# SIG-PSY 2026-10-08 — Session Notes

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

---

## 📖 The Reading
Arthur C. Clarke's essay on the hazards of prophecy (referred to as "perils of prophecy"), which distinguishes "failure of nerve" from "failure of imagination." It appears in a 1960s-era anthology (full manuscript located by Florian) that also contains essays by figures such as Paul Ehrlich and Margaret Mead. The essay includes a closing table contrasting "expected" and "unexpected" technologies.

## 🧭 Overview
The group did a short in-session read of Clarke's prophecy essay, then went around sharing reactions, with particular focus on how its framing maps onto the current AI moment. Discussion centered on the nerve/imagination distinction, the expected-vs-unexpected technologies table, and the problem of evaluating qualitative "narrative" predictions. The session closed with reflections on a recent symposium and planning for the group's "psychohistory"/prime radiant research agenda for 2027.

## 💡 Key Points & Themes
- **Aneesh:** AI has shifted the limits of knowledge — predictions once blocked by inaccessible information are now easier because AI can survey the latest literature. Short-term (~10 year) science-based prophecy is feasible, but longer horizons remain out of reach. Favors causality (an "if-this-affects-that" graph) as the project's backbone, since no single brain can hold all the relevant science; cited Clarke's ICBM example (rockets were abandoned until nuclear warheads made them worthwhile).
- **Aneesh:** The hard part remains asking the right questions. People now fear looking foolish far less (betting markets, social media), so prediction has become everyday and possibly overdone.
- **Florian:** Three takeaways — (1) the nerve/imagination distinction is useful and maps onto AI discourse, where facts are already available but people fail to connect the dots (failure of nerve); (2) the line "every long range prediction is an act of faith" is strong, best read as a secular statement; (3) the expected/unexpected technologies table is intriguing — expected items (helicopters, rockets) vs. unexpected ones (x-ray, spectrum analysis of planets) might be analyzable through a "world machine" lens, perhaps because merged/component world-machine views are harder for the establishment to grasp.
- **Kyle:** Found the ideas familiar from Venkatesh's prior writing. Self-analysis: he didn't predict current AI but didn't disbelieve it either — just didn't think about it. AI is the "most expected thing that's ever been expected," permeating sci-fi and myth. Yet LLMs are different from what was predicted: not a singular intelligence or personality, but more like "linguistic computing" — a CPU running linguistic operations. Sees much failure of imagination and notes nerve and imagination overlap (being locked into a lifelong narrative makes it take nerve to think differently). AI-safety people frame themselves as the "nervy" ones.
- **Sean:** Analogizes AI's current state to "post-Sputnik, pre-moon landing." Notes Clarke's era had many people "making their bones" declaring things impossible; today there seems to be a stronger pro-possibility bias (possibly his own bias). Broadly endorses Clarke's law about older people declaring things impossible being likely wrong. On the table: senses-accessible, envisionable technologies came to pass; those requiring unique instruments, sensors, or theories (where you see effects but not mechanisms, e.g. gravity) were far harder to foresee.
- **Venkatesh:** Failure of nerve is multidimensional — fear of looking stupid is only one cause; others include identity attachment and plain bad analysis. The failures came from very smart people (cited Vannevar Bush arguing in 1945 that V-2s wouldn't reach intercontinental range soon — ICBMs arrived within ~10 years).
- **Venkatesh:** The core challenge is the **eval problem for AI/prophecy**: easily-checkable predictions (stock prices, temperatures, game outcomes) are uninteresting; the interesting ones are narratives. Evaluating a 1967 narrative of the future is hard but not ill-posed, since consensus does form (Ehrlich's population-bomb Malthusianism is now largely dismissed, with worry reversed toward depopulation). More qualitative = more interesting but harder to eval.
- **Venkatesh:** AI largely *solves* the imagination problem — in his chats it generates interesting futures he'd miss (though both miss things). It may also indirectly solve nerve by making futures concrete — a "bravery prosthetic" that shifts us from vague "fear of fear itself" (Churchill) to fear of specific things. Expects the next decade to make us all more literate in AI-assisted futurism.

## 🔀 Questions & Disagreements
- **Can AI predict "shortcuts"/hacks? (Aneesh):** Many unexpected-column technologies were hacks — borrowed properties repurposed (carbon-14 dating from decay physics; also sound recording, lithography, transistors), whereas the expected column was all "big picture." Such shortcuts are highly localized in time and may be genuinely hard for AI to anticipate.
- **Florian** contrasted this with his artist friends' maxim that "there are no shortcuts in art," suggesting shortcuts may actually be more plentiful than people assume. **Sean** quipped that refusing a shortcut (fearing it's "cheating") is itself a failure of nerve.
- Open proposal (not resolved): how to build **evals for narrative prophecies** — e.g., each member writes a ~100–200 word narrative prediction of AI's state in 10 years and defines how it would be checked.
- No sharp disagreements surfaced; differences were mostly in emphasis.

## 🔗 References Mentioned
- Arthur C. Clarke — "hazards/perils of prophecy" essay; his law about elderly experts declaring things impossible.
- Paul Ehrlich — population-bomb / Malthusian prediction (an anthology essay now seen as a miss).
- Margaret Mead — also in the anthology (noted as likely a "miss").
- Vannevar Bush — 1945 skepticism that V-2 rockets would reach intercontinental range.
- Eliezer Yudkowsky — recent tweet defending his AI futures ("Yudkovskian futures"); Roko (of Roko's basilisk) arguing we're on a different path.
- Winston Churchill — "the only thing we have to fear is fear itself" (quote attributed by Venkatesh).
- An IBM CEO's remark about the world needing few/five computers (mainframe era).
- ChatGPT — Venkatesh's Iran-war scenario brainstorming and later narrative eval (got base Fed interest rate wrong; ~80% correct overall).
- Concepts/tools: "world machine" lens, "prime radiant"/"prime gradient," psychohistory, linguistic computing, Cloudflare analytics, effective altruists / AI doom.

## ✅ Action Items & Next Time
- **Venkatesh** to send a starter doc to flesh out the 2027 research agenda for the cycle-history project (a placeholder funding project for PI has been submitted).
- The group may each attempt a ~100–200 word narrative prediction of AI in 10 years plus proposed evals (floated, not firmly committed).
- **Aneesh** to pull material from the forthcoming "DocLake" and continue scoping the tech-building direction; noted he lacks time and that funding could support summer student researchers/interns to explore what's possible.
- **Sean** finishing his proceedings paper from the symposium.
- Consider ingesting the full 1960s anthology into the group's "Oracle" as historical documents (Venkatesh/Florian).
- Next meeting in **two weeks**; topic to be chosen in the channel.

## ⭐ Memorable Quotes
- **Kyle:** "Artificial intelligence is the most expected thing that's ever been expected."
- **Sean:** "For AI, we're in the post-Sputnik but pre-landing-on-the-moon sort of phase right now."
- **Venkatesh:** "The predictions that are easy to check after the fact are rarely the interesting ones."