ChatGPT thinks you are stupid
This is going to be a long post so:
TL;DR Never trust an AI at face value. It's a useful tool but don't expect miracles.
I've been working on a project where I have decided to print some long spars in polycarbonate (PC). There is a specific reason I chose this material over nylon (PA) or ASA-GF. I won't get into the debate over "Why didn't you use X filament instead?" because I've already been down that road elsewhere.
To cut to the story, this part is long and rectangular... the perfect candidate for extreme warping. So after eating through a $50 roll of PC and many failed prints I finally got everything dialed in: measured shrinkage, calibrated flow rate, insulated the inside of the chamber so I can get up to and hold 60C, used the correct sheet (G10), prevented the sheet from lifting off the bed magnets, used the correct adhesive (Vision Miner Nano) in the correct spots, getting the correct # of perimeters, wall thicknesses and using the correct infill type (gyroid)... I got a near perfect print, some minor warping but happy with the results, no complaints.
Anyway, so I thought to myself "Maybe I could have saved a lot of time and money if I just uploaded my STL to ChatGPT and ask it: 'What should I do?'"
So I did that.
I gave it a very clear description of my physical setup: exact filament, nozzle, layer height, type and thickness of print bed, printer modifications (insulation). I uploaded the STL and asked it to give me the PrusaSlicer settings for a successful print. After analyzing it for 30 seconds or so ChatGPT gave me the "perfect" recipe for a successful print. "Just follow these instructions and everything will work perfectly" (I'm paraphrasing). So for fun, I sliced my part with those exact parameters. But then I'm thinking "Some of these parameters are kind of weird, I would have never had chosen this approach. But ChatGPT must know what it is regurgitating because the answer seems so confident" I play a lot of chess and StockFish (which can easily beat every human on the planet including Magnus Carlson) makes moves that no human would ever consider. So I thought maybe ChatGPT is on to something?
Sent it off to the printer and watched a series of disasters unfold.
When I went back to the chat and said "The print warped in the center very early, I think the initial bed temperature you gave me is wrong" It responded with a "Of course, let's tweak it and make it better" (again paraphrasing). So it gave me some custom g-code to ramp the bed temperature down over the first 6 layers. Interesting.. I wouldn't have thought to do that. OK, that worked but then the edge started curling up half way. Mind you my original successful print took 1.5hrs. ChatGPT also had me drastically lower the volumetric flow rate so now the same print is taking 5hrs (more on that later). It also suggested that if I don't add mouse ears I'll get in trouble with the part lifting off the sheet. So I asked it, won't that create too much adhesion and lift the build plate off the bed? It confidently told me not to worry because since we lowered the bed temperature, the magnets will be stronger. Well, sure enough it warped the build plate like a potato chip.
Now to the fun part. I uploaded my very successful 3MF and told it. "This failed, can you tell what I did wrong?" It dutifully analyzed it and in the nicest terms told me I was an idiot (again paraphrasing but that was the gist) and suggested lots of tweaks which would have again taken me in the wrong direction. In one case when I insisted that substantially reducing volumetric flow would waste a lot of time especially since I'm using 0.2mm layer lines, it threw physics at me trying to explain how "faster flow rates significantly affects warping due to the fact that putting down more plastic creates more shrinkage force and internal stress." OK... this is definitely true up to a point. And it also depends on the material type of the plastic. But it said it in a way that exuded confidence that it was right and I was wrong.
I'll still probably use it from time to time but I'm very wary of taking what it says as face value. It did however give me an excellent recipe for carrot cake that was very delicious.
RE: ChatGPT thinks you are stupid
very wary of taking what it says as face value.
Absolutely... With "experience" as my central value proposition to employers, I embrace AI, make it work for me. And my impression so far, it actually widens the gap, lets me do stuff faster just weed out the bad ideas, thank you very much.
The verdict is still out but I suspect it may be a losing game: The internet gets contaminated increasingly with AI-generated cr*p, every second.
The stuff I know for certain is not necessarily on the web 🙂
RE: ChatGPT thinks you are stupid
It did give me an excellent idea concerning ramping down bed temperature over different layers. I'm definitely going to have to experiment with that a bit more.
RE: ChatGPT thinks you are stupid
First, glad to see there's a fellow chess player here!
Yes, what I "like" the most about AI are those responses "Oh sorry, of course what I just suggested doesn't work at all. Here's the correct answer". So why the f...k didn't you say that in the first place? And of course more often than not the second answer is wrong or unhelpful too.
Don't get me wrong, I don't use generative AI but I use AI searches instead of Google a lot if I want to avoid browsing through a bunch of pages for a factual result. But I wish AIs would just say "I don't know" instead of pulling crap out of their ass and then "talking" to us like our favorite demented aunt.
Formerly known on this forum as @fuchsr -- https://foxrun3d.com/
RE: ChatGPT thinks you are stupid
One thing to understand about generic AI models is that they are not trained engineers, they are trained storytellers. All they want to do is to find one answer that satisfies you, a good enough story but not necessarily correct, and they don’t really know if the answer is correct or what correct means.
I had my first encounter with the human version of storytelling engineers many years ago during the outsourcing era when I had a particular guy coming for some on the job training. He was given an entry level task and asked to present a solution a week later. One week later he presented a perfectly valid solution and I asked him one question, why did you choose this solution above the other options? The response stunned me, he did not seem to understand the question and asked to come back with an answer. Next meeting when I expected him to present his reasoning for why he chose that particular solution he started to present another solution. I again asked why this new solution was preferred over the old one and got the same response, let me come back with an answer. This continues for a number of iterations but never concluded.
later I discussed this with the hiring manager and he explained they faced a problem. When hiring some applicants where storytellers, they where experts at faking expertise and could present any solution as the perfect answer but did not know the topic. Some made it through the hiring process this way, an this one person made it all the way to the on the job training.
currently AI systems not trained for a specific purpose remind me a lot of storytellers. They are not really interested in giving a correct answer, they are just trying to give you a good enough answer with the token budget permitted to keep you happy so that you will continue to subscribe to the service. It is after all the business model of the company providing the service.
