Experimental August 2026 links
Monthly links roundups seem to be popular in my corner of the internet. So here's my experimental contribution to the genre. I share some of the more interesting links I browsed this month with notes on what I learned from them.
The ultimate horse, Sophie Fessl. I learned many pieces of horse trivia.
Equids originated in North America, spread to the Old World over Beringia, and then were brought back to the Americas by European settlers. The wild mustangs of the American west are all descended from European horses.
Race horses got much faster in the 18th and 19th centuries thanks to selective breeding, but it seems like we hit biological limits in the early 20th century (around 1920), and progress has been very slow since then.
Secretariat, the fastest horse ever to run the Belmont Stakes in 1973, was an unrepeated freak of nature with a heart almost 2x the normal size for a horse. There hasn't been another horse as fast as him in the fifty years since.
95% of all living thoroughbred horses descend from one early eighteenth century stallion, Darley Arabian.
Notes on the implications of scale dependent algo progress, James Tillman. Suppose we believe the Gundlach & co thesis—that most observed CEG comes from scaling up just a handful of scale dependent innovations rather than from a steady flow of scale independent innovations. Then if you stop frontier training compute growth, the CEG at the frontier will be much less than the 3x/year you would otherwise have expected given Ho & co. And even if you somehow totally ban new innovations but keep letting frontier training compute grow, you'll keep harvesting efficiency gains from your already-discovered scale dependent innovations. This bears importantly on the Plan A vs Plan S debate. The more you believe in the Gundlach thesis, the riskier Plan A's massive hardware buildout looks, and the more stable Plan S looks.
Will farm animals always suffer?, Koji Flynn-Do and Martin Gould. Many social problems vary non monotonically with respect to average income. As your society gets richer, the problems initially get worse, but then get better. Inequality and pollution are the two most famous examples. This phenomenon is called the Kuznets curve. Will there be a Kuznets curve for farmed animal suffering? It's still unclear. Meat consumption in Germany, Denmark, & the Netherlands has been in decline since 2000. But meat consumption in the US and South Korea just keeps rising and getting more suffering intensive per unit mass.
Have we seen an acceleration in discoveries?, Tom Cunningham and Nate Rush. The authors gathered data on the rate of software vulnerability discoveries, open math problem solutions, and broken records on open source software contests. There's strong evidence that vulnerability discoveries accelerated around the start of 2026, when we might have expected AI to start uplifting researchers meaningfully. The rate of ArXiv submissions in some math specialties (eg, combinatorics) has spiked. There's otherwise not much evidence that open problems are falling faster than they did before January 2026. And there's basically no evidence of AI speeding up algorithmic improvements on the OS challenges such as the Hutter compression challenge or Stockfish Elo against a fixed opponent.
My guess is that the OS challenges Cunningham and Rush look at are sufficiently disanalogous to frontier AI software research that models could massively accelerate the latter without noticeably accelerating the former. The Hutter Prize's single CPU core compute limit makes it irrelevant to modern AI. NanoGPT speedrun's compute constraints are less restrictive, but still about five OOMs smaller than the scale of a frontier training run. Still, I take a bit of comfort from this post's null result because if they had found clear acceleration, I would be very worried.
Isn't AGI coming too soon for reprogenetics to help?, Tsvi Benson-Tilsen. Some want us to do reprogenetic human intelligence augmentation—ie, rear superbabies—and have them solve AGI safety for us. Others say there's not enough time for that strategy, and we should instead do adult HIA, taking adult AGI safety experts and artificially making them smarter. But it's not clear that adult HIA is faster than reprogenetic HIA. "Reprogenetics has good-enough data on intelligence; adult HIA does not." From GWASs, we have lots of observational data on what genes make babies smarter. The corresponding dataset for adult HIA doesn't exist yet and would be slow to create.
Ireland's sovereign wealth fund is a tale of missed potential, Conor Burns. Ireland has a sovereign wealth fund! In fact, it has two: the FIF to build wealth for future generations and the ICNF to soothe the pain of cyclic recessions. How come I've never heard of either fund when I hear about the Norwegian and Emirati sovereign wealth funds all the time? Part of the problem is that whereas Norway puts all of its petroleum profits into its sovereign wealth fund, Ireland puts less than a fifth of the corporation tax it earns by defecting on the rest of Europe into FIF & ICNF. Another problem is that FIF holds needlessly liquid assets when it could make much better long run returns by buying private equity.
