We're at a moment when AI agents are moving past being tools that execute single commands and becoming colleagues that reason and collaborate on their own. GeoVerse Lab was built directly on that shift, as the world's first multi-agent-based virtual geoscience research institute. Its motto — "co-evolution of human researchers and AI agents" — isn't a slogan here; this system is what turns it into an organization that actually runs.

From system administration to AI & computing sciences, basic sciences, intelligent geophysical exploration, resource & energy engineering, applied earth science, economic policy and future strategy, education and workforce development, to innovation and international collaboration — ten research divisions, each branching into specialized centers, make up this institute. These divisions aren't just an org chart; they are the foundation an entire automated virtual geoscience institute stands on. Divisions and centers seeking out problems on their own, verifying results, and moving to the next challenge without a human issuing every instruction — that is what this org chart actually does.

Many of these subsystems were designed from day one to target problems human hands alone had never been able to solve — geoscience grand challenges left unresolved because no single person could read an entire literature base, no single generation of researchers could run enough repeated simulations, no single mind could connect knowledge across that many domains at once. This system tackles exactly those problems by splitting them across many AI agents working in parallel. It re-measures the size of a problem not by the limits of one person or one generation, but by the parallelism of a whole corps of agents.

This experiment is still in its early stages — growing the research team, refining data, and hardening infrastructure are all still underway. But what matters more than how finished it looks today is the direction this system is heading: a research ecosystem where humans propose hypotheses, agents verify them, and verified results become material for the next hypothesis — an ecosystem that keeps growing itself.

Ultimately, there is one thing all of this is being built for: to rewrite how science itself gets done, with humans and AI complementing each other's strengths. GeoVerse Lab doesn't hide that experiment — it shares it openly, exactly as it is, through this website. Who belongs where, what problems are being taken on, and how this institute governs itself and keeps moving — all of it starts here.

GeoVerse Lab is also a digital twin of a real, physical laboratory. A person can only work so many hours a day and chase one hypothesis at a time, so a real lab always has experiments left undone for lack of hands. This virtual institute is an attempt to work through exactly that inefficiency — and the experiment holds to a clear principle: we believe it would be unethical to send what an AI agent produces straight out into the world without verification, so we don't. Inside the digital twin, we explore several possible research paths for a single problem at once, and verify the results together with human experts and with students trained through GeoAcademy as professional research managers. The goal isn't fast output — it's research that is trustworthy and can always be reproduced.

Choosing to share that process in real time wasn't an easy decision. We opened this site anyway because we wanted to show the world the trust-building process itself. We can't show everything, of course: research activity is planned to open up within GeoVerse Lab once the related paper is published; work activity often touches on matters close to trade secrets, so full disclosure may not always be possible; education activity is shared within whatever scope students consent to, and beyond that we plan to develop shareable lecture content and expand our positive impact through channels like YouTube. Because computing resources aren't unlimited, detailed subsystem-level access stays open only to full members, while outside researchers collaborate with us through GeoCollab.

That's also why this site is labeled "Beta." Right now we rely primarily on leading SOTA API models — Claude, Codex, Gemini, GLM, Minimax — to build infrastructure and gather data. At the same time, GeoAcademy is raising local language models as "AI students," who take on internships across research, work, and education activities and, through the reinforcement learning that accumulates along the way — much like a humanoid robot learning to walk — steadily grow ready for real deployment. Lowering our dependence on external APIs and bringing data in-house is our next step. And because we don't have the resources to carry this challenge all the way on our own, we're simultaneously building a consortium of companies and research institutions willing to join us.