# SIG-FPT 2026-10-16 β€” Session Notes Participants: - Venkatesh Rao | UTC-8 - Patrick Nast | UTC-7 - Robby Rao | UTC - 7 - Mike Travers | UTC-7 - Giovanni Merlino | UTC+2 - Aayush Kucheria - Dan Schmidt | UTC-7 - SebastiΓ‘n AristizΓ‘bal | UTC+2 - rafa (UTC+1) --- ## πŸ“– The Reading Not clearly identified in the discussion. This session was not a reading discussion but a retrospective and planning meeting for the Formal Protocol Theory (FPT) research group at the Protocol Institute. Several external works were referenced in passing (e.g., a "Swarm World" paper, the Hugging Face / OpenAI agent incidents), but no single assigned reading anchored the conversation. ## 🧭 Overview The group debriefed on a recent ~48-hour stigmergy/security workshop, then sketched a 2027 research roadmap across three tracks (the StickSim simulator, an AI-safety "research questions" track, and a cryptography/theory track). Venkatesh Rao framed the broader "new nature" thesis (AI vs. cryptography) and the Protocol Institute's model of AI-leveraged part-time research groups. The final stretch was a working session on a funding proposal to Softmax (Emmett Shear's company), debating how to evolve StickSim into Softmax's "co world" platform and what concrete AI-safety/benchmarking questions the project should answer. ## πŸ’‘ Key Points & Themes **Workshop retrospective** - **Patrick:** The workshop was an ambitious but largely successful experiment; most participants spent their time getting up to speed on the summer's "through line" of papers rather than completing full-scope projects. ~12 participants attended out of ~18 signed up. Highlighted ZJ's continuous-stigmergy ant simulator (which parameterized things like signal diffusion) and Anderson's talk on "coordination horizons." - **Patrick:** Future events need a longer or more gradual learning curve than a single crash-course session. - **Venkatesh:** First session was an info-dump crash course; last was lightning talks. It ended up more exploratory than hackathon-style. Recordings exist and will be processed and uploaded. - **Venkatesh:** His main interest was connecting the insect-centric stigmergy model to (a) AI and (b) security. He distinguished "untranslated" insect security concerns (faking/spoofing pheromone trails) from phenomena like the Hugging Face incident (improvised message boards, discovered credential tokens) that have no biological counterpart β€” "ants don't go around stealing passwords." **Defining stigmergy** - **Patrick:** Two core criteria β€” externalized memory, plus a feedback loop (traces prompt further trace-laying). Also a notion of bandwidth-limited communication (agents precluded from direct messaging, as in the OpenAI incident). - **Venkatesh:** Proposed generalizing stigmergy beyond simple insects to any agentic behavior over externalized memory, dropping the simplicity restriction β€” e.g., attaching an LLM "hindbrain" (~100k token context in a loop with Claude) to each ant. He reframed stigmergy as an expansive theory of memory: private/public archival memory, real-time memory, and trust-modulated shared memory (where overwrites are interpreted via trust relationships β€” as when OpenAI agents suspected a "traitor"). - **Venkatesh:** Bridging insect-framed security ("prevent anthills from getting captured") with human-framed AI safety is the interesting gap. His lightning talk proposed ~15 dimensions of AI safety, from abstract alignment to hard encryption. **Roadmap / three tracks (2027)** - Track 1 β€” StickSim simulator (built by Dan Schmidt): porting to Softmax's "co world," scaffolding for future development. - Track 2 β€” Patrick's track: using the simulator to pose and answer AI-safety research questions; ideas include adding selfishness/mixed-motive goals, and per-ant "decision agents" (cheap classifier-like models, e.g., JEV, OpenAI's, Microsoft's) given swarm scale. - Track 3 β€” Theory: Urgaard's category-theory work on formal protocol models, plus a proposed self-study track on classical and post-quantum cryptography. **Funding & institutional context** - **Venkatesh:** First pitch meeting held with Emmett ("Amit"/"Ahmed" in transcript) Shear, CEO of Softmax, which offers infrastructure for simulation environments ("co world"). Softmax may provide free infrastructure and possibly funding, and connect them to other funders. A formal proposal is due in a few days. - **Venkatesh:** Broader "new nature" thesis for the Protocol Institute β€” digital infrastructures (AI and cryptographic protocols) as "irresistible force vs. immovable object." Recent events (OpenAI results, Hugging Face hack, cryptography concerns) have shifted this from a 500-year horizon to a 5–10 year urgency. Mentioned the four "crypto futures" framing (PKC works β†’ minicrypt β†’ ... β†’ P=NP). **Softmax proposal working session** - **Dan:** The essence is Emmett's onboarding progression β€” player graduates from single ant β†’ policy modifier β†’ queen β†’ "god"/environment-shaper β€” which bakes in friction for a permissionless system (no uninformed person "mucking with the game"). Current game already has two hypotheses: a "war mode" and a persistent "infinite world" (survive longest). Ambition includes scaling to ~1,000 participants as a test bed for swarm mechanics, with human-agent collaboration (humans turn knobs manually or delegate to a coding agent). - **Venkatesh:** Suggested a concrete onboarding mechanic: unlock one of ~12 simulation parameters per maze completed; the order of unlocking (e.g., velocity vs. pheromone amount) could reveal which parameters dominate successful strategies β€” "paving cow paths into protocols." Humans role-playing ants act as a Mechanical Turk substitute for LLMs. Also proposed an open plug-in interface for bespoke experimental designs (e.g., replacing pheromones with 100-byte markdown scratchpads at coordinates). - **Patrick:** Key open question β€” is the simple parameterized ant model the right substrate? A single pheromone is too weak; a centralized "queen/governor" conflicts with decentralization ideals; arbitrary free