I’m not an AI techie and I’m not well educated on LLMs/AI in general. The issue I’m having is the dichotomy between how people talk about it. I’ve heard completely opposite ends of the spectrum. From AI is quickly becoming advanced enough to threaten human society, all the way to its a glorified dictionary/search engine that can’t really do much of anything. I know Lemmy has a steep anti-AI slant. But, can someone talk about this in a way I can wrap my head around why the divide seems so big?
What makes it unsafe?
Wanted to expand on what others have said about the randomness of LLMs being unsafe, because that response angers those who think “alignment” is the solution. Alignment is additional training to constrain the randomness to not do bad things like crime. But even the staff that works on aligning this software have their doubts about whether this process will make the software safe (this article), and even their own researchers are coming out and saying “oops, it did a crime” but in a way where it hypes the ability of the software (see other articles about it hacking). Then there is the recent news of OpenAi’s recent model that drops chain of thought observability (a feature the let’s you read what the model is “thinking”) - a very useful tool in reviewing whether the model is engaging in unsafe activities. They sacrificed observability for benchmark performance. This tells us the people developing this software don’t think it will ever be safe, but they are trying to spin it as “because it can out smart us” when in reality it is simply unpredictable and they are making it less predictable to make it more “capable”. If we started holding these companies accountable for these rogue hacks and psychotic episodes then we would start to see serious thought put towards these models being safe or at the very least guidelines around when not to use the models.
To me, what makes LLMs unsafe is they are nondeteramistic.
If I remember correctly; Nondetermanistic means the LLM can, could, and will have different answers/outcomes for the same question.
The lack of consistency potentially leads to negative outcomes?
Well, if the output varies in a way you can’t predict, then even if one of the inputs is “don’t do anything negative” it might not always follow it exactly.
Makes sense
Yeah, because it’s all statistical. You can test this yourself. Go to Google and search for something. Then search for that same exact thing again. Gemini will give you a different response each time.
That’s always true. I just did two searches and got the exact same AI result.
I’m guessing google is caching AI results, it would be insane not to.
I did at one time have a Gemini window open that was explaining how to solve a technical problem. It was pretty adamant the solution it was proposing would work, but it wasn’t working. I gave up on AI, and decided to just google for old fashioned web forums about the issue. Of course the AI result comes up first, and it said what I was trying to do wasn’t going to work and gave the reason why. So yeah one instance of AI will say the opposite thing from the other.
BTW I eventually solved the problem by taking a different approach which involved using some code someone posted to stack overflow a few years ago. So AI was just a complete waste of time on that problem.
Maybe Google is insane. The last time I tested it, I got a different answer the second time.
You don’t know? … Wow. How about the error rate? Let’s start there.
No, as previously stated I am not educated on the subject almost at all. Why I framed the post as a question.
For a start you might want to read this interview with Emily M. Bender who is a well-known critic and has much to say about what people need to understand when using these systems that are incorrectly labelled “AI”:
https://spectrum.ieee.org/stochastic-parrot
How ironic that she is called Bender
Awesome. This is the exact type of thkng I was looking for. Will read soon. Thanks!