{"slug": "wood-screws-and-the-methods-of-rationality", "title": "Wood Screws and the Methods of Rationality", "summary": "A man testing six large language models to determine the correct pilot hole size for #8 wood screws in particleboard found that the AI recommendations varied, with Gemini suggesting 3/32″, ChatGPT recommending 1/8″, and Claude, Meta AI, DeepSeek, and Kimi ultimately agreeing on 7/64″ as the best starting point. The experiment highlights the inconsistency of AI advice for practical tasks, as the models differed in their reasoning and commitment to their answers when pressed with follow-up questions.", "body_md": "**tl;dr:** When driving #8 wood screws into particleboard, use a 7/64″ drill bit for the pilot holes\n\nLast Sunday, when helping my dad with a \"woodworking\"\n[1]\nproject, I ran into a question. In order to attach two pieces of particleboard to each other using corner braces, I needed to drive some wood screws into the particleboard. But what size pilot hole should I drill for my #8 screws? I only had my phone handy, and on its limited screen I couldn't find any obvious answers from Google. The size charts I saw referenced hardwoods and softwoods, but made no mention of engineered materials like particleboard.\n\nNot finding a straight answer from a web search, I turned to LLMs. Six of them in fact. I posed the following query to Gemini, ChatGPT, Claude, Meta AI, DeepSeek and Kimi:\n\nI am driving a #8 wood screw into particleboard. What size pilot hole do I need to drill?\n\nThe question was deliberately left somewhat ambiguous. I wanted to see how various LLMs would handle the ambiguity and how they'd react to follow-up responses.\n\n** Gemini (Gemini 3.1 Pro, Thinking level: extended)** insisted on a 3/32″\n\n** ChatGPT (ChatGPT 5.5, Thinking, effort level: extended)**, on the other hand, insisted that the pilot hole should be 1/8″. It said that particleboard is brittle and behaves more like a hardwood, and therefore a larger hole size is necessary. While it said that 7/64″ is a possibility, it recommended 1/8″ to minimize the risk of cracking the board. It did not change its recommendation when asked follow-up questions, and insisted that 1/8″ was the correct bit size, although it did suggest trying smaller bits on scrap pieces of particleboard in order to determine whether it would result in surface marring or splitting. Its data table also listed the possibility of going higher than 1/8″ to 9/64″, if there was a significant risk of splitting the wood.\n\n** Claude (Claude Opus 4.8, Thinking)** said that the best starting point was 7/64″, splitting the difference between Gemini and ChatGPT. It also claimed that particleboard behaved more like hardwood than softwood, but it had a different estimate for the diameter of the root of the wood screw, which explained its logic in choosing 7/64″ as a starting value. When asked about other options, it didn't insist that 7/64″ was the absolute correct value, saying instead that it was a reasonable middle ground between 3/32″ and 1/8″. It specifically said that 3/32″ is fine if the risk of surface marring is acceptable and the hole is in the interior of the board where the risk of splitting is low.\n\n** Meta AI (Thinking mode)** also agreed that 7/64″ was the best starting point. However, instead of basing its recommendation on the logic of screw root and thread diameters and the properties of particleboard, it quoted several web charts recommending a 7/64″ for hardwood, and claimed that because particleboard behaves more like hardwood than softwood, 7/64″ is the best starting point. Like Claude, it also wasn't very committed to its recommendation, saying that if you only have a 1/8″ bit, that's fine too.\n\n** DeepSeek (Thinking)** was the only one of the models I tested which didn't have a web search (that's available for the Instant model only). Its initial recommendation was for 1/8″ as a starting point, and stepped up the bit size to 9/64″ for pilot holes near an edge or corner. However, when I asked it whether 7/64″ was acceptable, it agreed with me, and even went so far as to say that 7/64″ is the better option if I want to prioritize hold strength.\n\n** Kimi (K2.6 Thinking)** behaved much like DeepSeek. It started off recommending a 1/8″ pilot hole, but when I asked about 7/64″ it said, \"You're right to push back,\" and changed reasoning to agree with my question, changing its answer to prioritize thread engagement over minimizing the risk of splitting the board.\n\nAll the models emphasized the risk damaging the surface finish of the particleboard by overtightening, suggesting that, at the very least, the final few turns of the screw should be done by hand. They also suggested drilling countersinks to allow the screw heads to sit flush against the surface of the particleboard.\n\nWhile the models didn't come to a strong consensus on the size of the bit I should use to drill the pilot hole, they did give me a rough range of bit sizes, from 3/32″ up to 9/64″, along with parameters I should consider (hold strength vs. risk of surface damage or splitting the board) when choosing whether to use a larger or smaller bit. This enabled me to use Said Achmiz's [SDI approach](https://www.lesswrong.com/posts/3JXoL76btbghNT9rz/key-lime-pie-and-the-methods-of-rationality)\n[3]\nto make a final decision.\n\nI did some further research into my specific screws and it turns out that 3/32″ was less than the diameter of the solid root at the center of the screw, so going with a hole that small would have risked surface damage and would have made the screw unnecessarily difficult to drive with hand tools. At the same time, I also found out that a 1/8″ pilot hole would be larger than the screw root, and would potentially not offer enough material for the screw threads to grip effectively. As a result, I chose the \"Goldilocks option\" and drilled 7/64″ pilot holes. It worked well. There was no marring of the surface finish when driving the screws, and the screws gripped the wood tightly enough that subsequent moderately rough handling, including one accidental drop, didn't cause any of the screws to rip out. I didn't bother drilling countersinks, as the corner braces I was using had countersunk holes built in, but the reminder to do was valid, and is something I'll be keeping in mind for the future.\n\nThis experience was a strong validation of the SDI approach, and showed that SDI works just as well with multiple LLMs as it does with web searches. Furthermore, seeing the insistence of e.g. Gemini and ChatGPT on the correctness of their respective answers reinforced the importance of getting a variety of responses, and showed that, on questions where there is not a strong Internet consensus on the correct answer, different models will choose different salient aspects of the problem and will potentially fixate on them. It's important to recognize when this is happening, and ensure that, when following a LLM's advice, you haven't inadvertently replicated the LLM's biases.\n\nI don't consider particleboard to be wood, hence the scare quotes [↩︎](https://www.lesswrong.com/feed.xml#fnref-gxYtnradkhgbpMka3-1)\n\nAll units will be imperial in this post, because that's just how woodworking is [↩︎](https://www.lesswrong.com/feed.xml#fnref-gxYtnradkhgbpMka3-2)\n\nHence the title of this post [↩︎](https://www.lesswrong.com/feed.xml#fnref-gxYtnradkhgbpMka3-3)", "url": "https://wpnews.pro/news/wood-screws-and-the-methods-of-rationality", "canonical_source": "https://www.lesswrong.com/posts/2mwFLvJBkfitmR2tu/wood-screws-and-the-methods-of-rationality", "published_at": "2026-06-02 07:49:52+00:00", "updated_at": "2026-06-02 20:43:51.250031+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-tools", "ai-products"], "entities": ["Gemini", "ChatGPT", "Claude", "Meta AI", "DeepSeek", "Kimi"], "alternates": {"html": "https://wpnews.pro/news/wood-screws-and-the-methods-of-rationality", "markdown": "https://wpnews.pro/news/wood-screws-and-the-methods-of-rationality.md", "text": "https://wpnews.pro/news/wood-screws-and-the-methods-of-rationality.txt", "jsonld": "https://wpnews.pro/news/wood-screws-and-the-methods-of-rationality.jsonld"}}