It was impossible for computers to beat chess and go masters when the computers were trying to play like humans -trying to model high level understanding of strategy and abstract values. The computers started winning when they got fast enough to brute force games - to calculate all of the possible outcomes from all of the possible moves, and to choose the best one.

This is basically the same difference between LLMs and ‘true’ general AI. The LLMs are brute forcing the next line of a screenplay, with no way to incorporate abstract concepts like truth or logic. If you confuse an LLM for an AI, then you’re going to be disappointed in its performance. If you accept that an LLM is a way to average past communications, and accept that a lot of its training set were fiction, then it’s an amazing tool for generating consensus text (given that the consensus includes fantasies and lies). It’s not going to write new code, but it will give you an approximation of all the existing examples of some algorithm. An approximation that may introduce errors, like copy-pasting sequential lines from every stackexchange answer.

Computer graphics, computer game opponents, they’re still doing the same things they were doing decades ago, and the improvements are just doing it all faster. General AI needs to do something different than LLMs and most other ML algorithms.