What is this?

One machine. Several futures.

The mechanism is treated as a rule, not merely as a metaphor. The stories vary who receives it, who controls it, what it costs and what survives its misuse. The point is not to make five arguments for the same conclusion. It is to discover how one invention becomes five different moral and practical problems when it enters different lives.

This first series is about the afterimage. In 2035, an overlay procedure can transfer selected memories from simulated futures into living people. Recipients know that the memories came from a model, yet retain them much like ordinary experience. Skill, loyalty, fear and grief can therefore arrive before the events that would have produced them.

A world model you can remember.

In AI research, a world model is a learned representation of how an environment behaves. It lets an agent anticipate changes and test actions without trying every possibility in the real world. Ha and Schmidhuber demonstrated the strange core of this idea especially clearly: a policy could be trained inside a model's generated “dream,” then carried back into the environment.1

Afterimage takes that transfer literally, but not as a prediction of any particular technology. If a simulated branch can teach an agent what to do, what changes when the lesson arrives as autobiography? A forecast can be doubted. A memory has already helped construct the person doing the doubting.

Is language enough?

One live dispute in AI is whether increasingly capable language models can grow into general intelligence, or whether general intelligence requires a different foundation. Bubeck and colleagues argued that an early GPT-4 displayed broad, if incomplete, signs of general intelligence, while also leaving open whether systems must eventually move beyond next-word prediction.2

Other accounts put more weight on grounded prediction, persistent goals, planning and action. Dawid and LeCun describe an autonomous architecture organized around learned world models,3while Chollet argues that possessing a large stock of task skills is not the same thing as being able to acquire and generalize new skills efficiently.4

These stories do not settle that argument. They borrow its machinery. They assume models rich enough to simulate consequential lives, then ask what happens when their output is not advice but experience. Is a remembered future evidence? Is borrowed expertise yours? If a memory is removed after you act on it, which consequences remain yours? And what if several coherent futures arrive in the same present and refuse to agree?