In March 2016, in the second game of a five-game match in Seoul, a Go program made a move that no human would have played. Move 37. The commentators went quiet, then assumed it was a bug. Lee Sedol, one of the best players alive, got up and left the room, needing time to make sense of it. It was not a bug. It was AlphaGo, and it was about to win a game that the experts had said was a decade away.
No human taught AlphaGo that move. There was no database entry that said "in this position, play here." It found Move 37 the same way it found everything else that startled the room: by running a loop, millions of times, until it understood the game in a way its makers could not have written down.
I bring this up because in the issue about the reinforcement loop I claimed the unit of an AI-native organization is a loop: Intent, Action, Signal, Learning, Adaptation. A reader could have filed that under "nice metaphor, Chuck." So let me close that escape hatch. At the model level, the loop is not a metaphor I borrowed from somewhere comforting. It is an actual mechanism that produced the AI you are now trying to deploy. From clever analogy to actual engine.
Watch (ok, read) how it maps. The intent is the objective: win the game. The action is a move. The signal is the result (did that line of play lead toward winning or losing). The learning is the part where the system updates its internal sense of which moves are good (the credit assignment that says "that early move mattered more than it looked"). The adaptation is an updated model that takes the next action with a slightly better sense of the board, and around it goes. Intent, action, signal, learning, adaptation. That is reinforcement learning, and reinforcement learning is the reinforcement loop wearing a lab coat (the name was not an accident on my part).
And it did not stop with games. The next version, AlphaGo Zero, threw out the human examples entirely and learned the whole game from nothing but self-play, the loop closing on itself with no training wheels. Then the same idea walked straight into the language models you use every day. (I am making the presumption that anyone reading this newsletter is already an AI user.) A major part of what turned a rambling text-predictor into something genuinely useful was not simply more facts poured in. It was a loop: the model acts (drafts an answer), gets a signal (a human, or another model, rates it), learns, and adapts. We call it RLHF, reinforcement learning from human feedback, but you can just call it the loop. One reason the reasoning models pulled ahead this past year was a cleaner signal, training against problems where the answer can actually be checked, so the feedback is true rather than flattering. (There is a reason Mercor, a startup specializing in the human feedback layer, became one of the fastest companies ever to reach a billion dollars in annualized revenue.) Better signal, faster loop, better model.
Now the turn, and it is the reason this sequence exists. In the model, the loop is literal. In the organization, the machinery changes, but the discipline does not. The same five moves are available to you one level up, at the scale of a whole organization. You are no longer updating weights and gradients. You are updating people and processes (slower, messier, and with considerably more meetings).
That hands you an uncomfortably clear diagnostic, because the failure modes transfer too. A model trained on a bad reward gets confidently good at the wrong thing, which is exactly what any organization does when it optimizes a vanity metric (or the metric a senior leader invented because something about this AI thing had to turn green). Garbage signal, garbage learning, in silicon and in boardrooms alike. The labs obsess over reward design because they learned the hard way that a model becomes whatever you actually measure, not whatever you say you want. (Quick show of hands: whose company has ever hit its number and lost the plot? Same bug.) The discipline that improves a frontier model offers a fairly good blueprint for improving your operating model, and almost nobody is treating it that way.
So the lesson from the machines is not "buy the machines." It is subtler and more useful. The advantage is rarely the raw model, which connects straight back to my point about systems beating models. The advantage is how fast and how honestly you close the loop. AlphaGo did not win because anyone handed it Move 37. It won because it ran the loop more times, on truer signal, than any human could in a lifetime. Your organization will not win because it bought the best model either. It will win, or not, on the same terms.
That is the loop that taught the model, and it is the bridge to the follow-on posts, where it stops being about software and starts being about people, organizations, and markets. They all run the same five steps and break them in their own special ways.
For now, one question, and it is the one the labs ask first about any system they are trying to make smarter. What are you actually rewarding? Because that, far more than which model you licensed, is what your organization is quietly training itself to become.