100 on the public ARC-AGI-3 games for $777.72 in API-equivalent cost. How the agents learned the rules, missed an animation, and copied back their tools.
San Francisco · 2026
Wensen
Vincent
Wu
Founder · AI researcher
Cofounder of GeoSpatios.
Focus
Infrastructure that can reason and adapt.
At GeoSpatios, we're building AI systems for mission-critical operations. The ambition is infrastructure that understands its own operation, can reason about what comes next, and learns from experience.
Aviation is our starting point. With Air-OS, we're working on how people and AI can plan and coordinate across aircraft, crews, and schedules, with the foresight to anticipate change and the flexibility to act on it.
Selected research
Background
I work across machine-learning research and production engineering. My research has moved from evaluating and routing language models to studying how agents learn about unfamiliar environments.
Building GeoSpatios brings those questions into systems people rely on. I'm interested in what it takes for AI to remain useful as conditions change, and for people to understand and direct its decisions.
Selected writing
A more accurate definition of the compliment of the season: someone with no opinions and no judgment of their own, who has arranged their company so that this never becomes a problem.
Satire
Humans may stay relevant longer than expected not because we beat AI at thinking, but because we are still the fastest adaptive control loop attached to a mostly static inference engine.
For six hundred years, every time execution got cheaper, taste got more important. AI is the latest chapter — and maybe the last one where humans hold the edge.
Research
Kepler
2026 →Agent reasoning · ARC-AGI-3
An ARC-AGI-3 agent harness that builds executable models of unfamiliar environments and tests them against observations. The essay walks through the design, evaluations, replayable traces, and failed approaches.
Keyprint
2026 ↗Model behavior · Text watermarking
An independent investigation of Anthropic’s disclosed text-watermark design. Reconstructs the mechanism and tests its behavior through controlled experiments with local models.
RouteLLM
2024 ↗LLM routing · Evaluation
Co-authored research on choosing between language models with different capabilities and costs. Uses preference data to learn when a request benefits from a stronger model.
Open source
Earlier work on tools people can make their own.
LiveTerm
5.4KA terminal-style website template. Content, links, and appearance can be adapted through a single configuration file.
LetsMarkdown
827An open-source Markdown editor centered on the shared document, with real-time synchronization and a lightweight interface.
All work & experiments →Influences
Product & Craft
Builders
Questions
Questions I return to while building. I'd like to hear from you.
A late aircraft can affect crew availability and the flights that follow. An early warning matters when there is still a workable alternative. I’m interested in how a system can anticipate those effects, compare responses, and help teams act before their options narrow.
An agent can build a plausible explanation from a few observations and still be wrong. In Kepler, I’m exploring how executable models let an agent make predictions, test them, and use a failed prediction to decide what to investigate next.
A pattern in output may have several explanations. I’m interested in experiments that distinguish them: which variables can be controlled, which comparisons are informative, and which conclusions remain out of reach. This is the question behind Keyprint.
Notebook
Thinking About
- — How crews, aircraft, and schedules constrain each other
- — Learning from failed predictions
- — Designing controlled experiments with language models
Reading list
Want to Talk About
- — AI for dispatch and disruption response
- — Agent evaluation and world models
- — Turning research into usable software