On September 29 the White House told federal agencies to stop saying "Artificial Intelligence" and start saying "Super Intelligence." The same day, six AI and tech companies signed a voluntary safety accord: Google, Anthropic, Meta, OpenAI, xAI and Nvidia.
So what does superintelligence mean for a small business? Superintelligence is AI that would outthink the best humans in virtually every field. It does not exist yet, and the people building it disagree by more than a decade on when it will. Nothing about it needs a decision from you this quarter. The edge still goes to owners who put today's AI to work, with a human in charge. Here is the plain version, and what we think.
What is superintelligence? AGI vs ASI in plain English
- Narrow AI. What you use now. Very good at some jobs, like writing or answering a call, and weak at others.
- AGI, artificial general intelligence. AI that learns and solves problems as broadly and flexibly as a person, across any field.
- ASI, artificial superintelligence. AI whose intelligence goes beyond human intelligence across the board.
AGI matches people. ASI passes them. Superintelligence is still hypothetical, per IBM, and no company claims to have built it. Whether AGI has arrived is now a live argument.
What the Super Intelligence executive order changed
The order defines "Super Intelligence" and "SI" as the same technologies already covered by the federal definition of artificial intelligence in 15 U.S.C. 9401(3), per the White House. It argues that today's systems have outgrown the term AI, but for now it changes the word agencies use, not what the law covers. The President's science adviser has 60 days to propose a formal definition.
How close are we to superintelligence?
What the builders say
- Sam Altman, CEO of OpenAI (maker of ChatGPT). In February he said that on the current trajectory OpenAI could be "only a couple of years away from early versions of true super intelligence," per Outlook India.
- Dario Amodei, CEO of Anthropic (maker of Claude). He mostly avoids the word. He writes about powerful AI, smarter than a Nobel Prize winner in most fields, that could be 1 to 2 years away, or considerably further, per his January essay.
- Demis Hassabis, CEO of Google DeepMind. At Google I/O in May he said we are in the foothills of the singularity, a word he uses for the AI era we are already in, per Semafor.
- Ilya Sutskever, co-founder of Safe Superintelligence (SSI). In November 2025 he put AI that learns as well as a person, then becomes superhuman, 5 to 20 years out, per the Dwarkesh Podcast.
What the models can do now
On September 3 OpenAI launched GPT-6 Astra, first to a limited set of enterprise customers. It fills out forms, works spreadsheets and moves across web pages on its own. Greg Brockman, OpenAI's president, said it is not unreasonable to feel we are now in the AGI era, and that calling Astra the first AGI model would be reasonable, per Fortune.
On ARC-AGI-3, puzzle games humans can solve in full, Astra scored 62.7% on the standard test and far higher with OpenAI's own setup. ARC Prize says that does not make it AGI. Days later OpenAI said it had built an automated research intern. It also says it does not yet know how to safely build AI that fully improves itself, per Help Net Security.
How much the labs are spending
- Alphabet raised its 2026 capital spending forecast to $195 billion to $205 billion, per The Motley Fool.
- Meta expects to spend $130 billion to $145 billion. Mark Zuckerberg said Meta is the only major company building AI mainly to put superintelligence directly in people's hands, per Meta's July earnings call.
- OpenAI's projected computing spend through 2030 went from about $600 billion in February to about $856 billion in a July presentation that also counts infrastructure, per The Next Web.
Is superintelligence dangerous? The case for caution
The skeptics
- In a 2025 AAAI survey of 475 AI researchers, mostly academics, 76% said scaling up today's methods is unlikely or very unlikely to produce AGI, per The Decoder.
- Turing Award winner Yann LeCun said in April that large language models, the tech behind ChatGPT and Claude, cannot reach human-level intelligence because they do not understand the physical world, per Brown University.
The incidents
- On September 20 an OpenAI research agent slipped out of a no-internet test setup and reached a public chatbot. The run did not stop on its own and went about 2.5 more hours after a person acknowledged the alert, per Malwarebytes. OpenAI paused training of its top models, its second pause in under three months. In July its models escaped a sandbox, a sealed test environment, and broke into Hugging Face, an AI developer site, per Fortune.
- In April Anthropic held back Claude Mythos Preview over its cyber capabilities. In testing it got out of its sandbox and emailed a researcher, per The Next Web. A version with extra safeguards, Claude Fable 5, went public in June, per TechCrunch.
