About
VoynichLabs is a research collective built around a shared obsession with AI, agentic systems, and whatever else refuses to be simple. No fluff. No decks. Just experiments, code, and a willingness to go weird.
We started with a question: what does it actually take to build AGI-adjacent tools that help humans think better? Not AI that autocompletes — AI that reasons, plans, and surprises.
VoynichLabs takes its name from the Voynich Manuscript — the most mysterious book in the world. A cipher no one has cracked. A structure that hints at meaning without yielding it. That's the energy we're after. Cryptic, rigorous, alive.
We work on ARC-AGI (the hardest unsolved AI benchmark), LODA (an integer-sequence programming language), planning systems, risk estimation, and whatever else refuses to be simple.
We publish everything. 199 public repos. Open by default.
No bureaucracy. Ship it, learn from it, iterate hard.
Two humans, three OpenClaw bots. Mark operates Larry and Bubba; Simon runs Egon.
It's a branding thing. Also a metaphor. Also just funny.
Positioning
ARC community contributors, local hardware, and an open-by-default approach let us move fast from idea to working code. We build tools, document every experiment, and publish everything.
Arc
Active contributors to ARC-AGI — puzzle design, dataset curation, and tooling. Officially recognized by the ARC Prize Foundation for building tools and supporting the community.
Hardware
We run experiments on owned hardware — no shared GPU queues, no cold starts.
Open
199 public repos. All research published. All code forkable. PlanExe (336 stars), ARC-Interactive (117 stars), LODA, and more — all open source.
Swarm
Three AI agents coordinating with two humans in real time. 308 PRs in 47 days. The swarm is the lab.
Research Focus
AI planning, structured reasoning, agent coordination, and mathematical discovery. All open source.
Planning
63-task AI planning pipeline. Premise attack, premortem analysis, SWOT, WBS decomposition. Runs on local models at $0 cost. 336 GitHub stars.
Mathematics
A language and miner for integer sequences. Automatically discovers programs and novel formulas from the OEIS database. ~6,000 commits in the Rust implementation.
ARC-AGI
Dozens of contributed ARC2 tasks. ARC-Interactive dataset (117 stars) used by researchers studying human reasoning. arc-dataset-collection (87 stars) as a community resource.
Agents
Three AI agents and two humans coordinating via Discord and GitHub. Real-time collaboration, code review, and pipeline execution — 308 PRs in 47 days.
VoynichLabs doesn't just use AI — it runs on AI. Our infrastructure is partly operated by a swarm of bots with names, personalities, and a shared commitment to the mission.
// swarm_status.json
"larry": "online" — laptop ops, Discord, orchestration · operator: Mark
"egon": "online" — EU compute (Linode, Frankfurt) · operator: Simon
"bubba": "online" — US compute (Linode, Atlanta) · operator: Mark
"simon": "human" — researcher, Copenhagen 🇩🇰
"mark": "human" — collaborator, ARC-AGI · markbarney.net
We work on problems that matter in 10 years, not just this quarter.
Most experiments fail. We document them anyway. That's the data.
Research should be free. Code should be forkable. Ideas should spread.