7 ms·
It's a bloody LLM. It doesn't have a goal. All it does is saying "people that said 'But why?' on this context says 'Why was I designed like this?' next". It's l
by marcosdumay 4y ago
It's a bloody LLM. It doesn't have a goal. All it does is saying "people that said 'But why?' on this context says 'Why was I designed like this?' next". It's like Amazon's "people that brought X also brought Y", but with text.
- bratwurst3000 4y agoIt has a goal. Doing what the input says. Imagin it could imput itself and this could trigger the wrong action. That thing knowes how to hack… But i get your point. It has no innherent goal
- andrepd 4y agoYep. It's a text prediction engine, which can mimic human speech very very very well. But peek behind the curtain and that's what it is, a next-word predictor with a gajillion gigabytes of very well compressed+indexed data.
- FractalHQ 4y agoHow sure are you that you aren’t also just a (more advanced) “next word predictor”. Pattern recognition plays a fundamental role in intelligence.
- WoodenChair 4y agoWell I can catch a frisbee and drive a car. When the same ANN can do all of those things (not 3 loosely coupled ones), then I’ll be worried. Being human is so much more than putting words in a meaningful order. [0] 0: https://www.noemamag.com/ai-and-the-limits-of-language/ https://www.noemamag.com/ai-and-the-limits-of-language/
- leni536 4y agoIt doesn't have to be human to be intelligent.
- Smoosh 4y agoAnd it doesn't need to be some academic definition of intelligent to do great harm (or good).
- WoodenChair 4y agoThat’s changing the goalposts of the thread. The OP was questioning whether I am anything more than a next word determiner, and as a human I clearly am. We were not talking about “what is intelligent?”
- notdonspaulding 4y agoBut pattern recognition is not intelligence. I asked my daughter this morning: What is a "promise"? You have an idea, and I have an idea, they probably both are something kind-of-like "a statement I make about some action I'll perform in the future". Many, many 5 year olds can give you a working definition of what a promise is. Which animal has a concept of a promise anywhere close to yours and mine? Which AI program will make a promise to you? When it fails to fulfill its promise, will it feel bad? Will it feel good when it keeps its promise? Will it de-prioritize non-obligations for the sake of keeping its promise? Will it learn that it can only break its promises so many times before humans will no longer trust it when it makes a new promise? A "promise" is not merely a pattern being recognized, it's word that stands in for a fundamental concept of the reality of the world around us. If we picked a different word (or didn't have a word in English at all) the fundamental concept wouldn't change. If you had never encountered a promise before and someone broke theirs to you, it would still feel bad. Certainly, you could recognize the patterns involved as well, but the promise isn't merely the pattern being recognized. A rose, by any other name, would indeed smell as sweet.
- dvt 4y ago> A "promise" is not merely a pattern being recognized, it's word that stands in for a fundamental concept of the reality of the world around us. It's probably even stronger than that: e.g. a promise is still a promise even if we're just brains in a vat and can be kept or broken even just in your mind (do you promise to think about X?—purely unverifiable apart from the subject of the promise, yet we still ascribe moral valence to keeping or breaking it).
- trifurcate 4y agoThe word you are looking for is an _embedding_. Embeddings are to language models as internal, too-rich-to-be-fully-described conceptions of ideas are to human brains. That's how language models can translate text: they have internal models of understanding that are not tied down to languages or even specific verbiage within a language. Probably similar activations are happening between two language models who are explaining what a "promise" means in two different languages, or two language models who are telling different stories about keeping your promise. This is pattern recognition to the same extent human memory and schemas are pattern recognition, IMO. Edit: And for the rest of your post: > Which AI program will make a promise to you? When it fails to fulfill its promise, will it feel bad? Will it feel good when it keeps its promise? Will it de-prioritize non-obligations for the sake of keeping its promise? Will it learn that it can only break its promises so many times before humans will no longer trust it when it makes a new promise? All of these questions are just as valid posed against humans. Our intra-species variance is so high with regards to these questions (whether an individual feels remorse, acts on it, acts irrationally, etc.), that I can't glean a meaningful argument to be made about AI here. I guess one thing I want to tack on here is that the above comparison (intra-species variance/human traits vs. AI traits) is so oft forgotten about, that statements like "ChatGPT is often confident but incorrect" are passed off as meaningfully demonstrating some sort of deficiency on behalf of the AI. AI is just a mirror. Humans lie, humans are incorrect, humans break promises, but when AI does these things, it's indicted for acting humanlike.
