My suspicion is that any size in this range would work just fine. You are asking for (and LLMs are gladly spending tokens on) more precision than is available from the sources and reasoning they have.
Did you try at least a few options to see if any of the warnings or predictions were correct at this granularity? I’d expect the differences to be very difficult to notice.
Later on, out of curiosity, I did try a 3/32″ pilot hole, on a separate scrap piece. I did see the phenomenon that ChatGPT and Claude were warning about, where the surface of the material mushroomed up and ruined the surface finish. A 1/8″ pilot hole seemed to work fine, but I didn’t conduct a rigorous stress test to determine whether it was significantly weaker than the 7/64″ hole. It’s entirely possible that I could have used a 1/8″ bit and it would have worked fine.
Yes, my usage of 6 LLMs for this was gratuitous. It was less about getting practical advice and more about seeing how different LLMs would handle a question for which there wasn’t a single “correct” answer based on internet sources, and how they’d respond to feedback from the user about other approaches. I learned that ChatGPT and Gemini tend to fixate more and push back against the user, while Claude is less committed to its answer. ChatGPT, Gemini and Claude all seemed to have a similar level of detail while the quality of the answers provided by Meta AI, Deepseek and Kimi was a step below.
My suspicion is that any size in this range would work just fine. You are asking for (and LLMs are gladly spending tokens on) more precision than is available from the sources and reasoning they have.
Did you try at least a few options to see if any of the warnings or predictions were correct at this granularity? I’d expect the differences to be very difficult to notice.
Later on, out of curiosity, I did try a 3/32″ pilot hole, on a separate scrap piece. I did see the phenomenon that ChatGPT and Claude were warning about, where the surface of the material mushroomed up and ruined the surface finish. A 1/8″ pilot hole seemed to work fine, but I didn’t conduct a rigorous stress test to determine whether it was significantly weaker than the 7/64″ hole. It’s entirely possible that I could have used a 1/8″ bit and it would have worked fine.
Yes, my usage of 6 LLMs for this was gratuitous. It was less about getting practical advice and more about seeing how different LLMs would handle a question for which there wasn’t a single “correct” answer based on internet sources, and how they’d respond to feedback from the user about other approaches. I learned that ChatGPT and Gemini tend to fixate more and push back against the user, while Claude is less committed to its answer. ChatGPT, Gemini and Claude all seemed to have a similar level of detail while the quality of the answers provided by Meta AI, Deepseek and Kimi was a step below.