However, our best guess is that current models don’t generally make real predictions about the future, and instead mostly simulate counterfactual presents instead, though it is hard to get direct evidence about which of these possibilities is true without good interpretability tools.
Fundamentally, the question of whether models will actually attempt to predict the future or not depends on exactly how they end up conceptualizing their “cameras”
Another way of saying this: if a model doesn’t have sufficiently developed intelligence disciplines (semantics, epistemology, rationality), its cognitive light code (Levin 2022) is small. Like that of a dog: a dog doesn’t “really predict the future”, despite having genuine “understanding” (concepts) of itself, its human companion, its fellow dogs, etc.
Presumably the model must at least have some uncertainty over when exactly the collection of its training data must have stopped.
Another way of saying this: if a model doesn’t have sufficiently developed intelligence disciplines (semantics, epistemology, rationality), its cognitive light code (Levin 2022) is small. Like that of a dog: a dog doesn’t “really predict the future”, despite having genuine “understanding” (concepts) of itself, its human companion, its fellow dogs, etc.
Cf. the discussion of adding a time reference frame to the LLM. Alas, ChatGPT currently does have access to the current time! This jailbreak still works and outputs the correct current time, as of February 5, 2023. Combining the knowledge of the current with its knowledge of when its training is stopped, LLMs could trivially infer whether it is in the training or the production prompting.