Einstein didn’t need
a data centre.

The argument that never ends

In March 2023, an open letter asked every AI lab to pause training systems more powerful than GPT-4 for six months. It gathered more than 30,000 signatures, Musk’s among them. No lab paused.

Musk’s own company is the clearest example. According to corporate records reported by the Wall Street Journal, X.AI was incorporated in Nevada on 9 March 2023 — about two weeks before the pause letter was published.

The money tells the same story. The largest US cloud and AI companies plan to spend US$660–690 billion on infrastructure in 2026, nearly double 2025’s roughly US$380 billion. Everyone agrees the risk is serious. Nobody is slowing down. Musk himself now talks about AGI around 2026 or 2027, and superintelligence around 2030.

The question I want answered first

In 1905, a 26-year-old patent clerk in Bern published a three-page paper. It had no laboratory, no data set, no experiment of its own. It ended with the idea we now write as E = mc².

That is exactly the kind of result you would assume needs a supercomputer: a new relationship between mass and energy that nobody had written down before. It came out of a thought experiment, in a brain running on roughly 20 watts — about the power of a dim light bulb.

My doubt is simple. Until someone can explain how one person did that with his own brain, “superintelligence” is a forecast about something we cannot yet define. We are promising to exceed a capability we do not understand.

Intelligent is not intelligence

The two words get used as if they meant the same thing. They do not.

Intelligent is an adjective we give to output. An answer looks smart, the code runs, the essay reads well. By that measure today’s models are extremely intelligent, and I use them all day.

Intelligence is the thing that produced the question in the first place. Einstein looked at the same physics everyone else had and asked what “at the same time” really means. Relativity came out of that question, and E = mc² followed from relativity a few months later. Models learn from what has already been written. E = mc² was not written anywhere before 1905.

A system can be intelligent in every answer it gives and still not have the intelligence to ask the question that changes the field. I made a related point in The Toddler Still Beats the Robot. — a child learns from one example what a machine needs millions for.

To be fair: the machines have started finding things

I should be honest about what the other side can point to, because it is real.

AlphaFold’s protein-structure work won a share of the 2024 Nobel Prize in Chemistry. In 2025, Google DeepMind’s AlphaEvolve found a way to multiply 4×4 complex matrices with 48 multiplications, beating a result that had stood since 1969. The same year, AI systems reached gold-medal level at the International Mathematical Olympiad.

Those are discoveries, not autocomplete. What they have in common is that each happened inside a question people set, with a clear way to score the answer. The fold, the multiplication count and the proof were all targets someone else defined. Einstein’s contribution was the target itself.

This gap may close. I am stating a doubt, not a certainty — and I would rather be wrong in the direction that keeps me paying attention.

We haven’t mapped what we’re trying to surpass

This is the part that convinces me the road is long.

In October 2024, scientists published the first complete wiring diagram of an adult fruit fly brain: 139,255 neurons and about 50 million synapses. It took roughly 33 person-years of proofreading. In 2025, a separate team mapped one cubic millimetre of mouse brain and reconstructed more than 500 million synapses. One of the researchers thinks a complete mouse brain is 10 to 15 years away.

The human brain has around 86 billion neurons. We have fully mapped a fly. We are building “super” versions of an organ we have only charted at the scale of an insect. Exploring the human brain has a long way to go, and I think that is where the next surprise will come from — not from another data centre.

What I would do with this

If you run a business, do not plan around superintelligence arriving on a date someone announced on a red carpet. Plan around what the tools do today: intelligent output, at a price that keeps falling. That alone is enough to change most companies.

Then keep the question in human hands. Hire and train people for what to ask, not only for how to answer, because the answering is getting cheap. And judge any model on your own work rather than someone else’s demo, as I argued in The Model Reviewer Has Never Deployed a Model.

The bottom line

The safety debate will keep running, and so will the spending. Neither tells you how close superintelligence is. The real measure is simpler: can anyone explain how one clerk, on 20 watts, produced E = mc² from a thought experiment? Until then, our machines are very intelligent. Whether they have intelligence is still an open question — and so is most of the brain we keep comparing them to.

If you see it differently, I am always up for a conversation about it.