AGI Scaling May Depend on System-1 and System-2 Co-Design
Original title: Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
Study proposes conditional interaction loop between System-1 and System-2 that can reduce execution compute by factor approaching 1/ρ_u.
Progress toward AGI has relied on scaling System-2 through larger models and more compute. The paper asks if co-designing with System-1 as a lightweight predictor could be the next stage. It studies a predict-gate-execute-update loop where System-1 estimates utility of candidates and System-2 executes only those admitted by an adaptive threshold. Results show potential compute savings under ideal conditions, with bounds on selection regret and factors like prediction errors that reduce benefits.
- Loop spans training, test-time reasoning, adaptive decision-making, and agentic tool use
- Execution compute falls by factor approaching 1/ρ_u under ideal screening
- Prediction errors, exploration, and distribution shift reduce savings
- Regret bounded by predictor error and utility margin
- ρ_u is proposed as measurable coordinate of scaling efficiency
From the original article
1 min read at huggingface.co
Progress toward artificial general intelligence (AGI) has largely relied on scaling a resource-intensive executive, System-2, through larger models, more data and compute, longer reasoning, and more agents.
By Qi Wang, Yun Qu, Lizhou Cai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.