Scaling toward AGI depends on what System-1 lets System-2 execute
agents reasoning
| Source: HF Papers | Original article
Researchers question whether future AGI breakthroughs will stem from integrating the fast, lightweight System‑1 with the resource‑heavy System‑2, rather than merely scaling the latter.
A new arXiv preprint challenges the dominant “bigger‑is‑better” road‑map to artificial general intelligence (AGI). Authored by Qi Wang, Yun Qu, Xiangyang Ji and ten co‑authors, the paper titled “Does Scaling Toward AGI Depend on What System‑1 Lets System‑2 Execute?” argues that the field has leaned too heavily on expanding the resource‑intensive executive component, dubbed System‑2, through ever larger models, more data, longer reasoning chains and additional agents.
Instead, the authors propose a co‑design approach that pairs System‑2 with a lightweight, semantically aware predictor they call System‑1. This predictor would evaluate candidate interactions before they are handed to the executive, effectively filtering and guiding computation. The manuscript introduces the notion of “informative compute” as a scaling yardstick, concluding with a “qualified yes” to the question posed in its title: the efficiency and capabilities of future AGI systems may indeed hinge on what System‑1 permits System‑2 to execute.
If the argument gains traction, it could reshape research priorities across academia and industry. By shifting focus from raw compute scaling to smarter pre‑selection of tasks, developers might achieve comparable or superior performance with lower energy and cost footprints—a point of growing relevance as AI compute demand strains both budgets and sustainability goals. The idea also dovetails with broader debates about how to make AI systems more controllable and interpretable, offering a potential pathway to safer, more efficient general‑purpose agents.
The community’s next steps will likely involve empirical validation of the System‑1/System‑2 framework, replication of the “informative compute” metric, and discussions at upcoming conferences on whether this co‑design paradigm can complement or replace existing scaling strategies. Watch for responses from major AI labs and any follow‑up studies that test the proposed architecture in real‑world settings.
Sources
Back to AIPULSEN