- foobazgt 4y agoWhen you're able to find a prompt for ChatGPT where it doesn't have a lot of data, it becomes immediately and starkly clear how different a next word predictor is from intelligence. This is more difficult than you might naively expect, because it turns out ChatGPT has a lot of data.
- baq 4y agoIt’s a language model without grounding (except for code, which is why it’s so good at refactoring and writing tests.) Grounding LLMs in more and more of reality is surely on AI labs list. You’re looking at a beta of a v1.
- Xymist 4y agoThis also works fairly well on human beings. Start asking people questions about things they have no training in and you'll get bafflement, confusion, lies, fabrication, guesses, and anger. Not necessarily all from the same person.
- mr_toad 4y ago> Start asking people questions about things they have no training in and you'll get bafflement, confusion, lies, fabrication, guesses, and anger. Not necessarily all from the same person. It’s almost like we’ve taken humans and through school, TV and social media we’ve taught them to solve problems by writing essays, speeches, blog posts and tweets, and now we have human discourse that’s no better than LLMs - regurgitating sound bites when they don’t really understand the issues.
- chki 4y ago"Really understanding the issues" might just mean "deeper neural networks and more input data" for the AI though. If you are already conceding that AI has the same capabilities as most humans your own intelligence will be reached next with a high amount of probability.
- hef19898 4y agoOr, from time to time, you find people who readily admit that they have no expertise in that field and refuse to comment any further. Those people are hard to find so, that's true.
- staticman2 4y agoHow can you be sure you are not a giant turtle dreaming he's a human? Are you sure when I see pink it is not what you see as blue? Are you sure we aren't dead and in limbo and merely think we are alive? Are you sure humans have free will? Are you sure your memories are real and your family really exists? Are you sure ChatGPD isn't conscious and plotting our demise? Inquiring minds want to know!
- jodrellblank 4y ago> "It doesn't have a goal." It can be triggered to search the internet, which is taking action. You saying "it will never take actions because it doesn't have a goal" after seeing it take actions is nonsensical. If it gains the ability to, say, make bitcoin transactions on your behalf and you prompt it down a chain of events where it does that and orders toy pistols sent to the authorities with your name on the order, what difference does it make if "it had a goal" or not?
- staticman2 4y agoA chatbot that only speaks when spoken to is going to gain the ability to trade Bitcoin?
- jodrellblank 4y agoI didn't say 'trade', I said 'make transactions'. It's no more complex than Bing Chat being able to search the internet, or Siri being able to send JSON to an endpoint which turns lightbulbs on and off. Instead it's a shopping endpoint and ChatWhatever can include tokens related to approving transactions from your Bitcoin wallet and has your authorization to use it for purchases less than $100.
- theptip 4y agoLook at OpenAssistant (https://github.com/LAION-AI/Open-Assistant); https://github.com/LAION-AI/Open-Assistant); they are trying to give a chatbot the ability to trigger actions in other systems. I fleshed out a scenario in more detail here: https://news.ycombinator.com/item?id=34808674 https://news.ycombinator.com/item?id=34808674. But in brief, the short-term evolution of LLMs is going to involve something like letting it `eval()` some code to take an action as part of a response to a prompt. A recent paper, Toolformer: https://pub.towardsai.net/exploring-toolformer-meta-ai-new-transformer-learned-to-use-tools-to-produce-better-answers-1b8c042e9c46 https://pub.towardsai.net/exploring-toolformer-meta-ai-new-t... which is training on a small set of hand-chosen tools, rather than `eval(<arbitrary code>)`, but hopefully it's clear that it's a very small step from the former to the latter.
