Hacker News new | ask | show | jobs
by hatthew 36 days ago
With a car, when I turn the steering wheel a quarter turn, the car immediately starts to turn at a known rate (I can't quantify it, but my muscle memory knows it). With an LLM, if I ask it to "turn the car", my "steering wheel" is natural language. I have to spend effort to figure out how to convey the rate at which I want the car to turn, and because the LLM has a noticeable delay, I also need to preemptively include some language anticipating how the expected rate should feel and what to do if it doesn't seem to feel right. And then the next time I say the same thing, the LLM decides to steer the car a bit differently anyways.
2 comments

I'm sure the first farmers said that when they drove a tractor. "I can't feel the land", but the ones that got over that could do 10x yield. The "feel" just moves up the stack.
Why would you use an LLM to steer a car? That doesn't make any more sense than heating a water pressure chamber to drive a piston to steer a car. Wrong tool.