Why this CEO thinks video games make better training data than the internet

The article argues that today’s large language models may lack core abilities for AGI—specifically understanding how things move through space and time—and suggests this gap could be addressed using gaming data. It frames the approach as the bet behind a company called General Intuition, but provides no quantitative results or performance metrics in the excerpt.

Analysis

This is more venture-thesis than public-market catalyst. The economic value, if any, accrues to owners of high-fidelity 3D interaction data and simulation pipelines, which points more toward platform/data-rich franchises like RBLX, U, and infrastructure spend tied to NVDA than toward an abstract AGI narrative vehicle. The immediate risk is that investors confuse a research claim with a monetizable product; that gap usually compresses multiples rather than expands them once the market asks who actually pays.

Second-order, the theme favors companies that can turn embodied environments into repeatable training assets: game engines, simulation software, robotics/autonomy data pipelines, and GPU-intensive model training. Over the next 1-3 months, the stock reaction is likely headline-driven and fragile; over 6-18 months, the key test is whether this produces disclosed contracts, usage-based revenue, or capex budgets from enterprises, not just benchmark gains. If no conversion shows up, the story reverts to a niche AI research angle with limited equity impact.

The contrarian view is that consensus may be overrating data scarcity. If synthetic environments can be generated cheaply, the moat migrates from the data itself to distribution, compute, and product integration, which is a better setup for incumbents than startups. That makes a standalone 'gaming data' beneficiary harder to underwrite unless it has unique proprietary environments, customer lock-in, and a measurable path to recurring revenue.

Key Decisions for Investors

• No immediate trade in TSTS; keep it on watch only until management can evidence a monetization path. Falsifier: no customer wins, ARR, or bookings lift within the next 1-2 quarters.
• Prefer a relative long RBLX vs. short an AI-concept basket over the next 1-3 months if the market starts paying up for embodied-AI optionality. Risk/reward is better if Roblox-like platforms are the ones actually producing proprietary interaction data.
• Use NVDA as the cleaner second-order expression only on weakness, and only if next earnings season shows rising simulation/training capex. Stop if data-center capex guidance or AI spend trends flatten.
• For a more cautious expression, avoid chasing pure-play AI/data startups until they disclose paying customers; the likely outcome is multiple compression once the market demands proof. This is an alert, not a conviction long.