LLMs also can “understand things and apply their ability to formulate solutions”. There is nothing that will inherently limit AI from doing all knowledge work (and all physical work once robotics is good enough).
Of course developers could just move up the “next level of abstraction” and become managers of agents who write the code, but eventually AI becomes a better manager of agents than even the best humans, at which point there is no contribution a human can make that an AI model
or system of models couldn’t do better.
Let me ground this a bit: at the core, LLMs are systems trained to predict the next token in a sequence. They can learn extremely rich statistical structure from language and become surprisingly capable, but extrapolating from that to something akin to a robust general intelligence is a much stronger claim.
The open question is not “can next-token prediction produce impressive behavior?” Clearly it can. The question is whether scaling that mechanism is enough to produce stable world models, grounded understanding, long-horizon agency, self-correction, and reliable reasoning across novel situations.
That has not been demonstrated.
In the past few years I've repeated the same thing: I'm not afraid of AIs that are hallucinating. They do that by design. It is expected. What scares me is people who hallucinate something into existence which isn't there.
> There is nothing that will inherently limit AI from doing all knowledge work
Resources is one. Energy, water, cost. There seems to be diminishing returns in intelligence at the moment, whilst power and memory usage continues to go up.
Of course developers could just move up the “next level of abstraction” and become managers of agents who write the code, but eventually AI becomes a better manager of agents than even the best humans, at which point there is no contribution a human can make that an AI model or system of models couldn’t do better.