# SIG-PSY 2026-08-13 — Session Notes

Participants:
  - Aneesh Sathe
  - Sean Stevenson | UTC -4
  - Spencer Norwick U-0700
  - Andre Comeau | UTC-5
  - Ellen Kaye-Cheveldayoff (UTC -4)
  - Kyle Mathews | UTC-6
  - Florian Lohse (UTC+0)
  - rafa (UTC+1)
  - Venkatesh Rao | UTC-8

---

## 📖 The Reading
Not clearly identified in the discussion. This session was not a reading-discussion in the usual sense; it was a demo and working session for the group's "World Machines" project tools (the Oracle, the Prime Radiant, the Scribe note-taking app, and related infrastructure). Aneesh also mentioned a piece he "published last night" outlining principles used in the Prime Radiant, which he shared in chat, but no specific shared reading was worked through.

## 🧭 Overview
Aneesh gave an extended walkthrough of the software he has been building for the group: a note-taking/wiki app ("Scribe"), the "Oracle" retrieval system built on the group's corpus of books and essays, and the "Prime Radiant" — a temporal knowledge-graph system that models history as interacting "machines" and generates extrapolations/"prophecies." The second half turned to planning the upcoming symposium/special session (involving Aneesh, Sean, and Florian), which Venkatesh framed as the project's public launch. Discussion touched on how notes and essays could eventually feed automatically into the modeling system, and on how contributions and funding might scale afterward.

## 💡 Key Points & Themes
- **Project context (Aneesh):** The group began as a book club on Venkatesh's Substack, evolved into members writing and cross-referencing essays, and became the "World Machines" project. The core framing (attributed to Venkatesh) is that history can be read as different "machines" in play — e.g., a modernity machine, a "versions" machine, a "liveness" machine.
- **The Oracle:** An AI system that ingests the group's material (books, members' essays, external essays) into a searchable database so members can ask questions or explore, ideally lowering the barrier to writing essays. It functions as retrieval-and-answer ("it doesn't do prophecy yet"), and behind the scenes searches a **knowledge graph**, taking several back-and-forth steps across related links before answering.
- **Two-tier vs. MCP access (Aneesh):** The portal AI is described as "fairly low powered" and uses a two-tier agent setup (the Oracle queries a second agent with access to raw notes, which returns summaries). A dedicated **MCP** option lets members connect their own, more capable AI directly to the corpus (including the raw text) via a generated access token.
- **Scribe app (Aneesh):** `scribe.worldmachines.org` — a lightweight web app (installable as a PWA on iPhone via "Add to Home Screen") for quickly adding notes to the corpus. Motivation: editing/uploading text files was too much friction and meant he wasn't writing anything. Notes can be quick thoughts, book notes, or "velocity notes" / term definitions (Venkatesh had wanted to define some terms).
- **Notes as corpus / "hive mind" (Aneesh):** The purpose of adding a note is that it becomes a source the Oracle can cite. Aneesh described it as a hive mind where members' thoughts enter the shared sourcing and others (and their AIs) can build on them.
- **Provenance/tagging (Aneesh):** There's no truth filter, but every note is attributed to a person or tagged as coming from a book — so a book source may be more trustworthy than a personal opinion.
- **Prime Radiant — "cards":** Each history "machine" (e.g., LLMs) is modeled as a **card**. Cards are generated when the AI does a deep search, writes a report, and **extrapolates** ("extract") how things might evolve. Cards sit in an ontology/knowledge graph and are treated as machines with inputs and outputs (themselves machines), operating within a landscape made of other machines that exert selection pressure.
- **Inter-machine couplings (Aneesh):** Machines influence each other, including surviving elements of the "modernity machine" (e.g., electricity) affecting LLMs today. Example: if AWS broke down, that would ripple into the LLM world. Cards also now include predicted end-of-life (when a machine is likely to stop working), which the earlier version lacked.
- **Atlas:** A visualization tracking machines across time and their interconnections, including long-lived machines predating the project's ~1600s starting point (e.g., agriculture). A "walk" feature runs time forward, showing machines appearing — including projected future machines and their intersections (Aneesh picked an example involving land use, "lake"/lease, protocols, and cultural effects, noting he didn't fully know what it meant).
- **Almanac:** Associated reading list / recorded future items (described as very large and interesting).
- **Oracle output types (Prime Radiant side):** Three modes — **ask** (a question), **story** (set a year and lens, get an invented person's life story), and **prophecy** (choose a type of oracle to generate a prophecy). Aneesh noted outputs aren't polished yet.
- **Design shift (Aneesh):** He originally planned a graph-neural-network-based AI to learn from existing cards; encouraged by Venkatesh, he relaxed that and pursued a looser "prophecy" approach that doesn't require being overly specific. The GNN piece is not yet built.
- **Symposium/special session:** Aneesh, Sean, and Florian will present. Proposed structure: a walkthrough/demo of the tools, then individual theses/talks, then possibly a hands-on mini-workshop (Aneesh worried about compute bandwidth for many simultaneous users). Florian suggested a **staged live demo** where they make a claim from their talks and show what the Oracle says.
- **Members' theses:** Sean is working on a "new nature" theme and an argument that AI works its way down the **pace layers** (toward infrastructure, governance, elements of nature); he's still working out the underlying theory. Florian is prepping his symposium talk, focusing on **time** ("timekeeping moves deeper into the infrastructure layer," leaving a gap possibly filled by a "new nature" angle), referencing his 2014 master's thesis on time.
- **Launch framing (Venkatesh):** Treat the special session as the public launch; afterward take a public posture (Substack articles), invite more people, and consider funding for more capable infrastructure. He noted the challenge of compressing ~two years of thinking into one or two hours.

