Post B4i43YxYNz8bfby7UG by tschundler@leds.social
(DIR) More posts by tschundler@leds.social
(DIR) Post #B3AlhsEXmwo9svENbE by hailey@hails.org
1 likes, 3 repeats
Computers can be understood https://blog.nelhage.com/post/computers-can-be-understood/
(DIR) Post #B4i43YxYNz8bfby7UG by tschundler@leds.social
0 likes, 1 repeats
@haileyThis is one of my biggest hurdles with LLMs. It's a machine that can't be understood.I have a predictable mental model of what I work on and the tools that I use. And if I find that it doesn't hold, I've found a bug that I can fix or need to update my mental model. I can build a model by playing with inputs and observing their effect of outputs.Meanwhile I can't predict an LLM, as the output changes with no input change. There's no way to model what effect a input has.