His concept of AI seems to be more AGI in the sense of abstract reasoning and creative problem solving applied to problems in one domain where the AI can then accumulate the lessons learned and apply them to another domain. As opposed to the current reality which is mainly stringing classifiers together.
I work as a consultant, and I get these sort of questions from my clients all the time. "Can't we just load the (unlabeled) data into the AI solution and just have it 'figure things out' without training it?" "Can't we take the model that we already trained (on one very narrow domain) and use it (on some completely different domain) without going through all that again?" People want unsupervised transfer learning with a whole lot of a priori knowledge baked into it.
I work as a consultant, and I get these sort of questions from my clients all the time. "Can't we just load the (unlabeled) data into the AI solution and just have it 'figure things out' without training it?" "Can't we take the model that we already trained (on one very narrow domain) and use it (on some completely different domain) without going through all that again?" People want unsupervised transfer learning with a whole lot of a priori knowledge baked into it.