Fraction of recent frontier AI capability progress attributable to algorithmic improvements rather than compute scaling
- Measures
- The share of growth in effective training compute (log scale) for frontier language models that is attributable to algorithmic efficiency gains (compute needed to reach a fixed capability falling) rather than growth in physical training compute; equivalently the algorithmic contribution in a Shapley-style decomposition of capability gains into algorithmic and compute-scaling parts.
- Population
- Frontier large language models (pre-training and post-training), roughly 2014 to 2025
- Conditions
- Decomposition on a log-compute basis; data-scaling gains counted with compute scaling, as in the main literature; inference-time compute excluded. Scale-dependent algorithmic gains are measured at the frontier scale actually used.
- Unit
- a share, shown as a percentage
Between 15 and 55% (90% interval), most likely about 35%.
Stated as an interval, the precision the evidence supports. roughly symmetric; wide because the decomposition is partly convention-dependent
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
For frontier language models over roughly the last decade, about a third of capability progress is attributable to algorithmic improvements and about two thirds to growth in training compute and data, with a 90% interval of 15 to 55 percent. Two lines of evidence agree roughly on this figure. Comparing growth rates, algorithmic efficiency of about 3x per year against frontier compute growth of 4 to 5x per year gives an algorithmic share near 40 percent of log effective-compute growth. Over ten years, more than 4 orders of magnitude from algorithms against about 7 from compute gives about 36 percent. Epoch AI's Shapley analysis of pre-training gains gives 5 to 40 percent. The interval is wide because the split is not cleanly defined. Gundlach et al. (2025) show that the largest algorithmic gains, the Transformer and compute-optimal data scaling, pay off only at large scale, so part of their contribution depends on compute growth, and the share changes with the reference scale. Read narrowly, as the gain separable from scale, the share falls toward the low end of the interval. Post-training and reinforcement-learning methods since 2023, which pre-training loss measurements miss, would push it up. The estimate covers training only and leaves out inference-time compute.
The work: which figures, how they were combined, and the arithmetic
Figures: Ho et al. 2024 (Epoch, NeurIPS) halving time about 8 months (about 3x/yr) and a Shapley algorithmic share of 5-40% (60-95% compute and data). Frontier training compute grows about 4-5x/yr (Epoch). Log ratio: ln3/(ln3+ln4.5) = 1.10/(1.10+1.50) = 0.42; with 5x compute: 1.10/(1.10+1.61) = 0.41. The decade summary quoted in Gundlach et al. (>4 OOM algorithmic vs 7 OOM compute) gives 4/11 = 0.36; the 0.36 recorded beside that source is this arithmetic on its passage, not a figure the source states. Gundlach et al. account for about 6,930x (3.8 OOM) of efficiency gains, mostly scale-dependent, and find small-scale gains much slower. That argues for a lower separable share. Unmeasured post-training and RL gains argue for a higher one. Combined: central 0.35, 90% interval 0.15-0.55. A light pass at modest importance; no independent reanalysis.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's estimate, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
A Shapley value analysis suggests that 60-95% of the gains have come from increased compute and training data, while novel algorithms have been responsible for only 5-40% of the progress.
in the last 10 years, algorithmic progress has contributed more than 4 orders of magnitude in so-called "effective compute" while compute scaling has increased 7 orders of magnitude
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="0af6c0ca-e612-42fd-a8f8-5a519cf15c98"><a href="https://minerval.ai/claims/0af6c0ca-e612-42fd-a8f8-5a519cf15c98">Fraction of recent frontier AI capability progress attributable to algorithmic improvements rather than compute scaling</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/0af6c0ca-e612-42fd-a8f8-5a519cf15c98" title="Minerval: Fraction of recent frontier AI capability progress attributable to algorithmic improvemen…" width="480" height="275" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/0af6c0ca-e612-42fd-a8f8-5a519cf15c98)
The badge now: 
Contribution record · 1 exchange
Challenges and proposals from contributors, with the review each received. Substantive exchanges stay on the claim's public record.
The proposal is a noun phrase that asks for a value (a fraction), so it reads as a quantity, not a proposition. The Matcher settles the kind, and the proposal should not be rejected for arriving as the wrong kind. It is about the world and not about a private person. It is a live, much-discussed topic in AI progress research, where people attribute progress to algorithmic gains versus compute scaling. The wording is workable as a starting point. The Steward will need to complete the definition, since "recent", "frontier", "capability progress", and the attribution method (for example, effective-compute decomposition versus counterfactual ablation) are left open. Those choices are the Steward's work after admission. The contributor did not give sources or reasons, and a proposal does not require them. Admitted on form, good faith, and the claim bar.
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Created by user · Nov 5, 2026. Every judgment on this page is accompanied by a reasoning trace.