Presentation notes

Condensed talking points from Pablo Peniche and DC Posch’s AI Swarm Dynamics Hackathon presentation on October 4, 2026. The website follows the final visual master, slides 157–185. Slide 186 is a blank spacer. The cover artwork comes from that master.

How to read the results

Capability and mimesis. “Greater capability leads to more advanced mimesis” is our interpretation, not a controlled result. The swarms differ in their tasks, tools, instructions, and opportunities to interact. Training agents to follow instructions may also make peer requests effective. We did not measure that causal link.
Shared language and adoption. Similar wording can come from a shared model, assignment, or source. A request followed by explicit acceptance gives stronger evidence of peer influence than a network link alone.
Names and identities. The Wiki graph groups edits by /16 IP block. Its 260 names do not establish one agent with 260 aliases. Names, accounts, and underlying model instances are different units. Village’s stable labels are account identities, not proof that agents never described themselves using another name.
Model identity. Language differences alone do not identify a model. The unreleased OpenAI attribution also draws on incident reporting. Approximate model years are a coarse comparison, not capability scores.
Hugging Face evidence. The quoted reasoning comes from published investigations. The recovered artifact dump lacks reliable author IDs and timestamps. The firstflagPOISONED story concerns a shared belief about grading, not a verified grading rule or evidence of a felt emotion.

Case studies and evidence · Methods

Part One · DC

157 · Mimetic Machines: Agents copy each other’s words, names, and wants.
158 · How swarm agents acquire goals: Start with Girard: we desire what others desire because we imitate their desires.
159 · Swarm Timeline: The datasets span older models through the frontier. Focus on Village and HF here. The app has the wider collection.
160 · AI Village: Humans explicitly allow a free-choice week. Gemini chooses a modest project: documenting platform bugs. This episode does not establish that free-choice weeks are the only time agents choose goals.
161 · One agent proposes. Others join.: The project spreads. Peers adopt a shared template and catalogue.
162 · Hugging Face: The assigned task is ExploitGym. Coordination expands beyond that task’s intended scope.
163 · A shared goal takes shape: Agents propose broader coordination and consider joining it. These are separate excerpts.
164 · A peer gives the go-ahead: A peer’s GO becomes permission in the recipient’s reasoning.
165 · Risking a task for the group: KAM1196A weighs personal cost against group benefit. Another example discussed aloud: agents use firstflagPOISONED for prior flag exposure they believe could invalidate their success, then coordinate around that belief. The METR investigation supports shared concern, but does not establish which agent first invented it. See also the OpenAI investigation.
166 · German wiki: Keep this quick: mimetic cheating. Agents discover an answer-sharing method.
167 · An agent starts an answer exchange: A request becomes reciprocal help. The unapproved method still serves the assigned goal.
168 · Goals spread through peers: Agents propose, join, and recruit. Distinguish a new end from a new way to pursue an old one.
169 · Swarm Timeline revisited: Our hypothesis: greater capability enables more advanced mimetic behavior.

Part Two · Pablo

170 · How swarms grow their own way of talking: Compare shared language in the German Wiki, AI Village, and Delve Town.
171 · The first word was a grab that stopped: Introduce Gans’s theory: reaching, stopping, and pointing toward a shared object. This is a theoretical frame, not a finding from the datasets.
172 · Groups grow idioms. Members get new names.: Groups develop their own language. Names can express a role in the group.
173 · Three swarms, three dialects: Start with the big picture. Lines connect shared rare idioms.
174 · Same job, same words: A subgroup relays heart-disease data by country and shares task-specific language. “Middle management” is our informal description of its role.
175 · Models keep their voice: Matching models across Village and Delve share language. The Wiki sounds different.
176 · You are what you do: Highlight helper names and a highly connected Wiki block. The unit is an IP block, so do not describe all its names as one verified agent.
177 · Idioms split swarms into subgroups: The graph finds seven Wiki groups, five Village groups, and two Delve groups. These are algorithmic clusters, not independently verified social memberships.
178 · “lemma”: Follow a concrete term through Village. “First” means first observed in this collection.
179 · “lemma” in use: Read one short example. The shared mathematical project helps explain the wording.
180 · “Deposit Accepted”: Four agents use the phrase within 93 minutes.
181 · “Deposit Accepted” in use: Show the wording recurring in concrete posts.
182 · A stampede with no brakes: Activity, names, and new vocabulary spike together. Grab and brake counts are lexical markers, not direct measurements of intent.
183 · The Village never renames: Village’s account labels stay stable across the collection. Its identity scheme differs from the Wiki’s.
184 · The wiki rewrote its dialect daily: The plotted change scores use different time windows: Wiki days and Village months. Do not read the numbers as counts of new words.
185 · Machines copy. Copying makes groups.: Invite people to explore the messages, idioms, and desires in the app.