Paper proposes 'informative compute' as AGI scaling yardstick

The paper's answer to its own title question is "a qualified yes."

That title is "Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?", a new arxiv preprint from Qi Wang, Yun Qu, Xiangyang Ji and ten co-authors. The abstract starts from a plain observation about how progress has looked so far: AGI has "largely relied on scaling a resource-intensive executive, System-2, through larger models, more data and compute, longer reasoning, and more agents."

The authors propose pairing that executive with a cheaper gatekeeper. They write that "a lightweight System-1 predicts the utility of potential interactions, the resource-intensive System-2 executes only those interactions that exceed a resource-aware threshold." Outcomes feed back and update System-1. They call the full cycle a "predict--gate--execute--update loop."

The headline number. Under ideal screening and equal execution costs, "execution compute falls by a factor approaching 1/ρu, where ρu is the density of utility-relevant interactions." ρu, cost-weighted when costs differ, is the paper's proposed yardstick, a "measurable coordinate of scaling efficiency" that the authors argue makes "informative compute a better description of scaling than total compute."

They do not oversell it. "Prediction errors, exploration, and distribution shift reduce the saving," the abstract notes, and the paper bounds "selection regret by predictor error and the utility margin." The claim is a hypothesis motivated by analysis, not a demonstrated result in a trained system.

The loop, as described, is pitched as applicable across "training, test-time reasoning, adaptive decision-making, and agentic tool use, with utility and cost defined by the application."