## 🔀 Questions & Disagreements
- **Ellen** asked clarifying questions as a newcomer: what the purpose of adding a note is, whether it's just to continue book-club discussion, and whether there's a human-readable view or if the corpus is only AI-accessible — concluding "it's a wiki."
- **Spencer** asked how the system handles **factual differences vs. perspective differences** in notes (e.g., if two people believe different things). Aneesh's answer: no truth filter, but provenance tagging (person vs. book) gives a sense of reliability.
- **Sean** raised several open questions: whether/how information can be added to the Prime Radiant to affect machines; how the "tension phases" for the Prime Radiant piece were derived; and how the Atlas's connections are "supposed to work." Aneesh explained that many patterns/labels *emerged* from the AI's modeling rather than being hand-specified, and admitted he hadn't fully thought through some applications (e.g., identifying "pinch points" in history) — an area he explicitly invited input on.
- **Florian** asked whether the system could incorporate "crisis moments" / eventful history discussed in a prior theory call, and requested a look at how the system is set up in the background (privately if needed). He also raised logistics about symposium slot deadlines and whether individual talk titles/descriptions need finalizing.
- No sharp disagreements; the main tension was Sean and others still working out the *theory* behind their contributions and how the tools fit that theory.

## 🔗 References Mentioned
- Venkatesh Rao's Substack / the original book club (origin of the group).
- Daniel Dennett — "cranes" for building up concepts (used as a sample Oracle query).
- "Pace layers" concept (invoked by Sean re: AI moving toward infrastructure/governance/nature).
- "New nature" — recurring group theme (Sean, Florian).
- Florian's 2014 master's thesis on time; two recent books on time he's been reading (titles not given).
- An article Aneesh published "last night" (shared in chat) on principles he used in the Prime Radiant (title not stated).
- Tools/infrastructure: the Oracle, Prime Radiant, Atlas, Almanac, "cards," MCP (Model Context Protocol) access, `scribe.worldmachines.org`, graph neural networks, GitHub repo/issues.

## ✅ Action Items & Next Time
- **Aneesh:** Continue building out the tech before the special session; make access easier and arrange more compute for that day; check whether the GitHub repository is public (offline). Wants to eventually auto-route new essays/notes through AI to detect new cards/machines and feed World Machines content into the Prime Radiant. Open to adding requested features and creating public GitHub issues to invite outside contributors.
- **Members with corpus access:** Try the tools (Scribe PWA, MCP with your own AI, Atlas, Almanac, Oracle modes) and give feedback; add notes as you go, even partial thoughts. Request a login/access from Aneesh if needed.
- **Special session (Aneesh, Sean, Florian):** Finalize structure — likely tools walkthrough → individual talks → possible hands-on mini-workshop; consider Florian's staged live-demo idea. Confirm/update individual talk titles and descriptions before the slot deadline.
- **Venkatesh:** Treat the special session as the public launch; afterward pursue Substack write-ups, wider invitations, and possible funding. Will post details about the *next regular session* once he's figured out its focus (not yet decided).
- **Kyle:** Was present but silent; no items noted for him.

## ⭐ Memorable Quotes
- **Aneesh:** "I think of it a bit like a hive mind where all of us are then exploring different things... and everybody's, at least their AI, knows that this thought has been had, and then you can build on top of that."
- **Venkatesh:** "We're trying to pull together almost two years of interesting thinking in one hour, so it should be an interesting challenge — or two hours."
- **Aneesh:** "Basically, just control that it doesn't do anything nuts... and then lots of interesting things happen."