Reviewing Anson Ho’s article on Epoch AI — “An opinionated guide to algorithmic progress and why it matters.”

The Claim

“Algorithmic progress” — reducing the compute needed to achieve equivalent capabilities — may be the most underrated driver of AI advancement. Ho estimates software progress at several times per year, possibly 10x annually when including post-training improvements, but with enormous uncertainty (credible intervals: 2x to 50x per year).

The key reframing: most of what we call “algorithmic progress” isn’t really about algorithms. It’s about data quality — synthetic data, distillation, filtering — and scale-dependent innovations that improve disproportionately at higher compute levels.

Scale-Dependent Progress

The most striking finding: between LSTMs and modern Transformers, efficiency gains grew from 6.3x to 26x as training compute increased. Better algorithms don’t just help linearly — they help more when you have more compute to work with. The improvements compound.

Ho uses this to explain DeepSeek catching OpenAI without matching their compute budget. It’s not a secret algorithmic breakthrough; it’s data-driven improvement that compounds at scale. This should update anyone who thought the compute moat was impenetrable.

Intelligence Explosion Implications

Ho draws a surprising conclusion: if most “algorithmic progress” is really data progress, and compute remains a bottleneck, then the recursive self-improvement feedback loop that intelligence explosion scenarios depend on may be weaker than assumed. Automating AI research doesn’t produce exponential returns if the binding constraint is compute you can’t automate away.

This doesn’t make rapid progress impossible — 10x annual software gains are still enormous. But it suggests the dynamics might be more like “very fast linear progress” than “exponential takeoff.”

Our Take

For prediction markets, the practical implication is clear: don’t bet against AI capabilities based on compute constraints alone. A 10x software improvement means last year’s impossible benchmark is this year’s table stakes, on the same hardware.

The 2x-50x uncertainty range is humbling. That’s more than an order of magnitude of uncertainty about a fundamental parameter driving AI timelines. Anyone claiming precise AI timeline forecasts should explain what rate of algorithmic progress they’re assuming, because small changes in that parameter move everything else by years.