I am working on a new project, a way to reduce costs from large LLMs such as Claude or GPT, and continue my math research program. A brief summary of how I am doing it.

I am building a smart AI friend in my room, and it helps me do math

AI-Team - Update after a month

stangeli.substack.com

Anthropic and OpenAI plans needed for my research are getting too expensive. 200 €/month for each, for just a weekend (then I hit the week limits and have to wait for 5 days), is not affordable any more. Looking for hints about how to work with open-weight models locally. Any help is welcome! Thx!

"In many systems, the situation is such that under some conditions chaotic events take place. That means that, given a particular starting point, it's impossible to predict outcomes.

This is incredibly amazing 💙👇

Time. Time is what I am missing. I have so much to do, and only 24 hours each day. I have a full-time job (that has nothing to do with CMH and with what I write here on Substack), and that leaves me with just a few hours a week for what I like to do: my research on CMH, learning math, 1/3

A deeper explanation of my new project to develop CMH (inspired by Asimov’s Psicohistory) with a new math and the use of my “AI-Team”. I love it when a plan comes together! stangeli.substack.com/p/ai-team-my...

AI-Team - my new project

Using adversarial multi-agent AI to help me in developing CMH

stangeli.substack.com

The new project is working in a completely different way compared to the way I worked so far on CMH. This time I started from the math itself. Not from the data available, not from the available theories or mathematical models.

Computational Macrohistory Bulletin | Stefano Angeli | Substack

I apply mathematical and computational methods to study large-scale historical dynamics: revolutions, political cycles, economic crises. Founder of the Foundations Institute of Computational Social Sc...

stangeli.substack.com

Summer time, vacation, lots of free time. Time to think, to imagine, to dream. So, while I was sitting under the hot sun in a Sardinia’s beach, I had an idea.

Congratulations!

I was seventeen, an Italian boy set down in the middle of Iowa as an exchange student, from August 1986 to July 1987. I lived on a farm outside Winterset, fields running to the horizon in every direction, an hour by school bus from town each morning.

Three cases, perfect classification. Impressive? Count it: a separating threshold exists by chance one time in three, the exact ordering one in six. So no validation claim. The win is narrower: definitions, weights, and a procedure that reproduces in a spreadsheet.

Always happy to connect with people working in computational social science, broadly defined. If that's you, say hello.👋

Why did Saudi Arabia hold while Tunisia and Egypt fell? Bigger youth bulge, same unemployment, more internet. New paper: a five-variable stress index separates the cases. The term that does it: full autocracy as suppression capacity. Three cases, no inference. The machinery runs on real data.

Equal weights sound like the safe choice for a composite index. They are a strong claim in disguise: that internet access matters as much as regime type. In our data that claim fails the classification. Weights are theory written as numbers. Choose them like it.

Twenty-five variables. Not fifty. Not two hundred and fifty. Every extra dimension is a new place for overfitting to hide. The model has to be large enough to capture the structure and small enough to stay testable. That tension is not a problem to solve. It is the art.

People think prediction means one thing: what happens, when. There are two others. Regime: stable, stressed, tipping point. Spectral: a slow cycle lives here. Different claims. Confusing them produces numbers that look scientific and read like astrology.

An equation fitted to data is a guess. An equation that traces to a mechanism in the literature is a claim you can test. Every drift term in CMH has a documented story. Know why it is there and you know what would break it. That is falsifiability.

You can add variables forever. But every new one costs: a parameter to estimate, a place for noise to hide, a thing you now have to measure for decades back in time. The discipline is knowing when to stop. What a model omits is not a detail. It is the part that actually took thought.

GDP is not wellbeing. A Gini coefficient is not inequality. A Polity score is not democracy. Every variable in a model is a proxy for something else. The number is never the thing. A model that is honest about where its data gets thin is harder to publish and works better in the long run.

WP-2026-002 is on Zenodo. 25 variables, 5 families, a coupled SDE system, and the machinery that turns eight axioms into testable numbers. The youth bulge definition is fixed. The prediction hierarchy is explicit. I wrote about what the rules actually build. On Substack. Link in first reply.

Interdisciplinary work means being a beginner again. You don't know the words, which questions are stupid, who to read. But beginners see things experts stopped seeing, not by being smarter, but by not knowing what to ignore. The dumb question is often the interesting one.

When I can't write clearly, it's not a writing problem. It's a thinking problem. Writing is how you figure out what you think. Most people stop at the first version that sounds okay. But "good enough" means you stopped thinking before you were done.

I keep a folder called "dead ends." 40 analyses, zero publications. It's the most useful thing I own. Every wrong turn taught me something real. But journals want findings, not maps of where not to go. Real research is a mess. The published record is not.

Nobody tells you how much of research is just sitting there. Staring at a wall. Walking the block again. The best ideas come from empty space. But institutions measure papers, not thinking. So researchers learn to hide the part where the actual work happens.

AI tools give you answers in seconds. Your brain, given time, gives you understanding. They are not the same thing. Speed costs depth. I use AI every day. But I try not to skip the part where I sit with a problem long enough to really feel it.

The foundational CMH paper just got a major upgrade. Same eight axioms as before. Now with four theorems: predictive decay, the Herding Threshold, reflexive fixed points, and calibration decomposition. The axioms were stated. Now they're proven. Substack link below 👇️. tinyurl.com/y2vscu2n

The work published on this account under the name Galen Fontaise, an anagram, was mine. I am Stefano Angeli. The Computational Macrohistory programme and its six working papers now carry my legal name. The research itself does not change.

Summer reading is packed. Four books this year: James Gleick's Chaos, Ian Stewart's Does God Play Dice?, David Bessis's Mathematica, and a 1910 classic, Silvanus Thompson's Calculus Made Easy. A short thread.

Every serious research project runs into the same wall: AI assistants forget. You spend hours establishing context, get useful work done, close the window, and the next session starts from zero. This is a guide to the setup I use for a long-running independent research programme. #AITools

Building a Research Environment That Remembers

A Practical Guide

galenfontaise.substack.com

From now on I write here under my legal name, Stefano Angeli. Earlier work in this series appeared under the anagrammatic pseudonym Galen Fontaise. The decision to publish under my own name reflects where the project now stands. The work itself (the framework, the papers) is unchanged.