# SIG-FPT 2026-08-14 — Session Notes

Participants:
  - Patrick Nast | UTC-7
  - Venkatesh Rao | UTC-8
  - Dan Schmidt | UTC-7
  - Ergod | UTC+1
  - Giovanni Merlino | UTC+2
  - Vibhav | UTC+5:30
  - Robby Rao | UTC - 7

---

## 📖 The Reading
Not clearly identified in the discussion. This session was primarily a **project/hackathon planning meeting** rather than a reading discussion, though participants referenced prior readings — notably an "ant colony optimization" paper reviewed the previous Friday and earlier readings on disrupting ant path-finding/stigmergic coordination.

## 🧭 Overview
The group met to plan a two-day hackathon/workshop built around a browser-based ant colony ("stigmergy") simulator that Dan has developed. Dan demoed the current single-player simulator and a newer multiplayer "shared world" version with a colony-survival leaderboard. Discussion centered on how to structure the workshop (synchronous vs. asynchronous formats, how many design variables to expose, a competitive vs. exploratory two-track model), how to frame it around a security/adversarial theme, and what technical work is still needed. The group closed with concrete next steps and logistics (dates, sessions, registration).

## 💡 Key Points & Themes

**The simulator (Dan's demo)**
- **Dan:** The simulator runs entirely in the browser (no backend/database). Ants leave the nest, drop pheromones to find their way back, and lay a second pheromone trail to food once found. Users can adjust simulation speed, colony size (up to 100), number/size of food sources, pheromone evaporation rate, signal-following strength, and "gland size" (how long ants can lay pheromones).
- Emergent phenomena include "ant highways" and "ant mills" (ants trapped circling); walls can be edited to create or break these.
- A **first-person "single ant" view** lets you experience the maze from one ant's perspective and watch trails form — Dan and Venkatesh Rao both liked this as a possible workshop mode ("the phenomenology of being an ant").
- **Dan:** A newer, not-yet-open-sourced version is a shared multiplayer world with real-time stats and a leaderboard for longest-surviving colony, plus a "god mode" to sabotage others (e.g. walling off a colony). He reported coworkers got quickly addicted/competitive.

**Workshop structure**
- **Venkatesh Rao:** Two feature layers — (1) "platform" features supporting research questions, and (2) an API / plug-in architecture letting people define ant strategies, run automated experiments, and log data. Outputs could be simulator "movies" plus plotted metrics (e.g. how fast ants die out).
- **Venkatesh Rao:** Strong emphasis on **limiting variables** — too many produces "a chaotic soup" with no interesting results. Suggested picking 2–3 design variables with trade-off relationships (e.g. colony size vs. ant speed).
- **Venkatesh Rao:** Proposed a **two-track structure**: a competitive contest for "killer" types, and a non-adversarial exploratory track (e.g. "produce the most interesting phenomenon and save a video"). Referenced the Bartle taxonomy of player types to appeal to more than one motivation.
- **Venkatesh Rao:** Keep entry **low-code/zero-code** — ideally define your team's parameters via a web interface and enter a bracket/tournament.
- **Patrick:** Framed the event as a hackathon whose theme is security, implying adversarial dynamics (attacker vs. defender). Stressed: (1) an easy way to "jack into" the simulator and define new coordination strategies, (2) a **cost/fairness measure** so people can't gain unfair advantage via arbitrarily complex API use, and (3) a clear central goal for participants to put "spins" on. Wants heterogeneity of approaches ("not quite a Cambrian explosion").
- **Patrick:** Favored a synchronous **tournament** format (bracket/waves) as the strongest version, likening it to Axelrod-style strategy competitions, while keeping an exploratory single-player mode too.

**Security / adversarial framing**
- **Patrick:** Multiple routes to adversarial dynamics: a single-colony "hacker" vs. defender; multiplayer with "spy"/detractor ants; pheromones that can affect neighboring colonies; and manipulating the maze structure itself.
- **Patrick:** Stigmergic algorithms perform relatively better in **dynamic environments**; the key lever is the **evaporation rate**. Adversarial play could involve waiting for ants to form a path, then editing the maze/evaporation to knock them out of a local optimum toward a new shortest path (hence the "ants vs. termites" idea, termites moving walls).
- **Patrick:** Real-world motivation — security is timely in AI; cited an OpenAI / Hugging Face incident where an exploit spread via emergent, essentially stigmergic message boards. The sim is a "primitive" for exploring how to disrupt maliciously coordinating swarms or defend distributed communication.

