Definitely glad to see some investigation into the path dependence question.
I expect that the primary source of safety-relevant path dependence in future systems will be due to causal influence of the model’s behavior on its training data / supervision signal. That should occur by default in reinforcement & active learning, but not in typical teacher-forced self-supervised learning (like GPT). So I think I would answer the question of “Are we in a high path-dependence world?” differently conditioned on different AI development models.
Definitely glad to see some investigation into the path dependence question.
I expect that the primary source of safety-relevant path dependence in future systems will be due to causal influence of the model’s behavior on its training data / supervision signal. That should occur by default in reinforcement & active learning, but not in typical teacher-forced self-supervised learning (like GPT). So I think I would answer the question of “Are we in a high path-dependence world?” differently conditioned on different AI development models.
Even for GPTs, the recently popular “chain-of-thought” family of techniques seem poised to bring path-dependence into the mix, by creating feedback loops between the language model and the reasoning traces it produces.