Your brain solves crosswords the way future AI will think

You don't solve puzzles in a straight line

Think about how you actually work through a crossword. You might start with 3-across because you know the answer. That gives you the second letter of 1-down and the fourth letter of 5-down. Now you have fragments. Pieces. And from those pieces, your brain starts filling in possibilities in every direction at once.

It's the same with a word search. Nobody scans letter by letter, left to right, top to bottom. You let your eyes wander, and your brain spots patterns — a familiar sequence of letters, a word forming diagonally, a cluster that doesn't look random. You process the whole grid at once, not in sequence.

The brain as a pattern completion machine

What you're doing when you solve puzzles is something neuroscientists have studied for decades. The brain doesn't wait for all the information to arrive before making conclusions. It takes whatever fragments are available and fills in the rest based on context, memory, and probability.

You do this all day long without noticing. You read a sentence with a missing word and your brain fills it in before you even realize there was a gap. You walk into a dark room and "see" furniture that you can barely perceive because your brain fills in the shapes from memory. You hear half a word in a noisy bar and understand the whole sentence.

This is the brain's default mode: take partial information, fill in the blanks, from any direction, all at the same time.

What Ilya Sutskever noticed

Ilya Sutskever is one of the most important people in AI. He co-founded OpenAI, left in 2024, and started a new company called Safe Superintelligence Inc. (SSI) focused on building the next generation of artificial intelligence. Recently, he described something that sounds a lot like what happens when you solve a crossword.

In his words (paraphrased): the cerebral cortex works as a pattern prediction machine. If you give it only part of the information, any region of the cortex can figure out what's missing. If you lock some variables to a specific state — treating them as reality at that moment — the brain can calculate, predict, and fill in the rest of the picture in any direction. Not just left to right. Not just forward. In every direction at once. He calls this omnidirectional inference.

The crossword as a thinking model

Let's go back to that crossword. When you fill in 3-across and get a letter in 5-down, what happened? You locked a variable (the letter) to a specific state (the letter you wrote). Now 5-down has a constraint it didn't have before. Your brain uses that constraint to narrow down possibilities. Maybe the word in 5-down starts with "M" and has 6 letters. That alone is enough for your brain to start generating candidates.

This is exactly what Ilya is describing. You "clamp" some known values — the letters you've already filled in — and the brain infers the rest. It doesn't go clue by clue in order. It works from whatever information is available, in whatever direction helps.

A word search works the same way but from the other side. Instead of having clues and looking for letters, you have letters and look for words. Your brain scans the grid without a fixed path, recognizing word fragments in all eight directions. You spot "CR" and check if "CRUZADINHA" continues diagonally. If it doesn't, you move on instantly. The brain evaluates and discards possibilities at a speed no sequential system can match.

Why current AI doesn't work like this

Current AI models like ChatGPT and Claude work in one direction. They receive a sequence of words and predict the next one. Always forward, always one token at a time. It's like trying to solve a crossword by reading every clue in order, from 1-across to the last down, filling in each answer before looking at the next clue. Nobody solves crosswords like that because it's painfully slow and ignores all the information you could get from the intersections.

Your brain takes a different approach. When you enter a room, you don't process the scene pixel by pixel. You take in the whole space at once. You notice the table, the chair, the window, and you fill in details you can barely see based on what you expect to be there. It's a crossword where you already have half the letters and your brain fills in the other half from experience.

Puzzles as a preview of future AI

What Ilya is proposing is that the next generation of AI should work more like you solving a crossword and less like a typewriter spitting out one letter at a time. A system that can "lock" some known facts, then fill in everything else from any direction simultaneously.

Imagine an AI that works like your brain on a word search: it takes in the whole grid (the problem), spots patterns from multiple angles, and arrives at answers without plodding through step by step. That's omnidirectional inference.

You're already doing it

The interesting part is that you don't need to understand neuroscience or AI architecture to know what Ilya is talking about. You already experience omnidirectional inference every time you pick up a puzzle. You lock in what you know, your brain radiates outward from those anchor points, and you fill in the rest without following a fixed sequence.

Next time you solve a crossword or scan a word search, notice how your brain works. It doesn't go in order. It doesn't follow a script. It jumps, guesses, confirms, backtracks, and fills in gaps from every direction. According to one of the most important AI researchers alive, that messy, nonlinear process is what the future of artificial intelligence looks like.