
“The object is not to retain the peculiar advantages which the inequality of conditions bestows upon mankind, but to secure the new benefits which equality may supply.” — Alexis de Tocqueville, Democracy in America
Last time I wrote, I said I’d probably be at the hospital when that issue went out. I was! Baby #2 arrived at the end of August, and between the newborn and everything else, I missed the September issue but happy to be back and will be going forward. Hope your fall/spring is going well. Mostly AI in this issue with a cameo from the Austin mascot at the end.
Where’s the “Intelligence Explosion”?
by Ramez Naam
My rough guess is that AI has made me about 50% more productive, but very unevenly depending on what I am doing. I am something like 1000% more productive at coding, 50% at project management, and maybe 20% at writing. (In some cases, I suspect it has made me materially less productive because it tends to overcomplicate things.)
That came to mind reading this post by Ramez Naam, a futurist who was early to call the solar and battery revolutions. Naam is skeptical of “recursive self-improvement,” the idea that once AI can do AI research, each generation will build a better successor faster than the last until capabilities explode.
His case is empirical - it just doesn’t seem to be happening. The popular benchmarks (standardized tests for AI models) imply that models can reliably handle coding tasks that take a person about four hours. With the exception of well specified and repetitive tasks (e.g. reformat 100 blog posts in the same way), leaving an agent running for more than about 20 minutes usually ends up in a mess for me.
In fact, this seems to be the case for the frontier labs as well: when OpenAI published data on how its models perform on its own research work, they said that they usually ran for about 15 minutes rather than four hours.
Where he does expect superintelligence is in narrow domains where answers can be checked automatically: chess, Go, formal math, parts of coding. Machines can generate unlimited practice problems there and verify every answer without waiting on people or the physical world.
That matches my experience. Coding is where my productivity jumped most, and it’s also the most machine verifiable thing I do. Most of reality is not so easily verifiable. Indeed, it would seem reality has a surprising amount of detail that standardized benchmarks lack.
Cognitive Mercantilism
by Venkatesh Rao
Before capitalism, the dominant economic idea in Europe was mercantilism: wealth was a fixed stock of land and gold, one country’s gain was another’s loss, and ordinary people were labor to be managed rather than customers with wants of their own.
This article argues we are in a period of cognitive mercantilism in which we view intelligence now the same way humans viewed wealth pre-capitalism. It is treated as scarce, something to hoard, and dangerous in the wrong hands.
An analogy of the AI fear to the Industrial revolution in this view would be: farming will be too productive, everyone will be unemployed, and what will all these dangerous peasants do with their free time?!?! Adam Smith must be stopped!
Indeed, farming did get much more productive and most people stopped farming and they went on to do other things. This was actually pretty great!
One historical fact I didn’t know was that even before factories, ordinary English people started working more days every year so they could buy goods like tea, sugar, tobacco, and cotton cloth that had been luxuries.
The AI analogy here: If intelligence becomes sufficiently cheap, how many uses can we find for it that we wouldn’t have used in the past?
A personal example I’ve given before: when I replaced my 27-year-old HVAC system, I had an AI agent read the contractor quotes, find the actual equipment models and manuals, and help me work out what to buy. A few years ago I would have picked a quote and hoped for the best because the cognitive effort to read a bunch of HVAC manuals simply wasn’t worth it. When that cost fell sufficiently, it suddenly was.
Of Termites & Tokens
by Tom Critchlow
How does customer research work today? A company might periodically interview customers, summarize what it learned in a deck, and use that to make some decisions about launching a new product or repositioning one. Then the research is done until someone has time and budget to do it again.
Many projects in companies work like this. They are “initiatives” that start and stop.
One way to think about how the cheapening of intelligence may impact companies is that lots of discrete project may turn into continuous flows.
If the work becomes cheap enough, you (or the agent) just keep doing it perpetually. Rather than having a discrete project where marketing talks to sales and reviews new customer conversations over the last 3 months to see what’s changing, those conversations are just continually analyzed by AI and transformed into a monthly report that marketing reviews and integrates.
Most of the conversation around AI is still very focused on efficiency - using AI to improve margins by doing the same work for less money. But this idea of using it in a more expansive way, to do work that was never worth doing before, a distinction I made in A Lever Made of Agents, is more interesting.
Teaching the Ocean to Swim
by Hollis Robbins
I’ve read a bunch about the Big Five personality traits (AKA OCEAN - openness, conscientiousness, extraversion, agreeableness, and neuroticism). It’s the most empirically validated personality profile in the psychology literature and has been pretty helpful to me for understanding myself and others.
However, it was a method developed in a world where computing was scarce. This is a sort of extension of the cognitive mercantilism point - a lot of the methods we inherited were built for a world where intelligence and computing were scarce.
Economics, for example, often leans on tidy equations where you could instead simulate millions of individual buyers and sellers interacting, what economists call agent-based models. I think to a large extent, this is just because computing was expensive when the field was started and adaptation is slow.
In some ways, the Big Five is psychology’s version of this. In 1936, two psychologists went through Webster’s dictionary and pulled out every word English uses to describe a person. They found 17,953. Over the next few decades, researchers compressed that list using factor analysis, a statistical method for boiling many correlated measures down to a few underlying ones. One researcher got sixteen factors in the 1940s. Later researchers kept finding five.
That compression made sense with 1940s statistics, but ignores quite a bit. For example, the Big Five model doesn’t account for context at all. Everyone has a different personality in different contexts, a point psychologist Walter Mischel made in 1968, when he found that trait scores didn’t predict how people would act in specific situations.
In fairness to the Big Five, a crude summary can still predict things: conscientiousness predicts job performance across occupations in a pretty robust and repeatable way. A credit score throws away almost everything about your finances, but is still a helpful compression. I think the more interesting point is that these compressions are a result of insufficiently cheap computing and intelligence, a condition which seems to be rapidly disappearing.
Strong Centers
by Andrew Conner
This post made an observation I hadn’t articulated but that resonated as soon as I read it: we’ve built a world optimized for individuals and crowds, but stripped out much of the scale in between. You can take your car from your nice house to a big stadium and watch a concert more easily than ever. A lot of the ‘middle’ though has deteriorated: pubs, coffee shops, and other places where small groups develop real gravity.
This is probably one of the most discussed things by people in my generation: a feeling of lacking communal centers. A place you go often enough to recognize people, where you can turn up and see someone you know without arranging an event.
To be fair, some of it has moved online (e.g. group chats), and I wouldn’t dismiss those relationships. But having a good group chat and having somewhere nearby to see the people in it are still different things.
Matthew McConaughey’s University of Houston Commencement Speech
Best Motivisation
As you would expect from Mr. McConaughey: this speech is a bunch of cliché but true life advice delivered with folksy charm.
There is plenty here that you’ve heard before: decide what success means to you, pay attention to what makes your life better or worse, find some enjoyment in the work itself. He is very good at saying it, though, and sometimes it just hits you at the right time, so it’s worth a watch if you’re in that state of mind.

