# SIG-FPT 2026-07-30 — Session Notes

Participants:
  - Venkatesh Rao | UTC-8
  - Aneesh Sathe
  - Sean Stevenson | UTC -4
  - Matthew McDowell-Sweet (UTC)
  - Kyle Mathews | UTC-6
  - Florian Lohse (UTC+0)
  - Chris R.
  - Andre Comeau | UTC-5

---

## 📖 The Reading

The session focused on the work of **Peter Turchin** and his program of "Cliodynamics" — the mathematical/quantitative modeling of history. Participants engaged with a short introductory text by Turchin (including material on his "Mirian/Myriad Empire model" of agrarian states vs. nomadic confederations) alongside a critical article invoking **Popper's** critique of historical prediction ("historicism"). The exact titles were not stated; participants referred to reading a PDF and a shorter critical article. This builds on a prior session (two weeks earlier) on "logics of history" and "eventful time."

## 🧭 Overview

The group examined whether history can be modeled and predicted mathematically the way physics or modern biology is. Most participants were sympathetic to Turchin's approach as a tool for *retrospective* historical analysis but skeptical of its *predictive* power, largely aligning with Popper's critique. Venkatesh developed an extended synthesis distinguishing "cyclical" phenomena (with underlying integrable quantities) from "eventful"/discontinuous phenomena, arguing the useful target is the *intersection* of the two — making focused "prophecies" rather than long-range trajectory forecasts. The conversation closed on the idea (from Florian, Aneesh, and Andre) of prophecies as remembered causal packages, and of mathematical modeling as valuable for taxonomy/mapping even if not predictive.

## 💡 Key Points & Themes

**Framing prompts (Venkatesh):**
- How do you model the fact that new ideas/inventions (e.g., vaccines) can drastically and suddenly change historical trajectories?
- Is requiring rich, exhaustive historical datasets just "looking under the streetlight"? Are there ways beyond working with large piles of data?

**Retrospective analysis vs. future prediction:**
- **Sean:** Sympathetic to Turchin for historical analysis of the "contained" past, but finds Popper's argument compelling that prediction is hard — vaccines, technology, and human free will make history unlike a clock or machine. Noted the analogy to biology transforming (over ~200 years) from descriptive taxonomy into an experimental, hypothesis-testing science, which is what Turchin wants to do for history. Also flagged Turchin's own claim that *developing theories* (not gathering data) will be the hardest future task.
- **Kyle:** Endorsed Popper — history is not like physics because "the law of gravity" keeps changing; humanity reconfigures itself, technology reconfigures. "The past is not just a foreign country; sometimes it's a completely alien planet." Any analytical framework is vulnerable to a core assumption being negated. Cited "generals fight the last war" / going into the future looking through the rearview mirror.

**On mathematics and formalism:**
- **Matthew:** Struck by the provocative "without mathematics we're doomed to vague statements and wrong conclusions" — thinks it needs a qualifier and only holds above a certain complexity threshold; curious how reasoning worked *before* formalized mathematics. Linked "prediction verified in the marketplace" to management thinking (couldn't recall the author). Raised the **blind spots** of Cliodynamics: indigenous/embodied ways of knowing it is inherently blind to, and the "empty space" of day-to-day "non-history" that rarely gets documented. Suggested the information environment increasingly leads the "atoms of the built environment," strengthening the case that ideas shape history.

**Turchin as productive even if wrong:**
- **Aneesh:** Amused that the critique disregards Lee/Larry Darwin. Argued we need someone like Turchin to attempt formalization even if wrong — a "stone soup" situation that mobilizes useful work. Key substantive point: humans see problems arriving on the horizon and *act to solve them* (e.g., food crises), so historical systems have an **active/adaptive element** — this differs from mere chaos. Feedback can't go back in time; it feeds back into an already-changed system. Used the potato/island example: population grew from a new food source but the economy didn't keep pace, producing externalities — the kind of causal dynamic a mathematical model could capture.

**Venkatesh's synthesis — types of cyclical phenomena:**
- Warned against "principle cyclical thinking" that just detects a cycle and invents a just-so narrative — like stock-market technical analysis / Elliott waves ("crackpottery"); this is why Popper saw such theories (and once evolutionary biology) as pseudoscience.
- **True cycles** exist and can be modeled with underlying physics/"integrable quantities": earthquakes (tectonic pressure builds and must release; e.g., the overdue Pacific Northwest quake, ~50% chance in a 50-year window) and **demographic cycles** (aging populations generate and absorb fewer ideas — e.g., Japan aging rapidly and not leading AI).
- **Eventful phenomena** (from the prior reading): revolutions, pandemics, tsunamis — self-contained, analyzable on their own terms, with a "billiard-ball impulse" impact.
- **Proposal:** The valuable move is the *intersection* — e.g., elite overproduction (cyclical) crossing a threshold triggers a revolution (event). Build models and predictive machinery around **prophetic events**, not long-range cycles or abstract models of "what a revolution is."
- Prophecies differ from forecasts/predictions: aim for *focused, precise* prophecies (e.g., "a particular kind of revolution in China within 20 years") that, if wrong, are "precisely wrong." Don't try to query arbitrary space-time ("what happens in Poland in 2067?").
- Concluded Turchin's basic program isn't workable (never a rich enough database; "the mule in psychohistory" always knocks history off course), but a viable version looks for **post-cyclical phenomena with integrable quantities** plus a good qualitative theory of which events have "eventful temporality."

