Meta's AI chief credits evaluation systems, not raw intelligence, for agents beating 100 engineers

The $14.3 billion Meta paid for Scale AI, set against the 100 engineers in the only comparison Wang has offered, works out at $143 million per engineer beaten ^[14]. The deal bought data operations and a research bench as well, so the division is unfair to it.

Saying that evaluation decides the outcome is saying where the money should go, and it comes from a founder whose previous company's business model revolved around making AI systems measurable and accountable ^[11]. Wang kept the boundary tight. According to CryptoBriefing's account of the conversation, he said the agents outperformed engineers "on specific tasks" with "the right evaluation system" ^[8].

Which model the loops call, how long they ran, and what the compute cost do not appear in that account ^[15]. Two readings fit the gap equally well. Either the scheduling and the grading are the difficult part, or a frontier model absorbs the difficulty and the loop keeps it aimed at something checkable.

The account credits the result to robust evaluation methods, the ability to run continuously, and a feedback architecture that lets the agents improve without human intervention ^[6]. Taken at face value, that sends the next increment of engineering attention to graders, logs and schedulers instead of another point of model capability. The plumbing is free, so the budget question is about everything upstream of it.

The narrow version of the claim is the one I would back: where a task has a written success criterion, a loop that runs all night and checks its own output will outproduce a queue of people who each need a ticket and a review. The wide version, that swarms beat teams, rests on one report of what was said on a stage ^[2]. Two results would tell against the narrow version too: a task list whose items turn out to be trivially checkable, and loops that stop working when the model underneath them is swapped.