By Jon Scarr
For years, some of the biggest stories around game AI were about machines learning how to beat us. Atari, Go, and StarCraft became testing grounds for increasingly capable systems. Google DeepMind's latest work is moving in a direction I find much more interesting. It is exploring AI that can look at a game, understand what's happening, reason about what to do, and learn inside worlds people already created.
SIMA 2 can work through the same basic interface we use, looking at the screen and controlling a game through a virtual keyboard and mouse. It doesn't need the game's source code to understand what is happening. EVE Online then gives DeepMind something far more complicated to learn from, a universe shaped by more than 20 years of human decisions.
That gets back to something I was thinking about in our recent look at AI and PS2-era creativity. I'm far more interested in AI that expands what can happen inside a game than AI that exists mainly to generate more content. DeepMind's work points toward a future where an AI might learn how to participate in worlds people still deliberately create.
Winning the Game Isn’t the Goal Anymore
DeepMind's history with games goes back well before today's generative AI discussion.
Its Deep Q-Network learned to play dozens of Atari 2600 games directly from what it saw on screen. Later systems tackled Go and StarCraft II. Those efforts were important because increasingly difficult games gave researchers new problems to solve.
SIMA changed the question. Instead of training an AI to dominate one specific game, DeepMind began working on a system that could understand instructions across different 3D environments. The first SIMA could perform more than 600 basic skills, including navigating an area, interacting with objects, and using menus.
The important difference is that SIMA isn't reading the game's underlying code to figure out what to do. It sees the screen and uses keyboard and mouse controls.
That's much closer to the way we interact with a game. DeepMind isn't simply asking whether an AI can reach a higher score anymore. It is asking whether an agent can understand a virtual environment well enough to take useful actions inside it.
SIMA 2 Treats Games More Like We Do
SIMA 2 takes that idea further by adding Gemini to the system.
It can reason about a goal instead of simply following a short command. It can explain what it intends to do, answer questions, and work through more complicated instructions. DeepMind says SIMA 2 can apply concepts it learned in one game to different activities and even environments it wasn't trained on.
That’s where this starts to get really interesting. A convincing game character shouldn't need a completely new set of instructions every time something unexpected happens. If an agent can understand enough about the world around it to apply something it learned elsewhere, the interaction starts becoming much more interesting.
SIMA 2 can also improve through its own experience. DeepMind has tested agents that continue learning through self-directed play after their initial human demonstrations. That experience can then contribute to later versions of the system.
There are still major limits. DeepMind describes SIMA 2 as a research preview. Very long tasks remain difficult, its interaction memory is relatively short, and low-level keyboard and mouse actions can still cause problems.
That's an important reality check. We're not at the point where you can drop an AI companion into any game and expect it to understand everything happening around you.
But the direction is much more interesting to me than simply asking an AI to generate another line of dialogue.
EVE Online Adds More Than 20 Years of Human History
EVE Online changes the problem because its world already has a history.
The developers created its systems and universe, but people have shaped what happens inside them for more than two decades. Alliances have formed and collapsed. Conflicts have changed political relationships. Trade and resource decisions have shaped an economy spread across thousands of star systems.
DeepMind specifically wants to study AI dealing with economic decisions, negotiation, cooperation, competition, and behaviour that emerges from all of those interactions. It is also looking at memory, continual learning, and decisions that can play out over much longer periods.
That's a very different challenge from beating an opponent.
An AI operating in EVE would eventually need to make sense of a world whose history wasn't scripted for it. The interesting part isn't replacing that history with something generated. It's learning how to function inside a world whose identity came from what people did with the systems they were given.
DeepMind is being cautious about that research. It is starting with an offline version of EVE Online rather than putting experimental agents into the live universe. Fenris Creations says that version runs on a local server in a controlled environment.
DeepMind plans to move through EVE Frontier before it would consider bringing mature capabilities into live EVE Online or EVE Vanguard.
That separation makes sense. EVE is interesting as a research environment because its human history has consequences. Experimenting on a copy means the research can learn from that complexity without interfering with the people who created it.
Aura Guidance Starts With What EVE's Community Already Knows
There's already a much smaller example of AI working inside EVE.
EVE Online introduced Aura Guidance in February 2026 to help people who are learning the game. Instead of replacing Rookie Help, the system starts with knowledge collected from real Rookie Help conversations and ISD volunteers.
EVE's team built a large Q&A bank from those conversations. When someone uses Aura, the system looks for similar questions and builds a response from existing answers. It can also consider context such as location, ship, recent deaths, and whether the character is docked.
When Aura can't find a reliable match, it redirects you to Rookie Help instead. I like that distinction.
Aura isn't trying to become the community. It's trying to make knowledge the community already created easier to reach. EVE's team has also been clear that the tool isn't generating game content or replacing creative work.
Aura Guidance is completely different from the agents DeepMind is researching, so the two shouldn't be confused. But it does show another way AI can fit into games without becoming the reason the game exists.
Human Creativity Still Has to Lead
DeepMind describes its long-term goal as using AI as a catalyst rather than a replacement. That's the part of this research I keep coming back to.
It's also easy to say and much harder to maintain once these systems become more capable.
I don't need an AI to replace a writer, designer, artist, or programmer to find game AI exciting. An agent that can understand the world around it, remember what happened, adapt to something unexpected, and interact naturally with what people created is already a huge idea.
That's also where this starts connecting to the PS2-era creativity discussion.
Technology can give developers new tools and take repetitive work off their plates. It still takes people to decide what kind of world is worth creating in the first place. More AI-generated content isn't automatically more creative.
EVE makes that distinction especially clear.
Its developers created the sandbox, but the stories people remember came from years of human decisions inside it. DeepMind is now asking what happens when an intelligent agent has to learn from that kind of environment rather than simply master a fixed challenge.
That's the future of game AI I want to see explored. Not AI replacing the people who make games interesting, but AI becoming capable enough to learn alongside us inside the worlds they created.
Comments
Post a Comment
Comments are moderated. Please stay respectful and on topic.