Comment by DiscourseFan

Anyone whose done adversarial work with the models can tell you there are actually things that LLMs get consistently wrong, regardless of compute power. What those things are, it has not yet been fully codified but we are arriving now at a general understanding of the limits and capabilities of these machines and soon they will be employed for far more directly useful purposes than the wasteful, energy-sinks of tasks they are called on for now like "creative" work or writing shitty code. Then there will be a reasonable market adjustment and the technology will enter into the stream of things used for everyday commerce.