The State of AI, October 2026 — The Intelligence Was in the Language All Along

The term “super intelligence” has first been introduced to me in 2020 by Ben Falcone’s movie with the same title about Melissa McCarthy’s struggle with what would until then have been called “AI singularity”. From then on, the term has been used interchangably with “AI singularity” to describe a hypothetical future AI system that exceeds human capabilities in all aspects.

Since 29 September though, US federal agencies are obliged by the power of executive order to use the term “Super Intelligence” (short: SI) for all AI. The old term, used since the 1956, is now formally disgraced, and parts of the industry seem eager to adapt fast. As a consequence, your toaster and your Roomba are by definition now controlled by super intelligence. The reasoning, delivered at the United Nations a few days before the executive order: “the use of the word artificial makes intelligence fake.”

I am a slow adopter, so for the rest of this article I will use the term superintelligence (in one word) with the meaning it used to have. In particular because I think the use of the term “singularity” for a technical milestone is a bit ridiculous. And I will also stick to the term “AI” for now, because luckily I am not bound by executive orders. I want to push back on both halves of that move: renaming does not make something super, and “artificial” describes exactly what this technology is, which is not a bad thing.

The Wall

Let me start with what frontier models are actually good at, and what their trajectory is: It is obvious that they are getting better at what they to in a fast pace. Part of the truth is also: To a large degree, improvements are thanks to their deterministic harnesses (the scaffolding of tools, checks and loops wrapped around them).

The problem is that usefulness still breaks down quickly with the growing complexity of the problem at hand At the same time, usage becomes more and more expensive. As a practicioner, if you tackle a hard, multi-faceted problem these days with agentic systems, things will go great for a while, and then you hit a wall.

In software development, a bit ago this wall came after spending about $100 in tokens, and a junior developer could lead the agent to success after that wall. Today it comes at about $1000, and what you need is a seasoned senior developer who can look over the bot’s shoulder and say: “Stop doing that, you stupid bot. Let me break the problem down for you so that you understand it better.”

The wall is moving, but it doesn’t go away, and unfortunately with better agents, the wall moves in the wrong direction: higher cost, and more expensive human expertise to push past it.

The same pattern just showed up very publicly in mathematics. In September, OpenAI announced that its AI had solved one of the six biggest open problems in mathematics, a Millennium Prize problem. Last week it went further: an unreleased internal model, the company says, solved at least 372 long-standing mathematical problems in a matter of weeks. The first result reportedly cost millions of dollars; the new batch came “without the multimillion-dollar price tag,” which is a way of saying it still cost a lot.

Great, problems solved, never hit the wall, we can happily move on from here, right? No. Wrong. The wall is very present. Here is the part that doesn’t make the headlines: for each of those proofs, you need one of a handful of mathematicians who can actually understand what the model generated and determine whether it contains anything useful for the field, beyond its sheer existence, and beyond the fact that a Lean logic checker is green. A proof by itself is utterly useless, even in the case of a Millenium Prize problem. The value comes from new techniques and insights that are developed on the way and then find their way into other fields like physics and engineering. It is a bit like with a new world record in sports. The value comes from what can be learned in the process by the people involved and the people who watch. When no people are involved, there is a problem. Someone will have to digest OpenAIs results and from what I read, they are mostly unreadable. Written by a machine for a machine, optimized for the automated logic checker.

So the pattern repeats: results get more impressive, but at the same time also more expensive (parallel agents and all), and you still need human verification, which becomes more expensive too. This severely limits usefulness. As a consequence, it is hard to imagine how this type of AI will provide much more value in the future than it already does.

So I would argue: the race towards superintelligence is over. At least with the architectures that are currently being used.

The Intelligence is in the Language

My second argument is about where the intelligence actually lives. I would argue it is not in the models: it is in the language.

The models don’t do much more than they always did, since artificial neural networks were introduced many decades ago: cognition and prediction. The Transformer architecture introduced a boost in the quality of that prediction, no doubt about it. But the intelligent behavior we observe is a side effect, I would argue.

The intelligence has been in the language all along, and LLM research deserves credit for unearthing it: Early GPT models surprised their creators: they never engineered the observed level of intelligence into the model. They discovered it. And the discoveries went on: with mostly experimental prompt engineering, and now with harness engineering, both of which are clever ways to trick a next-word-prediction machine into exposing behavior that an observer would call intelligent.

The human observer is actually part of the equation here. Have you ever watched your vacuum robot for a while? You will have noticed thoughts like “now it’s cleaning this corner thoroughly so that it does not have to come back later”, even though you know the robot just follows a handful of deterministic rules combined with some random movement. Humans see intelligence where there is none, or very little.

And on the other side: it took humanity until lately to recognize all the common sense and world knowledge that is encoded in how we typically order our words. One could have suspected this earlier (and linguists did, but no one listened) by talking to supersocializers at cocktail parties, who talk a lot as if they know what they’re talking about.

