We all have experienced the OpenAI “hallucinations”. Right now, if an OpenAI model gives you a response of which you aren’t sure, you have to turn around and Google search to verify the answer.
I do not see how it becomes a tool that can be trusted in business without the ability to provide verification at the same time it provides answers. And, the simplest way to achieve this is by listing the source link(s) along with the responses (if requested by user).
If this is possible now, I am asking how to do it. I asked ChatGPT3, but it keeps telling me I need to submit a 3rd key in my fine-tune dataset, which is not allowed. Another “hallucination”.
If anyone knows the answer to this, please educate us!
This is because OpenAI is GPT-based which generates text completions based on a large language model.
You must verify these GPT generated text completions as a human. This is true for all GPT based systems.
Regarding Google, or videos by YouTubers “suggesting Google has a GPT product more reliable or accurate” than OpenAI; I’m not holding my breath given given that there is currently no Google completion engine to test when writing code
I dunno. While I hear what you’re saying, and they haven’t shown anything publicly yet, Google doesn’t need to make a big splash? Or at least they don’t need to act too hasty. They do have a huge constantly updating index of a lot of the internet and have been doing that for a while.
I remember Blake the ex-Google guy said Lambda is actually hooked up to other system in the Google ecosystem, so… maybe they do have something?
As you said, we won’t be able to know for sure until they show us what we’ve got. The way things are moving so quickly, I’m betting we see something from them sooner rather than later.
So, there is no mechanism currently for making sure that the GPT responses return source links, even if you include them in your fine-tuning?
You must verify these GPT generated text completions as a human. This is true for all GPT based systems.
So, you need a ChatBot AND Google to be reasonably sure of the responses received? This can’t be Google OR Microsoft’s plan! Wouldn’t it be far easier to program the AI to return it’s source materials? Or have some mechanism to fine-tune with source links that could be returned in the responses?
These GPTs generate responses based on taking chunks of text and predicting the outcome.
These GPTs. are basically “auto-completion engines” They are not expert systems and they do not check references, they assemble weighted “puzzle pieces” as a language model and generate text, often “just nonsense”.
I am not sure you understand OpenAI’s “Terms and Conditions” where they state everything must be verified.
That is not how a GPT deep neural network works. It is not an expert, writing papers, it is generating text similar to your auto-completion when you type in an application with auto-completion text (often inaccurately and embarrassing you when you write email, haha)
You, @SomebodySysop, are asking a GPT engine to behave as an expert system, writing papers for you and checking all references. That is not how a GPT engine works. You want a tractor to behave as a sports car, and are (seemingly) not happy you are using a tractor and it is not performing like a sports car
You MUST check everything from GPT if you wish accuracy or technical correct information. This is a core “Terms and Condition” from OpenAI. ChatGPT is based on a large language model, and it is not an AI expert system. ChatGPT predicts text, it does not write technically accurate papers with perfectly referenced citations.
Thank you for the responses. I am not debating what this GPT can and can’t do. I am asking the question.
Now that I know the answer, I am saying that, for all it’s whizzbang-ery, ChatGPT is pretty useless for real business use, whether you are a dentist, doctor, lawyer, real estate agent, chef, baker or candlestick maker, if one of it’s chief selling points is that it is unreliable.
I mean, I’ve been watching technologies rise and fall and crash burn since the 1980s. No one really knows where this one is going to land for sure. But, as someone who would like to use and promote this technology to new users, primarily business people, having it be able to verify it’s sources – if queried to do so – is a very, very huge feature.
BTW, no one is more sold on this technology than me. I’ve been using it for the past couple of weeks to assist me in coding, and it has been a godsend. Finally, not having to wait on issue forums or wade through Google ads to find answers to difficult code problems, having my own assistant right at my fingertips. Dude! So, I’m pretty sure it has some solid uses right now as is.
I just struggle with seeing how business folk who would like to use it as a reference tool aren’t better off using Google at this point if they are going to end up having to use Google anyway.
I am a professional software developer and I use ChatGPT and the OpenAI codex to write code and it increases productivity greatly, without a doubt.
The difference is, that as a professional, I verify all the code completions and code generated.
So, I completely disagree with you @SomebodySysop that ChatGPT is “useless for real business use” as I use it daily as a tool. Developing code is “a business” and a good one at that
Verifying and confirming information is one of the key “ingredients” of a professional.
ChatGPT is not marketed as a “reference tool” (is it?) because it cannot be a final reference. A “true reference” is considered the “final authority” and it is well established in the OpenAI Ts&Cs that accuracy must be verified.
This is true of all GPT engines, BTW. These LLMs are hallucinating, autocompletion engines who reply “as an expert” but in fact, they are not “expert” anything. They must be used as a tool, a digital assistant, not as a final reference.
There are two possible approaches. 1. In your question, explicitly request that sources be provided. I have posted an example of using that approach. 2. Use GPT-3 as part of an NLP pipeline. There is documentation on how to do this as well as examples and tutorials.
It is my understanding that Microsoft is using ChatGPT4 as the engine for it’s Bing chat. And, the Chat search results in Bing contain the source materials. So, apparently, the ability to do this is already there.
If OpenAI is truly open, I am hoping they will share with us how to accomplish the same the the API.
I mean, it should be easy. Bing/Chat has to go out and search the Internet then bring back the specific pages it cites. If we are fine-tuning a model, it will already have the source citations for it’s responses because we will have already provided them.
Hi, for one, I’m not complaining at all. I have found chat GPT 3 to be extremely useful and have increased my productivity using it.
I also accept the fact that chat GPT3 is not designed to be optimized for search and functions such as returning source citations.
But, considering that chat GPT-4 appears to do just that, I am wondering if there are possibilities with chat GPT3 or will we be able to do it once chatgpt4 for is released for " open" usage?
With ChatGPT and GPT-3 currently this Chrome extension offers a working solution. It uses the DDG browser to present results of a web search to ChatGPT as part of the prompt; the result is a pretty good approximation of how WebChatGPT would perform with native web search capability.
I think GPT-4 does not provide the 'references` directly from the model. These come from the user’s search results, as I understand the news releases.
GPT-4 is advertised as a chatbot for guided searches on the internet, so you might be reading more into the capabilities for GPT-4 than what is available.
It’s premature to draw conclusions from marketing blurbs.
Yeah, so true.
I really wish it was possible to moderate the negative energy and complaiing here at times; but I guess that’s part of life.
It’s no wonder that some scientists fear future AIs will find humans an annoyance, LOL*5
I have used ChatGPT and the OpenAI API for around 6 weeks non-stop and written a lot of code around the API. I have found nothing to complain about and look forward to future releases and new products. Yeah, some error messages are vague, but that is typical for a beta product. Yeah, the system if overloaded. That’s a good thing. Developer can learn to trap network and API return calls and deal with errors.
Lots of folks are trying to get a text generating, auto-completion AI to perform as an expert system. GPT does not work as an expert system. GPT is not a rule-based AI. You must build components to complement GPT if you want expert system capabilities.
