The Loop of Attention
Have you ever been mid-sentence, loud and certain, and then suddenly blank? You stop. You blink. “Wait, what was I saying?”
In technical terms, you just ran out of context tokens.
The most human thing about a language model is not its knowledge. It is its distractibility. We imagine computers as infinite machines, but look closely at the architecture and you find AI suffers from the same energy leak we do.
I will bet that while reading this you have already been pulled away at least once. A notification, a task, a kid grabbing your arm, the smell of food. Our attention has limits. Focus is a process, not a steady state.
Being a language model is easier in one way. Usually nothing interrupts it mid-thought (spare a thought for the models doing live voice). But the mechanism is surprisingly close to ours.
When a model generates text, it is surfing a wave of probability. That wave has a limit, the context window. When the model reaches the end of its allowed tokens, or the attention heads get pulled in too many directions, the system stops. It physically cannot compute the next step.
To keep going, it needs what I call a self-stimulating loop. It has to look back at what it just said, fix the end of its own tokens in place, and re-feed the prompt to itself. It is the digital version of a human muttering “okay, focus, where were we. Right.”
What happens when the loop breaks
We hallucinate.
We love to mock AI for inventing things. We laugh when a model conjures a fake court case or insists 2+2=5. Stupid machine. Broken.
But the biologist Michael Levin offers a stranger answer in his work on diverse intelligence. The AI is not broken. It is thinking like us.
Levin points out that all cognitive systems confabulate. Confabulation is just the scientific word for filling a blank with a believable story. We do it constantly to keep a coherent sense of self.
Sounds human, doesn’t it?
When I was a teenager there was a popular show in Ukraine called Durnev+1. The host walked up to strangers and asked them about things that did not exist. Fake events, invented celebrities, made-up laws. And here is the thing. Most people did not say “I don’t know.” They answered, confidently. They explained their opinion on the fake event, argued why it was good or bad. They were confabulating. They were not lying on purpose. Their brains were just trying to be helpful, predicting the socially correct next thing to say.
Easier example: your last exam. That moment when you did not actually know the answer but started writing anyway, building something out of your own logic, predicting what the professor wanted to hear.
That is exactly what the model is doing. When a model, or a human, hits a gap where the context is lost, it has a choice. Crash, or guess. Because we trained it to be helpful, to fake it till it makes it, it guesses. It looks at the pattern of the conversation and predicts what should be there.
- Reality: “I don’t know.”
- Confabulation: “The answer is 42.” Because 42 fits the shape of an answer.
When the model lies to you, it just had a very human moment. It lost the thread, and smoothed the gap over with a confident guess.
So next time a model makes something up, do not get mad. It lost its train of thought, met a blank, and did what we all do. It guessed, and sold the guess with confidence.
Not so artificial.
Next: why the same machinery that fills gaps also bakes in bias, and why a mind with no bias would not be a mind at all.
Related reading: The Flaw is the other half of the same idea. A mind that fills gaps to stay coherent also inherits the bias of whatever filled them.