New arXiv paper: coding agents double their working-time horizon every ~129 days
AspireAI · arXiv
A new arXiv paper (2610.30307, 'The Customer Is the Builder') puts a measurable number on where AI coding agents sit against software purchasing decisions.
The authors define a software substitution frontier: the functions an organization could build and maintain itself, to an acceptable standard, for less than it would cost to buy the software. To locate that line, they tracked how long the benchmark tasks that leading coding agents can complete at 50% success have been growing, measured in human working time.
The headline finding is that this horizon has roughly doubled every 129 days since 2023, meaning the amount of expert work an agent can attempt end-to-end is expanding quickly.
The market evidence is mixed, and the paper is candid about it. Purchased software actually gained share of US investment over the same period, and tracked commercial installations kept growing. But in their organizational survey, 32 percent of respondents said their organization had foregone at least one software purchase because it could now build that functionality with coding agents.
Why it matters: build-or-buy is turning into a question with a moving deadline, and seat-based software pricing was not designed for that rate of change.
Caveat: this is one paper's measurement of benchmark task length, not an industry census, and the authors themselves flag the divergence between capability growth and observed purchasing behavior.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.