I used “live with dolphins” specifically to preclude training on corpuses of dolphin behavior data. Can the LLM simply observe dolphins acting a certain way using computer vision, then form predictions of how a different dolphin will behave under the same circumstances? Can it 0-shot dolphin behavior?

You talked about models not persisting weights. If so, it basically means there is no way to see if LLMs can adapt to a totally non-human world the way a human (presumably) could.