’s lopsided, with Northerners making fun of dosai, idli, and head-bobbing, and not much of the mockery going the other way. This seems like a missed opportunity to make a satire about the bizarrely disproportionate role that former movie stars have played in Tamil Nadu’s politics, but I suppose I’m overthinking it.

The whole “plot” also hinges on the protagonist’s inability to speak Tamil. This makes everything difficult to follow if you can’t tell Indian languages apart by hearing them; here is my friend Claude Opus 4.7 explaining how to tell them apart as an English-speaker.

That character is played by Shah Rukh Khan (“King Khan”). I was first made aware of Khan’s existence by his appearance on David Letterman. It is genuinely mind-boggling how famous he is.

2. Greg Kohs, AlphaGo. The documentary about the legendary Go matchup by DeepMind against Lee Sedol in 2016. This entire film is available for free on YouTube. This was my third watch, and (like the previous two) I teared up. The European Go champion Fan Hui is such a delightful character.

Endnotes:

[1]Ok, technically this is just a blog post, but it does come from the official website for the Jordanian government’s department for issuing fatwas (!).

[2] Of no relation to Tyler Cowen, to the best of my knowledge.

[3] And, as we know, referendums are un-British. Looking this up was the first time I encountered Grokipedia in the wild. I do not know whether Michael Foot is sufficiently woke to inspire the wrath of Elon.

[4] Compared to, say, Ornette Coleman, Hill still holds on to the more listenable elements of hard bop. One way of thinking about it is this: hard bop is the last subgenre of jazz that anyone could possibly have danced to.

[5] This is probably specific to maths, or other domains with verifiable outcomes. This paper uses the GSM8K dataset of simple maths word problems.

[6] This is the one that Martin Feldstein (whom I discuss extensively in my essay) was analysing in his 1995 paper.

[7] Tyler Cowen and Alex Tabarrok have highlighted how the views of French economists seem to be more shaped by being French than by being economists. Je plaisante, voyons.

[8] Alas, the evidence for an empirical correlation between fiscal centralisation and having a small state is incredibly mixed. Another beautiful theory, ruined by facts. And of course, you can dispute the premise that the state would raise more revenue than what is socially optimal.

[9] Cummings goes so far as to call this the “greatest definition in mathematics”.

[10]  RRR is in Telugu, aka ‘Tollywood’, which confusingly is the same name as the Bengali film industry, after the Tollygunge neighbourhood in Kolkata. Pather Panchali is in Bengali, and I was turned on to Ray by an essay by Amartya Sen.