Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
Most discussions of AI scaling focus on making the expensive part of a system larger: more parameters, data, inference-time reasoning, rollouts, or agents. The paper by Qi Wang and colleagues asks a different question. What if the next efficiency gains come from deciding which interactions deserve that expensive computation in the first place?
The authors propose a co-designed pairing. System-1 is a relatively lightweight, semantically aware predictor that estimates the utility of a possible interaction. System-2 is the costly executor that performs reasoning, planning, tool use, training rollouts, or physical actions. The central idea is to let System-1 screen candidates before System-2 spends resources on them.
The interaction problem
The paper treats an interaction as an operation that incurs a cost before its outcome is known. Examples include generating a chain of thought, running a reinforcement-learning rollout, calling a tool, or testing a task configuration in a simulator or on a robot. These interactions can have very different value. Some produce a useful learning signal or change a decision; others repeat information the system already knows.
This matters because scaling interactions indiscriminately can produce diminishing returns. In reinforcement learning with verifiable rewards, for example, a prompt whose sampled responses all receive the same reward may provide no relative policy-gradient signal, even though generating those responses still consumes compute. In physical learning, many simulated or real-world tasks may be too easy, too familiar, or otherwise unhelpful for adaptation.
Conditional interaction: predict, gate, execute, update
Wang and the other authors formalize a loop called conditional interaction:
- Represent a candidate prompt, task, query, or action.
- Use System-1 to predict its utility, and sometimes its uncertainty.
- Admit it only if its predicted value clears a threshold that accounts for the available budget and execution cost.
- Execute the admitted candidates with System-2.
- Feed the observed outcomes back into System-1 and System-2.
The architecture is broader than a static model cascade. A conventional cascade routes requests at inference time using a model trained separately. In the proposed co-design, System-1 also helps allocate training compute, replay storage, context, model capacity, and inference effort. System-2’s results become supervision for the predictor, while the predictor determines which experiences System-2 receives.
Utility is deliberately application-dependent. The paper discusses information gain, expected learning improvement, value of information, adaptation risk, and accuracy improvement as different possible objectives. Information gain is a useful example, but the authors stress that a surprising observation is not automatically a valuable one. A utility function must be declared for the task and measured against its relevant cost.
What the analysis shows
The paper’s main quantitative claims are theoretical results under explicit assumptions, not a single demonstrated AGI system. With a perfect, free predictor and equal execution costs, the ideal efficiency improvement approaches 1 divided by the density of utility-relevant interactions. If only a small fraction of candidates are worth executing, selectively paying for that fraction can substantially increase utility per unit of compute.
The authors extend this accounting to imperfect predictors. False positives waste execution cost, false negatives reduce the amount of useful signal collected, and predictor overhead can erase the advantage when screening is not much cheaper than execution. They also derive selection-regret bounds: when prediction error is smaller than the utility gap separating good and bad candidates, the selected batch can match the oracle’s choices; otherwise, regret grows with prediction error and the number of selected items.
Additional results concern screening and memory. Under a synchronous evaluation setup, predictive screening can avoid holding rollout state for the entire candidate pool, leaving memory determined mainly by the executed batch and the screening micro-batch. These results are conditional on the stated execution model; a streaming reactive system may already avoid some of the claimed memory growth.
The paper also proposes informative compute as a possible scaling coordinate: total compute weighted by the fraction that is actually utility-relevant. The authors prove a horizontal rescaling result in an idealized setting where uninformative interactions do nothing and selection is perfect. They hypothesize, but do not establish, that the density of useful interactions declines as systems improve and that measuring informative compute could explain part of conventional scaling-law behavior.
Evidence and what remains unproven
The authors point to application-specific work on predictive prompt sampling, task selection, and model-predictive sampling as supporting evidence that selective execution can reduce rollout use. However, they explicitly distinguish those results from a validation of the entire co-design framework. The broader claims about a superior cost-quality frontier, changing scaling exponents, and progress toward AGI remain hypotheses.
Important limitations
The most serious limitation is selective supervision. Once System-1 controls execution, it also controls which outcomes it gets to observe. A candidate predicted to be uninformative may be suppressed indefinitely, preventing the predictor from discovering that it was wrong. The paper recommends explicit exploration, random audits, or known-probability sampling of rejected candidates, potentially combined with inverse-propensity weighting.
Other limitations are equally practical. Prediction must be substantially cheaper than execution. Distribution shift can concentrate errors in precisely the cases the gate rejects. The chosen utility may omit delayed benefits, exploration value, safety information, or prerequisite actions. In agentic workflows, individually unhelpful clicks or tool calls may be jointly necessary. Safety-critical or irreversible actions may require permissive overrides rather than aggressive gating.
Practical implications
For AI engineers, the proposal suggests measuring more than tokens, rollouts, or wall-clock compute. Track admission rates, predictor cost, false positives, false negatives, yield, and utility per screened candidate. Maintain an independent audit stream so rejected cases remain observable. For test-time reasoning and tool-using agents, route extra thought, retrieval, or action only when a calibrated estimate says it can change the answer or decision enough to justify its cost.
The paper’s qualified conclusion is persuasive as an engineering direction: scaling the executor alone may be wasteful when interaction value is uneven and predictable. But the decisive test is empirical. Teams will need to measure utility density over time, compare systems at matched informative compute, and determine whether selective execution delivers merely a constant-factor saving or a deeper change in scaling behavior.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.