3 ms·
> I want to understand how things work. AI feels like a black box to me. The amount of papers I’d have to read and mathematics that I’d have to ingest to really
by mo_42 3y ago
> I want to understand how things work. AI feels like a black box to me. The amount of papers I’d have to read and mathematics that I’d have to ingest to really understand why a certain prompt X results in a certain output Y feels overwhelming.
First, no paper will make you understand why prompt X made result Y. But you can understand the architecture of these systems and try to understand the prompt answer relations somewhat intuitively (e.g., by learning your own models with various text collections).
I think we need to accept these AIs as creatures of their own kind. As complex actors that we don't fully understand.
Humans have been breeding dogs since tens of thousands of years and we don't understand them fully. But we understand enough to employ them in useful ways and mitigate the risks.
- hibikir 3y agoMost of us don't understand many complicated things we interact with in a regular basis. Back to even the simple cases of what happens when I type something in this form, and hit enter. Every keystroke is a bunch of work, by a bunch of devices talking to each other through complicated protocols, just to get to memory, which will get to my screen, eventually... and that's pretending we don't care about how physics works, because it's reliable. It's all overwhelming. And if we only focus on the unreliability, look at many bugs in our systems today. Why does this crash with some input? Why does a new enough processor have security vulnerabilities, so someone can steal passwords? All might as well be magic, even though a few humans, somewhere, might know the reasoning. And if AI feels difficult, imagine biotech. I could, theoretically, mess with weights and retry a prompt over and over again, seeing what changes. It's a lot of work, but it can be done. See how much fun we have figuring out what a single nucleotide polymorphism does phenotypically, and why it does it: It's a research project by an expert in the easiest of cases! We are only a little ahead of the new C programmer changing things at random to see where his pointer math went wrong. We don't understand how anything works. We are just sometimes satisfied with our degree of ignorance, and decide to stop asking questions.
- Workaccount2 3y agoAI is different because even when you understand all the steps in generating one token, it still doesn't help you understand how the next token works. I can press a key on my keyboard and work through the system to understand exactly how the physical press up to the letter on the screen works, and then apply it to all the other keys no problem. AI does not work that way at all. The steps are seemingly random, meandering, and nonsensical, yet it still ends up with these well structured chains of tokens on the output.