**Reframing prophecy and failure:**
- **Florian:** Treat prophecies/predictions as *material*; it's fine, even the point, that theories will be wrong. A **failed prediction can be as valuable as a correct one** because it reveals when reality diverges from a straight line (analogy to climate models). Suggested learning from other quantitative-but-uncertain sciences like astronomy. Noted German intellectual "historian fights" of the 1980s between quantifiers and the hermeneutical camp, and hoped modern tooling (which lowers the barrier once requiring a mathematical mind) could bridge them.

**Kyle's "portfolio of models" idea:**
- Use history-derived models not as causal laws but as a *set* of models evaluated in real time — figure out which cluster of models is "working" now, then infer which underlying laws/assumptions are currently in operation. Events can be significant enough to impose "physics-style laws" temporarily (like a magnet that aligns everything), then recede. Example: US wars usually enjoy support that gradually declines — but Vietnam didn't, which is significant.

**Prophecy as remembered causality (Florian, Aneesh, Andre):**
- **Florian:** Small-scope prophecies are causalities packaged to transfer over time via memory — e.g., "if the waters retreat, run for the hills." Many real prophecies (Excalibur/sword-in-the-stone) are conditional ("if X, then Y") rather than dated predictions.
- **Andre:** Even if mathematical analysis isn't predictive, taxonomy can lead to discovering a *topography* (analogy: pre-existing lineage taxonomy enabled **Alexander von Humboldt's** "invention of nature," revealing that geography, not just lineage, was the key variable). Perhaps the best use of an LLM-like model of history is not prediction but an **"Atlas"** we can read to see what the world might hold — mapping in a non-human-readable way that yields human-readable outputs.
- **Venkatesh** endorsed this: imagine AI-driven psychohistory making *trigger/condition-based* prophecies rather than time-based ones — e.g., "if these 100 variables reach a state, and 80 are already in place, open the vault and make the prophecy." Using causality to carve out "islands of physics-like predictability" within the generally unpredictable flow of history.

## 🔀 Questions & Disagreements

- **Predictive validity:** Broad skepticism (Sean, Kyle, Chris) that Cliodynamics can predict the future, versus a more constructive stance (Aneesh, Florian, Andre, Venkatesh) that the enterprise is still valuable — for retrospective analysis, taxonomy/mapping, or learning from failed predictions.
- **Does mathematics only help above a complexity threshold?** (Matthew) — an open qualifier on Turchin's strong claim.
- **Resolution problem (Chris):** Turchin's claims feel "unobjectionable at low resolution," but the more data and higher resolution you get, the more questionable the specific claims become — he was still "chewing on" this.
- **Sampling/source bias (Chris):** Our knowledge of nomadic confederations comes almost entirely from the literate agrarian states (Ming on Mongols, Tang on Turks). This is an extreme sampling bias undermining the neatness of models like the agrarian-vs-nomad dyad. (Illustrated via the "Punt"/baboon-teeth-mineral-analysis story about locating an unknown Egyptian trading partner.)
- **Is the adaptive/free-will problem fatal?** Sean and Kyle lean yes for prediction; Aneesh reframes adaptation itself as a modelable feedback element (distinct from chaos).

## 🔗 References Mentioned

- **Peter Turchin** / Cliodynamics; his "Mirian/Myriad Empire" agrarian-vs-nomad model and "elite overproduction" theory of revolutions.
- **Karl Popper** — critique of historicism / prediction in history.
- The **"mule" in Asimov's psychohistory** (Foundation) — the unpredictable disruptor.
- **Alexander von Humboldt** — "invention of nature"; role of taxonomy in enabling discovery.
- **Lee/Larry Darwin** (as referenced in the critical article — exact figure unclear from audio).
- **Peter Thiel** vs. **Paul Graham** startup approaches ("call your shots" vs. "call your customers") — invoked re: ideas shaping outcomes.
- **Albert Einstein**'s (attributed) "if the bees die, humanity has ~7 years" — as a causality-shaped prophecy.
- **"Excalibur" / sword-in-the-stone** — conditional prophecy example.
- Complexity-theory case studies (tobacco/energy companies leaning into such thinking); **high-frequency trading**; Elliott-wave technical analysis.
- **AI usage studies** (Anthropic's economic index; a Google study).
- The **2004 Indian Ocean tsunami** and the island preserving tsunami warning via oral song.
- German 1980s historian debates (quantitative vs. hermeneutical camps).
- Prior session reading on "logics of history" and "eventful time/temporality."
- "All models are wrong, some are useful" (Kyle, riffing: "some models are useful sometimes, but not at other times").
- The "Punt"/Egyptian baboon-skull mineral-analysis pop-science story (Chris).

## ✅ Action Items & Next Time

- **Venkatesh** will look for readings that unpack the "prophecy as remembered causality / conditional-trigger" thread that Florian and Andre developed.
- This track resumes **in two weeks**.
- **Next week:** a more practical/applied session led by **Aneesh**.

## ⭐ Memorable Quotes

- **Kyle:** "The past is not just a foreign country. Sometimes the past is just a completely alien planet."
- **Florian:** "Maybe a failed prophecy or a failed prediction is just as valuable or maybe even more valuable than a correct one because it enables you to see when something is diverging from a straight line of history."
- **Andre:** "Perhaps the best use would be to not have it be the predictive element, but be the Atlas that we could read to see what the world might hold."

*Note: The transcript was heavily interleaved (multiple speakers' lines were spliced together by the diarizer), so some attributions — especially where Matthew's and Aneesh's remarks overlap — are reconstructed from context and may be imperfect.*