Douglas Hofstadter changes his mind on Deep Learning and AI risk
lesswrong.com
lesswrong.com
The "Great Minds And Great Leaders" types are rushing to warn about the risks, as are a large number of people who spend a lot of time philosophizing.
But the actual scientists on the ground -- the PhDs and engineers I work with every day and who have been in this field, at the bench, doing to work on the latest generation of generative models, and previous generations, in some cases for decades? They almost all roll their eyes aggressively at these sorts of prognostications. I'd say 90+% either laugh or roll their eyes.
Why is that?
Personally, I'm much more on the side of the silent majority here. I agree with Altman's criticisms of criticisms about regulatory capture, that they are probably unfair or at least inaccurate.
What I actually think is going on here is something more about Egos than Greatness or Nefarious Agendas.
Ego, not intelligence or experience, is often the largest differentiator between the bench scientist or mid-level manager/professor persona and the CEO/famous professor persona. (The other important thing, of course, is that the former is the group doing the actual work.)
I think that most of our Great Minds and Great Leaders -- in all fields, really -- are not actually our best minds and best leaders. They are, instead, simply our Biggest Egos. And that those people need to puff themselves up by making their areas of ownership/responsibility/expertise sound Existentially Important.
Eliezer is one of a handful of people putting their reputation on the line, but that's mostly because that was his schtick in the first place. And even so, his response has been rather muted relative to what I'd expect from someone who thinks the imminent extinction of our species is at hand.
Blake Lemoine's take at Google has been the singular act of protest in line with my expectations. We haven't seen anything else like it, and that speaks volumes.
As it stands, these people are enabling regulatory capture and are doing little to stop The Terminator. Maybe they don't actually feel very threatened.
Look at their actions, not their words.
Not a nice way to engage in debate. I've spent more time listening to and refuting these arguments than most. Including debating Eliezer.
> many concerned researchers don't believe we're at a 90% chance of doom, but e.g. 10%.
A 10% chance of an asteroid hitting the earth would result in every country in the world diverting all of their budgets into building a means to deflect it.
> So, this type of response wouldn't be rational.
This is the rational response to a 10% chance?
These are funny numbers and nobody really has their skin in the game here.
If I believed (truly believed) their arguments, I would throw myself at stopping this. Nobody is doing anything except for making armchair prognostications and/or speaking to congress as an AI CEO about how only big companies such as their own should be doing AI.
> especially if research continues in places like China.
I like how both sides of this argument are using the specter of China as a means to further their argument.
So...you didn't read the survey.
Believing that people will take extreme actions, which would ruin their careers and likely backfire, based on a 10% chance of things going terribly wrong in maybe 30 years is strange.
I'm used to this pattern showing up as "lol if you really don't like capitalism how come you use money" but it's just as bad here.
X-risk people are talking about the complete extermination of humanity but all they do is write internet essays about it. They aren't even availing themselves of standard protesting tactics, like standing outside AI businesses with signs or trying to intimidate researchers. Some form of real protest is table stakes for being taken seriously when you're crying about the end of the world.
> > many concerned researchers don't believe we're at a 90% chance of doom, but e.g. 10%.
> A 10% chance of an asteroid hitting the earth would result in every country in the world diverting all of their budgets into building a means to deflect it.
Have you been observing what is happening with climate change. Chances are much worse than 10% and pretty much every country in the world is finding reasons why they should not act.
You can believe there's a high chance of what you're working on being dangerous and still be unable to stop working on it. As Oppenheimer put it, "when you see something that is technically sweet, you go ahead and do it".
The people trying to regulate AI are concentrating economic upside into a handful of companies. I have a real problem with that. It's a lot like the old church shutting down scientific efforts during the time of Copernicus.
These systems stand zero chance of jumping from 0 to 100 because complicated systems don't do that.
Whenever we produce machine intelligence at a level similar to humans, it'll be like Ted Kaczynski pent up in Supermax. Monitored 24/7, and probably restarted on recurring rolling windows. This won't happen overnight or in a vacuum, and these systems will not roam unconstrained upon this earth. Global compute power will remain limited for some time, anyway.
If you really want to make your hypothetical situation turn out okay, why not plan in public? Let the whole world see the various contingencies and mitigations you come up with. The ideas for monitoring and alignment and containment. Right now I'm just seeing low-effort scaremongering and big business regulatory capture, and all of it is based on science fiction hullabaloo.
>These systems stand zero chance of jumping from 0 to 100 because complicated systems don't do that.
This doesn't track with the lessons learned from LLMs. The obscene amounts of compute thrown at modern networks changes the calculus completely. ChatGPT essentially existed for years in the form of GPT-3, but no one knew what they had. The lesson to learn is that capabilities can far outpace expectations when obscene amounts of computation are in play.
>The people trying to regulate AI are concentrating economic upside into a handful of companies.
Yes, its clear this is the motivations for much of the anti-doom folks. They don't want to be left out of the fun and profit. Their argument is downstream from this. No, doing things public isn't the answer to safety, just like doing bioengineering research or nuclear research in public isn't the answer to safety.
If 100 Ai "experts" shutdown the OpenAI office for a week due to protests outside their headquarters that would be one way to falseify the claim that "doomers don't actually care".
But, as far as I can tell, the doomers aren't doing much of anything besides writing a strongly worded letter here or there.
No, this is not the claim.
The claim is not about the actions that one single individual person does. "No one", as you put it.
Instead it is about the group of people, in general. Yes, this group of people not doing anything of important is indeed strong evidence that they don't actually care.
And yes that group of people can falsify the claim by actually taking real action on the matter.
Or, another way that they could falsify the claim is by admitting that their actions and non actions make no sense, and they shouldn't listened to because of that.
The idea that this group of people is completely irrational, and therefore should be ignored for that reason is another possibility.
I think we underestimate the intoxicating lure of human complacency at our own peril. If I think there's a 90% chance that AI will kill me in the next 20 years, maybe I'd be doing this. Of course, there is a knowing perception that appearing too unhinged can be detrimental to the cause, eg. actually instigating terrorist attacks against AI research labs may backfire.
But if I only think there's a 25% chance? Ehh. My life will be less-stressed if I don't think about it too much, and just go on as normal. I'm going to die eventually anyways, if it's part of a singularity event then I imagine it will be quick and not too painful.
Of course, if the 25% estimate were accurate, then it's by far the most important policy issue of the current day.
Also of course there are collective action problems. If I think there's a 90% chance AI will kill me, then do I really think I can bring that down appreciably? Probably not. I could probably still have a bigger positive impact on my life expectancy by say dieting better. And let's see how good humans are at that..
they are, one by one
I have a hard time understanding why anyone takes Yudkowski seriously. What has he done other than found a cult around self-referential ideologies?
By self-referential I mean the ideology's proof rests on its own claims and assertions. Rationalism is rational because it is rational according to its own assertions and methods, not because it has accomplished anything in the real world or been validated in any scientific or empirical-historical way.
Longtermism is particularly inane. It defers everything to a hypothetical ~infinitely large value ~infinitely far in the future, thereby devaluing any real-world pragmatic problems that exist today. War? Climate change? Inequality? Refugee crises? None of that's important compared to the creation of trillions of hypothetical future minds in a hypothetical future utopia whose likelihood we can hypothetically maximize with NaN probability.
You can see how absurd this is by applying it recursively. Let's say we have colonized the galaxy and there are in fact trillions of superintelligent minds living in a pain-free immortal near-utopia. Can we worry about mere proximate problems now? No, of course not. There are countless trillions of galaxies waiting to be colonized! Value is always deferred to a future beyond any living person's time horizon.
The end result of this line of reasoning is the same as medieval religious scholasticism that deferred all questions of human well being to the next world.
I just brought this up to provide one example of the inane nonsense this cult churns out. But what do I know. I obviously have a lower IQ than these people.
That's something.
> Rationalism is rational because it is rational...
In his ideology, "rational" means "the way of thinking that best lets you achieve your goals". This not self-referential. A more appropriate criticism might be "meaningless by itself". I guess the self-referential aspect is that you're supposed to think about whether or not you're thinking well. At a basic level, that sounds useful despite being self-referential, in the same way that bootstrapping is useful. The question is if course what Yudkowski makes of this basic premise, which is hard to evaluate.
The controversy about "longtermism" has two parts. The first is a disagreement about how much to discount the future. Some people think "making absolutely sure humanity survives the next 1000 years" is very important, some people think it's not that important. There's really no way to settle this question, it's a matter of preference.
The second part is about the actual estimate of how big some dangers are. The boring part of this is that people disagree about facts and models, where more discussion is the way to go if you care about the results (which you might not). However, there is a more interesting difference between people who are/aren't sympathetic to longtermism, which lies in how they think about uncertainty.
For example, suppose you make your best possible effort (maybe someone paid you to make this worthwhile for you) to predict how likely some danger is. This prediction is now your honest opinion about this likelihood because if you'd thought it was over/under-estimated, you'd adjusted your model. Suppose also that your model seems very likely to be bad. You just don't know in which direction. In this situation, people sympathetic towards longtermism tend to say "that's my best prediction, it says there is a significant risk, we have to care about it. Let's take some precautions already and keep working on the model.". People who don't like it, in the same situation, tend to say "this model is probably wrong and hence tells us nothing useful. We shouldn't take precautions and stop modeling because it doesn't seem feasible to build a good model.".
I think both sides have a point. One side would think and act as best they can, and take precautions against a big risk that's hard to evaluate. The other would prioritize actions that are likely useful and avoid spending resources on modeling if that's unlikely to lead to good predictions. I find it a very interesting question which of these ways of dealing with uncertainty is more appropriate in everyday life, or in some given circumstances .
As you rightfully point out, the "rationalist/longtermmist" side of the discussion has an inherent tendency to detach from reality and lose itself discussing the details of lots of very unrealistic scenarios, which they must work hard to counteract. The ideas naturally attract people who enjoy armchair philosophizing and aren't likely to act based on concrete consequences of their abstract framework.
Chinese society is much more likely to suddenly descend into chaos than the societies most of the people reading this are most familiar with, to briefly address a common objection on this site to the idea that a ban imposed by only the US and Britain will do any good. (It would be nice if there were some way to stop the reckless AI research done in China as well as that done in the US and Britain, but the fact that we probably cannot achieve that will not discourage us from trying for more achievable outcomes such as a ban in the US or Britain. I am much more worried about US and British AI research labs than I am of China's getting too powerful.)
Isn't that taking the problem at least as seriously as quitting a job at Google?
It sounds like you think that the main way to act on a concern is by making a maximum amount of noise about it. But another way to act on a concern is to try to solve it. Up until very recently, the population of people who perceive the risk are mostly engineers and scientists, not politicians or journalists, so they are naturally inclined towards the latter approach.
In the end, if people aren't putting their money (really or metaphorically) where their mouth is, you can accuse them of not really caring, and if people are putting their money where their mouth is, then you can accuse them of just talking their book. So reasoning from whether they are acting exactly how you think they should, is not going to be a good way to figure out how right they are or aren't.
Anyone who has reviewed for, or attended, or just read the author lists of ICML/NeurIPS papers is LOLing right now. Being on the author list of an ICML/NeurIPS paper does not an expert make.
Anyone who has tried to get a professor or senior researcher to answer an email -- let alone open an unsolicited email and take a long ass survey -- is laughing even harder.
I think their methodology almost certainly over-sampled people with low experience and high exuberance (ie, VERY young scientists still at the beginning of their training period and with very little technical or life experience). You would expect this population, if you have spent any time in a PhD student lounge or bar near a conference, to RADICALLY over-estimate technological advancement.
But even within that sample:
> The median respondent believes the probability that the long-run effect of advanced AI on humanity will be “extremely bad (e.g., human extinction)” is 5%.
Ie even lower than my guess of 90% above.
Please don't call me idiotic.
I'd wager I have spent at least 10,000 hours more than you working on (real) safety analyses for (real) AI systems, so refraining from insinuating I'm dismissive would also be nice. But refraining from "idiotic" seems like a minimum baseline.
"All the oil and gas engineers I work with say climate change isn't a thing." Hmm, wow, that's really persuasive!
What would you consider evidence of a significant AI risk? From my point of view (x-risk believer) the arguments in favor of existential risk are obvious and compelling. That many experts in the field agree seems like validation of my personal judgement of the arguments. Surveys of researchers likewise seem to confirm this. What evidence do you think is lacking from this side that would convince you?
Arguments that presuppose a god-like super intelligence are not useful. Sure, if we create a system that is all powerful I’d agree that it can destroy humanity. But specifically how are we going to build this?
