AI doesn’t think. AI computes.

“AI thinking” is a semantic trap. Machines don’t reason in the human sense; they emulate it with statistics and layered mathematics. Understand that one distinction and the rest of the AI jargon, from LLM to superintelligence, stops being mystical and starts being useful.

The trap is built into the name

Artificial intelligence can be perceived, though not defined, as machines displaying intelligence similar to ours. Steve Wozniak drew the line that matters here: there’s Actual Intelligence,¹ and there are its symptoms. An AI chat produces the symptoms flawlessly. It talks, it sounds sharp, it holds an argument. Underneath, it’s guessing every next word from what billions of similar conversations taught it. The performance is intelligence-shaped. The mechanism is prediction.

Even the technical name carries the trap.

• In “Large Language Model”, “language” is a term from computational linguistics, not the humanities: a statistical sequence of tokens, with no claim to comprehension, reasoning, or anything the older ideas of language would promise.
• Same goes for “model”, which is a representation of a thing by definition, never the thing itself. That’s why a climate model isn’t the climate.
• And the third word in “LLM”, “large”, just describes the parameter count.

This is why the industry classifies AI by how closely it imitates human intelligence, not by how much of it a system actually has.

The scale runs from narrow AI (ANI), a specialist that does one task and often does it better than we do, through general AI (AGI), a machine that could match humans across the board, up to superintelligence (ASI), self-improving and beyond anything we can picture.²

The part worth remembering: everything running today, every chatbot, every search algorithm, every drone, is the narrow kind.

AGI hasn’t been built, because the human brain is staggeringly complex. ASI is probably centuries out (you may disagree here 😉), though the ethical debate about how far to push without losing control has already started.

What’s actually under the hood

If not thought, then what?

At the bottom sit algorithms: mathematical formulas turning inputs into predictions, classifications or decisions. Machine learning chains them into systems that find patterns in data on their own. The simplest way to hunt a pattern is to break information into elementary logical paths, a sequence of questions where each answer narrows the field until the system lands on a category or a score.

That works until the data stops being unambiguous, which is most of the time, so fuzzy logic adds fractional values between true and false and the messier cases become computable. And when relationships get too tangled for any ready-made rule, neural networks take over, processing information in layers where each level’s analysis feeds the next. That cascade is all “deep learning” means.

How the ladder learns depends on what it’s given.

Hand it examples pre-labeled by humans and it builds rules from them; that’s supervised learning, and it’s how a Support Vector Machine draws the cleanest possible boundary between categories. Hand it raw, unlabeled data and it finds the natural clusters itself. Hand it nothing at all and it learns the way game theory suggests: by acting, failing, scoring, and adjusting until the wins outweigh the losses over the long run.

Some problems break even this.

When the space of possible answers grows to billions of combinations, no analysis and no training finds the ideal one, so AI borrows from biology. Evolutionary methods breed a population of random, imperfect designs, test each for fitness, kill the weak, cross the strong, and let small mutations keep the process out of dead ends; enough generations surface the best result. Swarm methods drop the central search entirely and mimic bird flocks and ant colonies: hundreds of simple, independent particles roam the solution space, each remembering its own best find and signaling the group, until the whole flock converges on the richest region.

Counting, served convincingly

Follow the ladder from a yes/no question to a flock of particles and one thing never appears on it: understanding.

Every rung is mathematics scaled up. That’s the whole of it: AI doesn’t think, AI computes.

Not a revolution of consciousness, but a generative illusion of reality, served convincingly. Which, for anyone who wants to put it to work rather than worship it, is the more useful thing to know.