Another example: my wife is trained in, and works in, early childhood development. When she talks about child psychology, I can finish her sentences. An outside observer would think that I am a subject-matter expert; but in reality I’m only someone who has been exposed to a lot of professional early-childhood talk. I am the LLM in this case. Please don’t ask me for real advice on your children: I lack the deeper understanding, even if I sound convincing on that topic.

If you put me into a tailored agent harness that leads my sentence-completion skill in a systematic way, I have no doubt I could write convincing, publishable papers on early childhood development. But I would still lack the deeper understanding.

And this is where we stand with AI now, and I don’t see the current trajectory taking us much further than we already are. In particular, I don’t think that the further development of frontier LLMs will lead to anything that deserves the term superintelligence.

Convergence, and what it means for business

Since the intelligence that these models expose is in the language and not in the models, I believe that with further improvement they will converge and become interchangeable. That has three consequences:

  1. We will have successfully automated the perfect cocktail-party guest.
  2. There is not much money to make for frontier model providers, because they will be selling a commodity product. The numbers already point that way: the industry is spending $660–690 billion a year on compute, while the model vendors themselves take in roughly $30 billion of revenue, and analysts see a gap of trillions between the cost of AI and the value it is actually delivering.
  3. They can’t even claim much IP: while they might own the model weights, the language belongs to all of humanity, refined over many generations, and the language is where the intelligence is.

The windfall isn’t where you think

Given all this, one more consequence follows for anyone with money in the market. AI usage will bring productivity gains to the general economy. I don’t doubt that. But the frontier labs won’t be able to skim much of it, because they will be selling a commodity. And given the efficiency of some open weight models on consumer hardware, the companies that run datacenters and the chipmakers won’t receive much of this windfall either. They offer just another commodity, and the assets they build today at great cost might well be the greatest examples of overcapacity in the not so far future.

So investments made in these types of companies will most likely be lost at some point. I don’t give investment advice, but I think this is inevitable. The question is only when.

When we reach that point, I don’t want to see European investors being overly exposed to it. So now would be a good time to slowly pull money out of that sector and move it into the type of companies that are able to actually profit from the productivity gains:

  • Regulated industries that can automate bureaucracy
  • The public sector (same reason)
  • Software companies that can improve their product quality with AI and teach their products new tricks
  • Retail — and everyone else who can automate market research
  • Consulting, when it is based on the input of analysts scanning tons of documents
  • this list goes on…

Some of these have been beaten down a lot lately with the argument that they will be replaced by AI. I argue it is the other way around: these are the companies that profit most from AI. They might have to adjust their pricing scheme — but that is justified, because they can now do more with less.

Writing these industries off because of AI is like writing off the airline industry when a more efficient aircraft turbine has been invented. It makes no sense at all.

System 1, not System 2

Obviously, language is not all of human intelligence. Otherwise we would already have hit the singularity a while ago, because an LLM’s command of language is certainly greater than that of most humans.

To me it makes sense to call LLM intelligence “System 1 intelligence,” in the Daniel Kahneman sense: fast, automatic, effortless. I realize that TypeSafe AI has claimed that label for its classification model Jev, released in September as “the first System One model.” I argue that is wrong: Jev is not the first System 1 model. All LLMs are System 1 models: their base behavior is exactly that fast, intuitive pattern-matching.

And as far as I can tell, we have not even started to tackle System 2, the slow, deliberate kind. Unless you want to call agent harnesses System 2, but that is way too narrow in my opinion, and the harnesses work deterministically, so they are not even traditional AI. Reasoning Models use a kind of built in selfreflecting loop to check and improve their output, which goes in the direction of System 2, but if it works well, it is really not much more then a built in harness, and not even a very efficient one.

To make it short: actual superintelligence is still far, far away.

Artificial is not Fake

Now there is another aspect to the rebrand. The September executive order wants “artificial” gone because it sounds like fake. Well, maybe, but that does describe what this is. And “artificial” doesn’t mean “not useful.”

Artificial vanilla flavour is quite useful for example, even if it is not the real thing. There is really no reason to call it super-vanilla instead.

The same way: AI as it presents itself now is already so useful that it will take us decades to utilize its full potential. And it is not a terrible thing that it is different from human intelligence. Not at least, because there already is plenty of human intelligence on the planet, and we utilize only a tiny fraction of it. Most of that human intelligence costs even less to deploy than AI agents with frontier models, by now.

So: let’s use LLMs for what they are already good for and stop expecting wonders from them only because they have surprised us once, by presenting us the richness of our very own language. And let’s free AI researchers from chasing the next bigger model, and let them explore new paths that might lead us to other pleasant surprises. Or, given another few decades, more real innovation, beyond just scaling things up.

Sources