Telling me is that you don't know what these models are. If you do, then you're cherry-picking the data. Yes, AI will hallucinate. But this is not how you measure that. This is like saying ”Shoes are bad because 95% of them either don't fit or don't last," without accounting for size or use case.
I'm not trying to defend AI. I have plenty of issues with it. But I do think it's important to be accurate in an assessment. For example, AI models are often specialized. The general AI models are the worst (like ChatGPT). But there are non-general models which have extremely high accuracy.
I get that you're not defending it, but if there are models that hallucinate 15% of the time and that's the low end and there are models that hallucinate 80% of the time then that's the spread, I dunno how else you want to cut it.
Because it depends on what you're asking it. That's the reason it's about knowing how to use it. For example, some models don't make up anything and simply do consolidated searches for you. It's still on you to validate the results. I'm forced to use it often for research so I learned how.
Another way to look at it is to ask it information that's common. All of them should have no hallucinations for basic things like, "is the planet flat?" But there's generalized searches people make that some AI has either been isolated from the truth or trained (purposefully) to be wrong about.
That's called "guardrails". And all the models you can use without your own hardware are using lots and lots of guardrails. They do this to avoid getting sued by the government. Political narratives are the biggest reason AI can get so much wrong.
You're really not explaining this in a way that makes it any better but since you're not defending ai I guess that doesn't matter (not being sarcastic)
What I'm pointing out is that it's not AI that is the problem. It's humans either corrupting the info in the models or humans not using AI correctly. Nobody is complaining about hammers, but if I give a classroom full of 1st graders hammers, we call would agree that's going to end badly, right?
That's where we disagree.
Not that we shouldn't give hammers to kids lol but that humans are the problem not AI.
What is dangerous about hammers is giving them to kids without educating them about them and how to use them. There's really no difference between that and AI. If people are just given a tool they don't know how to use, of course the results aren't going to be great. See the point?
Guns are also tools. This is why we need gun control in the US & why other countries have already figured out it’s not worth it. You can use it to feed yourself, but it’s being loaded with hollow point bullets & it’s unregulated. As long as we’re using metaphors. Tools are what they are used to do.
Gun control opinions aside, yes, this is the point. Not everyone needs every tool. But if you're going to use a tool, you should know how to use it or go in knowing you're still learning it and not blame the tool if you can't use it right.
So you just don't give a shit about material reality in this discussion? People by and large will not learn how to use this tech "the right way" and people are already causing harm to the world and themselves by using it wrong. Your argument kinda turns flaccid at that point.
Are you trying to say AI only hallucinates when people "use it wrong" or am I misunderstanding?
No. I'm trying to say that the likelihood of a hallucination greatly increases when a user doesn't know how to use the AI. Thinking like a software engineer has, in my experience, brought better results. This makes sense when you consider that software engineers are very specific in wording things.
I'm curious what the source for that is.
So you were just lying earlier when you said you weren't defending AI. Thanks for wasting our time here.
It's still not a defense of AI any more than a defense of hammers. You're conflating "AI hallucinates because users don't know what they're doing" with "AI doesn't hallucinate". My point isn't a defense but keeping the record straight.
We *could all agree...
And that's my point. Right tool for the job or you're going to botch the job.