AI development could accelerate with each software improvement even on existing computers: Adam Marblestone’s model
AI development could accelerate with each software improvement even on existing computers: Adam Marblestone’s model
On November 7, Adam Marblestone, co-founder of the research organization Convergent Research, examined three conditions under which AI development could accelerate itself. With help from Claude, he connected computational savings, research team productivity, and the difficulty of finding new ideas.
The discussion was prompted by a report from the research organization METR: Anthropic’s AI models had begun improving on benchmark tasks at nearly twice their previous rate. Marblestone examines the conditions under which that rate could continue to increase.
In the model, the available computers remain the same. Researchers improve software efficiency, allowing the software to do the same work with less computation. The same computers can then run more copies of the AI or run them faster. If those copies develop and test new algorithms, successful improvements free up more computing capacity for research.
As the model itself becomes more capable, it may also become better at research tasks. Marblestone also accounts for inefficiencies as the team grows: “whether a million AI researchers working in parallel duplicate one another’s work and compete for the same computational resources for experiments.” In his model, all else being equal, doubling the number of researchers increases the rate of software improvement by less than a factor of two.
Testing a new algorithm may require training a model across many computers. The authors of an analysis on the research organization Epoch AI’s website explain that additional AI researchers may have to wait for hardware to become available for these experiments. Reducing the computational cost of the experiments themselves allows more ideas to be tested on the same hardware.
Suppose each successive doubling of software efficiency requires more research effort than the previous one. If the team’s productive output grows faster than the effort required for the next doubling, the intervals between doublings shrink. When the two grow at the same rate, the intervals remain constant; when finding ideas becomes harder faster than productive output grows, the intervals lengthen.
In one illustrative example from Marblestone, each doubling of efficiency halves the time needed for the next: six months, three months, one and a half months, and so on. This infinite series adds up to one year: in this model, infinitely many doublings fit within a year. If the first doubling still takes half a year but subsequent intervals shrink more slowly, the total exceeds one year.
The authors of the Epoch AI analysis propose giving researchers different amounts of computing capacity and comparing the resulting improvements. They also suggest testing whether improvements carry over from small training experiments to large ones. If they do, ideas could first be tested in small, inexpensive experiments, with large training runs reserved for selected approaches.