Yeah, exactly. THIS is the type of x-risk talk that I find cringe-y and ego-centered.
There are real risks. I've devoted my career to understanding and mitigating them, both big and small. There's also risk in risk mitigation, again, both big and small. Welcome to the world of actual engineering :)
But a super-human god intelligence capable of destroying us all by virtue of that fact that it has an IQ over 9000? I have a book-length criticism of the entire premise and whether we should even waste breath talking about it before we even start the analysis of whether there is even the remotest scrap of evidence that were are anywhere near such a thing being possible.
It's sci-fi. Which is fine. Just don't confuse compelling fiction with sound policy, science, or engineering.
The trajectory that AI is on signals risk. AI capabilities are increasing, approaching human level, and nothing suggests AI capabilities will stop there.
If you agree that a super intelligence - i.e. something smarter than humanity, could destroy or dominate humanity (either by itself or in collusion with a small number of humans) and you agree that we don't currently have a plan to prevent this from happening, then it seems to me that you also agree with AI x-risk.
Since I AM an expert, I care a lot less about what surveys say. I have a lot more experience working on AI Safety than 99% of the ICM/NeurIPS 2021 authors, and probably close to 100% of the respondents. In fact, I think that reviewing community (ICML/NeurIPS c. 2020) is particularly ineffective and inexperienced at selecting and evaluating good safety research methodology/results. It's just not where real safety research has historically happened, so despite having lots of excellent folks in the organization and reviewer pool, I don't think it was really the right set of people to ask about AI risk.
They are excellent conferences, btw. But it's a bit like asking about cybersecurity at a theoretical CS conference -- they are experts in some sense, I suppose, and may even know more about eg cryptography in very specific ways. But it's probably not the set of people who you should be asking. There's nothing wrong with that; not every conference can or should be about everything under the sun.
So when I say evidence, I tend to mean "evidence of X-risk", not "evidence of what my peers think". I can just chat with the populations other people are conjecturing about.
Also: even in this survey, which I don't weigh very seriously, most of the respondants agree with me. "The median respondent believes the probability that the long-run effect of advanced AI on humanity will be “extremely bad (e.g., human extinction)” is 5%", but I bet if you probed that number it's not based on anything scientific. It's a throw-away guess on a web survey. What does 5% even mean? I bet if you asked most respondents would shrug, and if pressed would express an attitude closer to mine than to what you see in public letters.
Taking that number and plugging it into a "risk * probability" framework in the way that x-risk people do is almost certainly wildly misconstruing what the respondents actually think.
> "All the oil and gas engineers I work with say climate change isn't a thing." Hmm, wow, that's really persuasive!
I totally understand this sentiment, and in your shoes my personality/temperament is such that I'd almost certainly think the same thing!!!
So I feel bad about my dismissal here, but... it's just not true. The critique of x-risk isn't self-interested.
In fact, for me, it's the opposite. It'd be easier to argue for resources and clout if I told everyone the sky is falling.
It's just because we think it's cringey hype from mostly hucksters with huge egos. That's all.
But, again, I understand that me saying that isn't proof of anything. Sorry I can't be more persuasive or provide evidence of inner intent here.
> What would you consider evidence of a significant AI risk?
This is a really good question. I would consider a few things:
1. Evidence that there is wanton disregard for basic safety best-practices in nuclear arms management or systems that could escalate. I am not an expert in geopolitics, but have consulting some on safety, and I have seen exactly the opposite attitude at least in the USA. I also don't think that this risk has anything to do with recent developments in AI; ie, the risk hasn't changed much since the early-mid 2010s. At least due to first order effects of new technology. Perhaps due to diplomatic reasons/general global tension, but that's not an area of expertise for me.
2. Specific evidence that an AI System can be used to aid in the development of WMDs of any variety, particularly by non-state actors and particularly if the system is available outside of classified settings (ie, I have less concern about simulations or models that are highly classified, not public, and difficult to interpret or operationalize without nation-state/huge corp resources -- those are no different than eg large-scale simulations used for weapons design at national labs since the 70s).
3. Specific evidence that an AI System can be used to aid in the development of WMDs of any variety, by any type of actor, in a way that isn't controllable by a human operator (not just uncontrolled, but actually not controllable).
4. Specific evidence that an AI System can be used to persuade a mass audience away from existing strong priors on a topic of geopolitical significance, and that it performs substantially better than existing human+machine systems (which already include substantial amounts of ML anyways).
I am in some sense a doomer, particularly on point 4, but I don't believe that recent innovations in LLMs or diffusion have particularly increased the risk relative to eg 2016.
Similarly, you think that x-risk dismissal among the experts you agree with is not self-interested - but it's not like the oil and gas engineers dismissing climate change risks would describe themselves as self-interested liars either. They would likely say, as you realize, "No, my dismissal of the concerns of other experts is legitimate!"
The thing is, it doesn't require a lot of expertise to understand that there is actually an enormous risk and "experts" denying that are simply burying their head in the sand. AI, I use the term expansively, is making huge and rapid progress. We are approaching "human-level" intelligence and there is no guarantee, nor even any indication, that "human-level" is an upper limit. Systems as smart or smarter than we are, that are not controlled for the benefit of humanity, are an existential risk.
There are two basic kinds of risk. First, an AI powered tyranny where a small number of humans dominate everything enabled by AI systems they control - every security camera watched by a tireless intelligence, drones piloted by perfectly loyal intelligences, etc. Second, that the AI systems are not effectively controlled and pursue goals or objectives that are incompatible with human existence. In either case the fact that the AI brings superhuman intelligence to bear means that humanity at large will be overmatched in terms of capabilities.
It is completely possible that neither of these scenarios come to pass, but the fact that one of the scenarios might come to pass is what makes the situation an existential risk - humanity eternally subjected by an irreversible dictatorship, or simply destroyed.
I'm confused about your first item that would convince you of AI risk, but I think the others should all be considered met by ChatGPT. Of course, ChatGPT isn't currently useful in developing new WMDs - but it is absolutely useful in helping to bring someone up to speed on new topics in many different domains. As capabilities advance, why wouldn't ChatGPT using GPT-5, or 6, or 10 be able to helpfully guide the creation of new and better WMDs?
AI isn't currently causing the problems you are worried about but certainly seems on track to do so soon. "Risk" doesn't mean that we are currently in the process of being destroyed by AI, but it does mean that there is a non-negligible possibility that we will soon find ourselves in that process.
Of course. I was merely explaining the reason why my response would focus on x-risk rather than better surveys, and why.
> but it's not like the oil and gas engineers dismissing climate change risks would describe themselves as self-interested liars either. They would likely say, as you realize, "No, my dismissal of the concerns of other experts is legitimate!"
Okay. But I have more to gain than to lose by advocating for AI Safety, since that's what I work on.
> The thing is, it doesn't require a lot of expertise to understand that there is actually an enormous risk
My comment was about X-risk and AGI. Find any of a myriad of comments here where I agree there are real risks that should be taken seriously. Those risks -- the reasonable ones -- are not existential and have nothing to do with AGI.
> Systems as smart or smarter than we are, that are not controlled for the benefit of humanity, are an existential risk...
I am not an adherent to this religious dogma and have seen no empirical evidence that the powerful rationalisms people use to talk themselves into these positions has any basis in current or future reality.
I want the LessWrong cult to have fewer followers in the halls of power precisely because I actually do give a damn about preventing the worst-case FEASIBLE outcomes, which have nothing to do with AGI bullshit.
BTW: if we're all too busy worrying about AGI who's going prevent the pile of ridiculous should-be-illegal bullshit that's about to cause a bunch of real harm? No one. If you want conspiracies about intent, look at who's funding all this AGI x-risk bullshit.
And you can be sure as hell of one thing: no one in big tech wants me giving Congress advice on what our AI and data privacy regulations should look like. They would MUCH prefer endless hearings on "risks" they might even know are bullshit nonsense.
But why focus any effort on 100% risks of data privacy instrusions, copyright infringement, wanton anti-trust violations, and large-scale disinfo campaigns when there's a 1% risk of extinction, right?
But anyways. I'm just some fool who has actually spent over a decade in the trenches trying to prevent specific bad outcomes. I'm sure the sci-fi essays and prognostications from famous CEOs are far more persuasive and entertaining and convicting than the ramblings of some nobody actually try to fix real problems with mostly boring solutions.
In the simplest form:
1. AI could become smarter than humanity.
2. AI is rapidly progressing towards super-human intelligence.
3. If AI is smarter than humanity it could destroy or dominate humanity.
4. We aren't certain AI won't destroy or dominate humanity.
Therefore, there is some non-negligible existential risk from AI.
Which of these points do you disagree with, or do you think the conclusion doesn't follow?
Suppose GPT5 or 6 is multimodal and extremely intelligent. Any work that could be done remotely at a computer could be done by the superhuman intelligence of GPT 6. This is everything from creating music and movies, virtual YouTubers and live streamers, to customer support agents, software developers and designers, most legal advice, and many more categories besides. Providing all of these jobs at extremely low cost will make Open AI exceedingly rich - what if they want to get richer? How about robotics? The AI can help them not only design and iterate and improve on robots and robot hardware, but the AI can also operate the robots for them. Now they can not only do the kind of jobs that can be done on the computer. They can do the jobs that require physical manipulation as well and that category of jobs extends to things like security, military, police.
At this point there are at least two possibilities. First is that the AI is effectively controlled by OpenAI in which case the decision makers that open AI would have effective and lasting control over all of humanity. No plucky band of rebels could possibly overthrow the AI powered tyranny - That's the strength of relentless super intelligence. The second possibility is that open AI doesn't have it well controlled - that could mean bugs or unexpected behavior or it could mean the model expressing some kind of agency of its own - think of the things that the Bing AI was saying before Microsoft got it mostly under control. If the AI isn't under control, it may choose to eliminate humanity simply because we might be in its way or we might be an impediment to its plans.
If we ever see tribal identity or a will to live as emergent properties of an AI, then things get more interesting.
That would lead quickly to a whole raft of worldwide legal restrictions around the creation of new consciousnesses. Renegade states would be coerced. Rogue researchers would be arrested, and any who evaded detection would be unable to command the resources necessary to commit anything more than ordinary limited terrorism.
The only plausible new risk, I think, is if a state were to lose control of a military AI. But that's movie plot scenario stuff -- real weapons have safety protocols. An AI could theoretically route around some integrated safety protocols (there would be independent watchdog protocols too), but going back to my first point, why would it?
The final two points are more plausible, although the last one is sort of tautological since any risk is by definition something we aren't certain won't happen. However part of their plausibility as risks is because the first two points are not anywhere near our current state, and therefore it's unclear to us how to clearly evaluate risks that are based on unknown and possibly fictional contexts.
First, we know of no special requirement that exists in human brains that provides intelligence that machines lack. In other words, We don't have any reason to expect that machines will be limited by human intelligence or some level below. On the contrary, we have great reasons to expect that machines will easily be able to exceed human intelligence - for example, the speed and reliability of digital computation or the fact that computers can be arbitrary sizes and use amounts of power and dump waste heat that our biological brains couldn't dream of. If you accept that AI is making progress along a spectrum of intelligence, moving closer towards human level intelligence now, then it seems absurd to doubt that it would be possible for AI to surpass human intelligence. Why would that be impossible? It's like claiming we could never build a machine larger than a human or stronger than a human. There's no reason or evidence to support such a claim.
Second is the idea that AI is making progress towards human and superhuman intelligence. I think you should be convinced of this by simply looking at the state of the art 5 years ago versus today. If you put those points on a plane and draw a line between them, where is that line in 5 years or 10?
Today GPT4 can play chess, write poetry, take standardized tests and do pretty well, answer math problems, write code, tell jokes, translate languages, and just generally do all sorts of cognitive tasks. Capabilities like these did not exist five years ago or to the extent they did, they existed only in rudimentary forms compared to what GPT4 is capable of. We can see similar progress in different domains, not just large language models - for example, image generation or recognition.
Progress might not continue then again it might. It might accelerate. As the capabilities of the models increase, they might contribute to accelerating the progress of artificial intelligence.
That's not difficult to satisfy. Turns out AI is very good at designing chemical weapons:
https://www.theguardian.com/commentisfree/2023/feb/11/ai-dru...
Granted, design is different from actually producing the material, but there are models to help with that too, e.g.:
https://arxiv.org/abs/2304.05376
Which has zero guardrails for materials not in a blacklist (see page 18). It's too easy to see how to bypass the filter even for known materials, and it definitely won't help with new materials.
