3 ms·
While I reckon there's at least 5 (and probably 10+) years before it's even possible for a sufficiently independent AI to start literally paperclipping us (most
by ben_w 4y ago
While I reckon there's at least 5 (and probably 10+) years before it's even possible for a sufficiently independent AI to start literally paperclipping us (mostly because of the speed/power needs of good image processing AI), the people who are trying to solve that don't know if we can actually do so in even double that time frame.
But the paperclip scenario was only ever an illustration that the AI doesn't need to hate us to destroy us, merely to want to use our atoms for something incompatible with our own desires.
For disasters to happen, there are many options which may come sooner, varying from not paying attention to what the AI is doing when it makes mistakes — the default, as we can imagine even today someone putting a Boston Dynamics Spot to work in a virology lab and it walking out with a sample because it doesn't know any better — to the malicious, for example one idiot running a more competent version of ChaosGPT: https://youtu.be/g7YJIpkk7KM https://youtu.be/g7YJIpkk7KM
Those other, lesser, disasters are sometimes called "fire alarms for AI".
- jiggawatts 4y agoThe cynic in me says that the #1 AI-related risk by far is that corporations will immediately apply it to the task of listening in on everything we say and write. This all reminds me of "A Fire Upon the Deep" by Vernor Vinge, where the Slavers enslave humans by essentially constantly monitoring them for misbehaviour. If I remember correctly, a plot element is that they use yet more slaves to keep the bulk of the slaves in line. With ChatGPT, you no longer need to do "sci fi things" like turning squishy meat brains into cyborg slavedrivers. You can simply ask the AI to do it, and it will. It will do it 24/7, without pity, or remorse. It will do vigilantly and faithfully. It will be a checkbox in Office 365 that any psychotic manager can turn on. And I guarantee they will turn it on. All of them.
- ben_w 4y agoI've not yet had time to read/listen to A Fire Upon the Deep yet, only its sequel A Deepness in the Sky, but that certainly sounds like one of the many bad endings we might get. That said: > With ChatGPT, you no longer need to do "sci fi things" like turning squishy meat brains into cyborg slavedrivers. You can simply ask the AI to do it, and it will. It will do it 24/7, without pity, or remorse. It will do vigilantly and faithfully. ChatGPT has all the vigilance and faithfulness of my late mother's mid-stage Alzheimer's, so if someone tried this today it would rapidly result in everyone getting million-dollar-no-work-required contracts the company isn't allowed to break. That LLMs can be socially engineered just by asking nicely is a pleasant surprise twist in AI alignment, but unfortunately I think it illustrates how primitive is our current state of the art AI safety and alignment.
- ChatGTP 4y agoWhat I think people who work in tech fail to realise is that non-tech people are learning about these dangers and I don’t think they’re impressed or care if you have a computer to play with or robots playing soccer. They’re worried about their families safety and their (which used to be important, you know children? Other biological life etc?) livelihoods. To a regular, nontechnical person these “talking computers causing damage” sounded like some far off nerd revenge dreams. But people know maybe they’re not just dreams, maybe they are realities. My honest opinion, and I don’t care what people think about this, shit is going to get wild regarding surveillance and regulation of computers, networks, software and the internet. I mean what else is the alternative if these systems grow more powerful ? I mean what are we to do ? Just let people make tsar bombs in their garage and detonate them as they see fit? Have an autonomous AI agent take down a lower grid and trigger a war for fun ? I really don’t understand how this is supposed to work going forwards ?
- ben_w 4y ago> What I think people who work in tech fail to realise is that non-tech people are learning about these dangers and I don’t think they’re impressed or care if you have a computer to play with or robots playing soccer. Doubtless. That said, when it comes to AI risk, normal people have no reason to know the boundaries between reality and sci-fi for the same reason that I, as a non-lawyer, have no reason to know the boundaries between law and Hollywood tropes. By extension, this means that I expect the general public to be aware of countless fictional examples of AI takeover scenarios, some where the AI was given a goal and blind logic made it bad for everyone (Days of Future Past, Colossus the Forbin Project), some where the AI is actually evil (Terminator, the character Lore in TNG/PIC), and some where the AI is a tool used by malevolent humans (many Dark Mirror episodes, some TOS), and some where it mixes several of the above (Westworld TV series). > I really don’t understand how this is supposed to work going forwards ? Does anyone?
- ChatGTP 4y agoWell, by going forwards, I mean practically by the way :) So here is where I don't even know then why it's being pursued so aggressively, especially in a business context. There is absolutely no way that Google or Microsoft can just put an ultra-intelligent "AGI" on the internet for hire with a credit card, to start with. It would likely be an existential threat to their own business, people, and or civilization. So yeah, even for a business like Microsoft, there is a limit to how good you can let this thing get before it actually destroys their own business or worse. Then, how would it work from a national security standpoint? Would it not just immediately provoke a war if America got an AGI first and China finds out? Even if someone started leaking rumors the Pentagon has an AI war general with an IQ of 900, that could possible make things escalate immediately?