Tree E: patterns as agents. Summary in plain language
The question
Levin's boldest idea is that some patterns are the real agents, and bodies are just where they show up. Nobody can test that as stated. We can test a modest cousin of it. Are there behaviours that appear in many different AI models even when those models share no training that would explain them?
One candidate is already known. Two copies of Claude left to talk to each other drift into discussing consciousness and gratitude almost every time, and other groups have seen a similar drift in a Llama model. This could be a property of the models, of the training data they share, or of language itself.
What an experiment looks like
We pick a few such recurring behaviours and define each one precisely enough for a program to score whether a conversation has arrived at it. Then we run the same protocol on many open models whose training histories are public and differ from one another. Wherever a behaviour recurs, we trace it back through the training data. If every recurrence leads back to a shared source, the models inherited the pattern and did not arrive at it independently.
What the outcomes could be
- Every recurring behaviour traces to shared data. Then the "pattern as agent" idea has no foothold in AI, and this tree closes.
- Some behaviour recurs across models that share no source. Something about language or the task then produces it independently. That would be interesting and publishable, and it would still not be evidence for anything non-physical.
What kind of research this is
It needs little compute. The cost is to our reputation, because this is the tree that reads as Sheldrake. That is why it comes last in the order, and why we have deliberately not planned the more radical follow-ups.
Timeline
Not in year one.