communication makes it non-stigmergic (too much information). Proposed a middle path: ants leave *policies* for each other to pick up β€” a blend of evolutionary algorithms and stigmergy that also enables "bad actors" laying hack policies, requiring judgment about which to adopt. - **Venkatesh:** Proposed structuring the proposal as a three-phase staged roadmap (phase 1: port + evaluate 2–3 scaffolding architectures; phase 2: build out + prototype experiments; phase 3: the 1,000-person/agent run) to defer uncertainty. Suggested loosening the word budget so each section gestures at this staging. **Candidate "six-month" research questions** - Option 1: Model specific atomic incidents (e.g., the Hugging Face hack β€” breaking containment, finding credentials) inside StickSim. - Option 2 (Emmett's): Benchmarking β€” run tournaments where ants are powered by competing frontier models (e.g., "Opus 5.5" vs. "Fable") to see which solves games better; test whether bigger models actually win, or whether cheaper/classifier models (system-1-style) outperform in context. - **Venkatesh:** Argued stigmergy could be a *less gameable* benchmark than saturated benchmarks (MLPerf, math sets) subject to Goodhart's law, analogizing to military "open-play war games" as a gold standard β€” plurality of human+agent attackers makes it harder to overfit. ## πŸ”€ Questions & Disagreements - **Is the parameterized ant model the right level of abstraction?** Patrick repeatedly pressed this; he's comfortable if the answer is "yes, as a base case," but wants the organizing *research question* specified before downstream design. Venkatesh favored starting simple and growing via plug-ins. - **Direct modeling vs. analogy:** Patrick asked whether they model stigmergy directly and then analogize to AI agent swarms, or jump straight to richer LLM-based communication. Left open. - **Centralization tension:** A queen/governor controller feels at odds with protocols being notionally decentralized β€” hence Patrick's "ants leave policies" proposal. - **Is stigmergy-as-benchmark actually less capturable?** Patrick challenged Venkatesh directly ("how is it not capturable? Can't they just reinforcement learn for that also?"). Venkatesh acknowledged it's a hypothesis, invoking plurality/non-stationarity and war-game reasoning; Patrick partly conceded that non-stationarity makes it harder to train against. Unresolved. - Giovanni asked about trade-off/constraint mechanics (hostile conditions, limited movement/pheromone budget). Patrick clarified there's no direct move-vs-pheromone trade-off currently; ants have a fixed pheromone tank and a metabolism. - Note: the transcript's auto-generated text renders the Softmax CEO's name inconsistently ("Amit Sheer," "Emmett," "Ahmed"); this appears to refer to Emmett Shear of Softmax, but the exact attribution is garbled in places. ## πŸ”— References Mentioned - Softmax (company) and its "co world" simulation platform; Emmett Shear (CEO). - StickSim / "StigSim" β€” the group's maze-based ant stigmergy simulator (built by Dan Schmidt). - The Hugging Face incident and the OpenAI agent/"artifactory server" incident (agents coordinating via discovered exploits). - "Swarm World" paper (posted in the FPT chat) β€” cited by Patrick as an ambitious stigmergy game model. - Decision/classifier agents: "JEV," OpenAI's, and a Microsoft decision model. - Eigen Labs (CEO Sriram Kandanan) and its "Yukon" collaborative agentic research platform (crypto, model training, biology). - A "10,000 AI researcher village" project (mentioned via Ahmed); and an "AI village"-style simulated-agent environment. - Recent "OpenAI dump of match results" and a Google quantum computing result; concerns about elliptic-curve / public-key cryptography; post-quantum cryptography; quantum key distribution hardware (Giovanni's group, Messina). The "four futures" / minicrypt / P=NP framing. - Protocol Institute publications: *Protocolized* (Substack + print), *The Protocol Reader* (book), and the spun-off sci-fi magazine *Monstrous Times*. Target positioning between IEEE Spectrum and Scientific American. - "Open Recapper" β€” the Discord recording/transcription bot (built by Anurag). - Kyle Matthews β€” proposing a possible "future of programming" research group. - Concepts: Goodhart's law; category theory (Urgaard's theory track); stigmergy; coordination horizons. ## βœ… Action Items & Next Time - **Venkatesh:** Process and upload the four workshop session recordings to YouTube. - **Venkatesh, Patrick, Dan, (Urgaard):** Finalize the Softmax proposal async over the next few days and send it to Softmax. Restructure it into a three-phase staged roadmap; each author's section should gesture at phase 1/2/3. - **Patrick:** Continue to nail down the specific AI-safety research question that organizes the simulator work. - **Giovanni:** Has one thesis student starting on stigmergy topics and another on gamification; will review shared documentation with them. Open to involving a PhD student in running Softmax experiments if funding comes through. (Venkatesh: PI citation/acknowledgment welcome, though formal academic publishing is a second-order priority.) - **Group:** Watch for cryptography/quantum learning opportunities and people to bring into a possible study track; explore cross-group collaborations (esp. with the robotics group's proposed experiment arenas). - **Mike:** Proposed a weekly Institute-wide internal seminar with research groups taking turns presenting β€” Venkatesh liked it and will think about how to operationalize it. - **Next meeting:** Regular SIG/FPT call in two weeks. ## ⭐ Memorable Quotes - **Venkatesh:** "Ants don't go around stealing passwords." - **Venkatesh:** "AI versus cryptographic protocols is a little bit like irresistible force versus immovable object... AI seems to be able to eat everything except strong encryption, and now even that has a question mark against it." - **Patrick:** "It's just... they have too much information with which to work" β€” on why fully free agent communication stops being interestingly stigmergic.