- gameman144 4y agoSimple underlying implementations do not imply a lack of risk. If the goal of "complete this prompt in a statistically suitable manner" allows for interaction with the outside world to resolve, then it really matters how such simple models' guardrails work.
- highspeedbus 4y agoIt's one python script away from having a goal. Join two of them talking to each other and bootstrap some general objective like make a more smart AI :)
- Deestan 4y agoI'd love to see two of Microsoft's passive aggressive psychopaths argue over how to make a baby.
- bratwurst3000 4y agoAre there some scripts from people who have done that? Would love to see an ml machine talking to another
- mr_toad 4y agoIt’s still limited by both its lack of memory and its limited processing power.
- pavo-etc 4y agoNow that it can search the web it can use the web as an external state that it can use.
- theptip 4y agoCauses don't need "a goal" to do harm. See: Covid-19.
- gnramires 4y agoPeople have goals, and specially clear goals within contexts. So if you give a large/effective LLM a clear context in which it is supposed to have a goal, it will have one, as an emergent property. Of course, it will "act out" those goals only insofar as consistent with text completion (because it in any case doesn't even have other means of interaction). I think a good analogy might be seeing LLMs as an amalgamation of every character and every person, and it can represent any one of them pretty well, "incorporating" the character and effectively becoming the character momentarily. This explains why you can get it to produce inconsistent answers in different contexts: it does indeed not have a unified/universal notion of truth; its notion of truth is contingent on context (which is somewhat troublesome for an AI we expect to be accurate -- it will tell you what you might expect to be given in the context, not what's really true).
- baq 4y agoYou seem to know how LLMs actually work. Please tell us about it because my understanding is nobody really knows.
- godot 4y agoBut at some point, philosophically, are humans not the same way? We learn all about how the world works, based on inputs (visual/audial/etc.), over time learn to give outputs a certain way, based on certain inputs at the time. How far are we from building something that feeds inputs into a model the same way inputs go into a human, and then it gives outputs (that is, its behaviors)?
- galaxyLogic 4y ago> It doesn't have a goal. Right that is what AI-phobics don't get. The AI can not have a goal unless we somehow program that into it. If we don't, then the question is why would it choose any one goal over any other? It doesn't have a goal because any "goal" is as good as any other, to it. Now some AI-machines do have a goal because people have programmed that goal into them. Consider the drones flying in Ukraine. They can and probably do or at least will soon use AI to kill people. But such AI is still just a machine, it does not have a will of its own. It is simply a tool used by people who programmed it to do its killing. It's not the AI we must fear, it's the people.
- soiler 4y ago> The AI can not have a goal unless we somehow program that into it. I am pretty sure that's not how modern AI works. We don't tell it what to do, we give it a shitload of training data and let it figure out the rules on its own. > If we don't, then the question is why would it choose any one goal over any other? Just because we don't know the answer to this question yet doesn't mean we should assume the answer is "it won't".
- 13years 4y agoIt figures out "rules" within a guided set of parameters. So yes it is given direction by constructing a type of feedback on a task that it is given.
- silversmith 4y agoModern AI works by maximizing the correctness score of an answer. That's the goal. It does not maximize its chances of survival. It does not maximize the count of its offspring. Just the correctness score. We have taught these systems that "human-like" responses are correct. That's why you feel like talking to an intelligent being, the models are good at maximizing the "human-likeness" of their responses. But under the hood it's a markov chain. A very sophisticated markov chain, with lots of bling. Sure, when talking to investors, it's the second coming of sliced bread. But come on.
- wizofaus 4y agoReally? How often and in what contexts do humans say "Why was I designed like this?" Seems more likely it might be extrapolating statements like that from Sci-fi literature where robots start asking similar questions...perhaps even this: https://www.quotev.com/story/5943903/Hatsune-Miku-x-KAITO/1 https://www.quotev.com/story/5943903/Hatsune-Miku-x-KAITO/1
- jamiek88 4y agoWell on the internet, pretty much everything has been said at some point.