7 ms·

AI alignment people are the group who are capable of bringing AGI to life, but want to do it responsibly. AI ethics people are the group who are not capable of

• nradov 4y ago
  There is no evidence that any group is capable of bringing AGI to life.
  • swyx 4y ago
    no evidence? none? Google Lambda and Bing/Sydney are not at least incrementally closer to evidence?
    • bonzini 4y ago
      That's not AGI. An AGI would be able to tell you they don't know things, ask for clarification as to why you think they're wrong, etc. ChatGPT just spouts out a different wild guess when you tell it it's wrong, but doesn't learn from its mistake—not even within a single chat. Sydney just goes full psychopath.
      • EVa5I7bHFq9mnYK 4y ago
        That's not AGI. AGI would be able to write its own code, and improve it exponentially.
        • nradov 4y ago
          That's a big assumption. No one has built an AGI yet so we don't know whether an AGI would be capable of improving it's own code exponentially. Any exponential growth function quickly runs into hard physical limits.
        • AstralStorm 4y ago
          Not necessarily, much like every human does not know pharmacology, genetics, robotics or other specialties that could be used for self improvement. Including social engineering and programming. What you're describing is a particular kind of superintelligence, recursively improving kind. Between that and general intelligence is another gap.
          • EVa5I7bHFq9mnYK 4y ago
            Humanity as a whole can improve itself indefinitely, there is little doubt about that. For robots, there is no distinction between a single robot and robotuity, as software can be replicated cheaply and quickly.
        • CuriouslyC 4y ago
          We're generally intelligent but we can't consciously rewire our brains.
          • pnut 4y ago
            We sort of can, through meditation, habit, study, and chemical alteration - obviously within constraints, barring future technological enhancements.
          • EVa5I7bHFq9mnYK 4y ago
            We can, this process is called "learning".
            • CuriouslyC 4y ago
              Practice and study are conscious behaviors. Learning is not.
        • hgsgm 4y ago
          That's Super Intelligent AGI, not merely AGI.
        • mcguire 4y ago
          None of the existing models of what we call 'general intelligence' are capable of improving themselves exponentially.
    • Arkhaine_kupo 4y ago
      > Google Lambda and Bing/Sydney are not at least incrementally closer to evidence? No, in the same way that boiling water in a kettle is not incrementally closer to inventing the fuel for a rocket ship. You can, in hidsight do a linear connection of water boiling -> steam engine -> fuel engine -> aeronautics -> space travel But the first time someone boiled water, if you said that got you closer to being on the moon I think you would be mad. There is a non trivial chance that all our current work in AI is functionally useless in terms of creating AGI, it might take a complete breakthrough in ML, tech not invented yet, or it might not be possible at all.
      • SturgeonsLaw 4y ago
        Except, in your example, boiling water did get us closer to the moon :)
        • Arkhaine_kupo 4y ago
          Yep, and there was a much mre promissing science on checking Ether, and the magical qualities of gravity defying physics. Who is to say that our current ML models are not Ether studies and utterly useless. If we crack AGI in 10 years, maybe our current models are the fuel engine. But if they are not even close, we might be boiling water. Or perhaps be wrong all together, thats why I said your observation of us getting closer thanks to those proyects is not right, at least not yet.
      • CuriouslyC 4y ago
        We have lots of examples of generally intelligent systems, they're common and they evolve by themselves, so the idea that it might not be possible at all just seems ridiculous. Large language models might not be capable of general intelligence by themselves, but a model network that includes a LLM as part of a generalized game playing model would certainly have the potential to get there.
        • RandomLensman 4y ago
          I think it is fair to say that "not at all possible" is perhaps too strong. On the potential of certain approaches I am less certain we have or know what we need. But we also have examples of things all around us we cannot make ourselves (e.g., a simple animal or lifeform from scratch).
  • baq 4y ago
    AGI happens slowly then suddenly all at once, to paraphrase a saying. There is no other way it can happen.
    • darkwater 4y ago
      Why not? I'd be interested in hearing the reasons.
      • turtleyacht 4y ago
        It would have to be reproducible: one second of a snippet of a video clip of something; the next, the thing happening. You rewind it and play again. Scrub through and try to determine a different branch. At one timestamp, separation, peace, unknowing. In the next, flames and smoke and a flashing blur. But no--the state of the world in that second versus the next could not have been predicted. It just happened, just so, and it would take longer to calculate the determinism than to adjust to the new.
      • mcguire 4y ago
        Singularity-related religious beliefs.
