• Another thread in [Obsolescence and Impatience

golden_lucky AI technology is evolving so rapidly that I have a sense of frustration that it will become obsolete no matter how much I study. I’m feeling frustrated because I’m afraid that even if I study it, it’s going to be obsolete. golden_lucky I think I get the general idea, but I don’t think I can bridge the gap between that and “what is realized and in front of my eyes”. I don’t think I can bridge the gap between that and “what is realized and in front of my eyes. I don’t think I can bridge the gap between the semiconductor principle and the CPU as a product, and I don’t think individuals can reach it with a single hand.

ajiyoshi Really this. I tried to implement the contents of “Deep Learning from Zero” in go instead of python, but all I got was “You’re making NN”. I didn’t feel that my understanding deepened even though I made it. ajiyoshi transformer, and I’ve seen a paper on it, but I’m sure I wouldn’t know anything about why this would work or anything if I read that and tried to make it. I have a hunch. (It seems that the person you are dealing with cannot be understood by the method of making your own products and trying to tweak them on your own.)

@nishio: If I were to use an analogy, I could do $.ajax with jQuery and say, “Wow! I’m sure that’s true” and “I’m just trying to catch up with the introduction of algorithms in C languages”. and “I read an introduction to algorithms in C to catch up, but…” to “What? Isn’t there something wrong with the layers?” I’m sure that’s what you’re thinking.

ajiyoshi I understand, but I want to build models, tinker with them, break them, and play with them. but there is not much that can be grasped in that way for DNNs. I have no choice but to play around with the API.

nishio If you don’t like the API, there’s an approach to running it locally yourself. I added more memory, 64GB, and tried a language model I couldn’t get to work - Kishida’s Hatena

ajiyoshi I think even if you do that, you’ll end up not understanding it well (= no matter how much you tweak it, it will look like a big NN), so I’m not going to dig into the internal structure. I think it’s OK if I don’t dig up the internal structure of the NN. I should just focus on what I can do with it.

relevance - Looking for reasons not to - Left half of [Attacking the unfamiliar without looking at it properly. - They don’t get closer, saying, “I wonder if getting closer won’t help me see better” when they could see better if they got closer.


This page is auto-translated from /nishio/Ajaxで盛り上がってる時にC言語の話をされてもレイヤーおかしい using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.