What is this?

How the anthology is made

Most futures arrive with a crate of convenient miracles, but this project uses a stricter allowance: one new mechanism per sequence, with ordinary law, money, family, ambition and bad timing left in place around it. The point is not to conceal a moral inside a gadget, but to look at how the same capability changes when it becomes a medical treatment, a weapon, a labor practice, an art form or a piece of public infrastructure.

All five stories in this first sequence were written by an autoresearch system driven by GPT-5.6 Sol. The system generated competing premises and scaffolds, selected among them, drafted against long-range plot dependencies, and repeatedly revised the stories after separate reviews of plot, voice, novelty, subtlety, clarity and consistency.

The mechanism here is the afterimage, a memory of a future that did not happen. An overlay system models several possible continuations of a person's life, identifies the neural changes produced by selected experiences, and induces an equivalent pattern in the living brain. The nearest software analogy is a Git repository: simulated futures branch from a shared state, overlay cherry-picks some of their changes back into the present, and the resulting mind has to resolve conflicts without having lived the history that produced the patch. Although the recipient knows that the source was a simulation, this provenance tag does not stop a practiced hand, an old attachment or a fear response from feeling acquired, so the distinction between an authentic memory and a correctly labeled artificial one remains intellectually clear while becoming less useful in every other respect.

From world models to remembered experience

A world model compresses the regularities that matter for prediction and action, allowing an agent to test possible behavior without trying every bad idea in the physical world. Ha and Schmidhuber demonstrated the approach by training a controller inside a generated environment and then transferring its learned policy back to the outside task,1while more recent work is making the underlying representations simpler to train and easier to scale. LeJEPA, for example, proposes a compact joint-embedding objective for learning manipulable representations of a world and its dynamics.2

Other parts of the path have already left the paper stage, with World Labs describing Marble as a multimodal model that can reconstruct, generate and simulate persistent three-dimensional worlds from text, images, video and rough spatial structure,3while AMI Labs is pursuing world models that predict in an abstract representation space, retain memory, and condition their forecasts on possible actions so an agent can plan what to do next.4Afterimage projects outward from current world models, agents and neural interfaces toward one possible descendant, in much the same way that labels such as “GPT-9” or “Claude Gigantamax” name a conjectured capability level rather than a product specification.

In the stories, a personalized neural model supplies the starting state, a world model supplies the physical and causal environment, and language systems supply much of the social texture through which people bargain, lie, cooperate and misunderstand one another. Rather than saving a recording of every perceived detail, the simulation saves changes in the modeled person, after which a neural interface recreates selected associations through timed stimulation and reconsolidation while the recipient sleeps. This extrapolation produces several specific failure modes: the modeled world can be wrong, the social agents can be persuasive without being faithful, the translation into a brain can omit context, and a perfectly accurate memory can still belong to a future that the import itself prevents.

Language, worlds and mechanized psychohistory

The rapid growth of language model competence has opened a disagreement beneath this fictional engineering, with some researchers arguing that sufficiently broad prediction over language may already contain much of the machinery required for general intelligence. Bubeck and colleagues made an early and deliberately expansive version of that case for GPT-4,5whereas the competing view holds that verbal competence still leaves out grounded prediction, durable memory, objectives, planning and action. Dawid and LeCun describe one architecture organized around those missing functions,6while Chollet distinguishes a large inventory of learned skills from the ability to acquire new ones efficiently.7

Afterimage assumes that this argument may end without a clean winner, with language models providing legible human behavior, learned world representations providing causal structure and persistent environments, and action-conditioned models searching the branches. Once a memory interface can return the result to a person, the combined system becomes more than a forecasting instrument because it can change the decision-maker with the forecast.

In Isaac Asimov's Foundation, Hari Seldon's psychohistory predicted the statistical behavior of very large populations and used that knowledge to steer history indirectly. Afterimage is the mechanistic and personalized version: simulate a population, select a branch, install some of its consequences in the people who will make the next decision, observe the altered population, and run the process again. A government could let a cabinet remember several wars before authorizing one, while a company might give workers the memory of an accident instead of paying to prevent it and a campaign might distribute grief from a future blamed on its opponent, complete with an audit trail proving that every corpse was imaginary. Courts would then have to decide whether a modeled confession is evidence, whether simulated practice creates a professional qualification, and whether removing an afterimage is treatment or the destruction of a part of the patient.

The unnerving feature of this feedback loop is that the model need not predict the untouched future perfectly, because an output that changes what people remember also changes the conditions against which the prediction will be judged. Forecasting then becomes intervention, allowing a model to help produce the society in which its own account appears correct. That possibility is fascinating for the same reason it is dangerous: the system would not need to seize control from human beings if institutions kept asking it which human beings they ought to become.

Five consequences of the same invention

An imported memory might be part of a person, a tool used by that person or a contaminant that happened to learn the person's name, and the stories follow those different answers into a criminal recovery operation, an international emergency, a municipal referendum, a music market and an inheritance dispute. Each setting exposes a different failure mode, although the most serious trouble usually begins when the machine works closely enough to make its users confident.

Across the sequence, the same technology has to bear the weight of incompatible motives and permanent consequences, leaving the reader to decide which remembered lives count and who is entitled to make that decision for somebody else.