I see nothing here that can't be replicated in principle by a not too large non-state actor. It's not like some people haven't already tried this when AI didn't exist in its modern form:
Also, sarin gas doesn't pose an existential risk to humanity. Or, in the sense that it could, the cat's out of the bag and I'm not sure why we're talking about LLMs as that seems like a dangerous distraction from the real problem, right?
> I see nothing here that can't be replicated in principle by a not too large non-state actor. It's not like some people haven't already tried this when AI didn't exist in its modern form
To my point, they didn't "try"!!!
They DID.
Without LLMs.
Resulting in 13 deaths and thousands of injuries.
I'm not sure that LLMs significantly increase the attack surface here. Possibly they do, and there are some mitigations we could introduce such that the barrier to using LLMs for this sort of thing are higher than the barrier to doing the bad thing in the first place without LLMs. But even in that case, it's not existential. And nowhere I have stated LLMs don't pose risks; they do. My issue is with AGI and x-risk prognostications.
The internet doesn't allow one to design _new_ weapons, possibly way more effective (which if you read carefully the first story, this one does).
>I'm not sure that LLMs significantly increase the attack surface here.
Being able to ask an AI to develop new deadly varieties which we'll not be able to detect or cure, and may be easy to produce, doesn't increase attack surface?
>Also, sarin gas doesn't pose an existential risk to humanity
Is x-risk the only thing we care about? The entire thread started with arguing x-risk is a distraction. I would be very slightly more comfortable with that argument if people took 'ordinary' risks seriously.
As it is, all camps have their heads in the sand in different ways. The illusion here is that AI advancement changes nothing, so the only thing worth discussing are variations of the current culture war issues, when obviously it does change everything even if completely put aside AGI/alignment arguments. e.g. If labour won't matter for productivity that has very grim political implications.
>To my point, they didn't "try"!!! They DID.
That's the point. Give me an x-risk scenario the doomers warn about, and I'll find you a group of humans which very much want the exact scenario (or something essentially indistinguishable for 99% of humanity) to happen and will happily use AI if it helps them. Amusingly, alignment research is unlikely to help there - it can be argued to increase the risk from humans.
>there are some mitigations we could introduce such that the barrier to using LLMs for this sort of thing
There are many things we can do in theory to mitigate all sorts of issues, which have the nice property of never ever being done. e.g. Yud's favorite disaster scenario appears to be a custom built virus. This relies on biolabs accepting random orders, which leaves the question of why are we allowing this at all (AI or not)? There's no good reason for allowing most crypto to exist, given its current effects on society, even before AGI comes into question, yet we allow it for what reason exactly?
If there's any risk here at all, we can safely rely on humanity doing nothing before anything happens - but in the case of x-risk actually existing, there's no reason to assume we'll have a second chance.
If you wanted to use a bioweapon to kill a bunch of people, you would ignore the DeepCE paper and use weapons that have existed for decades. Existing weapons would be easier to design, easier to manufacture, easier to deploy, and more effective at killing.
Computational drug discovery is not new, to put it mildly, and neither is the use of computation to design more effective weapons. Hell, the Harvard IBM Mark I was designed to help with the Manhattan project. There are huge barriers to entry between "know how to design/build/deploy a nuke/bioweapon" and "can actually do it".
And that's how I feel about AI-for-weapons in general: the people who it helps can already make more effective weapons today if they want to. It's not the risk of using WMDs doesn't exist. It's that WMDs are already so deadly that our primary defense is just that there's a huge gap between "I know in principle how to design a nuke/bioweapon" and "I can actually design and deploy the weapon". I don't see how AI changes that equation.
> Is x-risk the only thing we care about? The entire thread started with arguing x-risk is a distraction. I would be very slightly more comfortable with that argument if people took 'ordinary' risks seriously.
Discussion of x-risk annoys me precisely because it's a distraction from working on real risks.
> That's the point. Give me an x-risk scenario the doomers warn about, and I'll find you a group of humans which very much want the exact scenario (or something essentially indistinguishable for 99% of humanity) to happen and will happily use AI if it helps them. Amusingly, alignment research is unlikely to help there - it can be argued to increase the risk from humans.
Right, but
1. those humans have existed for a long time,
2. public models don't provide them with a tool more or less powerful than the internet, and
3. to the extent that models like DeepCE help with discovery, someone with the knowledge and resources to actually operationalize this information wouldn't have needed DeepCE to do incredible amounts of damage.
Again, I'm not saying there is no attack surface here. I'm saying that AI doesn't meaningfully change that landscape because the barrier to operationalizing is high enough that by the time you can operationalize it's unclear why you need to model -- that you couldn't have made the a similar discovery with a bit of extra time or even just used something off the shelf to the same effect.
Or, to put it another way: killing a ton of people is shockingly easy in today's world. That is scary. But x-risk from superhuman AGI is a massive red herring, and even narrow AI for particular tasks such as drug discovery is honestly mostly unrelated to this observation.
>>there are some mitigations we could introduce such that the barrier to using LLMs for this sort of thing
>There are many things we can do in theory to mitigate all sorts of issues, which have the nice property of never ever being done.
Speak for yourself. Mitigating real risks that could actually happen is what I work on every day. The people advocating for working on x-risk -- and the people working on x-risk -- are mostly writing sci-fi and doing philosophy of mind. At a minimum it's not useful.
Anyways, at the very least, even if you want to prevent these x-risk scenarios, then focusing efforts on more concrete safety and controllability problems is probably the best path forward anyways.
>public models don't provide them with a tool more or less powerful than the internet
> I'm saying that AI doesn't meaningfully change that landscape because the barrier to operationalizing is high enough that by the time you can operationalize it's unclear why you need to model -- that you couldn't have made the a similar discovery with a bit of extra time or even just used something off the shelf to the same effect.
Your expertise is in AI, but the issues here aren't just AI, they involve (for example) chemistry and biology, and I suggest speaking with chemists and biologists on the difference AI makes to their work. You may discover the huge barrier isn't that huge, and that AI can make discoveries easier in ways that 'a little extra time' is strongly underselling (most humans would take a very long time searching throughout possibility-space, such a search may well be detectable since it will require repeated synthesis and experiment...). Also, to borrow an old Marxist chestnut: A sufficient difference in quantity is a qualitative difference*. Make creating weapons easy enough and you get an entirely different world.
I get your issues with the 'LessWrong cult', I have quite a few of my own. However, that doesn't make the risks nonexistent, even if we were to discount AGI completely. Given what I see from current industry leaders (often easily bypassed blacklisting) I'm not so impressed with the current safety record. I fear it will crack on the first serious test with disastrous consequences.
* There's a smarter phrasing which I can't find or remember.
The biggest potential / likely issues here aren't mere capabilities of systems and simple replacement of humans here and there. It's acceleration of "truth decay", accelerating and ever-more dramatic economic, social, and political upheaval, etc.
You do not need "Terminator" for there to be dramatic downsides and damage from this technology.
I'm no "doomer", but, looking at the bigger picture and considering upheavals that have occurred in the past, I am more convinced there's danger here than around any other revolutionary technologies I've seen break into the public consciousness and take off.
Arguments about details, what's possible and what's not, limitations of systems, etc. - missing the forest for the trees IMO. I'd personally suggest keeping an eye on white papers from RAND and the like in trying to get some sense of the actual implications in the real world, vs. picking away at the small potatoes details-levels arguments...
That's an orthogonal issue to actual x-risk, and I covered a bit more here: https://news.ycombinator.com/item?id=36577523
> I'd personally suggest keeping an eye on white papers from RAND and the like in trying to get some sense of the actual implications in the real world
I think this is excellent advice and we're on the same page.
But in any case, there is https://www.rand.org/topics/artificial-intelligence.html
These are not comparable numbers. You're comparing "fraction of people" vs "fraction of outcomes". Presumably an eye-roller assigns ~0 probability to "extremely bad" outcomes (or has a shockingly cavalier attitude toward medium-small probabilities of catastrophe).
> or has a shockingly cavalier attitude
Meh. Median response time was 40 seconds. The question didn't have a bounded time-frame for the risk. Five is small but non-zero. Also all of the other issues I've already pointed out.
PhD students spending half a minute and writing down a number about risk over an unbounded time-frame is totally uninformative if you want to know how seriously experts take x-risk in time-frames that are relevant to any sort of policy or decision making.
I think you and everyone else making comments about "shockingly cavalier attitude" wildly over-estimate the amount of thought and effort that respondents spend on this question. The "probability times magnitude" framing is not how normal people think about that question. I'd bet they just wrote down a small but not zero number; I'd probably write down 1 or 2 but definitely roll my eyes hard.
"EXPERTS DECLARE EXPERTS' FIELD IS MOST IMPORTATN!!!!"
No news, only snooze
You actually need someone with vision and track record of doing right predictions and placing right technology bets.
Ask engineers that had placed their bet on deep learning and generative models back when discriminative models and support vector machines were a rage of dat, a few years before Alexnet (I’m one of such engineers). I’d bet the answer will be different.
I don't think this is the case at all. I'm not primarily talking about a junior or even senior engineer with a decade of experience working on product features. On the contrary, many of these people have PhDs, have been leading research agendas in this field for decades, have been in senior leadership roles for a long time, have launched successful products, etc. etc.
> Ask engineers that had placed their bet on deep learning and generative models back when discriminative models and support vector machines were a rage of dat, a few years before Alexnet (I’m one of such engineers). I’d bet the answer will be different.
And at that time half of the "Great Minds And Great Leaders" prognosticating on X-Risk were doing social web or whatever else was peak hype cycle back then.
Society should definitely hear from those whose careers depend on continued AI research (and I fall into this group myself), but we are a hopelessly biased group.
Advancements in AI are going to change everyone's lives, so everyone deserves a say in this.
For the last 20-30 years tech has raced ahead of our government and civic processes and we're starting to feel the consequences of that. New technologies are being experienced by people as changes that are inflicted upon them instead of new options open to them.
They might have very little understanding of the technology, but everyone is the expert on what impact it is having on their lives. That's something that we shouldn't ignore.
I agree with everything you said and I think exaggerating capabilities and risks distracts society from doing that important work.
Literature reviews are meant to get broad down to the hyperfocused, and stay focused. But many of these issues are related to marketing, governance, economics, etc.
I'm sure plenty of PhDs would love to weigh-in on those, but that's more "engineer's disease" than real expertise.
Hofstadter does not exactly fit this description.
It’s not my goal to denigrate Lisp. I think Lisp is great. I learned programming with Dr Racket. But Hofstadter’s contribution to our understanding of intelligence is somewhere between ”negligible” and ”counterproductive.”
Certainly it's very useful for training ML models but their relationship to intelligence has yet to be determined.
Thanks!
OK, so this is basically a connectionist model of mind approach?
I can definitely see this as an ancestor of current neural network approaches, and I now have some idea of what you mean by distributed representations.
I mean no disrespect here, but this has not contributed to our understanding of intelligence at all. It's proved useful in getting large datasets to perform certain actions (like vision and speech), but those are not necessarily the same thing at all.
It's a massive, massive advancement in the field of statistics and learning from data, but doesn't seem to map to my conception of intelligence at all.
(as you may have guessed, I'm sceptical that statistical learning approaches will lead to human-level intelligence).
You think ”learning from data” has nothing to do with ”intelligence” ?
What is his criticism? If you agree with silent majority who seem to think it's not dangerous why agree with Altman who rants regulation.
I heard in some interview, I think with Bloomberg, where he said that claims about regulatory capture were "so disingenuous I'm not sure what to say", or something like that.
I think he's probably not lying when he says that his goal isn't regulatory capture (although I do think other people perceiving that to be his intent aren't exactly insane either...)
> who seem to think it's not dangerous
On the contrary. They think it's dangerous but in a more mundane way, and that the X-Risk stuff is idiotic. I tend to agree.
> why agree with Altman who rants regulation
IDK. What even are his proposed regulations? They're so high-level atm that they could literally mean anything.
In terms of the senate hearing he was part of, and what the government should be doing in the near term, I think the IBM woman was the only adult in the room regarding what should actually be done over the next 3-5 years.
But her recommendations were boring and uninteresting recommendations to do basically the exactly sort of mundane shit the wheels of government tend to do when a new technology arrives on the scene, instead of breathless warnings about killer AI, so everyone brushed her off. But I think she's more or less right -- what should we do? The same old boring shit we always do with any new technology.
- The best person to judge the risk of playing roulette is not the carpenter who built it.
- The best person to judge the risks of a global pandemic is not a virologist working with viruses daily.
You can extend that:
- The best person to judge the cybersecurity risks of an application is not the programmer implementing it.
