Game data, not LLM: how startup General Intuition sees the path to AGI

Game data, not LLM: how startup General Intuition sees the path to AGI

Startup General Intuition, backed by Jeff Bezos investments, claims a fundamentally new approach to achieving artificial general intelligence. According to the company, modern large language models — from ChatGPT to Claude — are exhausting their potential on text data and cannot fully develop the ability to generalized intelligence without understanding the physical reality of space and time.

Why LLMs are insufficient for AGI

Text models demonstrate amazing abilities in language processing: creativity, analysis, reasoning based on knowledge. However, they remain fundamentally blind to one critical area — the physics of space and time. These models lack a built-in understanding of how objects move, interact under the influence of forces, react to proximity to each other, how causes generate consequences in the physical world.

This understanding is necessary so that an AI system can generalize its knowledge beyond the textual sphere and apply logic in a real, physical context. Text data, even the most extensive sets of descriptive texts, cannot replace the experience of observing how objects move, fall, collide, react to forces.

Game data as the key to physical thinking

This is where video games come into the picture. In game worlds, all physical laws are modeled and recorded in data: how a character jumps, how an object falls, how objects interact on collision, how sound propagates, how light reflects off surfaces. These high-frequency, multidimensional datasets contain billions of examples of spatial-temporal behavior — exactly the information that LLMs cannot extract from text and even from unannotated video.

General Intuition sees game data as an underutilized resource for training AI systems to integrate knowledge about physics, causality, and spatial relationships. If a model trains simultaneously on text and game data — with explicit semantics of physical interactions — it can develop the ability not only to describe the world in words, but also to predict, model, and reason about its physical structure, make causal inferences. This, according to the startup's conviction, is a critical step towards more universal and adaptive intelligence.

Bezos's investment and significance for the industry

Jeff Bezos, known for his support of ambitious, fundamental research and risky bets, directed investments to General Intuition. This signals that the company's hypothesis attracts the attention of experienced investors and influential figures in the artificial intelligence community. However, the details of the funding round itself — investment amount, company stage, other investors — remain unknown from public sources.

What this means for the future of AGI

If General Intuition's concept is correct, then AGI development requires not just scaling text models, but a fundamentally different architectural approach: the synthesis and integration of knowledge from many data sources. This shifts industry focus from the question "how do we make LLMs larger?" to questions like: what other types of data are critical for general intelligence? How can we more efficiently integrate different modalities into a single system? What role does active learning through simulation and game environments play? The discussion about the fundamental ingredients for AGI becomes only more acutely relevant and conceptually diverse.