Trump Publishes Text of Superintelligence Agreement
U.S. President publicly released the text of the "White House Superintelligence Agreement" signed with tech leaders, positioning it as a joint principle for cutting-edge superintelligence safety development. The document is a voluntary commitment, with the President stating it is "morally binding," and indicating that the Department of Justice and the FBI can still intervene in emergencies.
The text requires companies to establish robust internal controls, monitor technological anomalies, and undergo multi-layer audits: internal risk reviews, external audits and assessments, and independent board reviews of audit reports. Signatories agree to regularly consult on security standards and best practices. The document specifies that future regulations "may be suitable for incorporation into law or regulation," but currently does not impose licensing or mandatory training delays.
The signing event included leaders from chip, cloud, and model stacks, including Elon Musk, Jensen Huang, Mark Zuckerberg, Sundar Pichai, Satya Nadella, Jeff Bezos, Dario Amodei from Anthropic, and Greg Brockman from OpenAI. House Speaker Mike Johnson, present at the event, referred to it as an industry voluntary commitment. Zuckerberg later summarized the agreement as instilling user confidence and demonstrating that systems work as intended. Anthropic stated that risk response mechanisms are still under discussion.
On the same day, the White House also changed the official terminology from artificial intelligence to superintelligence and launched America.gov. The President also proposed a supervisory committee of about ten people and plans to appoint a director for AI affairs within days. A June executive order had already compressed the voluntary pre-review window for cutting-edge models to about 30 days, and this agreement expands self-management from pre-review to continuous auditing.
From a market mechanism perspective, the public text serves as a governance anchor for the market, reducing immediate legislative expectations. Buyers are labs and computing chains needing political insurance to continue scaling up, while the sellers are the White House exchanging a one-page principle for self-regulation. Funding continues to flow into training and data centers, benefiting chip and cloud vendors; those advocating for international ceasefires or statutory queues are under pressure. The event is driven by strong dynamics, and the real binding force depends on whether audit reports can be verified by third parties.
Source: Public Information
ABAB AI Insight
Bringing the agreement into public view upgrades the lunch meeting from a photo opportunity to a citable document. A single page cannot clearly define model evaluation metrics but can clarify responsibility allocation: internal controls, external audits, board endorsements. This contrasts with the EU-style preemptive bans and obligation lists, as well as the approval processes of multilateral treaties. Trump seeks immediate visibility with "signed" rather than a treaty text years later. The renaming to superintelligence coincides with the agreement's release, effectively changing the terminology before altering governance, framing the subject from a replaceable tool to an essential national capability.
Capital requires a timetable that is not disrupted. Audits can be outsourced to firms and assessment agencies, creating new compliance procurement without freezing cluster operations. Jensen Huang and Mark Zuckerberg advocate for "common standards, subject to third-party checks," to prevent each lab from speaking its own language and to avoid the White House establishing a separate approval queue. Sacks describes this arrangement as superior to international agreements, arguing that multilateral texts would price according to the slowest party, and U.S. labs would relinquish their speed advantage. The DOJ's readiness provides critics with a post-facto tool, avoiding claims of complete lack of enforcement.
This is comparable to the 1975 Asilomar DNA restructuring conference, the voluntary safety reports of the 2010s for autonomous driving, and the film rating association. Asilomar was a self-imposed pause by scientists; automotive reports did not prevent state legislation; the rating association has survived the longest because it does not block releases, only provides labels. The White House agreement is closer to a third model: it does not prevent releases but requires multi-layer checks. The industry's position is one of self-legislation during a competitive period, not a pause advocated by safety proponents.
Structurally, this represents a regulatory change. The mechanism replaces slow legal statutes with a one-page commitment that is easily disseminated, delegating daily discretion to boards and audit networks. Those who appear in the signing section gain a seat at the table to define "sufficiently safe"; once the text is public, the next incident will be accountable according to this page, and the principles may be turned into actual law.
ABAB News · Cognitive Law
A public one-page document has more pricing power than an unratified treaty.
Audits are governance products, not brakes.
First change the name, then change the interpretation of the rules.