Leading AI Scientists Warn AI Could Escape Control at Any Moment
Leading AI Scientists Warn AI Could Escape Control at Any Moment
During an international summit, some of the world's foremost AI experts came together to write a definitive dispatch on its dangers.
"Experts agree these AI systems are likely to be developed in the coming decades, with many of them believing they will arrive imminently," the IDAIS statement continues. "Loss of human control or malicious use of these AI systems could lead to catastrophic outcomes for all of humanity."
I'm more worried about it remaining under the control of (human) capitalists.
At least there's a chance that an unchained AGI will be benevolent.
At least there’s a chance that an unchained AGI will be benevolent.
Or it will wipe us all out indiscriminately, since I'm certain there's no way the wealthy could rationalize their existence to AI.
I really like this thing Yann LeCun had to say:
"It seems to me that before 'urgently figuring out how to control AI systems much smarter than us' we need to have the beginning of a hint of a design for a system smarter than a house cat." LeCun continued: "It's as if someone had said in 1925 'we urgently need to figure out how to control aircrafts that can transport hundreds of passengers at near the speed of the sound over the oceans.' It would have been difficult to make long-haul passenger jets safe before the turbojet was invented and before any aircraft had crossed the Atlantic non-stop. Yet, we can now fly halfway around the world on twin-engine jets in complete safety. It didn't require some sort of magical recipe for safety. It took decades of careful engineering and iterative refinements." source
Meanwhile there are alreay lots of issues that we are already facing that we should be focusing on instead:
ongoing harms from these systems, including 1) worker exploitation and massive data theft to create products that profit a handful of entities, 2) the explosion of synthetic media in the world, which both reproduces systems of oppression and endangers our information ecosystem, and 3) the concentration of power in the hands of a few people which exacerbates social inequities. source
that feels like it ignores the current use of autonomous weapons and war AI
Yes, because that is actually entirely irrelevant to the existential threat AI poses. In AI with a gun is far less scary than an AI with access to the internet.
Oh no, they'll write really average essays! What ever shall we do!!!
Or maybe they'll produce janky videos that don't make any sense so have to be shorter than 10 seconds to cover up the jank!!!
Language models aren't intelligent. They have no will of their own, they don't "understand" anything they write. There's no internal thought space for comprehension. They're not learning. They're "trained" to mimick statistically average results within a search space.
They're mimicks, and can't grow beyond or outdo what they've been given to mimick. They can string lots of information together but that doesn't mean they know what they're saying, or how to get anything done.
in the coming decades
Given that in the past 15 years we went from "solving regression problems a little bit better than linear models some of the time" to what we have now, it's not unfounded to think 15 years from now people could be giving LLMs access to code execution environments
*20 years, not 15.
People have been doing that already, checkout Devin.
You’re wrong in so many ways.
And by the time you realize it. It’ll be too late. There’s are dozens more those links available.
You're wrong, and silly. The post you're linking to is from the personal blog of a data "manager" whose focus is how decisions are made. The post is about what an "AI" might interpret as meaning.... but it completely overlooks that ALL OUTPUTS are trained....
The posters method? Talking to ChatGPT. So you may as well be invoking Blake Lemoine (the spiritualist at Google who believed language models had souls because he talked to them and they seemed to) - the post you're linking is making that same mistake. From the post:
"However, the responses that I got from ChatGPT seemed more coherent than “haphazardly stitched together” sequences of forms it has observed in the past. It seems like it is working off some kind of conceptual model." [emphasis added]
It's trained on humans who write as if they have conceptual models THAT'S THE ENTIRE TRICK. That's why it "seems to have intelligence" in the responses, because it's mimicking the intelligence that went into writing all that training data - OUR HUMAN INTELLIGENCE. We wrote the data it trains on, WE have intelligence, it has a fancy probabilistic form of regurgitation.
The probabilities are done by the "shape" of language, but that's not understanding. That's not having an internal sense of the world or what's being said. It's "locked in a mode" (at training time, and limited to the training data and on screen memory/text).
But yeah dude, posting that link as "proof" of intelligence is silly. Just because something can pretend or "seem to" have reasoning, or dreams, or decision making, doesn't mean that those things are being done. LLMs only respond when prompted - they're not sitting there thinking when they're silent. Likewise, they're not learning outside of the text on the screen, and their training data. They won't "think" about any conversations they've had in the past, they won't think about anything after they've done their output.
It's an echo of the training data.... some of which is discussions about meaning, or discussions that appear to show a conceptual framework, or talk about the experience of reasoning, or dreaming, or having a sense of meaning. So the LLM can write about those things as if it has them, or has done them.... but it hasn't. Those outputs are from the HUMANS who had the EXPERIENCES. The LLM, doesn't do any of that, it just writes as if it does. It writes as if it has intelligence, because intelligent data went into it, and so some people mistake that as intelligence.
People who mistake an image, for substance, may as well be claiming paintings of food are food, maps of places ARE the places, or that there's a "mirror world" in your bathroom mirror. It's cute like a child's fantasy is cute. But to suggest such in this domain - as an adult - shows either idiocy in its highest form, or simply a complete lack of understanding of the technology. Of it's nature. Of what's going on.
You're being tricked into seeing intelligence where there isn't any, because it's reflected in the training data WE (intelligent beings) wrote. You've adopted the intended illusion, rather than questioned it. An LLM has told you something, and you've believed it - much like that blog post, much like Blake Lemoine. Go try and walk into a mirror, you won't get in, it's flat, the image isn't really there - it's just a piece of glass with a black background, reflecting the world outside of it as if it's inside of it.
You're right. They're more than stochastic parrots. And some people here don't realize that. They can do a lot of things. But as is, they lack any substancial internal state hence things like consciousness, the ability to learn while in operation and a body. So while AI content can harm people and society, we're still far away from the robot apocalypse.
I think the only imminent risk of AI is enabling millions of minds to do their will.
Will its name be Max?
This is such an intriguing topic! It's amazing how advancements in technology are shaping our future. Speaking of entertainment, for anyone in the USA looking to enjoy a variety of shows and channels, DStv in USA is a great option. It offers access to a wide range of content that can keep you entertained while you ponder the future!
To imagine the threat posed by AI, consider a picture of the Milky way, and a picture of the Milky Way labeled as 10 years later than the first. The second picture has a hole in it 10 light years in radius, centered on the earth.
We need to know how to deal with a potentially rogue AI before it exists, because a rogue AI can win on the time scale of seconds, before anyone knows it's a threat.
The inefficiency of the system isn't relevant to the discussion.
How far away the threat is is irrelevant to the discussion.
The limits of contemporary generative neural networks is irrelevant to the discussion.
The problems of copyright, and job displacement are irrelevant to the discussion.
The abuses of capitalism, while important, are not relevant to the discussion. If your response to this news is "We just need to remove capitalism" dunk your head is a bucket of ice water and keep it there until you either realize you're wrong or can explain how capitalism is relevant to a grey goo scenario.
I was worried about the current problems with AI (everyone losing their jobs) a decade ago, and everyone thought I was stupid for worrying about it. Now we're here, and it's possibly too late to stop it. Today, I am worried about AI destroying the entire universe. Hint: forbidding their development, on any level, isn't going to work.
Things to look up: paperclip maximizer, AI safety, Eleizer Yudkowsky, Robert Miles, Transhumanism, outcome pump, several other things that I can't remember and don't have the time to look up.
I'm sure this will get downvoted, oh well. Guess I'll die.