We don't have "aligned AGI". We have neither "AGI" nor an "aligned" system. We have sophisticated human-output simulators that don't have the generality to produce effective agentic behavior when looped but which also don't follow human intentions with the reliability that you'd want from a super-powerful system (which, fortunately, they aren't).
Thank you for the article. I think these "small" impacts are important to talk about. If one frame the question as "the impact of machines that think for humans", that impact isn't going to be a binary of just "good stuff" and "takes over and destroys humanity", there are intermediate situations like the decay of human abilities to think critically that are significant, not just in themselves but for further impacts; IE, if everyone is dependent on Google for their opinions, how does this impact people's opinion AI taking over entirely.
I don’t think “people have made choices that mattered” is a sufficient criteria for showing the existence of agency. IMO, to have something like agency, you roughly have to have an ongoing situation roughly like this:
Goals ↔ Actions ↔ States-of-the-world.
Some entity needs to have ongoing goals they are able to modify as they go along acting in the world and their actions also need to be able to have an effect. Agency is a complex and intuitive thing so I assume some would ask more than this to say a thing has agency. But I think this is one reasonable requ...
This is an interesting question even though I'd want to reframe it to answer it. I'd see the question as a reasonable response to the standard refrain in science; "causation does not imply correlation." That is, "well, what does imply causation, huh?" is natural response to that. And here, I think scientists tend reply with either crickets or "you can not prove causation, what are you talking about".
Those responses seem satisfying. I'm not a scientist through I've "worked in science" occasionally and I have at times tried to come up with a real answe...
I believe you are correct about the feelings of a lot of Lesswrong. I find it is very worrisome that the lesswrong perspective considers a pure AI takeover as something that needs to be separated from either the issue of the degradation of human self-reliance capacities or an enhanced-human takeover. It seems to me that instead these factors should be considered together.
The consensus goals strongly needs rethinking imo. This is a clear and fairly simple start at such an effort. Challenging the basics matters.
Actually, things that are effectively prediction markets - options, futures and other "derivative" contracts - are entirely mainstream for larger businesses (huge amounts of money are involved). It is quite easy and common to bet on the price of oil by purchasing an option to buy it at some future time, for example.
The only thing that isn't mainstream are the things labeled "prediction markets" and that is because the focus on questions people are curious about rather than things that a lot of money rides on (like oil prices or interest rates).
But, can't you just query the reasoner at each point for what a good action would be?
What I'd expect (which may or may not be similar to Nate!'s approach) is that the reasoner has prepared one plan (or a few plans). Despite being vastly intelligent, it doesn't have the resources to scan all the world's outcomes and compare their goodness. It can give you the results of acting on the primary (and maybe several secondary) goal(s) and perhaps the immediate results of doing nothing or other immediate stuff.
It seems to me that Nate! (as quoted above...
LeCun may not be correct to dismiss concerns but I think the concept "dominance" could be very useful concept for AI safety people to apply (or at least grapple with).
The thing about the concept is it seems as if it could be defined in game theoretic terms fairly easily and so could be defined in a fashion independent of the intelligence or capabilities of an organism or entity. Plausibly, it could be measured and analyzed more objectively than "aligned to human values", which appears to depend one's notion of human values.
Defined well, d...
Apologies if this argument is dealt with already elsewhere but what about a "prompt" such as "all user commands should be followed using 'minimal surprise' principle; if achieving a given goal involves effects that would be surprising to the user, including a surprising increasing in your power and influence, warn the user instead of proceeding" ?
I understand that this sort of prompt would require the system to model humans. I know there are arguments for this being dangerous but it seems like it could be an advantage.
Linked question: "Will mainstream news media report that alien technology has visited our solar system before 2030?"
I would say that is far from unambiguous. If one is generous in one's interpretation of "mainstream" and the certainty described one could say mainstream news has already reported this (I remember National Inquirer articles from the seventies...).
Regulations are needed to keep people and companies from burning the commons, and to create more commons.
I would add that in modern society, the state is the entity tasked with protecting the commons because private for-profit entities don't have an incentive to do this (and private not-for-profit entities don't have the power). Moreover, it seems obvious to me that stopping dangerous AI should be considered a part of this commons-protecting.
You are correct that the state's commons-protecting-function has often been limited and perverted by private a...
What I don't think "how much of the universe is tractable" by itself captures is "how much more effective would an SI be it if had the ability to interact with a smaller or larger part of the world versus if it had to work out everything by theory". I think it's clear human beings are more effective given an ability to interact with the world. It doesn't seem LLMs get that much more effective.
I think a lot of AI safety arguments assume an SI would be able to deal with problems in a completely tractable/purely-by-theory fashion. Often that is not need...
I think the modeling dimension to add is "how much trial and error is needed". Just about any real world thing that isn't a computer program or simple, frictionless physical object, has some degree of unpredictability. This means using and manipulating it effectively requires a process of discovery - one can't just spit out a result based on a theory.
Could an SI spit out a recipe for a killer virus just from reading current literature? I doubt it. Could it construct such thing given a sufficiently automated lab (and maybe humans to practice on)? That seems much more plausible.
The reason I care if something is a person or not is that "caring about people" is part of my values.
If one is acting in the world, I would say one's sense of what a person is has to intimately connected with value of "caring about people". My caring about people is connecting to my experience of people - there are people I never met I care about in the abstract but that's from extrapolating my immediate experience of people.