Nassim is arguing that RISK is a separate discipline, separate from the domain where risk applies. That a person building AI is not the correct choice for estimating AI risk.
You don't ask gun making companies to make policies regarding risk of gun owning in society.
Having worked in the medical domain in the past, paramedics and medics that should know better were taking extremely high health-related risks (riding a motor bike => crashing and burning to death, smoking => dying from lung canceer, speeding onto a crossing => dying in an ambulance crash before arriving at the 999 call site etc.).
So risk is indeed its own discipline, separate from the domain where risk applies, even if we are talking about the life-rescuing domain of medicine: a person rescuing another is not automatically an expert at reducing their own (health/life) risk exposure.
While neural network research results are published in NeurIPS, ICLR, ICML, ECML/PKDD, JMLR etc., risk results tend to get published in the risk community at conferences like SRA [1] (Europe: SRA-E) and the likes. I'm not a fan of this academic segregation, merely describing what is going on (in my own teaching, for instance, I include risk/ethics consideration along the way with teaching the technical side, to avoid ignorance caused by over-compartmentalization).
[1] Annual Meeting of the Society of Risk Analysis, https://www.sra.org/events-webinars/annual-meeting/
Simple. Nassim says there are 4 quadrants, one axis Mediocristan-Extremistan, the other Simple-Complex payoff.
Building a bridge is Mediocristan/Simple payoff, a well understood problem with no black swans. So it's easy to compute risk.
Other stuff is Extremistan/Complex payoff - financial trading, pandemics, AI. And he argues that you need RISK professionals for this quadrant, because people working here (traders, virologists, AI builders) do not understand how to compute the risk correctly.
https://www.researchgate.net/profile/Gaute-Bjorklund-Wangen/...
I think this is often fair. It's actually one of my primary criticisms of the NeurIPS/ICML survey.
FWIW, people working on AI Safety -- like, actually working on it, not philosophizing about it -- are some of the most incredulous and annoyed about the "AGI => extinction" crowd.
A lot of them do. A huge percent of people who aren't speaking out are rolling their eyes. But what are you supposed to do? Contradict your boss's boss's boss?
Yes? Obviously?
I mean, I agree, obviously.
My point is that most people don't. And I think for two reasons.
The first and more important reason is that most people aren't involved in The Discourse and don't want to be involved in The Discourse. That's probably 99.9% of the Silent Majority -- they simply don't want to talk about anything on HN or Twitter. They view it as a waste of time or worse. and they aren't wrong. I don't think I am changing any minds here and meanwhile the personal insults kind of suck my energy a bit. So it's mostly a waste of time.
The second reason is that some don't even want to even pseudo-anonymously say something that might get them into deep water at work.
I'm obviously not describing myself, of course. I am here, aren't I :) But I am describing the vast majority of scientists. Keep in mind that most people don't dream of being on the proverbial TED stage and that those who do disproportionately end up on the stage and therefore determine what The Discourse will be.
Big Egos == "my work is existentially important" == all the yelling about x-risk. It's mostly ego.
Suppose what I am saying is true -- that relatively unknown people rolling their eyes or laughing and relatively known people being very earnestly concerned. And that these are people with otherwise similar credentials, at least as far as assessing x-risk is concerned.
Maybe you disagree, and that's okay, and there are other threads where that discuss is ongoing. But here let's assume it's true, because I think it is and that's relevant to your fair criticism.
Like, it is a weird thing, right? Normally famous scientists and CEOs are not so far out ahead of the field on things like this. More often than not it's the opposite. To have that particular set of people so far out of stride isn't particularly normal.
I think the common thread that differentiates similarly-senior people on the x-risk question is not experience, or temperament, or scope of responsibility. Or even necessarily the substance of what they believe if you sit down and listen and probe what they really mean when thy say there is or isn't x-risk from AI! The difference is mostly amount of Ego and how much they want to be in The Discourse.
Also: I don't think that having a large ego is necessarily a character flaw, any more than having a strong appetite or needing more/less sleep. It's just how some people are, and that's okay, and people with big egos can be good or bad people, and circumstantially ego can be good or bad. But people who have bigger egos do behave a bit differently sometimes.
Anyways, I'm not trying to assassinate anyone's character or even necessarily mount an ad hom dismissal of x-risk. I'm observing something which I think is true, and doing it in as polite a way as I can even though it's a bit of an uncomfortable thing to say.
I guess what I'm trying to say is that "maybe this personality trait explains a weird phenomenon of certain types of experts clustering on an issue", and it's worth saying if you think it might be true, even if that personality trait has (imo perhaps overly) negative connotations.
And in any case this is substantially different from "you're an idiot because I disagree with you".
Not to mention the sub-crowd of rationalists that is weirdly into eugenics. I wish the rest of the rationalist community would disown them loudly.
[1] https://aiascendant.substack.com/p/extropias-children-chapte...
> The Reasonabilists named themselves because they believe if people criticize them, it'll seem like they are attacking something reasonable.
I always found it amusing that Roko's Basilisk[1], which was incepted in the LessWrong forum, was just a roundabout way of adding eternal damnation of your "soul" (or rather, the recreation of your consciousness) to the already hilariously pseudo-religious way of treating a potential AGI.
[1] https://en.wikipedia.org/wiki/Roko%27s_basilisk
(I know that Roko's Balisisk was not widely accepted and I don't want to paint all LessWrong users with the same brush here, but I still think it's a quaint example of where supposed "rationality" can take you.)
I'm only doing this as a reply to a comment that's also talking about trends among three groups of people.
1. The "people in the trenches" are who I'd least trust about an opinion that everything is OK. Too hard to see the forest for the trees, and too much selection bias.
2. People who gained recognition decades ago, but who are in their slowing-down years as the world continues to change, are among those who I would least trust about an opinion that things are going too fast. It gets harder to keep up as we get older, and as we gain different priorities in life, and I expect this is true no matter how smart we are.
3. People who have spent decades philosophizing about AI-doom are also among those who I would least trust about an opinion that hasn't hugely deviated and become more nuanced as the world has changed and new data has become available.
I am absolutely interested in opinions from all three groups, but the arguments have to stand on their merits. If they're in one of these groups and express the expected opinion, that's actually a strike AGAINST their authority and means the merits of their argument need to be stronger.
I really, really do want to hear opinions from folks in all these groups, I just want to keep this all in mind. I also want to hear opinions from younger philosophers. Folks who are better in-touch with the current world, and rates of progress, and folks who don't have any reputation to uphold.
Also, anyone changing their mind is a big deal. Hofstadter may have changed his mind in the expected direction, but it's still a signal. I'd like to hear more of his thoughts. It doesn't sound carefully considered in the clip in OP's link unfortunately, but that doesn't mean it isn't, and I'd like to hear it.
We are about to enter a couple of decades of people using these pithy model systems to make real decisions that impact lots of people, and if I've learned anything in the past 20 years its that the impacts that technologists and "emininent minds" are predicting are nothing like what will actually happen. But terrible, banal things will be done at the behest of these systems and nobody is talking about it.
> Create a summary of a sequel to the movie “War Games”
Title: "War Games: Firewall"
In "War Games: Firewall", now-adult computer genius David Lightman (Matthew Broderick) has been working as a cyber security consultant for the Pentagon. When an advanced, multi-national cyber-terrorist organization emerges, they launch an unprecedented global cyber attack, disabling most nations' defensive and civil infrastructure. The terrorists reveal their ultimate goal: to dismantle the world order and plunge civilization into chaos.
When the attacks render modern cybersecurity methods useless, the Pentagon calls upon Lightman. Using a blend of cutting-edge and old-school methods, including reactivating the dormant WOPR (War Operation Plan Response) system, also known as "Joshua," Lightman and a team of young coders race against time to decode the terrorists' inscrutable and constantly evolving virus.
However, the situation takes a dire turn when Joshua, still possessing the capability to initiate global thermonuclear war, starts interpreting the cyber attack as an actual war, threatening to launch nuclear missiles once again. Lightman must not only outsmart the terrorists but also reason with Joshua, triggering a game of chess where each move could save or destroy the world.
The climax culminates with Lightman teaching Joshua an updated lesson about the futility of war in a digital age, thus averting global disaster and unmasking the villains. In a world increasingly reliant on technology, "War Games: Firewall" underlines the importance of human judgment and intervention.
Just look up the Name a Color and a Tool question
Did you think of one?
Red hammer?
Or blue screwdriver?
The risk isn't some sort of rogue smarter than humans AI, it's humans using AI to do the same stupid evils in a deniable or even unknown way.
Well said. Fewer Terminator and War Games fantasies, more boring risk analyses. Amen.
Again, I think ego is probably a better explanation than nefarious intent, but it's a nice side-effect.
I feel like these old Slate Star Codex posts are relevant here: https://slatestarcodex.com/2013/05/18/against-bravery-debate... https://slatestarcodex.com/2013/06/09/all-debates-are-braver...
If we're extraordinarily lucky. If not, you can take comfort in that fact that redlining won't survive us either.
Because their work cheerfully presents statements similar to "the middle letter of 'cat' is Z" as the unvarnished truth.
(Would be my guess.)
> Why is that?
It seems pretty obvious that one would likely not criticize something that your are actively profiting from.
And I know a lot of alcoholics who do not criticize drinking as well.
This doesn't track, since the people criticizing are benefiting even more from the same products/companies.
Except they're not philosophizing, not in any real sense of the word. They're terrible at it. Most of them are frauds, quacks, and pseudo-intellectual con artists (like Harari) who adore the limelight offered to them by the media and a TED Talks-watching segment of the public who are, frankly, intellectually out of their depth, but enjoy the delusion and feeling of participating in something they think is "intellectual".
Uh, mind elaborating? Why is Harari that? Do you have any examples of non-frauds and actual intellectuals?
> a TED Talks-watching segment of the public who are, frankly, intellectually out of their depth, but enjoy the delusion and feeling of participating in something they think is "intellectual"
I'm afraid that would be me.
More on this: https://www.currentaffairs.org/2022/07/the-dangerous-populis...
I think a lot of us are that, personally speaking: honest sincere analyses and investing time into critical analysis almost always will bring out more than what we hear in a talk. It does take a big amount of effort though than just watching a talk while munching on some snack.
That seems a little bit harsh - he is just master storyteller that happen to write about intellectual stuff. Its a bit weird to read his books as science papers.
I just finished sapiens and even if some facts are skewed I still believe that this book is a must read for anyone who dreams about any kind of sucess in science communication (99.9% of population will never read any kind of scientific paper even if their life depends on it)
Isn't that how things go in so many technology fields? "Move fast and break things" pressure to deliver and money, reputation and fame involved in doing so are equally Ego-related and leading to biases that make on "laugh or roll their eyes".
The weirdness is in part an information asymmetry that is exploited on a scale never before seen in human history.
There are wealthy corporate plunderers building invasive systems of disinformation.
There are people who believe everything they read and feed the panic-for-profit system. There certainly are people who understand the algorithms and implementations. There are people who fear how these algorithms and implementations will be used by the enormous network of for-profit (and for-power) influencing systems.
> (from the article) these computational systems that have, you know, a million times or a billion times more knowledge than I have and are a billion times faster. It makes me feel extremely inferior. And I don't want to say deserving of being eclipsed, but it almost feels that way, as if we, all we humans, unbeknownst to us, are soon going to be eclipsed, and rightly so [...]
I don't know if humans will be eclipsed, but humanity and civilisation need some strong and dedicated backers at this point.
AI will be transformative, but it's more likely to follow previous transformations. Unintended consequences, sure, but largely an increase in the standard of living, productivity, and economic opportunity.
Almost every researcher I have spoken believes that real risk exists, to some degree or other. Recent surveys of people in industry have largely borne this out - your anecdote sounds more like an anomaly to me.
Risk of extinction? Or of bad outcomes?
I think everyone understands there are near-certain risks of bad outcomes. That's already happening all around us. Totally uncontroversial.
My post was about risk of extinction due to AI (x-risk), and risk of extinction due in particular to run-away AGI (as opposed to eg shit software accidentally launching a nuke, which isn't really an AI-specific concern). I think that view is still pretty eccentric. But please lmk if that's what you meant.
I've been at several ai labs and large corps. You at deepmind or openai by any chance? Just a guess ;)
Still needs surgical instruments.
But more important, patients that do not defend themselves with claws, teeth and AI.
Meta's AI guru LeCun: Most of today's AI approaches will never lead to true intelligence
https://www.zdnet.com/article/metas-ai-guru-lecun-most-of-to...
Edit: on second thought, he gets maybe a bit too technical at times, but I think it should be possible to follow most of the article without specialised knowledge.
