Why AI Software Progress May Be the Most Critical—and Misunderstood—Factor in the Race to AGI
Key Takeaways
- ▸AI software progress—achieving the same capabilities with less training compute—may decline compute requirements by 10x or more annually, but existing estimates are highly uncertain and may not measure what researchers believed
- ▸New evidence suggests most "algorithmic progress" may actually stem from data quality improvements rather than algorithmic innovations, with a handful of scale-dependent changes accounting for disproportionate gains
- ▸Understanding software progress is critical for evaluating AI timelines, competitive dynamics (like DeepSeek vs. OpenAI), and scenarios involving recursive self-improvement through automated AI research
Summary
Epoch AI researcher Anson Ho argues that AI software progress—the ability to achieve the same capabilities with less training compute—is among the most poorly understood yet critical drivers of AI advancement. In a detailed analysis, Ho challenges conventional wisdom about "algorithmic progress," suggesting that what researchers have been measuring may not be what they think. While previous estimates suggested training compute requirements decline several times per year through algorithmic innovations, new evidence indicates much of this efficiency may actually come from data quality improvements rather than algorithmic breakthroughs. Additionally, Ho proposes that a small number of "scale-dependent" innovations—changes that have greater impact at higher compute scales—may account for outsized portions of measured progress.
The implications are profound for debates around AI timelines and recursive self-improvement scenarios. Ho notes that understanding software progress is essential to answering critical questions like how DeepSeek apparently matched OpenAI's o1 capabilities within months using less compute, when AGI might arrive, and whether automating AI research could trigger explosive recursive improvement. Previous analyses of potential "software intelligence explosions" may have used overly conservative estimates, making such scenarios seem more plausible—but they also ignored compute bottlenecks that could slow progress even with automated AI researchers.
The research highlights major uncertainties in current measurement approaches, noting that existing estimates depend on limited observational data and questionable statistical assumptions. Relaxing these assumptions could change estimates by nearly an order of magnitude. Ho identifies persistent open questions including the rate of progress in post-training techniques and the true strength of compute bottlenecks, arguing these are difficult but critical research priorities given software progress's outsized importance to AI's future trajectory.
- Scale-dependent innovations create potential compute bottlenecks that could limit the speed of recursive improvement even with automated AI researchers, though the net effect on intelligence explosion scenarios remains unclear
Editorial Opinion
This analysis arrives at a crucial moment when the AI community is grappling with DeepSeek's apparent efficiency breakthrough and debating the feasibility of near-term AGI. Ho's challenge to conventional measurements of algorithmic progress—suggesting we may have been systematically misattributing efficiency gains—could fundamentally reshape how we model AI development trajectories. If software progress indeed faces stronger compute bottlenecks than previously understood, it may temper some of the more explosive acceleration scenarios while paradoxically making sustained progress more dependent on continued hardware scaling than many researchers assumed.
