JournalEvery action starts somewhere vague. A dry mouth. A page you need on paper. A few seconds later there’s a glass in your hand, or a sheet coming out of a printer.
In between, something vague becomes precise. This piece follows that path, level by level, first in a person, then in a machine. Then it looks at what changes now that AI can start from the vague part too.
Where I started: clarity
My starting point was Wittgenstein. The Tractatus opens with a promise: what can be said at all can be said clearly.
Russell and the logicians of his time wanted exactly that. A language with no ambiguity.
In a way, that language exists now. We use it every day. It’s called code.
But intentions start before words, and before words everything is vague. Code only begins after someone has done the work of being precise.
How it works for a person
Take the simplest thing. You’re thirsty, and you drink a glass of water. It looks like one gesture. It goes through five levels.
- InteractionDry mouth
- IntentionI want water
- PlanThat glass, there
- ActionReach, adjust, drink
- FeedbackThirst fades. The goal ends.
These levels aren’t separate floors. They talk to each other all the time. The body corrects while it moves, and sometimes the intention only takes shape while you’re doing it.
The goal is yours. Shaped by your body, your culture, other people. But yours.
How it works for a machine
Now the machine version. You press Print, and a few seconds later a sheet of paper comes out of a printer.
The print dialog does something important before anything else: it makes you be precise. Which pages, how many copies, which printer. The vague part stays with you.
From there, each level translates for the next one. Code turns the document into instructions, the system sends them along, and at the end the printer moves.
The machine starts from a goal it received, and settles every ambiguity before it acts.
The ladder and the circle
Same levels. Two architectures.
Where AI comes in
With classic software, the hardest step is yours: turning a vague want into precise input. That’s what the print dialog is for.
A language model takes over that step. You say what you want in your own words, and the model does the interpretation that used to need a dialog, or a programmer.
It’s the first machine that starts closer to where people start.
The ladder below it stays the same, and the goal is still received. Agents are starting to close the loop in a few places: they try, check, and correct. What’s still missing is real learning during use, and goals of their own.
The levels, one by one
Scroll. The person and the machine go down together, level by level.
| Level | Person | Machine |
|---|---|---|
| 1 | Interaction. Your mouth is dry. Nobody asked you anything: your body did. | Interaction. You press Print. The dialog has already made you precise: which pages, how many copies. |
| 2 | Intention. I want some water. The need becomes a goal. | Interpretation. The app turns the document into instructions the printer understands. |
| 3 | Plan. That glass, on the table. The goal meets what’s around you. | Handoff. The instructions travel through the print queue and the network to the printer. |
| 4 | Action. You reach, grip, drink. Your hand adjusts all the way. | Execution. The printer moves the paper and puts ink exactly where it was told. |
| 5 | Feedback. Thirst fades, and the intention simply ends. Or it doesn’t, and you pour another. | Feedback. “Printed”, or “Paper jam”. It can try again. It can’t decide you needed a different page. |
An open question
So, in this view, AI is a new disambiguation tool: it takes on our ambiguity and translates it into the deterministic level of code.
For now, AI systems don’t generate their own goals. They only work through short chains of tasks, and the ultimate goal still originates from an interaction with us. We’ve also seen that they haven’t yet closed the loop.
But if, as is happening, we’re moving toward these systems being able to close the feedback loop at every level, making them adaptive and therefore more capable of generating genuine goals of their own, are we still talking about tools?
Further reading
- Ludwig Wittgenstein, Tractatus Logico-Philosophicus (1921): gutenberg.org/ebooks/5740. Where this piece started: what can be said, and how clearly.
- Kevin Klement, “Russell’s Logical Atomism”, Stanford Encyclopedia of Philosophy: plato.stanford.edu/entries/logical-atomism. Russell’s idea of a language built from simple, unambiguous parts.
- Edsger W. Dijkstra, “On the foolishness of ‘natural language programming’” (EWD667): cs.utexas.edu. Why machines have always needed us to be precise.
- Elisabeth Pacherie, “Action”, Open Encyclopedia of Cognitive Science (MIT Press, 2025): doi.org/10.21428/e2759450.3036d218. An accessible overview of how intentions turn into actions.
- Emanuel Todorov and Michael I. Jordan, “Optimal feedback control as a theory of motor coordination”, Nature Neuroscience (2002): doi.org/10.1038/nn963. Why the body corrects while it moves.
- Andrej Karpathy, “Software Is Changing (Again)”, YC AI Startup School (2025): youtube.com. Programming in plain language: the model as the new interpreter.
- David Silver and Richard Sutton, “Welcome to the Era of Experience” (2025), overview by TechTalks: bdtechtalks.com. Agents that learn from their own stream of experience.
- Figure AI, “Helix: A Vision-Language-Action Model for Generalist Humanoid Control” (2025): figure.ai/news/helix. A slow system that understands and a fast one that acts.