Tree C: discovered or built. Summary in plain language
The question
AI models trained separately end up organising what they know in surprisingly similar ways. Perhaps they are all discovering the same structure, which was already there. Or perhaps they were fed the same internet and built the same thing from it. Levin's idea is the first reading pushed to its limit. According to the 2026 evidence, about half of the reported similarity was a measurement artefact. What survives is agreement about how things relate to one another. The geometry of their internal spaces does not match.
A second question sits inside the first. Sheldrake says a pattern gets easier to form each time it has formed before. Levin, at the mathematical end of his idea, says it does not. With AI we can test this. Does a later model resemble an earlier one more when the earlier one's outputs were in its training data?
What an experiment looks like
We train many small models under controlled conditions. Some share their training data and some have completely separate data. Some share an architecture and some do not. Some are large and some small. For each pair we measure how similar their internal organisation is, using the corrected 2026 methods that do not inflate with size. Then we repeat the experiment with one change. An earlier model's outputs go into a later model's training data, and we check whether the resemblance jumps.
We also train on invented data whose structure we fully control. That makes the "same world" condition exact, because we know precisely what there was to discover.
What the outcomes could be
- Similarity tracks shared data and nothing else. In that case the models built the structure out of their data, and Levin's reading loses something concrete.
- Similarity is higher than shared data predicts, even across different architectures. Then something about the task or the world forces the same organisation. That would prove nothing non-physical. It would, though, be the strongest evidence a Platonist could ask for, and it would need independent replication.
- The history test gives an answer either way. If later models resemble earlier ones only through those earlier models' outputs, then "patterns get easier once formed" is a data effect. If the resemblance is the same regardless, history does not matter. Both answers would be new.
What kind of research this is
Standard machine-learning experiments with careful controls. The main cost is compute. Of all the trees, this one has the best chance of being published at a mainstream venue.
Timeline
Six to nine months once compute is available. Apart from data preparation, nothing here runs on the current server.