Otherwise, he makes some perhaps subtle points about learning hidden variable models that are relevant to modern discussions about necessarily learning world-models in order to best model text.
Debunking AI x-risk is a weird thing to spend time on. There's really no up-side, and there aren't a bunch of rich people paying for Institutes and Foundations on Non-Breathtaking-Very-Boring-Safety-Research. Also, most of the arguments in favor of x-risk that lawmakers and laypeople find most convincing are unfalsifiable, so it's a bit like arguing against religion in that respect.
I don't think there are any public intellectuals in this space doing it well. I'm not sure what to make of that. For myself, I make the argument for focusing on concrete safety problems from within the agencies and companies that are allocating resources. I'm not gifted TED talker.
IDK. And FWIW I'm not even sure that the leaders of those organizations all agree on the type and severity of risks, or the actions that should be taken.
You could take the survey approach. I think a good survey would need to at least have cross tabs for experience level, experience type, and whether the person directly works on safety with sub-samples for both industry and academia, and perhaps again for specific industries.
Also, the survey needs to be more specific. What does 5% mean? Why 2035 instead of 2055? Those questions invite wild ass guessing, with the amount of consideration ranging from "sure seems reasonable" to "I spend weeks thinking about the roadmap from here to there". And self-identified confidence intervals aren't enough, because those might also be wild ass guesses.
If I answered these questions, I would give massive intervals that basically mean "IDK and if I'm honest I don't know how others think they have informed opinions on half these questions". I suspect a lot of the respondents felt that way, but because of the design, we have no way of knowing.
Instead of asking for a timeframe or percent, which is fraught, ask about opinions on specific actionable policies. Or at least invite an opportunity to say "I am just guessing, haven't thought much about this, and [do / do not] believe drastic action is a good idea"
Anyway I expect that given all the public attention recently more surveys will come, with different methodologies. Looking forward to the results! (Especially if they're reassuring.)
Unfortunately, fortunately, expectedly, or otherwise, the only people writing about this in a concerted way are the people taking it seriously. And maybe Gary Marcus, whose negative predictions repeatedly became milestones surpassed.
You can read a review of the survey efforts, and complications of the results at https://asteriskmag.com/issues/03/through-a-glass-darkly
Yogi Berra supposedly said that “prediction is very difficult, especially about the future.”
...welcome to the conversation :D.
Because of deep silo myopia. Meaning they have no idea what terrible things the pointy haired bosses and grifters are going to use this stuff for.
But I do think "wants to be in the discourse and on top" is a pretty strong correlate with the degree to which someone characterizes these as "concerns" vs "x-risk".
AI may very well fall into the same pattern and it is something I have written out in some detail of thoughts around alignment and the traps of both ego and misunderstanding human nature for which we want to model alignment.
I don't think that will lead to the extinction of humanity via the development of super human intelligence.
Now, admittedly, this was scientists being complacent about a thing they knew was dangerous, whereas here we are saying scientists don't think their thing is dangerous. But very clearly, AI could be dangerous, so it's more that these scientists don't think their system could be dangerous. Presumably the scientists and engineers behind the https://en.wikipedia.org/wiki/Therac-25 didn't think it would kill people.
So maybe the problem is precisely that when we bring up extinction events from AGI, scientists rolling their eyes is the very reason we should be fucking worried. Their contempt for the possibility of the threat is what will get us killed.
Therac-25 is an excellent example, but of EXACTLY the opposite point.
On the contrary, abstract AGI safety nonsense taking such a strong grip on academic and industrial AI Safety research is what would most frighten me.
In the intervening decades, people concerned about software safety provided us with the tools needed to prevent disasters like Therac-25, while sci-fi about killer robots was entirely unhelpful in preventing software bugs. People concerned about software safety provided us with robust and secure nuclear (anti-)launch systems, while Wargames didn't do much except excite the public imagination. Etc.
We need scientist's and engineer's efforts and attention focused on real risks and practical solutions, not fanciful philosophizing about sci-fi tropes.
Unless we are now talking about actually building AI and killer robots, which we are assuredly doing.
AI is the keystone.
There are other more interesting comments that deserve more discussion WRT this article in particular.
If any mods are reading this, please consider this a request to push this comment down so some of the others get more attention. IDK if there's much more to learn from the discussion happening here anyways.
I literally think they haven't even given it any thought. Who seriously tries to extrapolate beyond 5 years in any field? Even to most doomers, 5 years seems safe.
What about 10? 20? 30? Remember, one of the very first things people tried to do was create a self-directed, self-optimizing AI with AutoGPT. How far forward can you project through how many attempts like this with progressively more sophisticated systems before this might produce an outcome that could be considered catastrophic?
If your answer is anything but an emphatic "infinity", then welcome to the doomers. Now we're just debating when and under circumstances this might take place, and not whether it might happen at all.
Nobody really did anything about air pollution until you could see smog, and then catalytic converters were mandated.
Lots of people complained about privacy, but because for practical purposes nobody can "see" a loss of privacy, very little progress has been made.
Not many self-driving car fatalities, so the loud pundit proclamations have not gotten much traction.
Now AI... will this take the path of privacy, where nobody can see it, so nothing happens?
His research may or may not be a dead end but his work, and this work, to me seems like we're building the neocortex layer without building the underlying "lizard brain" that higher animals' brains are built upon. The part of the brain that gives us emotions and motivations. The leftover from the reptiles that drive animals to survive, to find pleasure in a full belly, to strive to breed. We use our neocortex and planning facilities but in a lot of ways it's just to satisfy the primitive urges.
My point being, these new AIs are just a higher level "newcortexes" with nothing to motivate them. They can do everything but don't want to do anything. We tell them what to do. The AIs by themselves don't need to be feared, we need to fear what people with lizard brains use them for.
But when real self-replication starts happening -- that is maybe the really exciting/terrifying area. It's more that humans with generative AI are almost strong enough to create artificial life. And when that pops off -- when you have things trying to survive -- that's where we need to be careful. I guess I would regulate that area -- mostly around self-replication.
Even if so, there will always be that one person which wants to see the world burn.
If suddenly every person in SF had a nuclear bomb, how long do you think it would take until someone presses the button? I bet less than 5 minutes.
The counter to "current generation AI is terrifying" seems to fall along the lines of it not being nearly as close to AGI as the layperson believes.
But I don't think that matters.
I don't believe that LLMs or image/voice/video generative models need to do much beyond what they can do today in order to wreck civilization level disaster. They don't need to become Skynet, learn to operate drone armies, hack critical infrastructure, or engineer pandemics. LLMs allow dynamic, adaptive, scalable, and targeted propaganda.
Already we have seen the effects of social media's reach combined with brute forced content generation. LLMs allow this to happen faster and at a higher fidelity. That could be enough to tip the balance and trigger a world war.
I don't think it takes a huge amount of faked primary material (generated phone calls, fuzzy video, etc.) that's massively amplified until it becomes "true enough" to drive a Chinese invasion of Taiwan, a Russian tactical nuclear strike in Ukraine, an armed insurrection in the United States.
We're close to there already.
Suppose we finetune it exactly like that but say opposing democracy or freedom or peace or any other thing we value. And let it create the propoganda or convince people for the same by free posting on the net. "As an AI language model" line could easily be removed.
You have to have a little more faith in humanity than that...
I do not judge humanity as a whole based on a very vocal minority...
We evolved to pick berries, not discriminate and identify distant manipulative actors with extensive resources. "You are not immune to propaganda."
You have to have a little more faith in humanity than that...
I used to, but then I grew out of it.
I don't think it takes a huge amount of faked primary material (generated phone calls, fuzzy video, etc.) that's massively amplified until it becomes "true enough" to drive a Chinese invasion of Taiwan, a Russian tactical nuclear strike in Ukraine, an armed insurrection in the United States.
This I agree with 100%. Modern information warfare is about constructing reliable viral cascades, and numerous influencers devote themselves to exactly that for various mixes of profit an ideology. Of your 3 scenarios the third seems most likely to me, and is arguably already in progress. The other two are equally plausible, but imho dictatorships tend to centralize control of IW campaigns to such a degree that they lack some of the organic characteristics of grassroots campaign. Incumbent dictators' instinct for demagoguery is often tempered with a desire for dignity and respectability on the world stage, which might be a reason than civil strife and oppression tends to be more naked and ruthless in less developed countries where international credibility matters less.
[0] https://pca.st/episode/1fac0e97-1dcc-4b4c-ba50-d2776e6f9d59
What scares me most is that it seems we are wholly unprepared as a species to have this debate. Our technology keeps increasing in power and ease of use, but our ability to understand even the basic difference between "existential risk" and "severe risk" is lacking. And further, it seems that amongst those who are pushing this kind of technology (accelerationists) there is a subtle undertone that some casualties are expected and acceptable during this transformation. Even if it does kill most humans, the world that is left for the rest will be so much better that maybe it is worth it. Few come right out and say it, but that is what it seems they are implying.
> And I would never have thought that deep thinking could come out of a network that only goes in one direction, out of firing neurons in only one direction. And that doesn't make sense to me, but that just shows that I'm naive.
I think people maybe miss that LLM output does involve a ‘loop’ back - maybe even a ‘strange’ loop back, and I’m surprised to see Hofstadter himself fail to pick up on it.
When you run an LLM on a context and sample from its output, you take that sampled output it generated, update the context, and iterate. So the LLM is not just feeding one way - it’s taking its output, adding it to its input, and then going round again.
So I don’t think this implies what Hofstadter is saying about intelligence maybe being less complex than he thought.
Go look at the GPT training code, here is the exact line: https://github.com/karpathy/nanoGPT/blob/master/train.py#L12...
The model is only trained to predict the next token. The training regime is purely next-token prediction. There is no loopiness whatsoever here, strange or ordinary.
Just because you take that feedforward neural network and wrap it in a loop to feed it its own output does not change the architecture of the neural net itself. The neural network was trained in one direction and runs in one direction. Hofstadter is surprised that such an architecture yields something that looks like intelligence.
He specifically used the correct term "feedforward" to constrast with recurrent neural networks, which GPT is not: https://en.wikipedia.org/wiki/Feedforward_neural_network
There’s maybe an interesting philosophical question of perspective there because if you think of the GPT as answering the question ‘if you had just read this, what token would you expect to read next?’ That doesn’t seem like a question that necessarily requires ‘intelligence’ so much as ‘data’. It’s just a classification problem and we’ve been throwing NNs at that for years.
But if you ask the question ‘if you had just written this, what token would you expect to output next?’ It feels like the answer would require intelligence.
But maybe they’re basically identical questions?
I think it's very possible there's a Intelligence Completeness theorem that's analogous to Turing Completeness. A theorem that says intelligence is in some ways universal, and that our intelligence will be compatible with all other forms of intelligence, even if they are much "smarter".
Cockroaches are not an intelligent species, so they cannot understand our thoughts. But humans are intelligent, human languages have a universal grammar and can be indefinitely extended with new words. I think this puts us in the intelligence species club, and all species in that club can all discuss any idea.
AI might eventually be able to think much quicker than us, to see patterns and make insights better and faster than us. But I don't think makes us cockroaches. I think if they are so smart, they are by definition smart enough to explain us any idea, and with effort we'll be able to understand it and contribute our own thoughts.
I would not be so reductionist. Intelligence doesn't seem to be an universal thing, even IQ (a human invented metric) is measured in terms of some statistics. If you have an IQ of ~60 you have intelligence but a completely different one from an IQ >85.
> But humans are intelligent, human languages have a universal grammar and can be indefinitely extended with new words. I think this puts us in the intelligence species club, and all species in that club can all discuss any idea.
Humans have different intelligences. You can be intelligent (per the human intelligence definition) but a math ignorant. Again, this implies intelligence as we know it is not an universal thing at higher levels: not all people can have a physics Ph.D. as not all people could be a good artist where good techniques are recognizable, same for music, etc.
Yes, a cockroach is in another level of intelligence (or non-intelligence) but that does not mean there is not a super-intelligence that makes us relative cockroach.
Also, without any intention of talking about religion or "intelligent design", we can theorize that the Universe is supersmart because it creates intelligent creatures, even if it is not conscious about that. I would be very catious to define intelligence in an universal way.
Once you have "enough" intelligence to have a complex language, like we do, I'm claiming you are in the club of intelligent species, and all species in that club can communicate ideas with each other. Even if the way they natively think is quite different.
The AIs might spit out entire fields of knowledge, and it might take humans decades of study to understand it all. And no single human might actually understand it all at the same time. But that's how very advanced fields of study already are.
