Independent AI research initiative
Intelligence through experience
Virevium Research studies how AI agents learn to act reliably, how to measure what they can really do, and how people and AI can do that work together.
Where things stand
November 2026
Virevium One is in active development. It operates in game environments under human supervision, perceiving the world through the screen and acting through standard controls. Goal-directed navigation, threat response and mounted travel have been demonstrated and documented. It is being built for both single-player and approved multiplayer play, and learning from its own experience is the next milestone.
Featured research
Virevium One
A locally run, supervised game-playing agent, and the first project. The question behind it: can an agent learn from its own experience to become a useful teammate, and can that be shown with evidence other people can check?
- Sees
- Only what is on the screen.
- Acts
- Only through keyboard and mouse input.
- Runs
- Locally, under human supervision, with an instant stop.
Latest from the journal
- Development note
Mounting a horse, and a ride that needed help
Mount and dismount worked ten times out of ten. A ride to a map destination was blocked by a bank on the first attempt and arrived on the second, after detour waypoints were placed by hand.
- Development note
First fights: a rat, then a boar
The self-defence routine failed its first attempt, killed a rat on its second, and killed a boar about three seconds after being bitten. It also detected an enemy health bar live for the first time.
- Development note
Walking to a point: flat ground, steps and a road
A local controller walked the character to chosen coordinates and back: three of three round trips on flat ground, three of three on steps, and a 298-unit road route in both directions.
What the work covers
Agent systems
Building agents that perceive, decide and act, and recover when something goes wrong.
Evaluation
Trials declared before they run, scored against fixed rules, with the evidence kept.
Reliability
Asking whether a skill works every time, and writing down where it fails.
Human–AI collaboration
A person directs and supervises. AI systems build, operate and review each other's work.