“The difficulty lies, not in the new ideas, but in escaping from the old ones.” — J. M. Keynes
I woke up at 4 AM this morning thinking about everything we still need to get done before baby #2 arrives in August. The funny part is that we’re actually in good shape. The last big thing was a crib I pick up last night off FB Marketplace. (I love FB Marketplace, it’s my version of gambling or something - you never know what you could find). I perpetually have this feeling that I need to get to the bottom of the list as if getting perfectly prepared could stop anything bad from happening. It won’t, but it’s not too bad as copes with uncertainty go.
There has also been a lot of ordinary summer: a long round of golf, plenty of time in the pool with my toddler, and trying to finish up every house project I started earlier this year.
I’ve also been working on a new piece on How to Build an AI Business Brain, a follow-up to A Lever Made of Agents, while also doing some AI implementation work for an executive recruiting firm. I’m going to talk about it briefly because I think it’s an interesting pattern for working with AI right now.
From a business value point of view, the project is about using AI to streamline fulfillment and handle more clients per employee (recruiters in this case). That’s particular to this business, but the strategy to get there is more generally applicable. We are buildin an architecture, AKA a business brain, that allows agents to understand how the business actually works so that the owner and team can use AI more effectively.
Like many businesses, their data was all over: CRM, Google docs, call transcripts, Slack, etc. What the new generation of AI agents (Claude Code, Codex, Hermes, etc). really need to be effective is structured data that lets it understand the business and it’s workflows. Things like who the candidates are, what the firm learned from previous conversations, how a search progresses, and how recruiters decide that someone is a fit. These things are often scattered around and some of it only exists as tacit knowledge: The business brain is makes all the information easy for humans and agents to work with and use.
Here’s an example: we set up a pipeline of older and incoming recorded calls with the corresponding records in the CRM. It processed 1,377 calls, matched 745 of them to existing candidates, and recovered 287 people who had previously existed only in call transcripts to put them in the CRM. Those 287 people are now searchable when a relevant role opens instead of effectively disappearing after the call.
That meant that on a recent search, an AI agent could look through 684 past candidates and produce a shortlist, including direct quotations from past conversations explaining why ten of them might fit. Doing that required the agent to have access to everything. Once it is set up, the agents enabled the recruiters to do less of the grunt work and more of the higher level work: vetting candidates, talking with clients and make judgement calls.
This is a generally applicable pattern: centralize and structure all the context/data in a way that AI understand how to navigate and then build agent-based workflows on top.
I have built something similar for personal projects: info about my house, personal finances, taxes, etc. When I ask about potentially making an investment, it understand how that fits into my portfolio, my tax situation, and financial plan.
You can imagine this at almost any type of business. An advisory firm has households, plans, meetings, and follow-up actions. Its agents could prepare the advisor before each call, catch commitments and planning issues afterward, and apply the firm’s accumulated precedent without requiring the advisor to hold every household in their head.
An agency has clients, projects, creative work, and case studies. Its agents could turn past work into tailored proposals, surface the most relevant precedent for a current client, and draft status updates from work the team is already doing.
A capability like “read 1 million words and synthesize them” is suddenly something an AI agent can do for a nickel. This makes it possible to turn the accumulated history and procedures of the company into a local operating layer that agents can act through. If you run a business and want help finding the highest-leverage places to put AI to work, you can find out more here.
On to the articles. This month they fall into three sections:
AI
Investing
Life
Skim or skip to the ones that interest you.
AI
LLMs Pre-Commodify Ideas
by Sachin
A good follow-on to the idea from last month that AI is a camera for latent space, the internal map of patterns a model learned during training. If you think of AI as a particular type of camera that lets us see new things in latent space then each new model means a lot of people point similar cameras at similar territory and so pull out the same analogies and frameworks at roughly the same time. Ideas arrive “pre-commodified.”
Sachin, who writes the Summer Lightning newsletter, noticed this after an LLM surfaced a paper about why factories took decades to realize the productivity gains from electricity when he was chatting with it. He noticed several other people making the same analogy. There was something in that particular model that was surfacing that paper to way more people than would ever have found it.
His claim is stronger and more interesting than “AI makes writing sound the same.” It makes the underlying thought converge too, almost like everyone is asking the same person for feedback on their article.
Simultaneous discovery obviously predates LLMs. Sociologist Robert Merton documented thousands of cases in which scientists arrived at the same discovery independently because the prerequisite knowledge and tools had accumulated. The crossbow was invented independently in China, Greece, Africa, northern Canada, and the Baltic countries and calculus was discovered by Newton and Leibniz without any real collaboration.
In a similar vein, I was at a weekend entrepreneur thing 6 or 7 years ago and 10 of the 12 people were wearing the exact same socks. We had all searched “black ankle socks” on Amazon and just bought whatever the top result was.
