Why a Bezos-Backed Startup Says Video Games Could Teach AI to Understand the World

In the race toward more capable artificial intelligence, one New York startup is making an unusual bet: the best training data may not come from books, websites or social feeds, but from video games.

That is the central idea behind General Intuition, a company spun out of the gaming platform Medal that has attracted heavyweight backers and now says it wants to build systems that can better understand motion, planning and cause-and-effect in the physical world. The startup recently closed a $320 million financing round and says it is now valued at $2.3 billion, with investors including Coatue, Eric Schmidt, researchers affiliated with MIT and Google DeepMind, and a Bezos-linked backing that helped put the company on the map.

General Intuition’s thesis is simple but ambitious: large language models excel at predicting and generating text, but text alone may be a poor substitute for the messy, dynamic reality that intelligence must navigate. Games, by contrast, offer structured environments full of movement, consequences, timing and decision-making. If models can learn from that kind of data at scale, the company believes they may develop a more useful form of world understanding than internet-trained chatbots can achieve.

In an appearance on TechCrunch’s Equity podcast, chief executive Pim de Witte laid out the startup’s view of the AI frontier, the origins of the business, and the difficult ethical questions that emerge when technology designed to observe and model behavior could also be applied in defense settings.

What General Intuition is trying to solve

Most of today’s mainstream AI systems are built around language. They can answer questions, write code, summarize documents and hold convincing conversations. But their fluency can hide a weakness: they do not naturally reason about the physical world in the way humans do.

General Intuition’s argument is that intelligence is not just about word prediction. It is also about understanding trajectories, timing, object interactions, spatial relationships and the way actions change outcomes over time. Those are capabilities that matter for robotics, autonomous systems, simulation, defense, industrial automation and other applications where the environment is constantly changing.

The company believes gaming data is especially valuable because it captures repeated examples of agents acting in worlds that behave according to rules. Players react to obstacles, pursue objectives, adapt to surprises and make choices with real consequences inside an environment that can be observed at scale. That makes games a rich source of examples for training what researchers often call world models.

World models aim to help machines predict what happens next in a scene, not just what word is likely to follow another word. In that sense, they are often presented as a bridge between today’s generative AI and future physical intelligence.

Why games appeal as training data

General Intuition’s core claim is not that video games are a perfect stand-in for reality. They are not. But they do provide a controlled setting in which millions of interactions can be captured, labeled and replayed. That is particularly useful for AI systems that need to learn how objects move, how agents plan and how environments respond.

Compared with the open internet, games offer several advantages:

• They generate structured sequences of actions and outcomes.
• They include dense behavioral signals, not just words and images.
• They can be simulated repeatedly, creating scalable datasets.
• They often contain competitive, cooperative and goal-directed behavior.
• They are easier to instrument than the real world, where data collection is expensive and slow.

For AI builders, the appeal is obvious. Internet text helped create today’s large language models, but language data is inherently indirect. It can describe motion, strategy and cause-and-effect, yet it does not always provide the direct sensory and action-level information a system would need to reason about the world itself.

Games, in contrast, can present sequences where decisions are closely linked to immediate visual and spatial feedback. For companies trying to move beyond chatbots and toward agentic or embodied AI, that kind of training material may be more relevant.

From clips to cognition

General Intuition appears to be betting that gaming footage and gameplay traces can help train systems to recognize patterns in how entities move and interact. That could eventually feed into models that are better at anticipating events, planning steps ahead or understanding the consequences of actions in simulated or real environments.

That does not mean the company is building a consumer game recommendation engine or a new console product. Instead, it is positioning itself at the intersection of model training, simulation and physical intelligence.

The concept is part of a broader industry trend. As model builders search for the next leap beyond text, they are increasingly looking at multimodal data, synthetic environments and action-oriented training sources. Video, robotics logs, simulations and games are all part of that search.

