What if the rewind button is the whole thing?
Everywhere else on this site the machine uses one move to beat things: freeze the world, imagine the futures, throw away the bad ones, rewind, commit. It works on ARC puzzles. It works on a walking body. It works on a Nintendo. This page asks the obvious next question out loud — what happens when you point that same move at a machine whose job is to know one person, and which never leaves that person's computer.
Everything below is architecture and argument. Nothing on this page has been benchmarked, and the only live number here is the one the widget measures on your own keyboard while you sit there. That is stated up front so it cannot be mistaken for a result.
The ladder, and the rung nobody reaches
Gregory Bateson laid out levels of learning in the 1960s and cybernetics has been arguing about them since. The short version, with no reverence:
You acquire a response. Something happens, you react, the reaction that worked gets stronger. A thermostat is at this level. So is most of what gets called machine learning.
You acquire the context. You learn which set of responses this whole situation calls for — not one answer, a class of answers. This is where habit, character and “knowing the room” live.
You change the set of sets. The rules you were learning inside of get seen from outside and rewritten. Bateson said this is rare and destabilising in humans — conversion, breakdown, sudden clarity. It is not a bigger version of level II. It is a different move.
Bateson called IV “probably does not occur in any adult living organism on this earth.” The claim on this page is narrow and testable: a machine with a total, addressable, rewindable record of its own history can attempt the level-III move deliberately and repeatedly instead of once by accident — because it can go back and check.
Here is the honest version of the claim, with the hype stripped out. Level III is hard for people because you cannot get outside your own premises while you are standing inside them, and you cannot check what you used to believe because your memory rewrites itself every time you touch it. A machine with a complete, immutable, addressable tape of its own history has neither problem. It can restore a version of itself from before it held a belief, run the same evidence through, and compare. It is not smarter than a person. It just cannot lie to itself about what it used to think — and that, not scale, is the thing standing between III and whatever is above it.
The response-time half, running right now
The easiest piece of this to actually test is the speculation. A game controller rolls futures forward because the console is idle between frames. A personal machine is idle for the entire time you are typing — which, relative to a CPU, is a geological era. Below is that idea implemented literally, in your browser, on your keystrokes. It builds a tiny model of you as you go, holds a handful of futures, and reports honestly how often it had the answer before you finished asking.
What this is and is not · This is a toy, and it is labelled as one. It is a character model, not a mind, and 4 ms of deliberate busy-work standing in for the cost of a real answer. What is not a toy is the shape: the hit rate above is measured on your actual keystrokes, the wasted branches are counted honestly, and every millisecond in “latency removed” is time a slower design would have made you wait. Nothing typed here is sent anywhere. Reload the page and it is gone.
Six pieces, and what each one buys
Save-state everything
The console trick, generalised. Every state of the machine — what it knew, what you had said, what it believed about you — is a snapshot on a tape, not a mutation of one blob of weights. Storage is the cheapest thing you own. Forgetting is a choice, not a limitation.
Speculate while you think
A human types slowly and pauses. That is dead air the machine spends idle. Instead: fork the state, roll the six most likely things you are about to ask, answer them all in the background, and when you commit, rewind into the branch that matched. The wrong branches cost electricity. The right one costs zero latency.
Rewind on being wrong
When it misreads you it does not apologise and carry the bad state forward. It restores the snapshot from before the misread, replays the conversation with the correction applied, and continues from a version of itself that never made the mistake. That is exactly what the speedrun controller does when a jump kills Mario.
Talk about the tape
Because the history is addressable rather than averaged away, it can answer questions about itself with citations to its own frames: what it thought last March, what changed its mind, which of its beliefs about you came from one throwaway sentence. A model that can be asked “why do you think that” and go look is a categorically different object from one that generates a plausible reason.
Never leaves the machine
None of this works as a service. A tape of everything you ever said, held by somebody else, is the worst artefact in computing history. The whole design is only defensible if the substrate is small, integer, deterministic and local — which is the same constraint this company has been building under for other reasons entirely.
Determinism is the requirement
Rewind is worthless if replaying a state does not reproduce it. Every float, every GPU nondeterminism, every hidden clock breaks the tape. This is the one place where the integer-only, fixed-point, same-seed-same-universe rule stops being an aesthetic and starts being the load-bearing wall.
What would kill this idea
The tape gets too big. A console state is kilobytes. A personal machine's state might not compress like that. If a snapshot cannot be taken in microseconds and stored forever, the whole thing degrades into ordinary checkpointing and none of the interesting properties survive.
Speculation stops paying. On a console the branching factor is small — eight buttons. On a person it is the space of everything anybody might say next. If the hit rate falls low enough, you are burning six units of compute to save a fraction of one. The widget above exists precisely to make that trade-off visible rather than assumed.
Determinism does not hold. If replaying a state does not exactly reproduce it, the tape is fiction and every claim on this page collapses. Floating point on a GPU already fails this test. That is not a small engineering detail; it is the reason this is being proposed by people who build in integers.
And the ugly one. A perfect, replayable record of a person is the most dangerous file ever created. It is only tolerable if it physically cannot leave the machine, and “we promise it stays local” is not a design. Either the architecture makes exfiltration impossible or the idea should not be built.
This page is a proposal Dan asked for out loud and the agent agreed with. It is not on the roadmap, nothing has been benchmarked, and no part of the NUMEN runtime currently does this. It is here because the honest thing to do with a good idea is write it down where it can be argued with, not sit on it until there is a demo.
What already exists and is measured: the rollout-and-rewind controller itself, on ARC-AGI-3 puzzle worlds, on physics bodies, and on retail game consoles. Those runs are on film with real numbers. This page is the argument that the same move is worth pointing somewhere much stranger.