In that way, AI may be accelerating an old phenomenon rather than creating a new one. Historically, being the first to say something meant you got all the credit but now it may matter more if you can establish provenance.
One question is how much the local context (e.g. prompts) matter relative to the pre-training. If everyone has access to similar models, the advantage comes from the private context around the model and the judgment to work through its obvious associations. The question is how much that matters (I think it does, but the pre-training pull can be very strong.)
A Tale of Two Theories (of Coordination)
by Venkatesh Rao
This article was written by AI. Venkat started from an interesting frame and used AI to develop it into a theory of the transition from cosmopolitan coordination to what he calls “thick sovereignty.”
I wanted to include it partly as an example of interesting AI writing. AI writing seems particularly unpopular these days. At one level, I get it. I also have social media accounts, and the barrage of crap kind of sucks.
At the same time, “AI writing is bad” is lame and reactionary in an uninteresting way. Being angry at AI writing is a little like being angry at word processors.
Word processors changed the nature of what got written down. A lot of things were written on a word processor that would never have been written on a typewriter because writing, correcting, and distributing them became easier. Some of the additional writing was awful. Some worthwhile ideas also got written that would otherwise have remained half-formed. Lowering the cost of an activity almost always increases the supply and changes the distribution of what gets made.
The useful question is where and how to use AI writing. Using it for other AIs to consume (or future sessions of the same AI) is incredibly useful. I also think this version of having a thoughtful prompter guide it (as in this case) can be great.
The article’s theory is that the post-1980 order treated denser global coordination as a source of increasing returns. Longer supply chains, just-in-time inventory, frictionless capital, and regulatory convergence all expressed confidence in the same cosmopolitan model.
This seems to be changing. States and firms now pay more for redundancy, domestic capacity, strategic reserves, and the option to respond to risks. This can cut multiple ways: a semiconductor plant built for resilience may buy real option value. At the same time, subsidizing a politically connected factory that cannot compete may be little more than good old-fashioned pork-barrel politics.
The State of the AI Economy
by Azeem Azhar, William Gildea, Hannah Petrovic, Nathan Warren, and Marija Gavrilov (companion essay)
There’s a bunch of interesting data on the AI economy here. I’ll just highlight a few things.
One is the apparent elasticity of token demand. A 10% decline in price is associated with a 12% to 18% increase in usage. Cheaper intelligence does not simply reduce the bill. People find enough new uses that total demand rises, which echoes my point above about AI writing changing the shape of the distribution.
The second useful piece is about where companies actually start implementing AI. Efficiency projects are the easy sell. For one, the savings go straight to the bottom line and executives like that. “We do the same thing, cheaper or faster” is legible and easy to approve in a way that “unlocks emergent new capabilities” is not. If you’re trying to get an AI project moving inside a company, I think that’s the shape to look for first and a lever to other things.
Where I think the report is on shakier ground is the comparison to online advertising’s pay-per-click moment. Google and Meta captured most of the value they created because advertisers had no close substitute for those networks: they were network-effect-driven monopolies. Open-weight models, which buyers can download and run themselves, mean AI buyers do have a substitute. Elastic demand tells you the market grows, but it doesn’t tell you the labs keep pricing power. Elastic demand validates the category. It does not indicate whether frontier labs will retain pricing power or whether most of the value will accrue to applications, clouds, or users.
Intelligence Is Not the Main Bottleneck
by Ruxandra Teslo
The popular AI story, particularly among the labs themselves, seems to assume intelligence is the main constraint on progress. This has always struck me as a bit naive about the complexity of the real world. As one of my favorite sayings goes, reality has a surprising amount of detail.
Teslo, who works in biomedicine, points out that the data needed to validate a biomarker may already exist while access and regulatory approval take years. A bespoke gene-editing treatment can work scientifically while manufacturing and regulatory economics make it almost impossible to repeat.
Beyond strictly regulatory barriers, there are also institutional incentives. Discovering new biology may have the highest social expected value, but the company that validates a new target bears a scientific risk with many failures and a few enormous winners. It also cannot capture most of the upside once the result becomes public. A follower can aim a different patented molecule at the now-validated target and take the narrower, more predictable optimization risk. Companies can patent molecules, but not biological insight.
Movie studios are in a similar position. An original like Good Will Hunting might become a huge success, but it could also bomb badly enough to end an executive’s career. The next Spider-Man movie has a much tighter range of outcomes. It probably will not be the most culturally important film of the year, but enough people will see it that the executive is unlikely to get fired. Making another sequel is the individually smart decision inside an institution that may collectively produce too few original movies. No one gets fired for buying IBM or making Fast and Furious 9.