/Anders
RE: ChatGPT thinks you are stupid
Yep, I run into that at work a lot when engineers engage in "cargo cult programming". They copy/paste large blocks of code without any understanding of how it actually functions in the hope that it solved the issue that they are unable to analyze on their own. AI makes it worse because the user assumes that "it must be correct" which is why we generally ban AI at work for code generation unless its for something like existing code analysis or unit test automation.
I've used AI to write code for me for my non-work hobbies. I could easily write it myself, it's that the AI is faster and I 100% understand the output so I know if it is correct or not.
In the case illustrated above, had I been a beginner I think I would have given up if I kept following the AI's advice which would have resulted in failures without being able to correct them on my own. I most likely have wasted even more time and material than I did doing it on my own. The "confidence" in it's answers almost comes across as gaslighting. Especially when I ask it "what did I do wrong?"
RE: ChatGPT thinks you are stupid
I also use AI but selectively. I have thrown code analysis problems on it only to find out what I described above, it easily finds a code fix for an issue only to inject one or more new ones, and then down the slippery slope it goes.
The trick is to not ask it to make engineering decisions but to do simple tasks. One of my favorites is search and refining searches without having to dig through sponsored content or advertisements, that alone saves me a ton of frustration. It is good for tuning text but need to be corrected on content sometimes when it alters the meaning. One strategy for engineering tasks is to present the problem and ask it to list strategies for a solution, it still needs guidance to find the correct one and does come up with bogus ones so asking control questions on the response filters out the bad seeds.
It does generate code for simple tasks if you are precise about what you want. I generated an automated Rasberry PI image build for a GitHub repository in 3 commits yesterday, but my prompts where longer than the code and if I would have been satisfied with the first responses it would not have worked.
/Anders
RE:
I’ve learnt that Claude is much better than ChatGPT with coding….it’s ridiculously good, but irritatingly I get free enterprise access to ChatGPT so when Claude kicks me out for 6 hours I have not better option 😂
I don’t think any AI model is going to admit they don’t know something, it’s not a good look…they just churn out the closest answer they have to the question you asked…what’s worse is when ChatGPT just straight up hallucinates API endpoints that don’t exist, Ive had that a few times.
RE: ChatGPT thinks you are stupid
I’ve learnt that Claude is much better than ChatGPT with coding….it’s ridiculously good, but irritatingly I get free enterprise access to ChatGPT so when Claude kicks me out for 6 hours I have not better option 😂
I don’t think any AI model is going to admit they don’t know something, it’s not a good look…they just churn out the closest answer they have to the question you asked…what’s worse is when ChatGPT just straight up hallucinates API endpoints that don’t exist, Ive had that a few times.
Yeah, I've had ChatGPT start to hallucinate on me and produce garbage output. I definitely have to keep an eye on it. Problem is, like you, I have free access via work.
RE: ChatGPT thinks you are stupid
Yeah, I've had ChatGPT start to hallucinate on me and produce garbage output. I definitely have to keep an eye on it. Problem is, like you, I have free access via work.
I once made a fun experiment that started with my typo asking for a non-existing character in a novel. The AI locked onto the one I was actually referring to. When I realized the mistake I claimed just for laughs, no, not that one. They are not the same.
Results were hilarious, superficially credible, like a chapter I skipped by accident. But of course, total utter nonsense.
Maybe Copilot is less aggressive than ChatGPT, better "garbage in" detection?
RE: ChatGPT thinks you are stupid
I'm just wondering what a typical coding interview question will be 2-3 years from now when nobody remembers how to code.
RE: ChatGPT thinks you are stupid
I'm new to 3D printing so I asked ChatGPT for help. It constantly offers after an answer to write me a file, draw me a diagram or asks me to upload a 3D file. I found that never works, and the drawings are totally useless. What I did find it helpful was to upload a photo of a failure and ask what went wrong. It seems pretty good at describing what went wrong and advice to improve the print, especially when analyzing a photo. I don't think I'd trust it for a complete setup.
It also was helpful when I sliced a model for a long 14 hour+ print and showed it a screenshot for areas of the slice that seemed concerning to me. I asked it to let me know what problems might happen with a longer print and whether to use input shaping mode or not. Generally, I think it helped me get up to speed pretty quick and with not so many failed prints, so I would recommend it–but not totally trust it!
RE: ChatGPT thinks you are stupid
AI works from data, large amounts of data. If the data is inaccurate, so are the results. Top AI experts quit their jobs so they could warn us AI wasn't ready. We are seeing the results.
AI needs accurate data. If we could get enough intelligence, people, to review the data and cull out the garbage I would feel more optomistic about AI.
There is an old SciFi movie from 1970. "Colossus: The Forbin Project" I think it should be required viewing for everyone in AI. " I am Robot" was a good warning also.
RE: ChatGPT thinks you are stupid
AI works from data, large amounts of data. If the data is inaccurate, so are the results. Top AI experts quit their jobs so they could warn us AI wasn't ready. We are seeing the results.
AI needs accurate data. If we could get enough intelligence, people, to review the data and cull out the garbage I would feel more optomistic about AI.
There is an old SciFi movie from 1970. "Colossus: The Forbin Project" I think it should be required viewing for everyone in AI. " I am Robot" was a good warning also.