Having AI not bullshiting will require an entirely different set of algorithms than LLM, or ML in general. ML by design aproximates answers, and you don’t use it for anything that’s deterministic and has a correct answer. So, in that rwgard, we’re basically at square 0.
You can keep on slapping a bunch of checks on top of random text prediction it gives you, but if you have a way of checking if something is really true for every case imaginable, then you can probably just use that to instead generate the reply, and it can’t be something that’s also ML/random.
It was only two years, and it was basically half nornal computer science classes, and half working with engines, making a game with classmates and mentors from the industry throughout the year, and learning about rendering, AI behaviors (the videogame kind, not LLMs). The graphics part was about shaders, lighting, post-processing, global illumination, renderers and math, not modeling. It was mostly technical, but we had some game desing classes.