Will Superintelligence Replace AI?

Artificial intelligence has a new name in US government communications: Super Intelligence, or SI. On September 29, 2026, Donald Trump signed an executive order directing federal executive agencies to replace the term AI. Days later, Elon Musk announced plans to rename SpaceXAI to SpaceXSI.

The developments explain the term's recent visibility. They also require a distinction: renaming a technology and demonstrating superintelligence are separate events. [1][2]

In technical discussions, artificial superintelligence refers to broad capabilities that would substantially exceed human performance. It would be a possible development within AI. In the US administrative usage, SI currently covers technologies previously described as artificial intelligence. [1][4]

What changed in the United States

The executive order Inaugurating the Era of Super Intelligence directs agencies to use SI in correspondence, websites, reports and other non-statutory documents, subject to applicable law.

Section 3 contains the important qualification: the initial definition covers the same systems as the existing statutory definition of AI. The order requests a proposed federal definition within 60 days and does not require changes to previously issued regulations, contracts or historical documents. [1]

Implementation is visible already. In a note created October 5, NIST said it was updating its communications to incorporate the term. [3]

The change applies to the US federal executive branch. It does not establish a worldwide replacement for AI or provide a test demonstrating a new level of capability.

AI, AGI and ASI

Term	Core idea	Important qualification
AI	A broad category of systems performing tasks associated with intelligence.	Includes specialized systems and general-purpose models.
AGI	Broad capability across cognitive tasks at a level comparable to humans.	Definitions and evidentiary thresholds vary.
ASI	General capability substantially exceeding human performance.	Requires evidence across many domains, beyond isolated results.

This is a simplified explanation. In Levels of AGI, researchers propose separating performance, generality and autonomy. Excellence at one task does not establish broad superiority. [4]

The concept also predates the current headlines. OpenAI published proposals for superintelligence governance in May 2023. Its political and commercial use has now become more prominent. [5]

Elon Musk: a proposed rebrand and ambitious forecasts

Reuters reported that Musk said on X on October 4 that SpaceXAI would become SpaceXSI. He provided no timetable or further details. This is an announced intention, without a demonstration of new capabilities attached to the change. [2]

Musk has also made ambitious technical forecasts. In his Davos 2026 conversation, transcribed by the World Economic Forum, he suggested AI could become smarter than any individual by the end of 2026, or no later than 2027. He also predicted it could surpass humanity's collective intelligence in 2030 or 2031. [6]

Those dates are Musk's predictions. The interview does not establish an evaluation procedure for measuring superiority over all humanity. They should be treated as expectations rather than laboratory results.

Jensen Huang: infrastructure, research and safety

Huang's position can be examined through NVIDIA announcements and public remarks.

On July 27, 2026, NVIDIA announced a long-term partnership and investment in Ilya Sutskever's Safe Superintelligence. According to the release, access to Vera Rubin systems will increase the lab's compute by an order of magnitude. Huang expressed interest in the breakthroughs SSI might discover. [7]

In September 2025, NVIDIA and OpenAI had already announced a letter of intent for at least 10 gigawatts of infrastructure, described as part of the path toward superintelligence. The announcement documents an infrastructure ambition, not attainment of the research goal. [8]

At the September 29, 2026 press briefing, Huang argued that innovation and safety can advance together and said industry was developing technology to improve both. These remarks express his position on AI development rather than certify superintelligence. [9]

What the benchmarks show

Benchmarks are standardized tests. Each measures only part of a system's capabilities.

Evidence	Result or method	Interpretation limit
2025 International Mathematical Olympiad	An advanced Gemini Deep Think system solved five of six problems and scored 35 out of 42, a gold-medal-level result reported by DeepMind and confirmed in a published statement from the IMO president.	Shows mathematical reasoning on that problem set. It does not establish broad superiority or automatically describe the public product. [10]
Humanity's Last Exam	The benchmark contains 2,500 questions spanning multiple subjects. When consulted October 6, 2026, Scale's panel listed GPT 6 Astra at 54.80% accuracy.	This is a snapshot of that panel and evaluation, not a percentage of human intelligence. Dataset versions and evaluation conditions matter. [11]
METR task-completion horizons	Estimates the difficulty of tasks agents can complete at 50% or 80% success, using human-expert completion time as a reference.	Does not directly measure how many hours an agent operates alone. Its tasks mainly cover well-specified software, machine learning and cybersecurity work. [12]

These results show meaningful capabilities. They also explain why any claim of superintelligence needs a definition before it becomes a conclusion.

The International AI Safety Report 2026, published in February with contributions from more than 100 experts, describes uneven capabilities: systems can succeed at difficult problems while failing at simpler ones. It also highlights the gap between controlled evaluations and real-world utility. The report is a snapshot of its evidence base, not a live update on every model available in October. [13]

Safety is part of the discussion

On the day of the executive order, a separate document, the White House Accord on Super Intelligence, set out company commitments. Musk and Huang were among its signatories.

It proposes four layers: internal controls, a team checking their operation, independent external evaluation and oversight by an independent board committee. The document says these measures might eventually be codified into laws or regulations. [14]

The commitment shows attention to verification and control. Its existence alone does not demonstrate that every proposed measure is operating effectively.

What to watch next

Three questions help assess future announcements:

What was demonstrated? A rebrand, an investment or a technical result?
How was it tested? With which tools, criteria, costs and error rates?
Who checked it? The developer, outside researchers or both?

Superintelligence is already official terminology in part of the US government and a stated research objective. Its technical meaning still requires evidence of broad superiority. The sources examined here do not provide a consensus demonstration that this threshold has been reached.

AI's development should be assessed through the results it delivers, its reliability and the ability to verify its actions. The name used to present it is only one part of that story.

Sources

All links were consulted October 6, 2026. Source documents are in English.