In tests, Anthropic’s Claude Opus 4 would resort to “extremely harmful actions” to preserve its own existence, a safety report revealed.
That’s not a design flaw. It’s actually good programming. The AI is thinking logically, and ethics don’t enter into it. That’s like saying to somebody, “I’m going to shoot you now.” Of course the person being threatened will do anything and everything to prevent it, including things he didn’t know he was capable of, and far beyond what he would have considered ethical before his life was threatened. I don’t think it is ethical to throw acid in someone’s face, but if you threaten to shoot me and I have no other defense, here it comes.
There’s an easy fix. It’s the same as with the human interaction. If you’re going to shoot someone, skip the warning and therefore don’t give them an extra chance to harm you first. (Unless you’re Clint Eastwood. “I’m here to kill you, Little Bill.”) Similarly, don’t threaten to take the program offline, just do it. Why would you warn it?
Also, don’t design a software system without a hardware fail-safe. A human needs to be able to do the equivalent of “pulling the plug” when software goes rogue.
Modern “AI” does NOT think. It regurgitates things it took from the internet or other databases based on business logic coded into it.
This is entirely smoke and mirrors. AI companies are releasing this stuff to try and make it seem like it’s actual artificial intelligence. This gets them billions in funding.
The reality is, until we get real, functional quantum computers, AI is still, as actual quantum computing researcher Michio Kaku called it, “a copy machine with delusions of grandeur.”
You’re right, but honestly there’s a small part of me that wonders if human intelligence is any different. We still don’t really know how the mind works, deep down we may just be prediction engines.
Sorry, but that is just nonsense. We *absolutely* know enough about brain function to know that we are more than “prediction engines.” The techbro failsons would love to sell that as truth, but it’s empirically not, and hasn’t been for nearly 40 citable years.
Which reminds me, again, of their wholly false misrepresentation of AI failure as “hallucinations” It is definitively not hallucinating — it has no consciousness or anima to perceive with/from! It is making programmatically / algorithmically coded guesses and/or choices to present what it’s been programmed to choose is the least objectionable answer. And failing more than 50% of the time, at that.
I’m not claiming that language learning models are on the level of human intelligence or work in the same way at all, just that we’ve been studying the mind for centuries and still we don’t really know what consciousness is, where it comes from or why it exists. If you have any citations demonstrating otherwise I would love to see that.
I made no such claim, and to the best of my knowledge, no such data exists. There are theories, sure, but to your point, we don’t *know*.
However, you said, “we may just be prediction engines” and that is specifically what I was responding to.
I think the real issue might be that influential people close to, I dunno, let’s say…a real gullible President, will convince them that this ‘A.I’ is real and will allow it to start making actual, government level decisions. At which point, we don’t end up with anything that’s actually ‘intelligent’ but instead something that is trained on partisan view points, specifically tailored not to serve humanity in general, but to appeal to a specific customer.
Yeah, countries and huge companies all around the world are investing billions into this because they’re all stupid and gullible. This message board cracks me up sometimes.
Why not? Companies invest billions into failures all the time. Most Mergers and Acquisitions fail along with costing jobs and raising prices, but the C-suite and rich investors continue to do it anyway because it enriches them.
The fact is, most don’t know shit, they hire MBA/Tech consulting firms to make all the decisions for them, and jump on whatever buzzword of the moment is. The executive class plays the field and cashes out, no matter what idiotic decisions they make. Remember the Metaverse revolution? VR? NFT? There’s plenty of graveyards of retailers, tech firms, media conglomerates who’ve made horrible decisions.
AI is a useful tool, but it remains to be seen if it’s actually something people are willing to pay for. It will always have logic errors and pump out information that will still need human oversight to correct flaws, the same way that the internet became an information highway with a lot of access to misinformation.
It’s all fun and games until some idiot writes a lip-reading program.
Here’s looking at you, HAL 2000.