The Alien Intelligence of AI

It is a comfortable belief: AI is just a mirror. Critics often dismiss large AI systems, especially large language models, as statistical parrots that reflect our words, images, and ideas without any deeper agency. They argue that such systems can't come up with new ideas.

I think that is wrong. The question is not whether AI has consciousness or human-style understanding. The question is whether it can see associations and links that humans often miss. Human civilisation also advances through copying and recombination. We build on what came before. AI does something related, though in a different manner. It notices connections across vast bodies of data and turns those relationships into outputs that can feel surprising to us.

In this essay I will look at several examples — a Go match, a mathematical proof, a rocket engine, a strange portrait, and a drug candidate — where I think AI has done something novel and interesting.

Move 37: a strategy that broke human intuition

In March 2016, during the second game of the historic Go match between DeepMind's AlphaGo and world champion Lee Sedol, AlphaGo played Move 37, a shoulder hit on the fifth line. To human Go masters, the move looked wrong. Traditional theory tells players to avoid opening plays on that line, so many commentators first treated it as a mistake.

It was not a mistake. I see Move 37 as a sign that AlphaGo could recognise a pattern of play that human intuition had not seen. It was not simply memorising grandmaster games. It had found a structure in the game that was hidden in the space of possible moves.

Verified reasoning: a preview of the collaboration ahead

Mathematics is where the shape of human-AI collaboration is clearest. In 2024, DeepMind's AlphaProof and AlphaGeometry 2 reached a silver-medal standard at the International Mathematical Olympiad. DeepMind's FunSearch went further, producing new constructions for the decades-old cap set problem that mathematicians then checked and built on. What makes this special is that the results were verified, not just admired. In AlphaProof's case, proof assistants such as Lean let a proof be checked line by line, so a claim either holds or it does not. FunSearch's cap set constructions were verified differently: their properties could be checked computationally, and mathematicians could then inspect and build on the result. That distinction matters, because it answers the standard worry about AI: that a fluent answer can quietly be wrong.

This is a sign of where the gap is widening. No mathematician can master all domains — the field is too large for any one person, and the same is true of law, medicine, and materials science. AI does not share that limit, and because it is multimodal and multidisciplinary, it can move between fields in a single session. I think that gap between what human specialists can do and what AI systems can achieve will only widen.

Organic engineering: building rockets from physics

I also see this pattern outside formal domains, in engineering, where the verification is physical rather than logical: a rocket engine either fires or it doesn't. Traditional computer-aided design depends on human draftsmanship, geometric symmetry, and familiar manufacturing rules. LEAP 71 is changing that.

Using their Noyron Large Computational Engineering Model, LEAP 71 generated a fully functional copper-alloy liquid rocket engine from code alone. Instead of starting from a human blueprint, the software works more like a physics-based compiler. It grows hardware from the first principles of fluid dynamics and thermal physics. The result looks strange: skeletal, fluid, and organic structures that humans would rarely design by hand.

Hyperganic, a Munich company co-founded by Lin Kayser — who also co-founded LEAP 71 — has used a similar approach to generate its own 3D-printed rocket engine directly from performance requirements. That makes me think AI is not just accelerating design. It is making connections across physics, manufacturing constraints, and geometry in a way that is less bound by human habit.

Latent space anomaly: the discovery of Loab

In 2022, AI artist and musician Supercomposite uncovered a strange phenomenon in latent space. By asking an image generator to find the opposite of a given concept, she discovered a recurring figure she called Loab.

Loab kept appearing across many prompts and styles: a gaunt woman with a corpse-like complexion, heavy dark eyes, and unnaturally reddened cheeks. I find that striking because the figure was not directly requested and not copied from one known artwork. It emerged from the model's capacity to connect visual features across many associations and pull them into a recurring form.

Molecular discovery: speeding up pharmacology

I also see this non-human intelligence in the medical sciences. AI tools can predict complex molecular structures and run massive virtual screens. One of the clearest cases is baricitinib, an arthritis drug that AI-driven analysis flagged as a candidate for COVID-19 in 2020, ahead of the clinical trials that later confirmed it. That does not mean the human scientist vanishes. It means the process becomes faster, broader, and sometimes less dependent on traditional trial and error.

Embracing alien intelligence

Taken together, the five examples point to the same pattern: in each case, the result was surprising, not just because it worked, but because it lay outside what specialists in the field would usually think to try. Move 37 was a shape professional players had trained themselves to avoid. AlphaProof matched a silver-medal standard under contest conditions, and FunSearch went further, finding a construction for the cap set problem that Terence Tao and other mathematicians then checked and built on — genuinely new, not just fast. LEAP 71's engine geometry looks nothing like a human draftsperson would produce, yet it fires. Loab kept surfacing from a request for the opposite of an image, not the image itself. Baricitinib had been sitting in the pharmacopoeia as an arthritis drug for years before an AI screen connected it to a virus. In each case, the result sat outside the set of answers a human would have reached for first.

I think we miss the point when we judge AI only by human standards. Historian Yuval Noah Harari has used the term "alien intelligence" to describe AI precisely because it doesn't share our biology or our history — though he means the phrase mostly as a warning about a force we may not be able to control. I want to borrow his term but not his conclusion. I think the same alien quality that worries him is also what produced the Go move, the proof, the engine, the portrait, and the drug candidate above.

That matters because AI does not share our biological history, our emotional biases, or our cultural blind spots. It can combine patterns, constraints, and possibilities in ways that do not come naturally to human specialists shaped by training and habit. I see that as a reason to move from fear of imitation toward collaboration and discovery.