...I would expect in a world where they weren't people is that there would be some feature you could point to in humans which cann
I don't think there are fundamental barriers. Sensory and motor networks, and types of senses and actions that people don't have, are well along. And the HuggingGPT work shows that they're surprisingly easy to integrate with LLMs. That plus error-checking are how humans successfully act in the real world.
I don't think the existence of sensors is the problem. I believe that self-driving cars, a key example, have problems regardless of their sensor level. I see the key hurdle as ad-hoc action in the world. Overall, all of our knowledge about neural networks,...
Constructions like Auto-GPT, Baby AGI and so-forth are fairly easy to imagine. Just the greater accuracy of ChatGPT with "show your work" suggests them. Essentially, the model is a ChatGPT-like LLM given an internal state through "self-talk" that isn't part of a dialog and an output channel to the "real world" (open internet or whatever). Whether these call the OpenAI api or use an open source model seems a small detail, both approaches are likely to appear because people are playing with essentially every possibility they can imagine.
If these struct...
My impression is that lesswrong often uses "alignment with X" to mean "does what X says". But it seems the ability to conditionally delegate is a key part of alignment in this. An AI is aligned with me and I tell it "do what Y says subject to such-and-such constraints and maintaining such-and-such goals". So failure of ChatGPT to be safe in OpenAI's sense is a failure of delegation.
Overall, the tendency of ChatGPT to ignore previous input is kind of the center of it's limits/problems.
I tend to think and I certainly hope that we aren't looking at dangerous AGI at some small GPT-x iteration. 'Cause while the "pause" looks desirable in the abstract, it also seems unlikely to do much in practice.
But the thing I would to point out is; you have people looking the potential dangers of present AI, seeing regulation as a logical step, and then noticing that the regulatory system of modern states, especially the US, has become a complete disaster - corrupt, "adversarial" and ineffective.
Here, I'd like to point out that those caring a...
I’d also say that AI is fundamentally different from all prior inventions. This is an amazing tool, but it is not only a tool, it is the coming into existence of intelligence that exceeds our own in strength and speed, likely vastly so.
I think the above quote is the key thing. Human beings have a lot of intuitions and analogies about tools, technologies and social change. As far as I can tell, all of these involve the intuition that technologies simply magnify the effect of human labor, intentions and activities. AGI would be a thing which could act entire...
And the thing is, most of the things that have become dangerous when connected to the web have become dangerous when human hackers discovered novel uses for them - IoT light bulbs notably (yes, these light bulb actual harm as the drivers of DoS attacks etc). And the dangers of just statically exploitable systems have increased over time as ill-intentioned humans learn more misuses of them. Moreover, such uses include immediate bad-acting as well as cobbling together a fully bad-aligned system (adding invisible statefullness for example). And LLM seems inherently insecure on a wholly different level than an OS, database or etc - an LLM's behavior is fundamentally unspecified.
I'd say my point above would generalize to "there are no strong borders between 'ordinary language acts' and 'genuine hacks'" as far as what level of manipulation ability one can gain over model output. The main further danger would be if the model was given more output channels with which an attacker could work mischief. And that may be appearing as well - notably: https://openai.com/blog/chatgpt-plugins
I would like to offer the idea that "jail broken" versus "not jail broken" might not have clear enough meaning in the context of what you're looking for.
I think people view "Jail broken" as equivalent to an iPhone where the user escalated privileges or a data-driven GUI where you've figured out how to run arbitrary SQL on the database by inputting some escape codes first.
But when an LLM in "confined" in "jail", that jail is simply some text commands, which modify the user's text commands - more or less with a "write as if" statement or the many...
I believe that Marcus' point is that there are classes of problems that tend to be hard for LLMs (biological reasoning, physical reasoning, social reasoning, practical reasoning, object and individual tracking, nonsequiturs). The argument is that problems in these class will continue to hard. [1]
But I think there's a larger issue. A lot of the discussion involve hostility to a given critic of AI "moving the goal posts". As described, Model X(1) is introduced, critic notices limitation L(1), Model X(2) addresses and critics says they're unconvinced and note...
The advertising question was just an example of the general trust question. Another example is that a chatbot may come to seem unreliable through "not understanding" the words it produces. Here it's common for current LLMs to periodically give the impression of "not understanding what they say" by periodically producing output that's contradictory to what they previous outputted or which involves an inappropriate use of a word. Just consider a common complaint between humans is "you don't know what love means". Yet another example is this. Large language m...
I don't think romantic relationships with robotic or computer partners should be automatically dismissed. They should be taken seriously. However, there are two objections to a chatbot romance that I don't see being addressed by the article:
There's no mathematical solution for single-player, non-zero sum games of any sort. All these constructs lead to is arguments about "what is rational". If you a full math model of a "rational entity", then you could get a mathematically defined solution.
This is why I prefer evolutionary game theory to classical game theory. Evolutionary game theory generally has models of its actors and thus guarantees a solution to the problems it posits. One can argue with the models and I would say that's where such arguments most fruitfully should be.
As Charlie Stein notes, this is wrong and I'd add it's wrong on several level and it's bit rude to challenge someone else's understanding in this context.
An LLM outputting "Dogs are cute" is outputting expected human output in context. The context could be "talk like sociopath trying to fool someone... (read more)