Skip to content
← cd ../blog

2026-11-087 min read

Elsewhere: a world learned from within

The life you remember is one sequence of events. You can explain how you arrived here, sometimes confidently, but you cannot inspect the life in which you answered differently. That missing comparison interests me. How much of what we call understanding is a rule we have learned, and how much is a conclusion that happens to fit the path we took?

I started Elsewhere with the personal version of that question: a choice, another possible life, another version of you. As I built it, the question became more precise. I wanted a virtual world with an observer inside it, learning from what happens, and a way to compare that observer with a learner trained on several alternative histories.

The life scenarios are the first case study. The larger experiment is about learning under limited experience: what a machine can infer from its own record, what changes when its training includes alternatives, and whether that additional perspective helps it act in the world it currently occupies.

What many-worlds contributes

In Everett's formulation of quantum mechanics, the observer belongs inside the physical description. A measurement correlates the system, apparatus, and observer's record through the same quantum evolution used for other interactions; there is no separate collapse rule selecting one term of the combined state. That treatment of an observer as part of the world is the idea I want to work with. Everett's 1957 paper is the starting point.

Decoherence helps explain why alternatives can behave like separate, approximately classical histories. Interactions spread information into the environment, suppressing the interference accessible to a local observer. It does not, by itself, settle which interpretation of quantum mechanics is correct. Schlosshauer's review develops that distinction in detail.

An everyday decision tree is a simplification of this picture. Choosing a new job does not give us a known number of quantum branches, and drawing four possible futures does not make them equally probable. In quantum mechanics, outcome weights come from amplitudes through the Born rule. Counting the options on a screen cannot replace that calculation.

My original curiosity also reached toward consciousness and the idea of continuing in another reality. But the existence of another history would not establish that awareness transfers into it, or that experience must continue indefinitely. The old-age question exposes how much extra biology and philosophy that claim needs. Mallah's critique of quantum immortality addresses the inference directly. Elsewhere leaves it as a question that helped start the project, rather than making survival a rule of the world.

A life gives the experiment something concrete

You enter a name and a starting age, then encounter situations involving other people, commitments, opportunities, and limits. Your choices affect what follows. Earlier forks remain available, and you can inspect another continuation without erasing the one you already explored.

The scenarios are assembled locally from authored situations and rules. People, settings, actions, and subsequent moments vary with the path. This creates many combinations, although the underlying vocabulary is finite: open-ended branching does not mean an infinite supply of original writing. The computer generates the parts you inspect instead of holding every possible life in memory.

There is also a deliberate distinction between the person navigating the website and the learner being studied. As a visitor, you can stand outside the story and compare lives. A learner following one history receives only its permitted observations. It does not acquire the memories of an alternative version merely because those memories exist in the simulator.

The four indicators—connection, purpose, freedom, and vitality—make changes visible and measurable. They are authored game variables. They let me test whether a model learns the simulator's rules; they are not a validated definition of a good life. A starting age supplies a context for the scenes, and a step counts a decision, not a year of biological aging.

The hypothesis I can actually test

The first hypothesis is this:

With the same number of observed outcomes, a learner trained across alternative histories may predict held-out situations within the same simulated world more accurately than one trained along a single continuous history.

The current comparison gives each learner 48 observed transitions. One follows a continuous life. An external trainer revisits forks and collects examples across alternatives for the other. Both use the same simple learning rule: estimate an action's effect from the changes observed after taking it, with broader averages when an action is unfamiliar. This model learns associations between action identifiers and indicator changes; it does not yet understand the prose.

Neither learner gets the author's hidden outcome rules. After training, both are frozen and evaluated on the same 24 held-out scenes, generated separately from their training data. A baseline that predicts no change provides a useful minimum comparison. Prediction error measures how far the forecasts miss; a separate decision score measures how much expected value is lost by choosing an action using those forecasts.

When chance is enabled, each training choice contributes only the response the learner observes. Evaluation accounts for both possible responses and their probabilities; the learner does not receive both outcomes for free.

That decision score concerns an immediate, declared game objective. It cannot tell you which life to live.

Equal observation counts are only the beginning of a fair comparison. A long trajectory can push an indicator to its limit, where later changes become clipped. A learner visiting many shallow branches may encounter a different distribution of states. If it does better, I need to distinguish useful coverage from that boundary effect. Matching depth and comparing repeated short lives are important next controls.

The current held-out scenes also draw from familiar authored action families. Better performance would show learning within this environment; it would not establish understanding of entirely new human situations. The hypothesis has value only if the experiment can reveal those limits, including cases where alternative-history training loses.

Where the quantum computer enters

Elsewhere now connects a small quantum experiment to the story's consequences. A two-qubit circuit splits a signal into coherent paths, changes their relative phase, optionally leaves a path record in a second qubit, and recombines the signal before measurement. Changing those settings changes its outcome probabilities.

I use Qiskit to calculate the ideal circuit and have also run eight circuits, with 1,024 shots each, on IBM's ibm_marrakesh processor. The completed job is dboiqpgqtr0c738ctov0. Those recorded counts are retained with the experiment so their origin can be inspected.

In the connected story mode, you choose an action and the environment adds one of two responses: support or friction. For an action centered on connection, that could mean an opening for conversation or a misunderstanding that asks for more patience. The selected response changes the indicators and the context carried into the next scene. Both responses remain inspectable when they have nonzero weight.

The mapping from detector output to those consequences is a design choice I made. The hardware measures a circuit; it does not measure the likelihood that someone will understand you.

Visitors use saved ideal probabilities or saved hardware frequencies, with draws performed locally in the browser. Clicking a choice does not submit a new quantum job or replay an individual recorded shot. This makes the experiment available to everyone without spending scarce hardware time on every page interaction.

It also makes the classical comparison essential. For this small circuit, an ordinary calculation reproduces the ideal probabilities. A classical sampler given the same probabilities and random stream reproduces the same story outcomes. The present implementation demonstrates a connection from quantum measurements to a fictional environment, but it has not demonstrated quantum advantage. The learners themselves are classical. These measurements also cannot distinguish many-worlds from interpretations that make the same experimental predictions.

The Quantum reference view in Elsewhere lets you inspect the circuit settings and recorded measurements. The learning comparison shows the observation budgets and scores alongside its forecasts.

What I want to learn next

The next useful work is to make the environment harder in ways that expose the learner's assumptions: delayed effects, rules that depend on context, and whole families of situations withheld from training. I want to see whether knowledge collected across alternatives remains useful when the learner returns to one ongoing life, where consequences accumulate and resetting is no longer free.

There is a focused hardware question, too: can a model calibrated on one set of quantum measurements predict a fresh set better than the ideal circuit model? That needs independently collected test measurements and a comparison specified before the results arrive. Reusing the same saved counts cannot answer it. A successful result would concern calibration of this device, not a general advantage for quantum learning.

The ambition is to bring those strands together: a world with explicit rules, an observer with limited access, and learning that can be examined rather than merely asserted. We can compare what the observer believes with what the environment actually does, then ask why the gap exists.

I want someone to enter Elsewhere, follow a life, and encounter a consequence they did not expect. Then I want them to be able to return to the fork and see what changed: the action, the environment's response, the experience available to the learner, or the learner's understanding of it. That is the experiment I am building.