      • baq 4y ago
        1) progress doesn't happen linearly, it's very coarse stepwise function (remember alphago?) 2) acceptance of an AGI actually being created is also a non-linear function - look at current definitely-not-AGI LLMs: some people think it's close, some people say that absolutely not, these are big matrices of numbers regurgitating words in plausible combinations and that's all - and these people won't change their minds easily one way or the other, thus declaring delivering of AGI will be very controversial even if we could agree on the definition.
  • 6nf 4y ago
    Ok so what’s your point? We can just ignore it? Or what.
    • RandomLensman 4y ago
      AGI worries we probably could ignore until AGI exists but there is now an industry around them which has its own momentum.
      • Jcowell 4y ago
        This seems dumb to me. It’s like if the Japanese had the clairvoyant ability to know that nuclear bombs were being built and used and the immediate future and said “We can ignore it until it exists)”. No. You leave Hiroshima and Nagasaki now and not when the freaking bombs are above your head.
        • nradov 4y ago
          You are certainly welcome to leave Earth right now if you like. But the notion that we should waste time worrying about some hypothetical future AGI is just dumb.
        • RandomLensman 4y ago
          We don't have the clairvoyant ability to know that AGI will exist in any future near enough to worry about. It might or it might not - and even if it exists we could get its nature wrong and worry about the wrong things.
          • flangola7 4y ago
            This isn't congruent with the AI research community consensus. There's debate over many points, but that we will develop AGI is seen as a forgone conclusion almost universally. The most pessimistic place it closer to 2100, others say 5 years, most say <25 years. But nobody is saying it won't happen at all, or even that it won't happen beyond the lifetimes of anyone currently living. At best we would be kicking the can down the road and making it out children's or grandchildren's problem. Notably the people closest to the fire, those advancing SOTA AI capabilities and those deepest into alignment/existential risk research, both say it is coming this decade. They predict very different outcomes (crudely summarized: utopia vs extinction), but they both predict the event horizon is on our doorstep. You're right that we might get it wrong, but getting things wrong with a new technology really always means negative outcomes. When you're figuring out rocket science and you realize you worried about the wrong things your rocket doesn't somehow still succeed and make it to space, it blows up in a fireball. It might not blow up because of the specific thing you were paying attention to, but it still blew up.
• dragonwriter 4y ago
  > AI alignment people are the group who are capable of bringing AGI to life, but want to do it responsibly. AI alignment people are the group who believe “responsibly” means, centrally, bringing AGI to existence as fast as possible, and, critically, under the tight, opaque, authoritarian control of a narrow group of people whose ideology is, well, exemplified by figures like Sam Bankman-Fried, and doing everything possible to advance that goal, including fostering broad acceptance of sub-AGI AI under the same tight control for other purposes without concern for adverse social impact as a means of advancing and funding progress toward AGI. They believe that the ends (AGI) inherently justifies any means taken in its pursuit. AI ethics people are the group who believes “responsibly”, with regard to AI, isn’t restricted to pursuit of AGI (which is maybe a nice to have, but a non-essential goal), and means openness, transparency, and resolving bias and harm issues as rapidly as possible when, and ideally before, adopting AI – including AI far short of AGI – in roles with social impact (and most critically, resolving those for roles with the greatest social impact.) They believe that AI is, to the extent it at any level is desirable, is a means to improving concrete, present, immediate conditions, not an ends to be pursued in itself.
  • circuit10 4y ago
    > AI alignment people are the group who believe “responsibly” means, centrally, bringing AGI to existence as fast as possible From what I’ve seen it’s the exact opposite? They want AI development to slow down so we can get it right and make it safe. There are people who are trying to speed it up but that seems like the opposite of alignment to me
    • dragonwriter 4y ago
      > From what I’ve seen it’s the exact opposite? They want AI development to slow down so we can get it right and make it safe. The only way that the alignment faction wants to slow things down is in that they want to maintain tight control of models and gated access to assure “proper” use. This does have an effect of slowing certain kinds of development, as broader access to build around models without centralized controls fosters some aspects of development (the “Stable Diffusion moment”) but in ideal terms, the alignment faction is about fast-but-narrowly-controlled development. The ethics faction is more about slowing things down – though more about adoption, particularly in sensitive uses, than development – but is more oriented toward openness/transparency.
• visarga 4y ago
  > but want to make sure that those who can do it responsibly or not at all. Mostly not at all. They write papers like "Can Language Models Be Too Big?" I say then: can people be too afraid of the unknown?
• mcguire 4y ago
  AI alignment people are dedicated to fending off the gigantic PR disasters like Tay (https://en.wikipedia.org/wiki/Tay_(bot) https://en.wikipedia.org/wiki/Tay_(bot)) in order to keep their funding. Whether either is capable of "bringing AGI to life" remains to be seen, but seems very doubtful.