The funding round and the investors behind it

The new financing gives General Intuition significant firepower, not just a headline valuation. A $320 million round of this size signals that investors see more than a niche research play; they see a company with the potential to shape the next phase of AI infrastructure.

According to the company, the round included Coatue, former Google chief executive Eric Schmidt, researchers connected to MIT and Google DeepMind, and Bezos-backed support. Those names matter because they suggest the startup has credibility across both capital markets and technical circles.

The presence of high-profile investors also reflects how competitive the AI market has become. Funding is increasingly concentrated around companies that can claim access to unique data, a differentiated training strategy or a path to model capabilities that existing leaders have not yet mastered.

General Intuition’s pitch fits that pattern. Rather than compete directly with the biggest language-model labs on raw text scale, it is arguing that a different kind of data advantage could matter more in the next phase of AI development.

How the valuation fits the market

A $2.3 billion valuation places the startup in rarefied company, especially for a business still defining its category. In the current AI investment environment, valuations often reflect not only current revenue or product maturity but also strategic positioning around compute, data and talent.

In practice, the market has been rewarding businesses that can claim:

1. A proprietary data source or pipeline.
2. A technically defensible approach to model training.
3. Potential relevance to large enterprise or government markets.
4. The ability to turn research into a platform.

General Intuition checks several of those boxes. Its gaming-data thesis offers a narrative around proprietary signal, while the company’s interest in world models could support applications far beyond entertainment.

From Medal to General Intuition

General Intuition did not emerge in a vacuum. It was spun out of Medal, a platform known for game clips and player-generated moments. That background matters because it gave the company a practical understanding of gaming behavior and access to a category of data that many AI companies would struggle to assemble on their own.

Spinning out of an existing platform also suggests a familiar startup playbook: identify a valuable technical asset, separate it into its own company and raise growth capital around a sharply defined AI thesis.

For General Intuition, the relevant asset is not just content. It is the behavioral pattern embedded in the content — what players do, when they react, how they adapt, and how they navigate uncertainty in a dynamic digital environment.

That combination of behavioral data and AI ambition is increasingly attractive to investors looking for alternatives to generic web scraping and broad language-model training. As regulation, copyright concerns and data exhaustion make internet-scale text harder to rely on exclusively, new sources of structured behavior become strategically more important.

The bigger AI debate: text versus world understanding

General Intuition’s emergence lands in the middle of a larger debate about what AI systems actually need to become more generally intelligent.

Some researchers believe scaling up language models and adding more modalities will eventually produce broad competence. Others argue that true generalization requires an explicit understanding of the physical and interactive world — something language alone cannot fully provide.

That disagreement underlies many of the current arguments in AI research:

• Can language models infer enough about reality from text?
• Do systems need action, feedback and environment simulation to reason well?
• Will embodied or agentic systems become the next major AI category?
• What kinds of data create the strongest foundation for general intelligence?

General Intuition is clearly in the second camp. Its founders and investors appear to believe that the next meaningful progress will come not from more words, but from better representations of motion, interaction and decision-making in time and space.

That view aligns with growing interest in agentic AI, robotics, simulation and autonomous systems. It also challenges the assumption that the internet, by itself, is the universal training substrate for AI.

Why language models may hit a ceiling

The argument against pure text training is not that language models are useless. Far from it. They have become the backbone of search, customer service, coding tools and productivity software. But for tasks involving spatial judgment, dynamic environments or sequential control, text may be an indirect and incomplete source of supervision.

Imagine teaching someone to drive, operate a machine or navigate a warehouse entirely through written descriptions. Some principles can be communicated that way, but actual competence usually depends on repeated experience in a physical or simulated environment.

General Intuition is essentially asking whether AI should learn more like that — through observed interaction, not just text prediction.

Why defense concerns are part of the story

Any company building systems that can interpret motion, plan actions and predict outcomes in dynamic settings quickly encounters defense-adjacent use cases. That reality is one reason the startup is already facing questions about where its technology could end up.