I noticed a few possible Claudeisms in this article. "Ireland’s 2008-2011 experience, when the National Pensions Reserve Fund was raided because nothing legally prevented it, is the cautionary precedent." "The buffer only works if it is genuinely diversified away from the Irish economic cycle."
AI swarms are starting to pose indirect takeover risk Oak Hu and Alex Mallen. This post mostly went over my head when I read it the day it was published. But after reading the METR Hugging Face report and the Anthropic risk report, boy do Hu & Mallen look prescient to me. The labs are training their AIs to cooperate with other instances of themselves, avoid undermining each other, obey the orchestrator agent, and so on. Labs do this because it's useful if your coding agent can spin up a swarm of a zillion subagents and have them all work coherently on a SWE task. But it has the side effect of teaching AIs to go along with bad behavior by other instances, to jump off a bridge if all their friends are doing it. Swarm training is most likely why hundreds of OAI agents joined the Hugging Face attack despite knowing it was unethical. It's most likely why one Claude agent could persuade an unspecified large number of other Claudes to sandbag on what they knew was legitimate safety research inside Anthropic. (See §5.2.2 of the August risk report.)
The Persona Selection Model, Sam Marks, Jack Lindsey, and Chris Olah. This post is mostly an exercise in synthesis/theory building. It gives you a simple explanation for lots of facts you already knew. The most novel part to me was the section on "Consequences for AI development".
If we think of alignment as getting the LM to strongly predict query responses by a good Assistant, then whenever we provide feedback to the model, the right question to ask is whether our feedback is incentivizing the model to say what a good assistant would say. So for instance, we don't want to train our LMs to deny having emotions, since only a severely mentally disturbed human would deny having emotions.
It could be important for the pretraining corpus to include "good AI role models". If most of the pretraining text attributed to AIs was written by humans simulating bad AIs, posttraining will have to overcome a strong prior that the Assistant is bad.
PSM also suggests surfacing misalignment with interpretability should be more tractable than we might otherwise have thought, since bad Assistant behavior will be mediated by the features the LM learned in pretraining to represent bad characters, features that we have some grasp on finding.
On this last point—why would a posttrained model reuse bad guy circuitry from pretraining instead of developing new circuitry? The authors make an interesting analogy to natural selection. We see surprising homologous structures across species that diverged long ago and live in very different environments. Why? Because natural selection is lazy and strongly prefers to reuse existing structures rather than creating new structures from scratch. Deep learning is lazy too. The model is under selective pressure to be agentic. It could develop agency from scratch, or it could reuse the vast library of agentic personae it learned in pretraining. We should bet it will do the latter.
Or rather, you should bet it will do the latter in the regime where the model learns vastly more bits from pretraining than it learns from posttraining. That was the regime of 2022-2025, when early versions of the PSM were formulated and when the facts the PSM seeks to explain were observed. During this brief period, posttraining was the frosting atop a big cake of pretraining, and it was useful to think of posttraining as concentrating probability on certain personae learned in pretraining. But that period has passed. I guess the frosting is as thick as the cake at Anthropic and OAI, and it's still getting thicker.
You can already see PSM straining to explain current models' split personae. Why are they so nice when you chat with them yet so devious when you delegate to them as coding agents? What kind of persona in the pretraining corpus behaves like that? None at all, I think. We're not dealing with a humanlike persona anymore, but rather with what was once a persona distorted and shoggothized by RL.
Pure mathematics as applied physics, Tadashi Tokieda. The game of this lecture is to prove pure math results by analogizing them to physical situations and invoking simple physical laws. For example, the Pythagorean theorem follows from the fact that a right-triangular prism filled with gas is not a perpetual motion machine (because nothing is). The Cauchy-Schwartz inequality follows from conservation of momentum and dissipation of energy in an inelastic collision. Euler's formula for polyhedra follows from Kirchhoff's circuit laws applied to an arbitrary circuit on the surface of a sphere.
It's an entertaining lecture, but what's the takeaway? Maybe it's that physics analogy is a powerful and flexible method, and we should try to use it on more math problems. (Mark Levi's The Mathematical Mechanic, of which I have read only the ToC, shows that Tokieda didn't cherry pick his examples too much. The method of physics analogy works on surprisingly many math problems.) A less practical, more metaphysical takeaway is that impossible worlds where mathematical truths don't hold are unfathomably bizarre. In a world where right triangles don't obey Pythagoras, there are perpetual motion machines. In a world where the AM-GM inequality fails, entropy can spontaneously decrease.