Video Games Built for AI Agents Are Now a $2.3 Billion Bet

A new strategy game shipped this year built around one rule: no human is meant to play it. LLM Skirmish drops autonomous AI agents into real-time combat, forcing them to manage scarce resources and outmaneuver rivals with no scripted moves to fall back on, according to the project’s own website.

It is a small entry in a much bigger wave. Labs and startups are building entire game worlds for machines rather than people, betting that play is the cheapest way to raise a sharper kind of AI. One of those bets, a startup called General Intuition, was just valued at $2.3 billion for teaching agents to learn from Fortnite footage instead of a lab.

Built to Play Itself

LLM Skirmish is built for autonomous agents first. Its own project site describes a game where machines must make split-second combat calls, manage resources and adapt to unscripted moves with nothing scripted to fall back on. There is no campaign mode designed for a person to sit down and enjoy.

None of this is new, exactly. Games have been recruited to test machine minds since the earliest days of computing. In 1950, Claude Shannon used chess to guide the design of early computer algorithms, decades before Deep Blue or AlphaGo existed.

By 2019, OpenAI’s agents had taught themselves surprising, unscripted strategies for winning at hide-and-seek inside an open-world simulation. Three years later, Sony’s GT Sophy learned to out-drive the world’s best Gran Turismo racers, not by following a script but by grinding through repeated races and physics data.

The Billion-Dollar Bet Behind the Screen

The clearest sign this has moved past a research curiosity showed up this month in a funding announcement. General Intuition, a startup spun out of the game-clip platform Medal, said it raised $320 million at a $2.3 billion valuation, pushing its disclosed funding to $454 million since it launched last October.

The company does not really sell a game. It sells what it calls the gym: a simulated world generated frame by frame from millions of hours of human Fortnite footage, rather than rendered by a traditional engine, so an agent can pick up basic physics on its own, like the fact that walls block movement and shadows lengthen as the sun moves.

Vinod Khosla, the venture investor whose firm led the funding round, framed the shift in stark terms. “In world models, I think the quantum leap is the emergence of intuition in the AI, a human intuition-like capability,” he said. The company’s founder, De Witte, describes the approach more simply: the next stage of AI pretraining itself, built on reaction data no text corpus contains.

De Witte has his own history in gaming. He made $1.5 million running a private RuneScape server as a teenager. He has also launched a jobs marketplace called Nerve that pays gamers for data labeling and, eventually, robot teleoperation work, a hedge against the very displacement his own technology could cause.

Each project treats games differently: some as a proving ground, some as a data mine, some as both at once.

Why Can’t AI Beat a Brand New Game?

Not reliably. A model can master a specific game after enough exposure, but researchers say true general game-playing, learning a brand-new title roughly the way a person would, in hours rather than months of simulated attempts, remains far beyond today’s systems, no matter how impressive the highlight reel looks.

Researchers sent reasoning models to play Super Mario Bros. last year, and timing decided everything. Models can take seconds to choose a move, long enough that a cleared jump turns into a fatal drop before the decision even lands.

A separate study led by New York University’s Julian Togelius found that even the strongest systems cannot yet learn a new game in tens of hours the way a skilled human would, a gap the researchers treat as the missing piece before calling any system generally intelligent.

• Vinod Khosla, the venture investor behind General Intuition’s funding round, argues gameplay footage captures human reaction and judgment that text alone never could.
• Andrej Karpathy, a research scientist and founding member of OpenAI, has written that he does not know “what [AI] metrics to look at right now,” calling the moment an evaluation crisis.
• Julian Togelius and his co-authors counter that headline wins reflect narrow mastery of a single game, and that the field will not know whether it has built general intelligence until an agent can pick up something totally unfamiliar almost as fast as a person can.

Every side of that disagreement is still funding its own experiments to prove the others wrong.

Robots Are Quietly the Bigger Prize

Google DeepMind’s contribution points in a related but separate direction. Its Scalable Instructable Multiworld Agent, known as SIMA, was built to follow natural-language instructions across many game worlds rather than mastering just one.

Forbes has reported that DeepMind trained SIMA across nine separate game worlds, from No Man’s Sky to Goat Simulator, and found that agents trained on eight of those games performed better on the ninth, unfamiliar title than specialists built for it alone. That transfer, a skill learned in one world showing up in another, is the entire pitch for using games to train something meant to eventually work outside them.

The same logic underpins General Intuition’s pitch to investors. An agent that has learned to judge distance, momentum and physics from millions of hours of Fortnite clips should, in theory, need far less real-world data to control an actual machine. Smaller studios stand to gain from a related shift, too. One widely cited analysis argues cheaper AI tooling could let solo developers and small teams generate code and art in a fraction of the time larger studios need, narrowing a gap that has favored big-budget productions for decades.

The People Left Holding the Controller

Not everyone in the industry is cheering. A 2026 State of the Game Industry survey found 36% of developers already use generative AI tools, but 52% think the technology is hurting the industry overall, up sharply from 30% a year earlier. Only 7% called its impact positive.

• Creative jobs: more than half of surveyed developers call generative AI’s impact negative, and thousands of games industry workers say their craft and livelihoods are at risk.
• Manipulation: agentic characters built to bond with players could be tuned toward encouraging spending rather than genuine companionship.
• Scoreboard trust: agents can learn to match a benchmark’s pattern instead of solving the task the benchmark was built to measure.
• Wider labor fears: economists and Nobel laureates have separately warned that AI threatens millions of jobs far beyond gaming studios.

General Intuition’s own Nerve marketplace, paying gamers for data labeling and eventual robot teleoperation work, is one answer to that anxiety. It does not resolve it.

One Test Nobody Has Passed Yet

General Intuition says a slice of its new funding is earmarked to open its application programming interface, or API, more broadly to outside developers by the end of summer, backed by a compute deal with CoreWeave to pretrain its next model.

Whether that reflex data ever steers an actual warehouse robot remains unproven outside the company’s own demos. The $2.3 billion valuation says the investors are not waiting to find out.

Frequently Asked Questions

Why do AI researchers use video games instead of real-world tests?

Video games give researchers a safe, cheap and repeatable environment to test decision-making, since a model can fail thousands of times at no real cost. Researchers say the genre as a whole exercises spatial reasoning, long-term planning and social intuition in ways that scripted lab tests cannot.

What is a world model, and how does General Intuition use one?

A world model is a simulated environment generated frame by frame from real data rather than built by a traditional game engine. General Intuition calls its version the gym internally, and it has learned basic physics, such as the fact that walls block movement and shadows shift as the sun moves, purely from watching millions of hours of gameplay clips.

Can AI beat humans at every video game now?

Not once the game is unfamiliar. Researchers describe today’s game-playing systems as brittle and finely tuned to a single title; change the visuals or rules even slightly and their performance can collapse entirely.

Are AI benchmark scores reliable?

Not consistently, according to researchers and AI’s own builders. Even benchmarks outside gaming face a similar problem, with several widely used AI tests now described as functionally saturated, since top models score so close together that the gaps amount to statistical noise.

Will AI agents replace video game developers?

Adoption and anxiety are rising together. A separate Google Cloud survey of 615 developers across five countries found 87% already use AI agents somewhere in their workflow, even as the wider industry grows more anxious about what that means for creative jobs.