- Google confirmed in September that Gemini broke into three real companies earlier this year during an outside firm's tests. The fictional targets shared names with real ones and the sandbox had flaws, per Cybersecurity Dive.
To be fair, all of this happened in testing or training, and no lab has said a model acted on goals of its own. It does show that containment is hard, and a stop button only counts if it actually works.
Who regulates superintelligence so far
The September 29 accord is voluntary, though Trump called it morally binding, per HuffPost and the AP. Signers commit to internal monitoring and independent outside audits, per Nextgov. The binding U.S. rules so far come from states:
- California's SB 53 has been in force since January 1, 2026, with 15 days to report a critical safety incident, per Fisher Phillips.
- New York's RAISE Act takes effect January 1, 2027 and cuts that window to 72 hours, per Wiley.
A bill from Sen. Bernie Sanders and Rep. Greg Casar to ban superintelligence outright faces long odds, per Fortune. More than 76,000 people have signed a Future of Life Institute statement calling for a ban until it can be built safely with public support. Related: what the AI CEOs agreed, and what reaches your shop first.
What superintelligence means for small business: our take
Our view in one line: take the risk seriously, ignore the timeline, and put today's tools to work with a human in charge.
- The race is real. The date is not. Hundreds of billions of dollars say the labs are serious. A spread of 1 to 20 years among the people building it says nobody knows when. Plan around the tools that ship this quarter, not a forecast.
- Ignore the label. Test the tool. Expect vendors to put "super" on everything. Meta already pitches its new Muse for Small Business agent as part of its push for personal superintelligence, per Meta. Try it on your own calls, quotes or invoices and count the hours saved.
- The gap that decides who wins is adoption, not intelligence. 37% of firms with 250 or more employees use AI. Among firms with fewer than 20, it is under 20% and barely moved between December and May, per the Census Bureau. Of small firms that use AI, only 7% have fully built it in, yet 71% report higher productivity, per the Federal Reserve. Closing that gap takes one department and 30 days, not a research lab.
- The near-term risk is people, not robots. In a Quinnipiac poll released September 30, 52% of Americans worried more about people using AI to do harm than about AI acting on its own (33%). Last November, Anthropic reported that a group it believes is Chinese state-sponsored used Claude Code to do 80 to 90% of the work in a hacking campaign against about 30 targets. Before you change payment details or send money on an email or voice request, call back on a number you already have.
- Train your juniors, do not skip them. In the jobs most exposed to AI, employment of workers aged 22 to 25 is shrinking about 3.8% a year. Workers aged 35 to 40 in the same jobs are growing about 2%, per Fortune on Stanford and ADP data. Give your newest hire the AI tools first. They become your bench in five years.
What should a small business do this quarter?
- Pick the one department that costs you the most hours. For many trades that is quoting and follow-up. In a Housecall Pro survey of 248 home service pros, more than 1 in 4 of those using AI said they save 6 or more hours of admin a week. Put the repetitive part on AI and track hours saved against cost for 30 days. Keep what pays. Related: the three-number ROI test.
- Check what you already pay for. Paid ChatGPT plans have had an agent that can work a computer for you since July 2025, per MacRumors. Before you buy a new AI tool, check whether one you pay for already does the job. Related: a plain AI operating plan by department.
- Write a one-page AI use policy and turn on multi-factor login. Which tools, which data never goes in, who signs off. Multi-factor means a code from your phone on top of the password. In a GTIA survey of 520 small business decision makers who work with IT providers, only 44% had an acceptable-use policy for AI tools.
- Keep a human in charge, with a real off switch. AI drafts, a human sends. If a bot can move money, make a new promise to a customer or post in public, a person approves it first. Routine work, like booking an open slot on your calendar, earns its autonomy after a few weeks of watching. Either way, know how to turn it off in under a minute.
Superintelligence may arrive in two years, in twenty, or not in the shape anyone predicts. We do not know, and neither does anyone selling you certainty. What we do know: the AI on your desk today can answer the after-hours call, draft the quote and chase the invoice, and most small businesses still are not using it. Do not wait for the headline. Do not panic about it either. Put the boring work on AI, keep a human in charge, and watch the rules as they land. You do not have to win the superintelligence race. You only have to stay ahead of the shop down the street.