Total Automation of AI R&D: A Leap Without Singularity

The Impact of Automation on AI R&D

The total automation of research and development (R&D) in artificial intelligence (AI) could generate a significant speed gain, even without reaching what is known as a "software singularity." This singularity is characterized by exponential advancements in AI, where algorithms improve at an accelerated pace, overcoming diminishing returns. Such a phenomenon would require these advancements to be sustainable enough to represent, for example, four years of progress in a single year.

Even without this singularity, the complete automation of AI R&D could accelerate progress for two main reasons. First, automation itself offers a notable speed gain. With median parameters and a factor of r=0.7, it is estimated that one could achieve 3.5 years of progress in just one year after automation, even without increasing computing power during that period.

Second, once AI R&D is fully automated, the increase in available computing power could generate higher returns than before. AIs, having become the primary workforce, could utilize this additional computing to conduct experiments and training, thereby optimizing their efficiency and reducing execution costs.

Feedback Loops and Progress

This situation creates a feedback loop where improved AIs lead to better experiments, in turn producing even more capable AIs. Even if this loop remains subcritical, each increase in computing could now lead to further progress. This effect could potentially double, triple, or even quadruple the rate of progress observed without automation.

By analyzing progress trajectories, an AI Futures model shows that after complete automation, it is possible to achieve more than two years of progress in a single year. This could allow an AI to surpass the best human experts in less than a year.

Indirect Acceleration Factors

Several indirect factors could also accelerate AI progress at the time of R&D automation. Increased AI capability could attract investments and revenues above the current trend, enabling the acquisition of more computing power. A company with a net advantage could more easily obtain computing from other lagging firms.

Moreover, AIs could accelerate hardware R&D by developing better chip designs and speeding up the research and construction of new manufacturing plants.

However, it is important to note that the rate of increase in computing power could be lower than today, which could reduce the rate of AI progress, making the acceleration relative to a lower baseline.

Historical Context and Outlook

Historically, progress in AI has been driven by two factors: the increase in computing power and the increase in labor. Computing for algorithms and training has been multiplied by about four times per year, while the number of employees in companies has tripled each year. This dynamic could change with the complete automation of AI R&D.

Conclusion

The complete automation of AI R&D could make moderate advantages more stable and predictable, as the labor component of AI R&D would be commoditized and similar across companies. However, maintaining a competitive edge would require retaining an advantage in computing power. A lagging company could catch up if it had more computing, as labor would be commoditized after automation. Investors might push a lagging company to sell its computing power to a leading firm, even if management is reluctant to do so.