> Quick summary before we dive in: the core of the claim that games provide better AI training data than the internet — why, and why this is now a topic the AI industry is watching closely
- Internet data is static — just fixed text/images, but games create “decisions with consequences” that let AI see real cause-and-effect in real time
- Games are a controllable sandbox — want to test a scenario? Simulate it as many times as you want, unlike internet data which can’t easily be expanded
- The AI industry is running short on quality data from the internet (a problem that’s been discussed for a while), so games are increasingly becoming an alternative worth watching
Note: this article doesn’t include specific numerical specs for this topic, so the content below focuses primarily on conceptual analysis
The image that sparked this discussion
This image directly conveys what the CEO was talking about: game worlds that aren’t just “pretty pictures” but systems with physics, rules, and outcomes that can be verified every single time.
What’s interesting is the comparison itself: the left side might represent overflowing, unstructured text data from the internet, while the right side shows an in-game scene where AI walks in, interacts, experiments, and gets measured results instantly. A single image that fully tells the story of the difference between “static data” and “interactive data.”
Where this discussion started, and why it’s relevant now
Many people who work with AI daily have probably run into the same case: ask a model a simple spatial reasoning question, like “if I turn left and walk straight, what will I find?” and it answers with something garbled or self-contradictory.
This problem isn’t a coincidence. Most models are trained on internet text, which is static data — there’s no feedback loop, no one confirming whether the sentence the model read is actually “physically true” or not.
This is exactly what makes this CEO’s idea interesting. He believes games have something the internet doesn’t: a physics engine that enforces consistent rules, and outcomes that can be verified repeatedly every time AI takes an action.
The question that follows is: if data from game worlds can take models further than reality can, is the AI training industry about to shift direction?
Where this idea fits into the company’s plan
This statement isn’t just the CEO’s personal opinion — it’s directly tied to the company’s roadmap, which is actively developing AI agents and robotics. Models that will control robots or agents in the real world need to understand physics and sequences of events, not just be good at stringing together text.
So data from games has become an important piece of the product pipeline, not just something to play around with. It’s training data fed to models to practice decision-making in environments with clearly defined rules, before being applied to agents that must make continuous, multi-step decisions.
Put simply: if this company is building AI that needs to “take action,” not just “answer questions,” games are a far more targeted training ground than the real internet.
Comparing this to the traditional way models have been trained
The traditional approach is to scrape text from across the internet. Sure, the volume is massive, but most of it is “descriptive knowledge,” not “action” — a model can read about how a ball falls, but it’s never actually practiced catching one.
Game data is different: every player action comes with immediate feedback (collision/no collision, success/failure), making the signal cleaner and directly relevant to spatial reasoning tasks. The cost of collecting data is also fundamentally different — the internet side means crawling then filtering, while the game side comes from gameplay logs that are already structured.
| Factor | Internet Text | Game Environment |
|---|---|---|
| Signal quality | Descriptive, no immediate feedback | Clear action-result pairing |
| Data collection cost | Heavy crawl + quality filtering | Comes directly from gameplay logs |
| Spatial/Physical Reasoning | Weak, because it's text-based | Strong, because it simulates real physics |
| Raw data volume | Massive | More limited |
How this could actually be applied in different scenarios
Simply put, it’s used to reinforce the areas where text-based data falls short.
Robot movement — In-game physics (gravity, collisions, balance) teaches robots to simulate walking and grasping more accurately than reading a physics manual from the web.
Multi-step decision-making AI agents — Quest design in games is pure multi-step planning: collect item A first before unlocking level B, which closely matches the workflows agents actually need to perform in real tasks.
Self-driving cars — Understanding of 3D space, distance, and perspective from open-world games helps reinforce spatial reasoning that text can’t fully explain.
Cooperation/competition — Multiplayer games train AI to read the game state and adjust strategy based on opponents or teammates — a skill that internet datasets barely provide.
Comparing this to other approaches in the same space
The AI industry right now is split into several camps, and not everyone agrees with the idea that “games are better training data than the internet.”
The internet-scale data camp (like typical GPT models) still believes that a massive volume of text data is enough to generalize across all tasks. Its strength is scale and availability; its weakness is a lack of spatial reasoning and real-time decision-making.
The simulation/synthetic data camp (used by some robotics teams) builds custom simulated environments specifically to train agents for particular tasks. It’s more precise than commercial games but far more expensive to build.
| Factor | Video Game Data | Internet-scale Text | Custom Simulation |
|---|---|---|---|
| Spatial reasoning | Strong | Weak | Strong |
| Data creation cost | Low (already exists) | Low | High |
| Multi-agent/strategy | Yes | Barely any | Depends on design |
Pros and cons of this idea
This viewpoint has both strengths and points worth questioning. Let’s look at both sides in full.
Pros
- +Clean signal, without the noise that internet text is full of junk data
- +Has a clear reward function — you can immediately measure whether the model 'won' or 'lost'
- +Can be repeated safely for experimentation — mistakes in a game don't affect the real world
- +Trains spatial reasoning and multi-agent strategy more precisely
Cons
- −Game worlds have physics and rules that are much more simplified than the real world
- −Game data licensing issues are still legally unclear
- −Risk of ending up with a model that's 'only good in-game' and can't generalize to the real world
- −Lacks the linguistic, cultural, and general knowledge context that text data provides
Honestly, both sides have solid reasoning behind them. The final answer might not be choosing one over the other, but combining both together.
The trade-offs of fully committing to this idea
The idea that “games are better training data than the internet” sounds cool as a tagline, but the real-world version comes with costs nobody talks about. It starts with engine and game licensing fees — you’d have to negotiate with studios one by one, and not every studio will agree to let you scrape or re-run their game millions of times.
Next is compute — simulating millions of game episodes to generate training signal consumes enormous resources, no different from training a large language model. On top of that, there’s legal risk around player data embedded in online games — the more a game has a community that logs player behavior, the more careful you need to be about privacy.
The heaviest issue is the generalization question: a model that excels in a game world with fixed rules might get completely lost when it encounters the real world, which has no clear rules. This is a gap nobody has been able to answer clearly yet.
Looking ahead: what to watch next
The real question is whether this is a genuine research breakthrough or just a narrative for the next funding round. We need to wait for real evidence, not just statements.
Signals worth watching: first, deals with major game studios to gain access to scale-level gameplay data. Second, papers or benchmarks that show agents trained on games can actually be applied outside the game — not just performing well in the same game. Third, funding round numbers tied to these milestones — if valuation moves before there’s proof, that’s a sign it might be more about selling a dream than delivering results.
If, over the next 6-12 months, no agent clearly proves generalization, this claim will remain just a hypothesis that sounds good on paper.