Why I rejected AI for the core of FixMyPrint

Why I rejected AI for the core of FixMyPrint AI is everywhere in 3D printing tools right now. Upload a photo, describe your problem, get an answer. It sounds like exactly what the hobby needs. I thought so too, until I actually tested it. Before I wrote a single line of the FixMyPrint settings engine I spent time experimenting with AI-generated print advice. I ran the same problems through repeatedly. Changed the wording slightly. Asked on different days. The results were inconsistent in ways that mattered. The same problem would get different recommended temperatures, different retraction distances, different root causes depending on how I phrased the question or when I asked it. For something people rely on to fix expensive failed prints that is not good enough. The trust problem with AI is not unique to 3D printing. It shows up everywhere people have tried to use it for technical diagnosis. The model generates the most statistically likely answer, not the most correct one. For 3D printing those two things diverge constantly because the correct answer depends on a specific printer, a specific filament, a specific slicer, and a specific failure mode. A general statistical answer trained on the entire internet is not equipped to be specific in the way that actually helps. I also watched how the community responds to AI-generated advice when people post it. The experienced makers spot it immediately. The answers are plausible and generic in a way that real troubleshooting experience is not. They hit the most common suggestions without accounting for the actual setup. Dry your filament. Check your Z offset. Level your bed. These are real fixes for some problems some of the time. As a default answer to every problem they are lazy and they do not help. That default answer is actually what bothered me most. The 3D printing community is enormous. Millions of people print, millions of prints fail, and the accumulated knowledge of what actually fixed those failures exists somewhere in all those threads and posts and follow-ups. The problem is that it is buried. The signal is there but it is hidden under an enormous volume of generic advice that never got confirmed as working. The gap I wanted to close was not between the user and an AI. It was between the user and what has actually worked for people in their exact situation. That is a data problem not a language model problem. So instead of building an AI that generates plausible-sounding answers I built a system that finds confirmed ones. I scraped tens of thousands of community posts and tracked which suggested fixes got confirmed as working by the person who had the problem. Not upvoted. Not repeated most often. Confirmed. Then I ranked everything by those confirmed outcomes scoped to the user's specific printer and filament. The settings engine went the same direction. Every value it outputs traces back to a rule derived from manufacturer specs, filament data sheets, and verified community knowledge. Same input, same output, every time. No sampling, no temperature, no hallucination. If you ask for Cura settings for a Bambu P1S printing PETG you get the exact field names Cura uses with values that have been verified against what actually works for that combination. Not a best guess. A known answer. AI does have a role in FixMyPrint. It classifies natural language problem descriptions into the diagnostic system and it powers photo analysis for users who want to upload a picture of a failed print. Those are tasks where AI genuinely helps - understanding language and identifying visual patterns. Generating specific technical values is not one of those tasks and I am not going to pretend otherwise. The other thing AI cannot do is learn from real outcomes. FixMyPrint can. Every fix confirmed in the app, every setting that gets marked as working, feeds back into the data. The next person with the same printer and the same problem gets a result that is informed by everyone who came before them and actually said what worked. The system gets more accurate over time not because a model was retrained but because real people confirmed real outcomes and that signal compounded. The photo diagnosis side is heading the same direction. Right now Photo Diagnosis uses GPT-4o Vision to identify failure modes from uploaded images. That works but it still relies on a general model interpreting a specific problem. The long term goal is to build a visual dataset from real user-submitted failed prints so that a photo of stringing gets matched against thousands of confirmed stringing photos rather than relying purely on what a language model thinks stringing looks like. Every photo uploaded in the app is a potential contribution to that dataset. The more confirmed failures we collect the less we need to lean on AI inference and the more we can lean on verified reference. The 3D printing community has spent years building an enormous knowledge base through trial, error, and the occasional person who came back to say what worked. That knowledge base is the most valuable resource in the hobby and almost none of it is organized or searchable in a useful way. FixMyPrint exists to change that. Not by generating new answers but by finding the ones that already exist, strengthening them with every confirmed fix, and putting them in front of the right person at the right time. That felt more honest than building something that sounds smart but guesses. fixmyprint3d.com - free to try, no card required.