AI summary
- A paper titled 'The Last AI Humanity Will Ever Create', covered by Live Science on 30 October, sorts self-improving AI into five levels, and nobody has shown the top one.
- Real examples exist but they're modest: the Darwin-Gödel Machine rewrites its tools rather than its model, and AlphaEvolve cut Gemini's total training time by 1%.
- A US deadline for a draft definition under the executive order that renames AI as 'superintelligence' falls at the end of November.
- The name may get settled by policy before the science settles it, so a .si name is still mostly speculation.
- Only buy a .si when the .com and .ai are taken or expensive, the words are generic and people actually search for them.
A paper about the "last AI" humans build
On 30 October, Live Science covered a paper from a Chinese research team with a dramatic title: "The Last AI Humanity Will Ever Create". I found it through a write-up by Fukuoka Koji on note. He says he started reading it annoyed by the headline, but the content was "surprisingly grounded".
The paper is about recursive self-improvement, or RSI. That's AI deciding its own training and development methods and passing what it learns from one improvement to the next. Repeat that often enough and, in theory, new generations arrive with almost no human help. It was posted to arXiv on 10 September and it hasn't been peer reviewed yet.
Why should a .si buyer care? Because self-improvement is the step most people mean when they talk about super intelligence, and the US is about to try to put a definition on paper.
The five-step ladder, and where we really are
The paper sorts systems into five levels, based on how much of the improvement process the AI handles itself. Fukuoka's chef analogy makes it easy to follow:
- Level 1: the AI follows a recipe a human wrote.
- Level 2: it chooses how to improve its own skills.
- Level 3: it decides what to learn from next.
- Level 4: it feeds what it learns from real customers back into its sourcing and tools.
- Level 5: it refines the way it improves, and even retrains its own taste testers.
Nobody has shown level 5. The coverage concludes that partial automation is moving forward, but a complete fifth level hasn't been demonstrated.
The examples are real, though modest. The Darwin-Gödel Machine, built by Sakana AI and the University of British Columbia, lets an AI rewrite its own program and keep the better versions. On SWE-bench it improved from 20.0% to 50.0%. But it's rewriting tools and work procedures; the underlying language model stays fixed. Google DeepMind's AlphaEvolve, announced in 2025, found a way to do 4x4 complex matrix multiplication in 48 operations, breaking a record set in 1969. That finding reportedly sped up one Gemini operation by 23% and cut total training time by 1%.
Sakana AI announced an RSI Lab in Tokyo in June, and Jürgen Schmidhuber joined it as Chief Scientific Advisor in September.
Not everyone buys it
Professor Peter Bentley of University College London calls the idea that AI will spontaneously create one better AI after another "nonsense". He argues the hard part is an AI doing the research to build its successor without depending on human expertise and physical computing. Even the paper's authors admit there's no reliable way yet to tell whether an AI has truly got smarter or is just gaming the test.
Kyla Blomquist of the University of Oxford reads the paper as describing responsibility being handed over little by little, not a sudden switch. That matters for naming: if there's no clear line, there's no clean date when "AI" becomes "SI".
Why the labs are nervous about the same thing
This isn't only an academic debate. According to Migalhas, Anthropic CEO Dario Amodei published an open letter on 12 September asking the leading labs to slow capability gains together, and named recursive self-improvement as an extra worry. On 18 September, four consumers filed a class action against Anthropic, OpenAI, SpaceXAI and Google, arguing that coordinated slowing would breach the Sherman Act. On 29 September, the White House signed the White House Accord on Superintelligence with the heads of the main AI companies.
Quartz reports that Jacob Coxon, who'd spent three years on pre-training research at OpenAI and then Anthropic, resigned on 8 September. He said neither company was acting responsibly and both were racing towards self-improving superintelligence. Quartz also reports that on 10 February Anthropic ruled Claude Opus 4.6 hadn't crossed its threshold for AI that can do AI research by itself, while admitting that ruling it out was getting harder.
The end-of-November definition deadline
Here's the part that's directly relevant to .si. Fukuoka notes that in the US, the deadline for submitting a draft definition under the executive order that renames AI as "superintelligence" is approaching at the end of November. His point is a fair one: even if the name is decided first, we'll still need a separate yardstick to measure which step we've actually reached.
I read that two ways. A formal definition could make "SI" feel more official in government writing. But this paper shows the technology hasn't reached the top rung, and experts don't agree it will. The word may be settled by policy before it's settled by science, and policy can change.
Which .si names fit this story
If you're building a reasoning tool, a research project or commentary on what self-improving AI means for people, a few names in my inventory fit:
- greatmind.si, $2,000: a brandable name for a reasoning model, tutor or mentor. It shows 33,100 searches a month, the .com is listed at $500,000 and the .ai is for sale with the price on request.
- futurehuman.si, $1,500: the "what happens to people" question in two words. It shows 5,400 searches a month, the .ai is listed at $290,000 and the .com has been held since 1999.
- brainbank.si, $1,500: a memory bank for machines and the people who use them. Searches are low at 1,600 a month, the .ai is listed at $399,999 and the .com has been held since 1995. Several startups already trade as BrainBank, so check trademarks before you build on it.
None of these has a guaranteed resale value. I've priced them for people who'll actually use them.
Should you buy a .si name because of this?
Not because of this paper alone. A preprint that hasn't been peer reviewed and a definition that's still a draft don't change the basics. Most .si names are speculation, and the aftermarket is small. If the .com you want is available at a fair price, buy the .com. It's still the default people type.
A .si makes sense only when it passes three questions:
- Is the .com and .ai taken or expensive? If either is cheap and free, that's usually the better buy.
- Is it a generic name, not a brand? Plain words like "great mind" or "future human" are fine. A copy of one company's name isn't.
- Do people search for it? Look at real monthly searches, not the headline of the week.
If a name fails any one of those, I'd skip it, even if it's one of mine.
Names from this store that pass the test
Questions
What is recursive self-improvement?
It's AI determining its own training and development methods and passing the lessons from one improvement to the next. The paper covered by Live Science on 30 October sorts systems into five levels based on how much of that process the AI handles itself.
Has any AI reached the top level of the ladder?
No. No examples of the fifth level have been shown, and the coverage concludes that partial automation is progressing but a complete fifth level hasn't been demonstrated.
When is the US draft definition of superintelligence due?
According to Fukuoka Koji's write-up on note, the deadline for submitting a draft definition under the executive order that renames AI as 'superintelligence' is approaching at the end of November.
Which .si names fit the self-improving AI story?
greatmind.si is $2,000 with a .com listed at $500,000, futurehuman.si is $1,500 with a .ai listed at $290,000, and brainbank.si is $1,500 with a .ai listed at $399,999.
Sources
- note (Fukuoka Koji), 2026-10-31
- Quartz (French edition), 2026-10-31
- Migalhas, 2026-10-29
This post was researched and drafted with AI help from the sources above, then checked by me before publishing. Facts are as reported on the dates shown.