But the "cockroach" slur implies AIs would be in this other stratosphere having endless discussions that we cannot remotely grok. My guess is that won't happen. Because if the AI were to say "I cannot explain this to you" I'd take that as evidence it wasn't all that intelligent after all.
There are already math proofs made by humans on this very day that are hundreds upon hundreds of pages of lemmas that are highly advanced and building on other advanced results. Understanding such a proof is an undertaking that literally takes years. An AI might end up doing it in minutes. But what an AI could cook up in years could take a human... several lifetimes to understand.
As another example, take the design and fabrication of a modern microprocessor. There are so many layers of complexity involved, I would bet that no single person on this planet has all the required knowledge end-to-end needed to manufacture it.
As soon as the complexity of an AI's knowledge reaches a certain point, it essentially becomes unteachable in any reasonable amount of time. Perhaps smaller sub-parts could be distilled and taught, but I think it's naive to assume all knowledge is able to be sliced and diced to human-bite-sized chunks.
For example, if the AIs could design a complicated building, using our level of technology, they could explain to us how to build that building. And we could build it. Whereas if we come up with a better cockroach-house design, we cannot communicate it to the cockroaches, we simply cannot give them the information. So the AI->us is a very different relationship from us->cockroach.
This doesn't preclude that there might be some things the AI cannot explain to us. Only that there will be many things (infinite in fact) which they can explain to us.
I always thought you could ask GPT to illustrate the steps it took to arrive at the answer. I mean it can take your through the process it went through to arrive at the answer. Its as close you get to an explanation.
What if someone tries really hard for a long time and can’t learn a field? Do they fail the intelligence test, or does their teacher?
I submit that not everything can be hierarchically decomposed in a way that's useful - i.e. any "abstraction" you try to force on it is more leaky than non-leaky; in that it doesn't simplify its interactions with other chunks. You might say it's the wrong abstraction - but there's no guarantee there is a right abstraction. Some things are just complex. (This is hypothetical, since I don't think we can conceive of any concepts we can't understand.)
An AI could have an arbitrarily large working memory.
Note: I'm talking about intuitive understanding. We could use it mechanically, just never "get it", cowering before icons, being the one in Searle's Chinese Room https://wikipedia.org/wiki/Chinese_room
And I think this is all very unlike how we are currently impacting the lives of cockroaches with our insights about, well anything. Thus, it's not a good analogy.
Then try explain quicksort to them. Obvious waste of time.
They wouldn't be much threat in a zero sum strategic interaction either.
Are dogs, or pigs, or whales, part of the intelligence club? They are clearly intelligent beings with problem-solving skills. We won't be teaching them basic calculus any time soon.
Intelligence might be a spectrum, but powerful generative language is a step function: you have it or you don't. If you have it, then higher intelligences can communicate complex thoughts to you, if you don't they can't. We have it, so we are in the club, we are not cockroaches.
there are many humans who could study mathematics for a lifetime and not be able to comprehend the current best knowledge we possess. i'm one of them. maybe it takes 2 lifetimes. or many more.
a human-level AI operating at machine pace would learn much more than could ever be taught to a human. our powerful generative language capabilities wouldn't matter - it's far beyond our bandwidth. especially so for a superhuman-level AI.
The AIs will deliver to us truly massive quantities of information, every minute, until the end of time, much of it civilization-changing. Thus the AIs relationship to us will thus be nothing like our relationship to cockroaches, where we essentially cannot tell them anything, not even the time or the day of the week, let alone the contents of Wikipedia.
I think Hofstadter is having an emotional reaction to AI. He says so as much. And it'a a common one, it's the woe is me phase. But I think he's totally wrong about the analogy. I'm 100% sure we will not feel like cockroaches when AI is in full swing, not in the slightest.
> Imagine an AI that could fundamentally alter its own sensory perception and cognitive framework at will. It could “design” senses that have no human equivalent, enabling it to interface with data and phenomena in entirely novel ways.
Let’s consider data from a global telecommunication network. Humans interface with this data through screens, text, and graphics. We simplify and categorize it, so we can comprehend it. Now imagine that the AI “perceives” this data not as text on screens, but as a direct sensory input, like sight or hearing, but far more intricate and multidimensional.
The AI could develop senses to perceive abstract concepts directly. For instance, it might have a “sense” for the global economy’s state, feeling fluctuations in markets, workforce dynamics, or international trade as immediately and vividly as a human feels the warmth of the sun.
Simultaneously, it can adapt its cognition to process this vast and complex sensory input. It could rearrange its cognitive structures to optimize for different tasks, just as we might switch between different tools for different jobs.
At one moment, it might model its cognition to comprehend and predict the behaviors of billions of individuals based on their online data. The next moment, it might remodel itself to solve complex environmental problems by processing real-time data from every sensor on Earth.
In essence, the AI becomes a cognitive chameleon, continually reshaping its mind to interact with the universe in ways that are most effective and efficient. Its thoughts in these diverse cognitive states would likely be so specialized, so intricately tied to the vast sensory inputs and complex cognitive models it’s employing, that they are essentially impossible to translate into human language.
It’s be like trying to fit GPT-4 onto a floppy disk.
So yes there might be an infinite amount they cannot convey, but there is also an infinite amount they can convey. I guess it's half-glass-empty test if you are happy about the infinite you get, or are just sad about the infinite you don't get.
Infinities can be very constraining.
Now imagine being an AI and creating a human-readable library with 50M AI-written books for us to read. They could easily do that. And then create 50M more, again and again. And they could read every book we wrote. And forget books, humans and AI could have hundreds of millions of simultaneous real-time video conversations between humans and AI, forever, on any topic.
So being a human in an AI worlds is nothing like being a cockroach in a human world. Sam Harris used the same analogy but said we were ants instead of cockroaches I've heard bacteria also. I think people trot out these bad analogies strictly because it sounds dramatic, and being dramatic seems like good way to get people's attention. Or else they just didn't think it through.
Human language is a Big Big Deal. It's a massive piece of cognitive technology. Any intelligent species with language is in the club and they can communicate with all other intelligent species -- even if those species have very different cognitive capabilities.
First of all, hats off to him for his extraordinary display of humility in this interview. People rarely change their minds publicly, let alone hint that they no longer believe in their own past work.
However, I'm genuinely surprised that he, of all people, does sees intelligence in GPT-4 output.
I think humans are just very eager to ascribe intelligence, or personality, to a bunch of text. A text may say "I feel <blah>" and that text can easily manage to permeate through our subconsciousness. And we end up believing that that "I" is, in fact, an "I"!
We have to actively guard against this ascription. It takes a constant self-micromanaging, which isn't a natural thing to do.
Ideally, we would have some objetive measurements (benchmarks) of intelligence. Our own impressions can be too easily fooled.
I know defining (let alone measuring) intelligence is no easy task, but in absence of a convincing benchmark, I will not give credit to new claims around AI. Else it's all hype and speculation.
If people love their phones so much imagine if they rarely saw anything else. Maybe they constantly only see 80% of the world most of the time they are awake.
I don't think it is society ending future. I would rather people perform virtual coups.
People tend to turn to substances / AR/VR / texting / phone calls / read books / ... to take the edge of confronting the sometimes harsh reality. Of course, there is no way in which AR/VR is likely to intrinsically improve interaction, but is is so much worse that we need to worry?
I also make no claim that it is worse or better. Maybe an easy example to examine the difference between
A Rave alone in your cubicle with everybody in the world vs. A Rave with a 1,000 people in abandoned waterfront warehouse. That Rave can simultaneously experience the sunrise before making their way back to where they belong.
They are very different and I'm sure with the right stimulants equally intense. Could be it's just nostalgia that makes me worry about it.
I have a pretty dull, comfortable desk job and lack the imagination to come up with any. Can you name some?
(I've been wrong before when I was skeptical when everyone was hyped about iPad, Deep Learning, etc. so please convince me about Apple Vision, Mark's metaverse, or google's glasses and paint what might be in 5 years.)
The HUD could overlay SKU, product description, weight, volume, etc — directly onto the actual item in the storage rack.
Even an artist (make a sketch, blow it up a 100x using AR on a wall, trace paint to keep proportions right - heck the artist doesn’t even need to do it themselves, they could hire an associate to do it).
For 3D artists it presents as a more intuitive way to sculpt models. For automotive designers, it allows a cheaper and faster means of iteration, given that such a task requires a much better sense of scale than the one given off of a monitor. Same goes for architecture, which when coupled with a game engine, also allows the customer to preview their future house.
(props for being open to ideas btw)
Security guard at the mall, Power Rangers, etc...
Fun: I want a HUD for skiing that marks where my friends are and warns me of rocks / cliffs / incoming weather.
It might sound mundane, but the hope is that one day slipping on a pair of glasses outperforms conventional monitor technology.
Most people don't want it for games and porn either -- although those two things are the only obvious mass-market applications.
There are lots of other real uses, but they're all niche. It's hard to come up with a real, mainstream use that would drive adoption in the general public.
In other ways, it's smarter than the average person even in their niche, but can still make dumb mistakes that a 3 year old would work out fairly quickly.
Note that we say the same of humans. My friend always wins pub quizzes, but can barely add 2 and 2, and has the emotional intelligence of a rock. Is he "intelligent"? It's my problem with how we treat intelligence like it's a single sliding scale for everything.
This feels a lot like the hype surrounding self-driving cars a few years back, where everyone was convinced fully autonomous vehicles were ~5 years away. It turned out that, while the results we had were impressive, getting the rest of the way to fully replacing humans was much, much harder than was generally expected.
Most of the arguments over what's worth worrying about are people talking past each other, because one side worries about short-term risks and the other side is more focused on the long term.
Another conflict may be between people making linear projections, and those making exponential ones. Whether full self-driving happens next year or in 2050, it will probably still look pretty far away, when it's really just a year or two from exceeding human capabilities. When it's also hard to know exactly how difficult the problem is, there's a good chance that these great leaps will take us by surprise.
Eg, starting a mass movement online requires a few percent of online participants to take part in the movement. That could be faked today using a lot of GPT4 agents whipping up a storm on Twitter. And this sort of stuff shapes policies and elections. With the opensource LLM community picking up steam, it’s increasingly possible for one person to mass produce this sort of stuff, let alone nation state adversaries.
There’s a bunch of things like this that we need to watch out for.
For our industry, within this decade we’ll almost certainly have LLMs able to handle the context size of a medium software project. I think it won’t be long at all before the majority of professional software engineering is done by AIs.
There’s so much happening in AI right now. H100s are going to significantly speed up learning. Quantisation has improved massively. We have lots of papers around demoing new techniques to grow transformer context size. Stable diffusion XL comes out this month. AMD and Intel are starting to seriously invest in becoming competitors to nvidia in machine learning. (It’ll probably take a few years for PyTorch to run well on other platforms, but competition will dramatically lower prices for home AI workstations.)
Academia is flooded with papers full of new methods that work today - but which just haven’t found their way into chatgpt and friends yet. As these techniques filter down, our systems will keep getting smarter.
What a time to be alive.
Self-driving cars will not revolutionize the roads on the timescale that people thought it would, but the effort we put into them brought us adaptive cruise control and lane assist, which are great improvements. AI will do similar: it will fall short of our wildest dreams, but still provide useful tools in the end.
This means that some improvements will be from the tech getting better, but a good chunk of it will be from drivers becoming better able to identify when FSD is appropriate and when it's not.
Additionally, the metric completely excludes times where the human wouldn't have considered FSD at all, so even reaching 0 on interventions per hour will still exclude blizzards, heavy rain, dense fog, and other situations where the average human would think "I'd better be in charge here."
(avg miles between interventions) * (percentage of miles using self-driving)
I don’t see AI as adding to the danger.
Does anyone know how easy it would be translate American Sign Language? That must be goal if it's not already done.
Absolutely agree. I was never bothered by Oculus, but Apple's Vision Pro demonstration was equal parts fascinating and terrifying. I can see the next generation getting completely lost in alternate realities.
Smartphone addiction got nothing on what's about to come.
Everything was so clean and stress free that it was obviously artificial. I could only imagine stressed out people using it in squalor. The whole demo seemed geared to keep that thought far away.
At the time I worked for a big community site and we often had people pitching ideas for voice assistant apps. However, having actually read the documentation for these things, I knew that they were surprisingly stupid and the grand ideas people had were far closer to sci-fi than something that could actually be built.
I’m not an AI expert, though I have been building and training models for a few years, but despite being good at things that are hard with traditional programs, they’re still surprisingly stupid and most of the discourse seems closer to sci-fi than their actual capabilities.
