Rube Goldber Machines, Marbles, and AI<\/p>\n\n\n\n
Context drives behaviour, it\u2019s true. There. If you were a contemporary safety professional who has read a few books, my title was successful in getting you this far, but it was a trap. This article isn\u2019t about the ways in which the environment shapes the worker behaviour on our job sites, though it obviously does so; please just quit talking about it so much. No, this is an article about artificial intelligence.<\/p>\n\n\n\n
But don\u2019t leave yet (please?).<\/strong><\/p>\n\n\n\n There are analogies between how we think about how context drives behaviour on the job site and how it drives behaviour when we interact with a large language model. What we commonly refer to as AI, you know: ChatGPT, Claude, Copilot, and so on. If you\u2019ve dabbled with AI for any amount of time, you\u2019ve no doubt realized that what you put in has a lot to do with what you get out. Similar to how the amount of effort we put into designing a system of work to enable the smooth flow of labour and cognition, the better our outputs will be.<\/p>\n\n\n\n Imagine in your mind\u2019s eye a rube Goldberg machine, you know, one of those things that you design, and you place a marble at the top, and the marble works its way through a series of ramps, obstacles, gadgets, switches and the like, which makes its way down to turn on a toaster or perform some other task. In the same way, when we design the context for an LLM, we need to consider what our intelligence will need to bump up against, slide down, run into, or turn on in order to complete the task.<\/p>\n\n\n\n For example, say we want a large language model to build an entire safety program manual. It would be a mistake to simply prompt a large language model with, \u201cbuild a safety program manual,\u201d and expect that we would get anything useful. This would be akin to building a Rube Goldberg machine with a single ramp. Yes, the marble will get to the bottom, but that was never the point. We need to get the marble to collide with a few more objects along the way.<\/p>\n\n\n\n What does this look like?<\/strong><\/p>\n\n\n\n The first step we might take is to ask the LLM to research our industry and local regulations. Next, we might have it ask us questions to further clarify what it is we are trying to design. We might grant it access to tools so that it can review files and photos, and to folders containing some of our project-related files. We will want it to present its output in a particular format and style, so we will provide templates and instructions.<\/p>\n\n\n\n Context drives behaviour, it\u2019s true. There. If you were a contemporary safety professional who has read a few books, my title was successful in getting you this far, but it was a trap. This article isn\u2019t about how the environment shapes worker behaviour on our job sites, though that\u2019s also an important concept. No, this is an article about artificial intelligence.<\/p>\n\n\n\n But don\u2019t leave yet (please?)<\/strong><\/p>\n\n\n\n There are parallels between how we think about context driving behaviour on the job site and how it drives behaviour when we interact with a large language model. What we commonly refer to as AI: ChatGPT, Claude, Copilot, and so on. If you\u2019ve dabbled with AI for any amount of time, you\u2019ve no doubt realized that what you put in has a lot to do with what you get out. Similar to how the effort we put into designing a system of work to enable the smooth flow of labour and cognition correlates with better buildings, products and widgets.<\/p>\n\n\n\n The Rube Goldberg Machine and the marble of synthetic intelligence<\/strong><\/p>\n\n\n\n Imagine in your mind\u2019s eye a rube Goldberg machine, you know, one of those things that you design, and you place a marble at the top, and the marble works its way through a series of ramps, obstacles, gadgets, switches and the like, which makes its way down to turn on a toaster or perform some other task. Similarly, when we design the context for an LLM, we need to consider what our marble of synthetic intelligence will need to bump into, slide past, run into, or turn on in order to complete the task.<\/p>\n\n\n\n For example, say we want a large language model to build an entire safety program manual. It would be a mistake to simply prompt a large language model with \u201cbuild a safety program manual\u201d and expect to get anything useful. This would be akin to building a Rube Goldberg machine with a single ramp. Yes, the marble will get to the bottom, but that was never the point. We need to get the marble to collide with a few more objects along the way.<\/p>\n\n\n\n What does this look like?<\/strong><\/p>\n\n\n\n The first step we might take is to ask the LLM to research our industry and local regulations. Next, we might have it ask us questions to further clarify what it is we are trying to design. We might grant it access to tools so that it can review files and photos, and to folders containing some of our project-related files. We will want it to present its output in a particular format and style, so we will provide templates and instructions.<\/p>\n\n\n\n As with building Rube Goldberg machines, our AI projects never work the first time. It\u2019s only through careful, consistent, and observant iteration that we begin to predict what the marble is going to do with some reasonable degree of confidence. But even when we have everything lined up, they\u2019re still a chance that, for whatever reason, the whole thing goes off course. But the beautiful thing about working with artificial intelligence is that it is self-reflective. We can\u2019t ask the marble how to design its own Rube Goldberg machine, but we can ask a large language model how to help us get unstuck from a particular problem it’s involved in.<\/p>\n\n\n\n So next time you go to work with a large language model, consider the problem you\u2019re trying to solve. Then imagine that you are unleashing an intelligence that will pursue an answer to your question with the persistence of a marble being pulled downward by gravity. Your job is to learn the best ways to guide that intelligence towards solving your specific problem. Then, press Enter and observe what happens so you can further refine the process through trial and error.<\/p>\n\n\n\n Good luck.<\/p>\n\n\n\n So next time you go to work with a large language model, consider the problem you\u2019re trying to solve. Then imagine that you are unleashing an intelligence that will pursue an answer to your question with the persistence of a marble being pulled downward by gravity. Your job is to learn the best ways to guide that intelligence towards solving your specific problem. Then, press Enter and observe what happens so you can further refine the process through trial and error.<\/p>\n","protected":false},"excerpt":{"rendered":" Rube Goldber Machines, Marbles, and AI Context drives behaviour, it\u2019s true. There. If you were a contemporary safety professional who has read a few books, my title was successful in getting you this far, but it was a trap. This article isn\u2019t about the ways in which the environment shapes the worker behaviour on our […]<\/p>\n","protected":false},"author":1,"featured_media":8,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-10","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-insights"],"featured_media_src_url":"https:\/\/pragmaticsafety.ca\/blog\/wp-content\/uploads\/2026\/01\/nanobanana_generated_2057_1768276965_0_hbuege-1024x572.png","_links":{"self":[{"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/posts\/10","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/comments?post=10"}],"version-history":[{"count":1,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/posts\/10\/revisions"}],"predecessor-version":[{"id":11,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/posts\/10\/revisions\/11"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/media\/8"}],"wp:attachment":[{"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/media?parent=10"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/categories?post=10"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/pragmaticsafety.ca\/blog\/wp-json\/wp\/v2\/tags?post=10"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}