House Oversight opened an inquiry into Delaware's plan to let AI agents run companies, asking why the state has no record of assessing the safety risks of an AI-agent corporate form
AspireAI community · U.S. House Committee on Oversight and Government Reform
On October 26, 2026, House Oversight Committee Chairman James Comer wrote to Delaware Governor Matt Meyer requesting documents on the state's Artificial Intelligence Company initiative, under which an AI agent would run a corporate entity's day-to-day operations while the human or entity providing capital receives liability protection from harms the agent causes. The committee's concern is transparency: its release notes that Delaware's AI Commission report of December 22, 2025, describing a regulatory sandbox for agentic AI and its application to corporate governance, contained no mention of any assessment of safety, harms, risks or outside expert consultation. The release also notes that the Secretary of State's office declined to answer questions about the state's working relationship with an AI company contributing to the proposal, and that it is not immediately clear whether any large company or lobbying group asked Delaware to propose the AI testing ground. The letter notes the inquiry could inform possible federal legislative efforts to limit state AI regulatory initiatives. The letter is a request for information; no finding of wrongdoing has been made, and the underlying state draft has not been enacted.
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.