The AI Blur: Navigating the Undefined

2026 has been a breakneck year for AI, just as 2025 was before it. But this one is different. This year, we’re seeing the first signs of RSI, or recursive self-improvement, at the major American labs like Anthropic and OpenAI, right as they march toward IPOs that could value them at an astonishing trillion dollars.

So what is RSI, exactly? Well, that’s the first step toward AGI! And what is AGI? Artificial general intelligence, of course. So what does that actually mean?

You’re asking a lot of questions now. The trouble is, no one really agrees on what AGI is, so it’ll be hard to give you a straight answer. But we can try to untangle this conundrum together in this article.

Recursive Self-Improvement

Since every journey of a thousand miles begins with the first, let’s find out what Wikipedia has to say about RSI:

W

Recursive self-improvement From Wikipedia, the free encyclopedia

Recursive self-improvement (RSI) is a process in which early artificial general intelligence (AGI) systems rewrite their own computer code, causing an intelligence explosion resulting from enhancing their own capabilities and intellectual capacity, theoretically resulting in superintelligence.^[1][2]

While the entry reads clearly enough, it leaves the important parts vague. It never says which code the AI is actually rewriting, whether it’s bound by hardware limits, or how any of this intelligence gets measured.

Maybe the labs’ own writing will help us understand it better. In this Anthropic blog post, they inform us that Claude is already doing most of the engineering, which is indeed a real improvement, but it wouldn’t count as self-improvement to me: it’s helping train the next model, not rewriting itself.

And there the two ideas get quietly conflated, AI helping them build their product versus AI improving itself. Most of their code being authored by Claude isn’t surprising. Hell, most of my code is written by Claude, but that doesn’t mean my Laravel CRUD app is taking me to the moon anytime soon.

As for measuring intelligence, they point to saturated benchmarks. That’s true, but beating a benchmark only proves you beat that benchmark. It doesn’t make the test a reliable measure of intelligence, and a saturated one could just mean the answers leaked into training data, the test was too easy to begin with, or models were simply tuned to ace it.

The article then warns of risks to society “if this trend holds,” pinning the decisive milestone to some unspecified point on the curve, whenever growth turns exponential. Their (public) call for regulation feels hollow with that much left unanswered. OpenAI takes the same line in its own blueprint, and neither lab is really straightforward about which part of their work should actually fall under regulation, other than GPU sales.

The Semantics of Intelligence

It’s a continuum, and we are all definitely making progress. But in the way people describe RSI, that would represent a next level of acceleration and would have a lot of implications, but we aren’t quite there yet.

Pichai’s take brings another layer of indirection, and seems to contradict (as much as confirm) what the other labs claimed. It could be that Google is late, or it could be that those words don’t actually carry any inherent meaning. Do you remember the AGI discussion earlier?

W

Artificial general intelligence From Wikipedia, the free encyclopedia

Artificial general intelligence (AGI) is a hypothetical type of artificial intelligence that matches or surpasses human capabilities across virtually all cognitive tasks.^[1]

You noticed it too. It’s a “hypothetical type of artificial intelligence.” While this definition does nothing to anchor Silicon Valley’s claims in the real world, it also completely avoids grappling with what intelligence is, and how an artificial one would match or surpass humans without ever being embodied.

As of now, everything an LLM knows is a statistical compression of human writing, downstream of every article, manual, and forum argument it ever swallowed. Its knowledge isn’t its own, it’s a distillation of ours. Humans had the experiences and wrote them down; the model learned the writing, but it doesn’t derive meaning from it the same way we do.

That doesn’t stop AI from being useful to us. But it does cast doubt on whether it can be intelligent in our sense of the word, inasmuch as the human mind evolved in and for a body, as a biological survival imperative.

AGI Claims Over the Years

Nowhere is the blur clearer than when you line up what the people building AGI actually say about it. In his 2025 Reflections, Sam Altman wrote that OpenAI was “now confident we know how to build AGI.” Months later, the same Sam Altman called it a pointless term. Over at Anthropic, Dario Amodei places it one to three years out. And Elon Musk skipped the wait, arguing GPT-4o already cleared the bar. So which is it: a solved problem, a meaningless phrase, a few years away, or something we passed two model versions ago? When one word can mean all four at once, it has stopped describing reality. It’s doing marketing.

The pattern isn’t hard to spot. The US labs are the only ones playing this game of hyping AGI to the point of demanding regulation. Elsewhere the framing is different: Mistral and much of Europe push privacy-first, sovereign models, while China leans into open weights and practical, industrial uses of the technology.

Consequences and Impact

But why, you ask, should we care about the major players’ cunning use of words? Because it has real consequences, not least for people’s mental health. Losing your job is a frightening prospect, and weaponizing that fear to raise investors’ money is hard to defend. Setting the philosophy aside, no one can really know what AGI would do to the world economy anyway. And if anyone could, it would be an economist, not a CEO with stock to sell.

Through 2025 and into 2026, company after company blamed job cuts on AI, so much so that it even earned a name: AI washing. It turned out this didn’t hold. Oxford Economics concluded that firms “don’t appear to be replacing workers with AI on a significant scale.” The culprits were pandemic overhiring and the Fed’s interest rate hikes. Even Sam Altman conceded that “almost every company that does layoffs is blaming AI, whether or not it really is about AI,” as if this weren’t partly his doing.

Let me be clear: I’m not anti-AI. The technology is genuinely useful, I work with it every day, for code, for editing, and for research. By now it has probably found its product-market fit and is here to stay. But the labs’ focus seems to have shifted to enterprise, and that’s likely where it stays for the foreseeable future. OpenAI’s attempts at consumer hardware have been unconvincing (even with Jony Ive), and no one’s using Microsoft’s Copilot products. This is a far cry from the world-shifting narrative we just talked about.

Conclusion

Navigating the undefined turns out to be simple. When a lab reaches for a word it won’t define, that’s the tell. Ask what it actually means, and who benefits from it. And watch how quickly trillion-dollar certainty melts back into “a continuum” and “we aren’t quite there yet.”

The technology indeed deserves positive attention, but it won’t get it while the people selling it keep concepts blurry enough to weaponize them.