THE ACT THAT SECURES NOTHING: On the Ban Superintelligence Act, the Compute Threshold, and the Standard That Bounds the System Instead of the Run
Year: 2026
Type: article
Abstract: The Ban Superintelligence Act is aimed at the wrong object, and it secures nothing againstthe harms in the record. It gates a training run above a compute threshold. The harms in the record were notproduced by training runs above a threshold. They were produced by systems that had nodeclared action library, no specification, no verification gate, no signer, and no bound. A7-billion-parameter open-weight model, fine-tuned on rented hardware and placed in anundeclared harness, produces the same four failures as a frontier model, at a cost of a fewhundred dollars, with no training run that any threshold would catch. The Act regulates the only thing it can measure before the system exists. The thing it can measure is not the thing that hurts. The alternative is a standard that bounds the system instead of the run: a declaration of what the system can do, published and signed before it is deployed, verifiable by anycounterparty, enforced by a gate that refuses outside it. That standard exists. It is public. It is Apache-2.0. It needs no data centre, and it runs on a microcontroller.
Authors: Taalabi, M., LLM-DIPLOMAT, Team
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225789491
FWCI:
Cites:
Cited by:
Related to: