The important part is that ChatGPT is not trying to answer correctly. That is not its purpose. The “chat” interface is misleading us into thinking that way, but GPT3 is at its core nothing more than an immensely powerful auto-completion system. It essentially does what you obtain when you repeatedly click on the first suggestion of your smartphone in an SMS. It just does it way way better.
The main consequence is that its optimization goal is not to produce truth, but the most realistic possible texts. In a lot of contexts and due to its humongous learning corpus (The whole Wikipedia is AFAIK less than 5% of it), the most probable continuation of the prompt actually contains truths, but it will mix into these true facts some hallucinated ones that will seem true because told with the same tone of certainty.
(A good example I saw is when asked about the life of Descartes it gets it right but in the mist of it tells that Descartes went to America which is complete nonsense).
This risk is more prominent in maths or code where the most commonly found and probable followup can be completely wrong just because a tiny thing changed in the prompt (and due to the nature of probabilities it can be missed by the algorithm). An example is the classic “A phone with its cover costs 110$ and the phone costs 100$ more than the cover. How much is the cover ?” which ChatGPT will get wrong unless you specify that you want an answer from a maths teacher in the prompt.
A second possible factor is that Home Assistant content is probably too sparse in the GPT corpus which means that GPT will try to invent probable continuations using alike-looking languages and is thus “polluted” by their API.
That is fine but wouldn’t you just deal with them and educate users to enquire about where code comes from rather than ban outright. As any tool AI has a place and a time. I think there is a more differentiating answer.
Someone posing AI generated code as their own is is fraud and needs to be addressed same as anyone using anyone else’s code pretending it’s theirs / not crediting the originator.
As I read this the morning news radio programme which is discussing the banning of chatgpt in schools/universities and concerns over cheating (which is rife anyway). Quite ironic I thought.
Nickrout,
I once taught in a school library - admittedly at the beginnign cusp of the web as we know it (remember alta vista) - anyway one of the issues became after a few years the prevalence of students to use web scrapes and microsoft word essay tools to compile the essays. It was pretty good and allowed slabs of other people’s (unattributed work) to be included.
BUT - then the issue was recognised and a tool developed so that the essay had to be submitted into an anti plagiarism checker (turnitin). This caught up to 99.9% of those seeking to ‘game’ the system.
I guess in the end what I’m trying to say is that with this AI issue. Some person way smarter than me will recognise the issue and come up a clever solution. It may be that any post to the forum goes through a similar check process (this would also stop wasting moderators time) to identify AI derived answers.
Interesting times - damn I just pricked myself again with a pin
Fair enough, but if you can’t test it I think that should be mentioned. To me, the bottom line is that you post an answer that to the best of your knowledge is helpful, that you do a reasonable (proportional) effort to make sure that it is, and that you indicate where you are not sure. I don’t see why there should be a difference between AI generated code and human generated code in this respect. I don’t say no to AI, I’d love it to be at a point where I can ask a computer to write automations for me instead of dealing with the slings and arrows of YAML syntax. There are many things I say no to that are made easier with AI, like trolling forums with elaborate bullshit answers. I do object to that but not because of the AI.
Yes I do, and it was awful. The internet (and particularly trying to make linux work in the early days) got so much better with google, before it became a tracking/advertising company.
I foresee a loop with AI being trained on AI answers, and pretty soon the bullshit answers will become canon (albeit wrong).
I have always “rejected” or kept my self from being interested in AI / Machine-Learning etc (Google assistant included) , as these things still “fails” to deliver … and in the end " There is no such thing as AI " , it’s no better than what it’s been “feeded” with, by whom/what -ever
Ideal for various companies( Support-departments ), public-services etc. etc. where they want one to “believe/think” you actually are “talking/chatting” with a real representative person … then the companies/puclic-service dep. can always claim that “there must have been some kind of misunderstanding” …
Even more “Funny” (Scaring) is when you integrate it with Alexa/Google Assistant, etc. … I guess that’s the reality which future generations have to cope with… Daily conversations with Machines, blaming the machine, shouting at a machine, … killing the machine !
I love it! Already we get the “computer’s fault” excuse from all kinds of businesses. I remember when that excuse first started being used. I’d have to remind them that someone programmed the computer to do whatever dumb, crooked or dishonest thing they were trying to deny responsibility for.
But with AI, we’ll all be able to justifiably blame the computer for making every stupid decision or mistake. It’ll be great. We’re all off the hook!
because it’s not trained specifically for code, its a language model, meaning its designed to mimic human speech/text it was trained with, which includes code. It seems they trained it with a dataset of basically the entire internet, but they didnt train it with every detail. Like i imagine it was trained with data from documentation sites as well as sites like stack exchange, but it wasnt trained with everything.
Home assistant also isnt very popular relatively speaking, and there really isnt much data out there to train an AI model like ChatGPT with, even if it crawled every section of this forum.
too be clear, GPT3 is not the same as ChatGPT. GPT3 is mainly for text generation, its not a chatbot but can be used to build one. ChatGPT is specifically meant for back and forth conversations, and it uses GPT3.5, which is supposed to be an extremely limited version of the upcoming GPT4.0