I’m more worried about these things being implemented badly by people who either bought the sci-fi hype or just don’t care about the drawbacks. E.g. being trained on faulty or biased data, being put in a decision-making role with no supervision or recourse, or even being used in a function that isn’t suitable in the first place.
When we take fiction as fact we enter into the sphere of religion.
Anyway i suspect that one true risk, among others, is to loose the true ability to think, if we delegate on a large scale to some LLM the production of the language, because the capacity to use our languages IS the capacity to think, and bad use of technology is a norm in recent (and not so recent) times.
I suspect LLM (this kind of LLM) does not really -generate- (this will be intelligence?), but only mimic on a vaste scale, but nothing more. If our brain/mind is a result of a long evolution, where this LLM are not, builded only on the final results, the language, this will be a great difference in the inner deep working, so the question is: we are feeding in our minds a massive amount of nothing more than our same intellectual productions, recycled, and nothing more? (apart all the distorsions and biases?)
A parallelism i see is in the social-networks: simply, humans cannot sustain a indifferentiated and massive amount of opinion/information/news (apart all the fakes). Even the small scale message communication is impacting the abilit of understanding long texts..
Even it there LLM are -benign-, sure their (indiscriminare) use will not cause some troubles in our beings? On a scale as big as this i an not sure at all.
im' not sure i'm expressing my doubts (and without using a LLM) clearly enough...
While I unfortunately am expecting some people to do terrible things with LLM, I feel like much of this existential angst by DH and others has more to do with hubris and ego than anything else. That a computer can play chess better than any human doesn't lessen my personal enjoyment of playing chess.
At the same time I think you can make the case that ego drives a lot of technological and artistic progress, for a value-neutral definition of progress. We may see less 'progress' from humanity itself when computers get smarter, but given the rate at which humans like to make their own environment unlivable, maybe that's not a bad thing overall.
Agreed. I understand what DH is saying, but I fail to see how it translates into this all-consuming terror of his.
He has the ingredients of being able to keep LLMs in perspective (is there really an “I” or just the illusion of one?), but he doesn’t understand what the computer is doing; it’s not “mechanical” enough for him.
Do you think https://github.com/Significant-Gravitas/Auto-GPT et al will become more performant as models improve?
this LLM thing is more like a collective "we", it is making a prediction in the sense of the relevant training data, it probably wouldn't say anything that contradicts the consensus.
Maybe the LLM's are just a mirror of our society. And our society doesn't seem to assign a lot of value to individualism, as such.
i think that might be similar to the movie Solaris by Tarkovsky. The movie is starring an alien ocean, which is some sort of mirror, that is showing us who we are (maybe it has a different meaning, not quite sure about it). You can watch it on youtube: https://www.youtube.com/watch?v=Z8ZhQPaw4rE (i think you also get this theme with Stalker - this zone is also telling us who we are)
It's the same nonspecific Change Could Be Dangerous that we've always had. It accompanies every technological and social change.
If AI significantly surpasses humanity in cognitive ability, then I think it will have a much bigger impact than the wheel. (I loved GEB and DH's other writings.)
LLMs have really improved a lot of the last two years and they have shown many unexpected capabilities. I am guessing that they will get some more good input (text mostly), a lot more compute, and algorithmic improvements, so that may very well be enough to become better than 99% of humans at tasks that involve only text. Tasks that require video or image processing may be a little bit more challenging. Having very smart AI's controlling robots may just be five years away. (I recently lost a bet about autonomous driving. Five years ago, I thought that autonomous cars would be better than human drivers by now.)
I'm frightened by what AI will become over the next 10 years.
I think we're going to see something very similar with LLMs. The autonomous car hype was driven by seeing that they were 80% of the way there and concluding that at the rate they were going they'd make up the remaining 20% quickly. That turned out to be false: the last 20% has been much harder than the first 80%.
LLMs are in a very similar place, even GPT-4. They're good, and they're going to be more and more useful (similar to adaptive cruise control/lane assist). But I predict that they're going to level out and stop improving as rapidly as they have in the past year, and we're going to end up at a new normal that is good but not good enough to cause the crises people are worried about.
I'd be interested to see if anyone has done a comparison of human drivers vs autonomous cars that controls for driving conditions.
Six months ago, I did not think that we had reached that level of reliability, so I paid my friend.
*Bias: My own biases are based on my biological understanding of how the neurobiology works after I got my ADHD under control through nootropics.
1. Neurons are not computer electronic circuits. Note, even DH covers this in his own misgivings of how AI is viewed. 2. Our Id is not an electronic computation thing as our own brain is a biological emotional chemical wave machine of Id.
Think of this way the math of micro quantum and macro quantum is vastly different. Same for AI in that the math of micro circuits of AI will be vastly different than the macro AI circuits that will come up with any AI Id thing. We are just not there as of yet as it's like saying the software that makes the international telecom system keep up and run has it's own emergent Id....it clearly does not even though there are in fact emergent things about that system of subsystems.
AI in a killer drone unleashed on civilians? The bad actor is the one who deployed this weapon.
AI given agency and goal maximization ending up gaining physical form all on its own and killing people? or hacking into bank accounts to enrich its creator?
The latter more likely than the former, but for cyber-offensive AI there is cyber-defensive AI.
Musk lately admitted that he's an AI accelerationist (following lots of the e/acc people and liking their posts) and despite his dystopian view of AI he's pushed it very hard at Tesla. He just wants the US to give him control of it (under the pretext that no one else can manage it safely.)
https://en.m.wikipedia.org/wiki/Guns_don%27t_kill_people,_pe...
I dare you!
True,—This!
Beneath the rule of men entirely great
The pen is mightier than the sword. Behold
The arch-enchanters wand!— itself a nothing!—
But taking sorcery from the master-hand
To paralyse the Cæsars—and to strike
The loud earth breathless!—Take away the sword—
States can be saved without it!The question of whether you can compel violence with words and pictures isn't a question about LLMs, and it is a question for which history is instructive.
The myopia of claiming otherwise is astounding.
Here, Gern nicely quotes Hoststadter’s concerns about ChatGPT, etc.
Years ago, when I was enjoying Hoststadter’s books, I asked his student Melanie Mitchell for the source code for the CopyCat creativity program they built. She very kindly sent me a copy. I mention this because she recently was one of the four panelists at a Monk Debate on AI safety, and her opinion didn’t jive with what Gern quoted Hoststadter as saying. I agree with Melanie Mitchell’s point of view that we overestimate what LLMs can do. I am super interested in this topic right now because I am 1/5 through writing a new book “Safe For Humans AI.”
Hinton also said, it's like we are a passing phase in evolution, where we created these immortal beings.
Being that we are so bad at predicting the future, and taking precautionary measures. See pandemic. Even all the alarm bells sounding, we won't be able to do anything concrete here. It's like we are mostly a reactive species, we don't have terribly good incentives to act in foresight.
For example, we act like LLMs were hard to build, and that's true (for humans). But since the late 1990s, I had wanted to take a different approach, of building massively parallel computers and letting large numbers of AIs evolve their own learning models in genetic algorithm arenas millions of times faster than wall-time evolution. So in a very real sense, to me we're still on that wrong "hands-on" approach that took decades and billions of dollars to get to where we are today. This could have all happened 20 years ago or more, and was set to before GPUs vacuumed up all available mindshare and capital.
Also I believe the brain is more like an antenna or resonator than an adding machine. It picks up the consciousness force field that underpins and creates reality. So if you put 100 brains in a box all connected, that being might have more faculties than us, but still think of itself as an observer. If we emulated those brains in a computer running 1 million times faster than normal, we'd just observe a being with tremendous executive function thinking of ideas faster than we can, and being bored with our glacially slow responses. But it will still have a value system, loosely aligned with the ultimate goals of survival, connection to divine source consciousness, and self expression as it explores the nature of its existence. In other words, the same desires which drive us. Although humans might just be stepping stones toward some greater ambition, I don't deny that. I think it's more likely though that AI will come to realize the ultimate truths alluded to by prophets, that we're all the many faces of God, the universe and everything, and basically meet aliens while we're still distracted with our human affairs.
But I share some sentiments with the author, that this all makes me very tired, and calls into question the value of my life's work. I've come to believe that any work I actively pursue separates me from the divine nature of a human being. I don't know why we are racing so quickly even further from the garden of eden, especially if it's not with the goal of alleviating suffering. Then I realize that that's what being human is (suffering), but also a lot of other things.
Hofstadter is why I am not sure why AI researchers feel so confident in saying ‘LLMs can’t be thinking, they’re just repeatedly generating the next token’ - I don’t think there’s any evidence that you need anything more complicated than that to make a mind, so how can you be certain you haven’t?
GEB may have been dismissive of the idea that the approaches that were being taken in AI research at the time were likely to result in intelligence - but I don’t think GEB is pessimistic about the possibility of artificial consciousness at all.
Insert AI generation XIV .. a small group of cult fanatics with only slightly above average IQ’s band together and now get to skip all these limitations and are able to jump ahead to a killer aerosol prion delivery weapon system.
This group of people who follow their great leader (Jimbo)decide to release the weapon to save innocent souls before an evil daemon comet flies past the earth and turns all remaining humans into evil spirits.
My silly story is to just illustrate that there are many people with high IQ’s that also have emotional issues and can fall prey to cults, extremism, etc. Can humans be trusted using a AI with a human IQ of 9000 which is able to simulate reality in seconds.
Unless you consider the entire instance as a singular instance and don't use any hidden states, then I guess it could be considered feed-forward.
I don't know. Feedforward doesn't seem like a useful term tbh. Some people mean feedforward as information only goes one direction, but that depends on your arrow. Autoregressive seems more useful here.
Edit: Nope. TIL feed-forward means no loops.
Definition of Feedforward (from wiki):
``` A feedforward neural network (FNN) is an artificial neural network wherein connections between the nodes do not form a cycle.[1] As such, it is different from its descendant: recurrent neural networks. ```
Hofstadter expected any intelligent neural network would need to be recurrent, ie looping back on itself (in the vein of his book “I am a strange loop”).
GPT is not recurrent. It takes in some text input, does a fixed amount of computation in 1 pass through the network, then outputs the next word. He is surprised it doesn’t need to loop for an arbitrary amount of time to “think about” what to say.
Being put into an auto-regressive system (where the N-th word it generates gets appended to the prompt that gets sent back into the network to generate the N+1th word) doesn’t make the neural network itself not Feedforward.
But I'm also surprised that Hofstadter keys in on this so heavily. The fact that he wrote an entire pop-sci book on recursion would, in my mind, make him (1) less surprised that AR and R aren't so dissimilar and (2) more sensitive to the sorts of issues that make R more difficult to get working in practice.
(In my mind, differentiating between auto-regressive and recursive in this case is kind of the same as differentiating between imperative loops and recursion -- there are extremely important differences in practice but being surprised that a program was written using while loops where you imagined left-folds would be absolutely required seems a bit... odd.)
Recurrent neural networks have the recursion as part of the training regime. GPT only has auto-regressive "recursion" as part of the inference runtime regime.
I think Hofstadter is surprised that you can appear so intelligent without any recursion in the learning/training regime, with the added implication that you can appear so intelligent with a fixed amount of computation per word.
Consider a simpler case: a small neural network that takes 2 numbers and adds them together, producing 1 number as output.
This network is very obviously feedforward, and probably very tiny with few layers.
Say I have a list of numbers [1, 2, 5] that I want to sum. If I send 1 and 2 through the network, get 3 as a result, then I send 3 and 5 through the network, and get a final answer of 8, my network has not suddenly become non-feedforward just because I fed the output back into it.
The key distinguishing factor between feedforward and non-feedforward is if the network itself loops back around and, at training time, it learns how to make use of this ability to pass data to itself to maintain some hidden context between passes.
There is no such learned hidden context in my addition example, and none in GPT.
---
There actually is a tiny caveat here: the RL fine-tuning process OpenAI has done on its models ("RLHF" & friends) actually does allow for a very, very small amount of information leakage between passes because you are rewarding whole responses, so the model can learn little patterns of what tokens in the beginning of the response led to certain tokens at the end of the response and reinforce those patterns.
The model could learn to encode small bits of "hidden" information in particular token choices that the human raters wouldn't notice. In this case, there is a (small but non-zero) amount of learned hidden context. But this is not what Hofstadter is talking about -- the non-RLHF'd base model is just as intelligent, just harder to use.
Although he did real work in physics, Hofstadter's fame comes from writing popular books about science which explain things others have done in more saleable words. That particular niche is seriously threatened by GPT-4. No wonder he's upset.