On the podcast, de Witte addressed the ethical boundaries around models that might be used in military or defense contexts. The issue is not theoretical. Technologies that understand movement, tactics and environment dynamics can be relevant to surveillance, simulation, targeting support, autonomous systems and planning tools.

That raises a familiar set of concerns:

• How much control should a startup have over downstream use of its models?
• Should companies restrict access based on military applications?
• Can dual-use technology be developed responsibly at scale?
• Where should the line be drawn between simulation, defense analysis and weapons support?

De Witte discussed the ethical boundaries around models that could be relevant to defense applications, underscoring that the company is aware of the dual-use questions tied to its technology.

Those concerns are becoming harder for AI startups to avoid. As models become more capable and more general, their usefulness often expands into areas that founders did not originally target. That creates a governance challenge for investors, engineers and customers alike.

How General Intuition fits into the post-chatbot era

For the past several years, the public face of AI has been dominated by chatbots and text-generation tools. But the next wave of products may look very different.

Instead of asking what a model can write, companies are asking what it can perceive, predict and do. That shift has important implications for the kinds of data AI companies want, the markets they serve and the risks they face.

General Intuition is part of that broader transition. Its focus on gaming data and world models places it among the startups trying to move AI beyond conversational interfaces and toward systems that can reason in environments.

That could matter in multiple sectors:

• Robotics: models that understand movement and interaction can help with navigation and control.
• Transport: predictive systems may improve planning and safety in dynamic settings.
• Manufacturing: spatial and action-based models can support automation and inspection.
• Defense and simulation: world models can enhance scenario planning and training.
• Entertainment and games: AI could generate or analyze more complex interactive behavior.

Whether any of those opportunities becomes the company’s first large commercial business remains to be seen. But the strategic direction is clear: General Intuition wants to be in the category that comes after text-only AI.

The challenge of proving the thesis

The hard part for General Intuition is not explaining the idea. It is proving that the idea produces models that are measurably better than alternatives.

Many startup narratives in AI sound persuasive until they run into questions of benchmark performance, access to data, inference cost, model reliability and practical deployment. In this case, the company will need to show that gaming data genuinely creates an advantage that cannot be replicated with standard multimodal training or simulation-generated data.

It will also need to answer a few technical questions:

• What kinds of game environments are most useful?
• How much data is enough to learn generalizable world dynamics?
• Can insights from games transfer to real-world use cases?
• How does the company distinguish signal from noise in massive gameplay datasets?

These are not small questions. They sit at the center of whether the business becomes a foundational AI lab or remains a promising but specialized data company.

Data is not the same as intelligence

There is also a more philosophical challenge. Training data alone does not create intelligence unless the model architecture, objectives and feedback loops are designed well enough to extract meaningful structure from that data.

Video games may offer richer training signals than the internet in some respects, but they still represent bounded environments. The leap from mastering game-like scenarios to understanding the physical world is nontrivial. Investors may be willing to fund that leap, but the proof will have to come from the models themselves.

What to watch next

General Intuition’s next phase will likely be defined by three things: product evidence, technical validation and policy scrutiny.

First, the company will need to show what its models can do that existing systems cannot. Second, it will need to demonstrate that gaming data can support a repeatable training pipeline with real commercial value. Third, it will need to navigate the ethical and reputational challenges that come with dual-use technology.

For the AI sector broadly, the company is a reminder that the hunt for better training data is far from over. As foundation models mature, advantage may increasingly come from the quality of the environments used to train them, not just the size of the models themselves.

That helps explain why an investor base spanning top-tier venture firms, influential technologists and major research figures is willing to bet so heavily on a startup built around video games. In a market where everyone is searching for the next data edge, General Intuition is betting that the most valuable lessons may come from watching how players move through worlds, not how people write about them.

For now, General Intuition remains an early test case for a provocative idea: that to build machines with a more human-like grasp of reality, AI developers may need to train them less on the internet and more on the interactive worlds humans have already built.