ChatGPT recommended an Arduino library that doesn't exist.
Vložit
- čas přidán 6. 09. 2024
- ChatGPT is great at fabricating convincing-looking but totally-false information, and it looks like Arduino code is no exception! I asked it about doing computer vision with an Arduino and it recommended using a computer vision toolbox that as far as I can tell, doesn't exist. Do I have this wrong? Leave a comment! Check out this playlist for more videos about using ChatGPT for Arduino and electronics: • ChatGPT and Electronics
---------------------------------------------
My other channel: / @ithacanature
My day job: / @science.buddies
That false information could very well be related to "garbage in/garbage out". In voice recognition, at least for a while, an editor successfully fixed multiple errors in the results of VR, than after another/others edited during another shift, the VR system reverted back to incorrect results. Pretty clearly related to incorrect information input.
the same type of scenario could be happening with ChatGPT, if malicious actors are entering bogus junk it could be learning that and until someone who is smarter tells it otherwise it spouts that incorrect info.
At least that is my supposition about what is happening.
BTW, thanks for very clear and concise videos. They'll help me immensely!
It's not necessarily "garbage in," it's using the internet as training data (well, arguably a lot of the internet is garbage) but that doesn't mean 100% of what it generates will be factually correct. There are lots of other videos on CZcams that will explain how large language models work better than I can.
@@BenFinio The "garbage in " I was referring to is the garbage fed AI by these bad actors. Possibly maliciously, as a joke, or just plain ignorance of what they are presenting to ChatGPT. Thus "teaching" it bad information. As in humans being taught by propaganda, or conspiracy theories, or just plain wrong information.
The training data ended on September 2021. It is reasonable to assume that information was taken down, censored, or otherwise changed in two years. It is limited by design which makes it even more humanlike, similar to chess engines programmed to randomly make errors to accomodate lower level human players. The desire for AI to be more human is what is horrible about 2001:A Space Odyssey.
Because GPT-AI renders its responses based on statistics, not on understanding the request or the response. And every statistic has errors. These "hallucinations" are nothing more than statistical errors either due to the source texts, the training algorithm or the response rendering algorithm. Only use GPT-AI if the responses are allowed to have statistic errors. If you need responses without statistical errors, do not GPT-AI technology.
I know - this video is intended as more of a warning for people who might go ask ChatGPT for Arduino code advice as opposed to Googling it. Google might give you forum results with bad advice, but in general at least it won't give you things that don't even exist.
@@BenFinio The true power of GPT-AI unfolds if you use it to uncover correlations between different topics, since it has all the knowledge embedded in its data and the statistics how their correlate to each other. "Evalute the ASPECT of the following CONTENT" (like "Evaluate the legality of the following contract: CITATION" or "Evaluate the ethics of the following statement" or "Evaluate the psycholigical state of the writer of the following letter") truly shows how powerful GPT-AI is, because it acts like an expert consultant - but a consultant that is an expert in everything and can correlate all fields with each other.
everyone be hiding their threads lmfao
Yep.