なぜAI R&D 自動化はシングルささえなく進歩を加速させるのか

Full automation of AI research and development (R&D) could deliver 3.5 years of progress in a single year, even without a runaway software-only singularity. This acceleration stems from a massive one-time speedup and a fundamental shift in how compute scales relative to research output.

The massive one-time acceleration effect

Even if the software-only feedback loop—where smarter AIs drive increasingly fast rates of algorithmic improvement—remains subcritical, the transition to AI-led R&D provides a significant jumpstart. In a scenario where the feedback loop is not self-sustaining (modeled with $r=0.7$), the AI Futures Model indicates that the first year after full automation yields 3.5 years of progress, assuming no increase in compute.

This jump occurs because AIs can execute the research cycle with far greater efficiency than human teams. While earlier AIs will have already plucked much of the low-hanging fruit, the transition to full automation still represents a massive leap over the existing baseline. Even when accounting for the gradual boost of preceding AI capabilities, the model suggests a conservative estimate of roughly 2.5 years of progress in that first year of full automation.

Higher returns on compute scaling

Once AI automates the R&D process, the relationship between compute and progress changes fundamentally. When humans are the primary source of R&D labor, increasing compute allows for more experiments and larger training runs. However, once AIs dominate the labor force, additional compute yields a dual benefit: it powers more experiments and it improves the AI labor force itself.

This creates a powerful feedback loop. More compute can be used to build smarter, faster, and cheaper-to-run AI researchers. These improved AI laborers then conduct better experiments, which in turn yield even better AIs. Even in a subcritical regime, this effect likely doubles, triples, or quadruples the rate of progress compared to a human-centric model.

As long as the system has not reached absolute algorithmic limits, every increase in compute will drive more progress than it did when human researchers were the bottleneck. This shift suggests that the era of compute-driven scaling will become even more potent as the research labor force becomes commoditized and automated.

This transition toward automated R&D likely sets the stage for a period of intense, compute-driven acceleration that could redefine the timeline for achieving top-human-expert-dominating AI (TEDAI).

Source: Full automation of AI R&D probably yields a large speed up even without a software-only singularity
Domain: alignmentforum.org