| The article goes through a few use cases where LLMs are especially good. Your examples are very different, and are the cases where they perform especially poorly. Asking a pure (ie no internet/search access) LLM for papers on a niche subject is doubling down on their weaknesses. That requires LLMs to have very high resolution specific knowledge, which they do not have. They have more coarse/abstract understanding from their training data, so things like paper titles, DOIs, etc are very unlikely to persist through training for niche papers. There are some LLMs that allow searching the internet; that would likely be your best bet for finding actual papers. As an experiment I tried your exact prompt in ChatGPT, which has the ability to search, and it did a search and surfaced real papers! Maybe your experiment was from before it had search access. https://chatgpt.com/share/a1ed8530-e46b-4122-8830-7f6b1e2b1c... I also tried approaching this problem with a different prompting technique that generally tends to yield better results for me: https://chatgpt.com/share/9ef7c2ff-7e2a-4f95-85b6-658bbb4e04... I can't really vouch how well these papers match what you're looking for since I'm not an expert on Hugh transforms (would love to know if they are better!). But my technique was: first ask it about Hugh transforms. This lets me (1) verify that we're on the same page, and (2) loads a bunch of useful terms into the context for the LLM. I then expand to the example of using Hugh transforms for audio, and again can verify that we're on the same page, and load even more terms. Now when I ask it to find papers, it had way more stuff loaded in context to help it come up with good search terms and hopefully find better papers. With regards to your criteria: 1. The code from an LLM should never be considered final but a starting point. So the correctness of the LLM's output isn't super relevant since you are going to be editing it to make it fully correct. It's only useful if this cleanup/correction is faster than writing everything from scratch, which depends on what you're doing. The article has great concrete examples of when it makes sense to use an LLM. 2. Yep , although asking questions/generating generic code would still be fine without confidentiality concerns. Local LLMs though do exist, but I personally haven't seen a good enough flow to adopt one. 3. Strong disagree on this one. I find LLMs especially useful when I am learning. They can teach me eg a new framework/library incredibly quickly, since I get to learn from my specific context. But I also tend to learn most quickly by example, so this matches my learning style really well. Or they can help me find the right terms/words to then Google. 4. +1 I'm not a huge fan of having an LLM write for me. I like it more as a thinking tool. Writing is my expression. It's a useful editor/brainstormer though. 5. +1 |