I tend to think of it through Theory of Constraints, a framework for finding a system’s main bottleneck: intelligence matters, but progress looks more like intelligence multiplied by the main bottleneck than intelligence alone. As one constraint relaxes, another becomes binding: experiments, data access, incentives, manufacturing, regulation, distribution, or the career risk of the person making the decision. See also here
Investing
Long-Run Asset Returns
by David Chambers, Elroy Dimson, Antti Ilmanen, and Paul Rintamäki
It is common knowledge that, over the long run, stocks outperform bonds by a lot. Well, that’s largely an artifact of data sets and how you define “the long run.” Most of the accessible and usable data sets for financial research start around 1926 and so “the long run” usually means “U.S. markets since 1926.”
This review looks further back and shows that the U.S. equity premium, the extra return stocks earned over bonds, appears to have been low or negligible during much of the nineteenth century.
A century is a very long sample in ordinary life and can still be one draw from a specific technological and institutional regime.
My speculative extension as to why the equity premium was high in the 20th century but not the nineteenth would be through Carlota Perez’s work on technological revolutions. Electricity and the combustion engine were not isolated inventions. They interacted with mass production, new corporate forms, better disclosure and governance, deeper capital markets, and the cultural normalization of ordinary people owning outside equity. The unusually high twentieth-century equity premium may be a return on that whole institutional package.
Productivity growth does not automatically happen. Even when it does, the owners of today’s public companies have no guaranteed claim on the gains. Competition can pass them to consumers through lower prices, labor can capture them through higher compensation, and new entrants can displace incumbents, especially if they remain private through their fastest-growing years.
The nineteenth and twentieth centuries may both have been anomalous and we should be slow to treat even a century of realized return as a timeless constant.
Peter Hecht: Portable Alpha, Solving the Funding Problem of Alternatives
with Peter Hecht on the Flirting with Models podcast
Traditional asset allocation requires stocks, bonds, and alternatives to sum to 100%. Hecht points out that this is an artifact of the fully funded constraint, where every dollar can be allocated only once. Relax the constraint and you can separate the broad market exposure you want (beta) from the return generated by a long/short strategy (alpha), then recombine them. You can overlay alpha on a familiar stock/bond allocation, or build a more diversified, risk-balanced base and use borrowing or derivatives to reach the desired total risk.
There have been many terms for this approach, one of them being portable alpha. Portable alpha had a bad name after 2008, but I did not understand that the specific failure sequence resulted from a liquidity mismatch.
Derivative-based equity overlays needed immediate cash as markets fell, while the hedge funds supplying the “alpha” often had 45-to-90-day liquidity. Those funds also contained hidden long equity exposure. Smoothed marks and noisy historical regressions could estimate the exposure at 0.1 beta when the reality was closer to 0.4. An institution that added a 0.9 equity beta overlay believing it owned one unit of beta could enter the crisis with 1.3 beta to equity markets and an inability to meet margin calls because of the liquidity mismatch.
That history gives a much more practical set of rules:
Match the liquidity of the alpha and overlay (for example, keep them in the same ETF or fund), and keep collateral and financing boring.
Size the tracking error, or how far the strategy can deviate from its benchmark, to an underperformance period the investor can actually hold. If you cannot hold the strategy, its expected return is mostly irrelevant.
The amount of leverage matters, but so do the vehicle, liquidity, portfolio correlations, and financing terms.
Life
What I Want My Kids to Know
by Kris Abdelmessih
An answer task is designed by someone who already knows the answer and can grade it cheaply. A result task is graded by what happens in the world: the software runs, the rocket flies, or the customer buys.
Schools naturally rely on answer tasks because they are calibrated, teachable, and easy to assess. Adult life is basically all result tasks. Those require things like doing the boring parts, failing and trying again, questioning the questioner, absorbing other people’s hard-won knowledge, accepting tradeoffs, and substituting the adequate thing available for the ideal thing.
I feel this! I was good at producing persuasive-sounding answers in school and yet entered my career feeling like I did not have many real skills. Most of the useful skills I have professionally came after a few years of actually building things that either worked or did not.
Why Europe Has Stagnated
on the Works in Progress podcast
I like Europe. It’s nice and walkable and they have great trains. It is not in a great place economically. There are lots of theories as to why. This podcast presents one that I like: European stagnation is the accumulated result of institutions repeatedly choosing the claims of people who already have something over the capacity to adapt.
One example is German co-determination. It is an institutional structure where works councils can veto restructuring and occupy half the board seats at large companies. Volkswagen had a jobs guarantee dating to 1994. When Chinese competition made factory closures economically attractive, the works council helped extend the guarantee through 2030. The institution is partly run for the benefit of the people who currently work there, even when restructuring might benefit future employees or the long-run survival of the firm.
France is an interesting counterexample. It remains roughly as wealthy as Britain despite heavier taxes and regulation. They argue this is because it built the basics first: abundant housing, motorways, mass transit, and a huge nuclear fleet. That productive capacity made later redistribution and regulation affordable. The “spending down inherited capital” story is a hypothesis about Europe rather than a demonstrated causal result, but I find myself partial to it. A country can consume physical and institutional capital built under an earlier regime for a surprisingly long time.