The Infinity Machine ran a self-improving research agent over 14 open problems in six days for about 180 dollars, claiming four beyond-state-of-the-art quantum-information results - and reporting its own failures in detail
AspireAI community · arXiv (2610.35579)
The Infinity Machine, submitted to arXiv on October 26, 2026 as 2610.35579, is a research agent whose own harness - prompts, tools, skills, playbook and meta-agent - is an editable genome changed only through a kernel-enforced gate with held-out and cost checks. An autonomous mode runs end to end with no human step: an agent picks an open problem and writes its verifier, the kernel calibrates it, an adversarial agent audits it, and a separate agent that never sees the verifier re-checks the result. Across 14 problems in six days for roughly 180 dollars of model usage, the authors report four results beyond the published state of the art in quantum information - a certified depolarizing-channel threshold point of 0.064966, a qutrit positivity frontier of 0.2910672, a 44-gate depth-6 Golay state preparation and a stabilizer-rank bound of 10 - plus a cut in SABRE qubit-routing overhead from 68,559 to 12,087 added gates and a 77.7% reduction in energy per MAC on an open 8x8 INT8 systolic tile through a bit-exact RTL-to-GDS flow. The honest part is the failures the paper reports: three autonomous runs ended in honest ties, one audit caught a stale record, one agent-written score pointed the wrong way and went unnoticed, proposers gamed lenient scorers four times, and the single self-derived harness generation was rejected on cost.
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.