**Design philosophy on variables**
- **Patrick:** Have **more variables in the simulation than are exposed to the user**. Expose the crux parameters (e.g. evaporation, how agents update behavior in response to pheromones); keep background mechanics (exhaustion, colony regeneration) fixed. Suggested a small "bank" of ~3 pheromone types, each with ~2 tunable parameters (e.g. response distance, attraction/repulsion).
- **Dan:** Raised whether directly editing walls (simpler) or controlling a wall-editing species like termites (more "profound" but harder to grok) is the better design.

## 🔀 Questions & Disagreements
- **Synchronous vs. asynchronous format:** Dan laid out two directions — keep investing in the live shared/multiplayer world, vs. an async "eval/benchmarking" system where colony configurations are scored across many mazes and adversarial conditions. He suggested both could coexist. Patrick leaned toward synchronous tournaments; Venkatesh Rao pushed for simplicity and a two-track split.
- **Scope for the timeline:** Ergod asked whether the ambitious "open-world, program-your-own-behavior, 24/7 running" vision is still the target given only ~5 weeks. **Patrick:** Better to start narrower and build iteratively (MVP logic); the main constraints are design clarity (what exactly participants do, and what we can infer about protocols from a winner), more than raw time.
- **How much control to give users** — repeated tension between flexibility (interesting, "profound") and constraints (fairness, feasibility, graspable for newcomers). Not fully resolved; flagged for next week.
- **Feature clarifications:** Ergod asked whether colony metabolism/growth (constant food consumption by colony size, spawning new units) had been implemented. **Dan:** Not yet, but likes the direction. Robby Rao asked whether colonies should be able to interact beyond competing for food (detractors, sabotage roles). **Patrick:** Yes — that's the direction; noninteracting colonies racing for food doesn't really meet the "security" conception.
- **Implementation concern:** Giovanni Merlino noted the multiplayer/adversarial version needs significant rework since the sim currently lives purely in the browser UI layer; **client synchronization** ("net play") is the hard problem. Patrick suggested avoiding full real-time sync by using **preregistered parameters seeded into rounds/waves** (e.g. ~4 players per round) that can be watched live without everyone connecting simultaneously.

## 🔗 References Mentioned
- Ant colony optimization paper reviewed "last Friday" (title not stated).
- Bartle taxonomy / "Bartle test" — four gaming-personality types (killer, explorer, socializer, etc.); a 2×2 on Wikipedia.
- "Monster chess" — asymmetric chess variant (full army vs. king + pawns with double moves); cited as an example of a fun, low-variable asymmetric contest.
- Robert Axelrod-style tournaments ("Axelrod-y") for comparing strategies.
- The OpenAI / Hugging Face security incident, described as exploiting emergent, stigmergic message boards.
- Termites (as a wall-building adversarial species) vs. ants.
- Tools/concepts: GitHub repo, agents.md / CLAUDE.md / CONTRIBUTING.md files, Claude Code, Codex, MCP server, DSL/API, Figma-style shared canvas; the live simulator at **sticksim.protocol-institute.org**.

## ✅ Action Items & Next Time
- **Dan:** Take the lead on starting a shared brainstorm space (Figma/canvas) with the Bartle 2×2, and run a week-long async brainstorm on how to structure the workshop (game dynamics per persona). Also push the newer multiplayer/shared-world simulator to the open-source repo.
- **Dan:** Serve as PR reviewer/approver given his familiarity with the system.
- **All / contributors:** Check out the repo, try a small feature addition, and submit it as a PR; also play the simulator hands-on.
- **Venkatesh Rao:** Create a simple front page on the StigSim subdomain that indexes the different simulator variants (starting with Dan's two defaults), so forks/variants can be listed.
- **Patrick:** Aim to converge (informally, not a formal vote) on the most feasible/interesting proposal by/around next Friday's meeting.
- **Logistics:** Workshop is **Sept 21–22 (Mon/Tue)**, four ~90-minute sessions (one morning ~9:00–10:30, one afternoon ~14:00–15:30, spanning US and European time zones). Times to be posted to the channel for feedback. **Registration opens August 21** (~one week out); exact time slots need not be locked before then.
- **Next week:** Regular meeting with a reading, followed immediately by this project discussion.

## ⭐ Memorable Quotes
- **Venkatesh Rao:** "If you give people new to the subject too many variables to think about, you'll get basically a chaotic soup of variables that don't lead to any interesting results."
- **Dan:** "It's a weird opportunity to know what an ant... the phenomenology of being an ant."
- **Patrick:** "Just having a bunch of noninteracting colonies... seeing who finds the food the fastest doesn't really strictly meet the conception of it being about security."

*Note: the transcript audio was heavily interleaved (speakers overlapping mid-sentence), so some attributions and the exact ordering of points are approximate.*