Jensen Huang declared we've achieved AGI on the Lex Fridman podcast last week, timing his claim with ARC-AGI-3's launch—a benchmark that immediately contradicted him. The new test presents interactive environments with no instructions, rules, or goals, requiring agents to explore and adapt. Humans solve 100% of these tasks. The best frontier models managed 0.37%. This isn't a minor gap—it's a chasm that exposes the fundamental limitation of current architectures.
As I noted after Huang's initial claims, this represents the industry's core definition problem around AGI. The timing makes it more pointed: the CEO of the company supplying compute for every major AI lab claims superintelligence while simultaneously, rigorous testing shows these systems cannot handle basic novelty. Current models excel at pattern-matching within their training distribution but collapse when faced with truly novel scenarios requiring genuine reasoning.
The market seems to agree with the data over the hype. This week's $25 billion in deals targeted infrastructure and specialized applications, not foundational models. IBM's $11 billion Confluent acquisition focuses on real-time data streaming—the pipes between models and reality. Physical Intelligence raised $1 billion for robot control systems. Eli Lilly bought Insilico's drug discovery pipelines for $2.75 billion. Smart money is betting on specialized systems that work within known constraints, not general intelligence.
For developers, this clarifies the immediate opportunity: AI excels at tasks with clear patterns and defined domains but fails at open-ended problem-solving. Build systems that leverage what current models do well—classification, generation within training distributions, structured reasoning—while keeping humans in the loop for novel situations and adaptation.