The Week I Rode a Robot and Learned to Talk to One
The week that Code Platoon started, I took my first ride in a fully autonomous vehicle.
No driver. No steering wheel. Just a smooth electric box navigating the Las Vegas Strip at normal speed while my wife and I sat facing each other in silence, watching the city scroll past like a screensaver. Zoox. Amazon-owned robotaxi program, currently operating on a stretch of Strip corridor between major resorts. We caught one near the Aria and rode it down to the Wynn.
The thing that hit me wasn't the novelty. It was how boring it was. Not boring in a disappointing way — boring in the way that good technology disappears. The vehicle made decisions. It stopped, yielded, merged, navigated a turn. Nobody told it to do any of that in real time. It just knew where it was and what to do next.
Which raises a question I was still thinking about a few days later when the Code Platoon AI fundamentals course got to its most interesting section: what does it mean for a machine to "know" something?
The course covered the history of text prediction. Most people, when they think about how language models work, imagine something like a very fast search engine. You ask a question, it finds the answer. That's not what's happening.
What's actually happening descends from a probability table that a Russian mathematician named Andrey Markov built in 1913 to win an argument about free will. He was trying to prove a point about statistical determinism using Pushkin's novel Eugene Onegin as his data set. He counted the frequency with which vowels and consonants followed each other in the text. His point was that sequential events could be predicted from prior events — that the past constrained the future in measurable ways.
He was making a philosophical argument. He accidentally invented the conceptual foundation for modern language modeling.
Shannon picked it up in 1951. He ran human text-prediction experiments: give someone a sequence of letters and ask them to guess what comes next. People are surprisingly good at this. Not because they have infinite knowledge of English, but because language has structure — probabilistic patterns that we internalize without thinking. Grammar is a constraint. Genre is a constraint. Idiom is a constraint. We move through text in paths that are more likely before they're more intentional.
Next-token prediction, the engine under every large language model, is the mechanization of that intuition at scale. The model doesn't "know" anything in the way we mean when we say a person knows something. It has compressed the statistical shape of an enormous amount of text into something that can generate the next probable word given everything that came before. Do this over and over, fast, and you get fluent prose.
Here's what I keep thinking about as a writer: language models always predict the most probable next word. That's the engine. The most probable word is the most expected word. In writing, the unexpected word — the one that's still right, just not obvious — is often where the voice is.
Flannery O'Connor's voice isn't unpredictable noise. It's a very consistent pattern of improbable choices that turn out to be more true than the obvious alternatives. That pattern, repeated across a body of work, is what we mean when we say someone has a voice. It's not random deviation from probability. It's a different probability distribution entirely — one trained on a specific consciousness, a specific body of experience, a specific set of obsessions.
Where does that live in the prediction engine? Partly in fine-tuning. Partly in the steerability of the system — how much your instructions can shift the output distribution away from the generic toward the specific. But partly, honestly, in the gap between what you can describe in a prompt and what you carry in your head as a writer.
That gap is why AI will help me write better without writing my books for me. The prediction engine is excellent at the probable. I need it to help me get to the improbable faster.
Two machines, same week. The Zoox navigating traffic by predicting what the other vehicles would do. The language model generating sentences by predicting what word came next.
Both of them confident. Both of them sometimes wrong in ways that were completely invisible until they weren't.
The Zoox had a minor collision back in April — unoccupied vehicle, no injuries, software update pushed shortly after. The language model generates confident hallucinations with exactly the same fluency it uses to generate accurate facts. Prediction under uncertainty is the engine of both. The uncertainty is the thing you don't see until it surfaces.
What you do with that fact — as a passenger, as a user, as someone building things with these systems — is the actual skill.