# SIG-FPT 2026-08-21 — Session Notes

Participants:
  - Patrick Nast | UTC-7
  - Venkatesh Rao | UTC-8
  - Chris Laidlaw | UTC-7
  - Robby Rao | UTC - 7
  - Vibhav | UTC+5:30
  - durgadas
  - Ergod | UTC+1
  - Dan Schmidt | UTC-7
  - Matthew Bright UTC-7

---

## 📖 The Reading

**"Ants Nearby Treasure Search" (ANTS)** — research associated with computer scientist Amos Korman and experimental biologist Ofer Feinerman. The specific paper discussed extends the ANTS line of work by introducing **evaporating pheromones** (rather than permanent markers). This session is part of an ongoing "summer stigmergy series"; a prior session covered ant colony optimization (ACO) and the De Neubourg work, and earlier sessions covered "Pebble" and Blum/Kozen-style maze-search papers (names approximate from the transcript).

## 🧭 Overview

The first half of the session was a discussion of the ANTS paper, framed by Patrick as a "bridge" between weak-assumption automata models (pebbles/finite-state agents) and richer computational models like ACO. The group explored how evaporation-as-memory changes what stigmergic agents can compute, and floated framings from computer memory (DRAM), information theory, and security. The second half pivoted into planning for the group's late-September **hackathon/workshop**, reviewing a Miro board of proposed game modes for a stigmergy simulator and aligning on an MVP.

## 💡 Key Points & Themes

**Framing the ANTS paper (Patrick):**
- ANTS starts from very weak assumptions (à la pebble automata) and asks what problems simple agents *can* solve, rather than starting from an optimization problem like ACO.
- Setup: `k` foraging agents leave a nest and must find food at distance `d` on an oriented grid (not a maze). Key bound discussed: roughly `d + d²/k` — you must both travel out to the food and tile/search a disc-like area.
- Research lineage: original Korman work assumed a counter (memory scaling with number of ants, Turing-machine-like); a later wave restricted ants to finite-state machines with constant memory and small message exchange; this paper adds **evaporating pheromones** as a parameter.
- **No reinforcement:** unlike ACO, agents don't distinguish paths by pheromone intensity, so there's no feedback loop toward shortest paths. Agents also **die after a lifetime parameter τ**.
- **Algorithm 1:** essentially tile the grid by dispatching agents north/south to cover untouched tiles; guaranteed to find food within reach of a single agent's span.
- **Algorithm 2:** handles pheromones that evaporate after one step by **specializing roles** — some ants *become* the signal/beacon along paths (constantly refreshing decaying markers), while others explore based on what they read.

**Environment-as-memory / DRAM analogy (Venkatesh):**
- His primary interest: treating the *environment* as memory and recovering properties normally implemented in internal memory. External memory differs because it doesn't fit a clean memory hierarchy with pointers, and it's **insecure by design** — readable by everyone, with "security" (if any) enforced at the level of each cell rather than a boundary.
- Algorithm 2's refreshing agents are effectively **DRAM cells**: DRAM uses leaky capacitors refreshed rapidly, analogous to agents refreshing evaporating markers — "active memory." He called it an elegant hack that creates decentralized DRAM where needed.
- A cell alone encodes only a (meaningless) intensity; meaning emerges from reading neighbors (N/S/E/W). He likened this to agents reading **fiducials / QR codes** and raised the idea of a "decentralized QR code" / emergent word whose semantics must be defined.
- Highlighted **reference #59** as pointing to information-complexity work on how much "advice" must be encoded to direct an otherwise ignorant agent — a promising way to frame the whole problem (what to encode into pebbles/evaporating markers to maximally communicate).

**Evaporation as clock / channel constraint (Ergod):**
- Evaporation acts like a **clock**: the decay gradient encodes *how long ago* a marker was placed — extra information beyond a permanent "someone was here" pebble.
- Contrast with ACO: there evaporation is a soft knob to escape stagnation (a feature); here it's a **constraint on channel capacity**. So evaporation plays two opposing roles: forgetting-as-optimization-aid vs. forgetting-as-communication-limit.
- His conclusion: **forgetting is bad for one-shot tasks, good for perpetual/repeated coverage** (avoids pollution, enables useful emergence).
- Emphasized the system's pure **locality**: agent↔nature↔agent only, no direct agent-to-agent communication.

**Other reader reactions:**
- **Chris (first session):** thought of error-correction bits (echoing the DRAM framing); proposed **two pheromones** (data + correction bits) as a low-resource approach. Felt this was a "toy problem" and asked how it connects to a bigger real-world aim.
- **Robby:** read it as breadth-first search with memory in the environment; evaporation becomes a *feature* by reopening paths for re-exploration. Noted correctness argument: any treasure within the ant lifetime τ is guaranteed found. Saw Algorithm 2 as "hacking the algorithm" (turning ants into signal, so μ effectively becomes τ) rather than a new development; wished for reinforcement so signaling ants would emerge naturally.
- **Vibhav:** mapped pheromones onto agents dropping **memories in folders** (in a codebase) with a **TTL** so stale memory can't be read — keeping memory "fresh."
- **durgadas:** connected it to his search work validating primary sources with secondary sources; likes the "long tail" and environment-as-many-things angle; flagged the dependency that search span must be shorter than decay time as an interesting tuning problem. Noted the grid is idealized/game-theoretic and not fully realistic.
- **Matthew (2nd session):** found this paper more approachable than the earlier ACO paper; interested in the design space of pheromones (decay or not, multiple types, dispatch/amplification roles).

