I don't get this sentiment `AI is wrong! OMG WE'VE BEEN LIED TO!`
Is anyone REALLY thinking that AI will respond with 100% correct facts? Because in my mind it's the same as google/microsoft/etc search results. The response is just a summary of the "knowledge" it has based on the content it was trained on.
Also, is it realistic to expect AI to be correct about everything it spits out?
This reads the same as "what did you expect of full self driving? Of course there are technical limitations." A lot of people have an expectation that AI is "smart" based on popular culture, relatively very few realize what we have currently is word prediction not intelligence of some sort
> relatively very few realize what we have currently is word prediction not intelligence of some sort
Honestly this impresses me most. The fact that we have word prediction and it can still be very useful (in my experience). You're right people don't understand it; hell i don't either, fully. Nevertheless it is shockingly useful for being nothing more than an autocomplete over existing information.. in my view.
Even if all we could add to the service is the ability to cite sources of information - ie to validate for ourselves statements more easily - then i think it would be a huge leap in interfaces.
Being able to put a chat interface over content seems wholly unique to me. I'd love to explore the depth of complex inputs like the entire LOTR series as a chat. To discuss, learn, and grow on existing content. I don't care about intelligence; i just want the ability to mitigate confidently-incorrect. If that's even possible.
That's because these things are way more than word prediction.
Their job is 100 percent to predict the next word, but saying that hides what you have to do in order to predict the next word
In a factual database? You have to repeat facts.
In a logical database? You have to learn logic.
In a database containing only good chess moves? You have to learn to make good chess moves.
Train any AI model in a strong enough way and it will pick up donation knowledge along the way. Yes, it's predicting words, but it's using that knowledge to do it.
It's by no means perfect, but do not underestimate the potential here, and remember that reductive definitions (it's just x) can make most great things seem bad or mundane.
Right, it’s (very roughly speaking) an approximation to an algorithm which maximises the probability of the entire response by factorising the distribution as a sequence of conditional distributions.
If the approximation works, and the training data is coherent, it will produce coherent responses.
The surprising thing to me is that this does work so much of the time.
Self driving cars are supposed to drive themselves. That's their entire goal.
LLMs are not supposed to answer knowledge questions. They are not knowledge bases. The fact that everybody, including their developers insist on using them that way is just absurd. Are we in a Kafka novel?
That's the equivalent to complaining that the self driving cars you rented didn't deploy your marketing message effectively, so their AI must be bonkers.
I'm not saying its not impressive or doesn't contain logic and facts and knowledge, i'm saying people generally understand the word intelligence to mean one thing, and this doesn't meet that understanding.
AI doesn't even pretend to get anything 'right'. As far as I understand it produces statistically common phrases related to the subject words.
I asked for a chemical bond description of caffeine. It began with the formula C8H10N4O2, then proceeded to blather about the number of bonds of this type and that. Getting absolutely everything wrong, egregiously wrong.
I believe it was just spouting chemistry-sounding verbiage from other descriptions of chemical bonds of other substances.
AI doesn't 'look things up' or 'disgorge something it's heard'. It makes shit up wholesale from fragments. Always spouting it out with complete confidence.
Microsoft appears to believe they can market with a message extremely close to "Our AI technology will provide you with 100% accurate facts and summaries", and then apologists (like your post) come along and say "well, come on, nobody really thinks that".
No, it's not realistic to expect AI to be correct about anything, especially when that "AI" isn't built to give a fuck about correctness. But tell that to the proponents of "chatgpt everywhere".
Fair enough, I missed the line about work on misinformation.
However, AFAIK, there is no AI-field wide consensus about sensible ways to address this, and no conviction from anyone that there are any reliable techniques. So it's nice that they have "teams working" on it, but IMO that doesn't justify deployment of clearly flawed technology for this purpose.
Data from users ("It was wrong") is hard to incorporate. A scheme like the one you propose basically implies using users as literal testers, which wouldn't matter so much if this was a UI/UX question. Instead, users will be given garbage, a few of them will find out, a few of those will leave feedback. This is not a sane model for improving the behavior of a language model.
> there is no AI-field wide consensus about sensible ways to address this, and no conviction from anyone that there are any reliable techniques
Oh no, an open problem! Better just give up. Certainly don't allow anyone in the public to be involved in finding a solution. Much better to have an internal team stumbling around in the dark for years then force their product out when a different company comes along who was willing to develop in the open and has moved much faster accordingly.
> A scheme like the one you propose basically implies using users as literal testers
Users are literal testers, that's why the product is out now. Some users are being given garbage, some of them are finding out, and some of them are leaving feedback. That user flow is occurring at a measurable rate.
In the future, different models will be made. They will be given access to different "oracles" (in the computability theory sense), and these oracles will change their behavior. They will be able to do things like query {the web, wolfram alpha, python, prolog, etc.} and provide cited sources in responses. However, it's not enough to add the oracle. You must also verify the oracle improves the user's experience. This is done by comparing the measured feedback rate with/without the oracle(s).
A lot of ChatGPT responses are presented as factual, with confidence. And it's ChatGPT, not "search random results GPT". And it's AI, which has the word "intelligence" in it.
When you deploy a tool as such, you're asking for it.
Is anyone REALLY thinking that AI will respond with 100% correct facts? Because in my mind it's the same as google/microsoft/etc search results. The response is just a summary of the "knowledge" it has based on the content it was trained on.
Also, is it realistic to expect AI to be correct about everything it spits out?