| I like LLMs. I really do. But my experience with them is very different from the Chicken Little folks. Let's park coding to the side for a bit. Case 1: I am collaborating with a friend to build a graded Sanskrit reader for beginners using the Aesop's fables. As a precursor, I asked Gemini 2.5 Pro if it had access to all the stories. Yes, it said. The three popular PD ones? Yes. I asked it to print all three versions of a particular one, and it did. One of them was not the version it confidently claimed it was. We argued about it for a while. It shut up when I provided actual evidence. I then decided to upload the three Gutenberg text files and asked it to use them as the source of truth to give me a list of unique stories putting variant plots, variant titles etc under the main heading. I gave it certain formatting requirements so that I could later verify if all 600-odd tales across the three books were properly accounted for. Gemini tied itself into knots trying to do this. It could not guarantee that all the tales were present in the list it generated. It didn't know how to accomplish the task. Finally, I gave it a series of steps, an algorithm based on an n-branched tree. Only then did it manage to generate the list for me. This took me four hours of wrangling across three different sessions. Case 2: I have been buying TASCHEN editions of impressionists and other classical artists. I wanted Gemini to compare various editions, give me the pros and cons so that I could pick a good edition to buy. By the time we came to Michelangelo it went nuts, hallucinating editions, ISBN numbers, page counts, authoritative urls, worldcat searches ... This took about two hours. There are more such amusing anecdotes. Some from DeepSeek as well. I have tried LLMs with python, and typst and a few other things. Sometimes they work, sometimes they don't. They definitely do not write code the way I want them to. They will use OOP even if I specifically warn them not to. LLMs are VERY good at translation and languages. I will give them that. But reasoning? I am not convinced. I will believe that LLMs are good enough to replace programmers when the Amodei siblings can operate their company only using LLM developers. |