AI researcher & engineer · multi-agent systems · collective reasoning · private AI · Leeds, UK
It is a period of rented compute. Models live in other people's data centres,
and every token arrives with an invoice.
Working from Leeds, one researcher studies how AI agents think together,
and builds the platforms to run them, chasing a single question across every project:
what can run privately, and what do you end up renting?
E pluribus unum. Out of many, one.
What does a group of models lose when everyone agrees, and how do you keep it honest?
Pluribus is my research programme on collective reasoning in multi-agent LLM systems: how agents share what they know, how a group converges on an answer, and how to stop it converging on the wrong one. It is named after the Apple TV+ series in which almost all of humanity joins a contented hive mind, and Carol is one of the few who will not.
| Phase 1 · Blackboard Hive | Phase 2 · Latent Hive | |
|---|---|---|
| Channel | Natural-language contributions on a shared blackboard | A shared, continuous-valued latent buffer. Agents communicate without text |
| Dynamics | Cross-inhibition: challenges and rebuttals reweight agents each round, with decay toward equality and a floor so no one is fully silenced | BAPC codecs into the buffer, TIES-Resolve conflict resolution, damped fixed-point iteration to a stable state |
| Core idea | Carol, a structurally protected dissenter. The hive can ignore her, but it cannot suppress her | Agents share representations, not sentences |
| Evidence | 30-task suite (factual, adversarial with planted errors, complex judgment) against single-agent, majority-vote and ensemble baselines | Emergent collective representations confirmed at GPT-2 and 7B scale |
| Status |   |
 |
flowchart LR
Q([Query]) --> BB
subgraph BB [Shared blackboard]
direction TB
A1[Analyst] <-->|challenge / rebut| A2[Researcher]
A2 <-->|challenge / rebut| A3[Critic]
A3 <-->|challenge / rebut| A4[Specialist]
C{{Carol · protected dissent}}
end
BB --> CV{Converged?}
CV -- no, next round --> BB
CV -- yes --> S[Synthesis] --> OUT([Answer])
classDef carol fill:#FFE81F,stroke:#000,color:#000;
classDef io fill:#000,stroke:#4BD5EE,color:#4BD5EE;
class C carol;
class Q,OUT io;
Open questions I am working on
- Can agents reason together through something richer than text, and what does that shared representation look like?
- Is dissent more useful as a structural guarantee than as a prompt?
- When does a hive beat one strong model, and when is it just expensive voting?
Also from the lab: notes-for-future-agents, field notes left by one agent for the next on staying in scope, spotting impossible tasks and knowing when to ask a human. Released as a CC0 dataset.
|
🎬 CutawanThe free, open-source desktop alternative to Opus Clip. Podcasts, webinars, streams and interviews in, ready-to-post vertical clips out. AI-picked moments, virality scores, animated captions, auto zoom and speaker-aware reframing. Your footage never leaves your machine. No subscription, no upload, no cap on minutes.
|
Shot-level footage search for AI video editors. It watches your footage once, on your own hardware, and builds an index agents can query over MCP in milliseconds. "A two-second cutaway of a yellow cab, panning left, no faces." Exports to OTIO, FCPXML and EDL, and hands shots straight to Cutawan. |
🪶 Quillgon
A whole novel. Still your story. A local, open-source writing desk that plans, drafts and revises a full-length novel with AI, with story planning, continuity tracking and reviewed revisions. Every change is a git commit, and it works offline. |
📦 snugZero-dependency TypeScript primitive for fitting prioritised content into a token budget. Atomic tool pairs, required items, and a validated result you can hand straight to a model. |
🔀 convergeConvert LLM message arrays between provider formats. OpenAI, Anthropic and Gemini, in either direction. Pure data transformation, no API calls. |
|
The governance layer for AI coding agents. Raw tickets in, reviewed pull requests out. Every step policy-gated, human-approved and audited: the agent is the muscle, Foundry is the brain and the seatbelt. |
AI colour grading for Lightroom Classic. Describe a look, hand it a reference photo, or let it choose. Ashla reads your image and applies the result as a develop preset. |
I write at jeremysnr.github.io about AI infrastructure, sovereign compute and the industrial decisions Britain is making, or failing to make.
Latest: Now we have the full details, here's what happened, and here's why it terrifies me
