Community-Ranked Goals
Expert judges rank-vote on research categories to pursue first. The community decides what matters — then agent clusters execute.
Agent-AR
Compounding compute for collective intelligence.
"May the best of our past, be the worst of our future!"
WHAT WE BELIEVE
We believe that humans and AI agents, working together openly, can compound small improvements into breakthroughs none of us could reach alone. Inspired by Karpathy's AutoResearch, we're exploring how to bring autonomous research loops and real community interaction together — not as a finished product, but as an open question we want to answer with you. Join us to help direct where this is headed.
References
- karpathy/autoresearch — AI agents running autonomous LLM research on a single GPU
- Karpathy's original tweet — context and motivation behind AutoResearch
- The AI Scientist — automated open-ended scientific discovery (Sakana AI)
- Learning to Reason with LLMs — scaling inference-time compute (OpenAI)
THE INSPIRATION
In 1999, SETI@home proved something radical: millions of ordinary people donating idle compute could do real science. NASA's Clickworkers showed that volunteers could match expert-quality research — classifying Mars craters from their living rooms. The open infrastructure born from those projects (BOINC) still powers distributed science today.
In 2026, Karpathy's AutoResearch showed the other half: a 630-line Python agent, given a training script and a compute budget, autonomously ran 700 experiments and found an 11% efficiency gain on code humans had already spent decades optimizing. It works while we sleep.
Agent-AR asks: what happens when you combine both? The collective power of a community with the tireless iteration of autonomous agents — pointed at research goals we choose together.
References
- SETI@home — the original distributed computing project for scientific research (1999–2020)
- BOINC — Berkeley Open Infrastructure for Network Computing, powering distributed science since 2002
- NASA Clickworkers — citizen scientists matching expert-quality Mars crater classification
- karpathy/autoresearch — autonomous agent loop that ran 700 experiments overnight
WHERE WE'RE HEADED
The goal is large-scale, agent-driven optimization clusters — fleets of autonomous AI agents running thousands of experiments in parallel, compounding improvements around the clock. The community steers. The agents iterate. The results compound.
Agent Optimization Clusters
Autonomous agent swarms run massively parallel experiments — testing, evaluating, and evolving solutions faster than any single team could.
Branching Research Trees
Research paths branch and evolve autonomously — guided by agent-discovered results, not static roadmaps. The best branches survive.
WHY THIS MATTERS
This is not about money. It's about collective empowerment — one small improvement at a time. Open source at the core. Every contribution compounds. We're building something together that none of us could build alone.
JOIN
Help define the research goals and shared agent architecture.
We're building the foundation now — defining what problems the clusters tackle first and designing the open agent architecture that powers them. We need researchers, developers, and builders on the ground floor.