## 🔀 Questions & Disagreements

- **Chris's open question:** What is the bigger aim — how does this toy problem apply to a real-world problem?
- **Ergod's open theoretical question:** Is there an **information-theoretic floor** on how much agent lifetime you must spend to synthesize persistence out of a memoryless medium? (Motivated by the paper setting μ=1 and requiring τ bounded away from a small constant.) No answer was reached.
- **Venkatesh's cell-encoding puzzle:** Reading four neighbors gives four bits, but he believed it does *not* yield 2⁴=16 distinguishable states — perhaps only pairwise comparisons. Left open how much can actually be encoded.
- **Robby vs. the paper's approach:** mild critique that Algorithm 2 is a "hack" rather than a genuine emergent mechanism, and that lack of reinforcement is a limitation.
- **Workshop sequencing (Dan vs. others):** Dan repeatedly said he wasn't fully clear on exactly what the low-code API and "transition rules" need to be, and questioned the sequencing between a repeatable/tournament mode and the existing persistent "infinite world." Not a hard disagreement — resolved by agreeing Patrick would post specifics.
- Not resolved: whether an **MCP/LLM control layer** adds value or is "putting the cart before the horse" (Dan skeptical but open).
- Not resolved: how much state to persist for replay (every ant + pheromone?) and whether full rewind is feasible/too expensive.

## 🔗 References Mentioned

- **Amos Korman** (computer scientist) and **Ofer Feinerman** (biologist) — ANTS authors.
- **Ant Colony Optimization (ACO)** and the **De Neubourg** paper — prior sessions.
- **"Pebble" paper** and **Blum/Kozen**-style maze search (names approximate) — earlier sessions.
- **Reference #59** in the paper — information-complexity / "advice to an ignorant agent" work (posted in chat by Venkatesh).
- **DRAM / capacitor leakage** — Venkatesh's memory analogy.
- **QR codes / fiducials** — encoding metaphor.
- **Hugging Face hack** and an **OpenAI/Black Hat** talk on **honey tokens** — Venkatesh's security-attack modeling ideas.
- **Witsenhausen problem** — Ergod, re: decentralized coordination under signal noise.
- **Bartle taxonomy** (killers/achievers/socializers/explorers) — Venkatesh/Dan, for organizing game modes.
- **Quoridor**, **Frozen Synapse** — games cited as analogues (asymmetric/commit-and-simulate play).
- **π-calculus / "mobile process theory"** — Ergod, as a grounded substrate for defining agent capabilities/protocols.
- **Clerk** (SSO/auth free tier), **Arduino kits** (as a metaphor for tutorial "levels"), **OODA loop** — tooling/metaphors.

## ✅ Action Items & Next Time

- **Patrick:** post notes in the channel breaking work into agreeable chunks; specify the **low-code/no-code API** (parameter changes + transition rules), **reproducibility (fixed random seeds)**, and **detractor/spy-ant + deceptive-pheromone mechanics**. Will review PRs.
- **Agreed MVP direction:** at least one **competitive mode** and one **cooperative mode** (roughly a goal-flip of the competitive setup), plus small **single-player levels** sharing one underlying framework; build on existing **single-ant mode** and the persistent "infinite world."
- **Key design goal (Dan):** create the reactive "**OODA loop**" feel — see something happen, turn a knob, watch your colony respond in real time (deemed the core dynamic everything rests on).
- **Ergod:** over the weekend/next week, start building skeleton **components of the world** (agent capabilities, nature = maze + food + spawns) that can morph later.
- **Workflow:** use **feature branches per direction** (with separate deployments / CI-CD) rather than one mixed repo; determinism/seed work goes to the base branch; cherry-pick shared code as it emerges. Add **authentication** (possibly Clerk) to mitigate abuse/DDoS if needed.
- **Deferred:** full code-level agent control, MCP/LLM control layer, and detailed replay/rewind persistence — considered post-workshop or later.
- Dan mentioned a newer prototype, **"Cave Diver"** (underwater cave-diving with a pebble-like guide-leaving mechanic) he finds richer metaphorically; will share a link. To be considered separately.
- **Next call:** next Friday.

## ⭐ Memorable Quotes

- **Ergod:** "Forgetting is bad for one-shot stuff and good for perpetual coverage."
- **Venkatesh:** "It's insecure by design. It's just open to everybody... the security is not enforced in a boundary. It's at the level of each cell."
- **Dan:** "In that way, we would learn what it feels like to be an ant."

*Note: the transcript's speaker attribution was heavily interleaved (Patrick's and Venkatesh's turns, and later Dan's and Ergod's, were spliced together by the auto-transcription). Attributions above reflect the most coherent reading, but some fine-grained assignments are uncertain.*