On September 29, 2026, Donald Trump’s White House meeting with AI industry leaders concluded with a voluntary accord on controls for frontier models. House Speaker Mike Johnson presented it as a statement of industry principles and commitments. His remarks linked technology safety with maintaining the US advantage over China. Source: Mike Johnson’s remarks after the meeting.
The president signed a separate executive order that day changing the terminology used by the administration. Buyers of AI services should examine the documents separately: one describes proposed company controls, while the other governs terminology within the federal executive branch. Source: the Super Intelligence executive order.
What the AI industry leaders signed
The signed scan names Trump, Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Elon Musk (xAI) and Jensen Huang (Nvidia). It describes four oversight layers: internal controls, a team checking those controls, independent external assessment, and an independent board committee receiving reports. It specifies neither completion deadlines nor named auditors. Source: the signed accord, hosted by PBS.
In our view at Lazarus Systems, that structure gives buyers a basis for specific questions. What system did the auditor examine? Did the auditor have access to the environment where the model operates? How does the company record problems and verify fixes? A signature on a commitment does not answer those questions.
When selecting a service, we would separate marketing material, the vendor’s stated policies and evidence that checks were completed. An audit report should identify the model version and operating conditions it covers. Without those details, judging its relevance to a particular deployment is difficult.
What the “Super Intelligence” executive order changes
The order directs the federal executive branch to replace AI and Artificial Intelligence with SI and Super Intelligence in current communications and non-statutory documents, within the limits of the law. It retains a reference to the existing statutory AI definition and does not require changes to earlier contracts or actions. Source: the White House text.
A terminology change provides no technical evidence of capabilities exceeding human abilities. A product labelled SI still needs testing on defined tasks. We propose recording the model name and version, data access and evaluation conditions in procurement documentation. That gives buyers a basis for comparing offers regardless of the label.
Lazarus Systems commentary: control over the process
We recommend planning AI deployments with a clear division of responsibility. A model can prepare a proposal, while the application determines which data and operations it can access. Writing to a production system requires checking the result against the rules of that process.
Consider an assistant that retrieves machinery manuals. When the model provider is unavailable, the user should still be able to open the source document and search by equipment number. If the system cannot find instructions for the correct version, we would route the case to an employee. An answer based on a similar machine may look credible while failing the task’s requirements.
Local AI deserves consideration where data, availability or response-time requirements justify operating a company-controlled environment. We would budget for hardware, updates, monitoring and someone responsible for outages. Some stages could use an external API if the data being shared and the dependency on that service are acceptable.
Our guide to local Qwen models provides a starting point for model selection. The Claude Sonnet 5.5 analysis covers costs and integration for agents using an API. These are two options to test against the same tasks.
When preparing an AI project with Lazarus Systems, start with one process: name its owner, input data, acceptable error rate and operating procedure during a provider outage. That description gives us a basis for assessing what the business needs from models and infrastructure today.