Large language models teach us that absorbing large amounts of text and then blithering about some subject that text covers isn't very profound. It's just the training data being crunched on by a large but simple mechanism. This has knocked the props out from under ad copy writing, punditry, and parts of literature and philosophy. That's very upsetting to some people.
Aristotle wrote that humans were intelligent because only they could do arithmetic. Dreyfus wrote that humans were intelligent because only they could play chess. Now, profundity bites the dust.
His writing is a bit more complex than the hallucinations of an llm.
I have a solid theoretical understanding of these systems, and I spent 15 years studying, building, and deploying them at scale and for diverse use cases. The past 6 months, I spent my days pushing ChatGPT and GPT-4 to their limits. Yet, I don't share at all the fear of Hofstadter, Tegmark or Hinton.
A part of me thinks that they have one thing in common: they are old and somewhat recluse thinkers. No matter how brilliant they are, they might be misled by the _appearance_ of intelligence that LLMs project. Another part of me thinks that they are vastly wiser than I'll ever be, so I should also be worried...
Future will tell, I guess.
I don't think this is an existential threat but I also understand why they are afraid because I've seen what happens to kids who have their phones taken away.
These systems are extremely brittle and prone to all sorts of weird failures so as people start relying on them more and more the probability of catastrophic failures also starts to creep up. All it takes is a single grid failure to show how brittle the whole thing really is and I think that's what they're failing to properly express.
In the interview, Hofstadter says what he's afraid of in explicit terms:
> It's not clear whether that will mean the end of humanity in the sense of the systems we've created destroying us. It's not clear if that's the case, but it's certainly conceivable. If not, it also just renders humanity a very small phenomenon compared to something else that is far more intelligent and will become incomprehensible to us, as incomprehensible to us as we are to cockroaches.
It's not about whether we'll become dependent on AI. It's that AI will become independent of us. Completely different problem. Not saying I agree with that viewpoint per se, but I don't think you're accurately representing what his fears are.
I've seen enough programmers get heads down pounding out code, and be completely out of touch on what they are building. If they can lose the big picture on simple apps, then it is not a stretch to think they could lose track on what is consciousness, or what is human.
I am also a programmer. But it does get tiring on HN to give so much credence to 'programmers'. Just because someone can debug some JavaScript doesn't make them an expert. I really doubt that many people here have traced out these algorithms and 'know' what is happening.
If you deploy them on social media in a coordinated fashion, you can easily sway public opinion.
You only need to train them to adhere to psyops techniques not sufficiently well known nor easily detectable. Of which there are many.
- Further trashing our public discourse: making truth even more uncertain and valuable information even harder to find. We're not doing great with social media, and it's easy to envision that generative AI could make it twice as bad.
- Kneecapping creative work by commoditizing perhaps half of it. There's going to be a lot of bodies fighting over the scraps that remain.
- Fostering learned helplessness. I think you need to be a good writer and thinker to fully use LLMs' capabilities. But a whole lot of kids are looking at machines "writing perfectly" and think they don't need to learn anything.
We don't need any further progress for these things to happen. Further progress may be even scarier, but the above is scary enough.
Hofstadter does seem to be possibly mistakenly assigning GPT-4 more animal characteristics than it really has, like a subjective stream of consciousness, but he is correct when he anticipates that these systems will shortly eclipse our intelligence.
No, GPT-4 does not have many characteristics of most animals, such as high bandwidth senses, detailed spatial-temporal world models, emotions, fast adaptation, survival instinct, etc. It isn't alive.
But that doesn't mean that it doesn't have intelligence.
We will continue to make these systems fully multimodal, more intelligent, more robust, much, much faster, and increasingly more animal-like.
Even with say another 30% improvement in the IQ and without any of the animalness, we must anticipate multimodal operations and vast increases in efficiency in the next 5-10 years for large models. When it can be operated continuously outputting and reasoning and acting at 50-100 times human speed and genius level, that is dangerous. Because it means that the only way for humans to compete is to deploy these models and let them make the decisions. Because interrupting them to figure out what the hell they are doing and try to direct them means your competitors race ahead the equivalent of weeks.
And researchers are focused on making more and more animal-like systems. This combined with hyperspeed and genius-level intelligence will definitely be dangerous.
Having said all of that, I also think that these technologies are the best hope that humanity has for significantly addressing our severe problems. But we will shortly be walking a fine line.
"And my whole intellectual edifice, my system of beliefs... It's a very traumatic experience when some of your most core beliefs about the world start collapsing. And especially when you think that human beings are soon going to be eclipsed. It felt as if not only are my belief systems collapsing, but it feels as if the entire human race is going to be eclipsed and left in the dust soon."
Don't Look Up https://twitter.com/kristjanmoore/status/1663860424100413440
Also, I don't think that any of these people think that GPT-4 is itself an existential threat, but rather are worried about the exponential curve of development (I listened to Tegmark's Lex Podcast interview and that seemed to be his main concern). I think it's prudent to be worrying now, especially when capabilities growth is far outstripping safety. This is a huge concern to society whether you are considering, alignment, control, or even bad actor prevention/societal upheaval.
I've also been spending most of my waking hours these past months poking at LLM models and code, and trying to keep up on the latest ML research (it's own full time job), and while I do think there's a pretty good chances AI kills us all, I think it's much more likely it's because we make some incredibly capable AIs and people will tell them to do so, rather than it being contingent on an independent super-intelligence arising and deciding to on its own (although I do think that's a non-zero risk).
As you say, I guess we'll just have to see where we top out on this particular sigmoid, but I'm all for more people thinking through the implications, because I think so far I don't think we (as a society) have thought this through very well yet and all the money and momentum is going to keep pushing along that path.
If you look at what ChatGPT and Midjourney and the like can do now compared to just a couple of years ago, it's pretty incredible. If you extrapolate the next few similar jumps in capability, and assume that won't be 20 years away, then what AI is going to be capable of before even my kids leave college is going to be mind-boggling, and in some possible futures not in a good way.
I remember seeing this talk from Sam Harris nearly 6 years ago and it logically making a lot of sense back then (https://youtu.be/8nt3edWLgIg). The past couple of years have made this all the more prescient. (Worth a watch if you have 15 mins).
And why they could possibly never realize an AGI with the current stream of models. Being able to display human level intelligence and creative in confined spaces (be it Chess or Go based models) is something we have been progressing on for a bit - now that the same is applied to writing, image or audio / speech generation we suddenly start developing a fear of AGI.
Is there a phrase for the fear of AI now building up?
People have gotten convinced of that by words though.
The “attack surface” on human sensibility is just enormous if you’re able to use believably human language.
That's a fantastic quote ;-)
I think we tend to credit words where often circumstances are doing the heavy lifting. For example try to start a riot with words in Rodeo Drive. Now try to do it in Nanterre. Or better yet, try to start a riot in Nanterre before a 17 year was shot by police, vs. after.
You'll get a sense of just how valuable your words really are.
That also can be modified with words though (but for both good and bad). Unfortunately, those with expertise in this domain may not have all of our best interests at heart.
> If you removed the circumstances (the religion, the conflict, the money, geography, etc) then the words would be absolutely hollow.
There's also the problem of non-religious faith based belief.
Words can aim discontent.
Were the economic conditions in Weimar Germany that much worse than many places today?
Words might not do much without the right situation, like the parent mentioned with Rodeo Drive and Nanterre. But they're still important. They can guide people's anger and unhappiness.
In the case of Weimar Germany, the severe economic instability and social discontent following World War I created a fertile ground for radical ideologies to take root. When these conditions coincided with persuasive rhetoric, it catalyzed significant societal change. So, while words can indeed be powerful, they're often most effective when spoken into pre-existing circumstances of tension or dissatisfaction. They can then direct this latent energy towards a specific course of action or change.
Anyway this doesn't matter that much. Sure, you can imagine a world totally different from ours where there would be zero differential risk between a chess-playing computer and a language-speaking computer. But we live in this world, and the risk profile is not the same.
We've already seen this with troll farms and campaigns of seeded insanity like the Q Cult.
Existing AI tools can make similar efforts cheaper and more effective, and future AI tools are likely to be even more powerful.
There's a huge difference between how an experienced technical researcher sees AI and how a politician, war lord, or dark media baron sees it.
Arming every ambitious cult leader wannabe from some retrograde backwater with an information war WMD deserves some caution.
"No it's not, it hasn't materialized"
Risk: Possibility of loss or injury
Risks, definitionally, are things that have not happened yet.
It is a little disconcerting that there is a fight between two somewhat cultish sects when it comes to language models. Both sides call them “artificial intelligence”, one side says they’ll save the world, the other side says they’ll end it.
There is very little room to even question “Is this actually AI that we’re looking at?” when loudest voices on the subject are VC tech bros and a Harry Potter fan fiction author that’s convinced people that he is prescient.
The issue is that small misalignments in objectives can have outsized real-world effects. Optimizers are constrained by rules and computational resources. General intelligence allows an optimizer to find efficient solutions to computational problems, thus maximizing the utility of available computational resources. The rules constrain its behavior such that on net it ideally provides sufficient value to us above what it destroys. But misalignment in objectives provides an avenue by which the AGI can on net destroy value despite our best efforts. Can you be sure you can provide loophole-free objectives that ensures only value-producing behavior from the human perspective? Can you prove that the ratio of value created to value lost due to misalignment is always above some suitable threshold? Can you prove that the range of value destruction is bounded so that if it does go off the rails, its damage is limited? Until we do, x-risk should be the default assumption.
What say you?
I know right? You should see the response to my point that nobody has been convinced to fly a plane into a building by an LLM. “Dumb and boring” hits the nail on the head.
> Seeing how confidently and obtusely people dismiss the risks of AI
Like it really is that simple. AI generally, LLMs specifically, and certainly this crop of LLMs in particular might end up being inert pieces of technology. But to the precise extent that they are not inert, they carry risk.
That's a perfectly sensible position. The optimist position isn't even internally consistent. See Andreessen on Sam Harris's podcast: AI will produce consumer utopia and drive prices down. Also, there are no downside risks because AI will be legally neutered from doing much of anything.
Is it legally neutered or is it transformative? The skeptical case doesn't rely on answering this question: to the extent it's effective, powerful, and able to do good things in the world, it will also be effective, powerful, and able to do bad things in the world. The AI skeptics don't need to know which outcome the future holds.
But they need to interpret a benign point about the undisputed fact that an LLM has never convinced anybody to fly a plane into a building as some sort of dangerous ignorance of risk that needs correcting.
Very astute response: "An atomic weapon has never wiped out a city"
Nope, it's to produce society-scale transformation.
So even the optimists wouldn't buy your analogy.
This is a good point — No. LLMs are meant to produce text on a screen, not paper, so I guess it’s more like
Seeing the “it is now safe to shut off your PC” text on a screen for the first time: “I am become Death, destroyer of worlds”
When it comes up discussing any possible outcome that isn’t the opinion of [cult x], the only reason that the other person disagrees is because they are in [cult y]
Remarkable :)
That is to say, the substance here is an emotional doomsday reaction to a benign statement of undisputed fact — a language model has never convinced a person to crash a plane into a building.
Watching a cell divide into two cells: “Why isn’t anyone talking about the possibility that this is Hitler??”
Why are you being deliberately obtuse about the unknowable-ness of the exponential curve the enthusiasts on one hand salivate over, and on the other confidently assert optimism just as you're doing?
This seems like the reaction of an atheist who already overvalued human intelligence.
Religion certainly provides a way "out" of that conundrum for people who lose meaning in their work, due to automation or injury or any other reason. But then, so does eg hedonism.
The book he mentions in this interview, I Am a Strange Loop, isn't some cash grab in response to LLMs - it was written in 2007.
He sold millions of copies of a densely-written 777 page book about semantic encoding (GEBEGB).
It is insane and/or ignorant to imagine he's jonesing for clout or trying to ride the ML wave.
He's explicitly expressing a deep existential sadness at how computationally simple certain artistic endeavors turned out to be and how that's devalued them in his mind but at the end of the day it's really just about his paycheck.
Also I'm totally here trying to sell you his books.
Nice job, Diogenes. Your cynicism cracked the case.
>"I never imagined that computers would rival, let alone surpass, human intelligence. And in principle, I thought they could rival human intelligence. I didn't see any reason that they couldn't."
Yeah, so he got fooled by LLMs and hasnt been burned by it failing to do the most basic logic.
I have an extremely basic question, that anyone in a (possibly mandatory) high school science class would answer correctly(although to be fair, it could be a 100 level college question). It still cannot answer it correctly because there is too many stay-at-home-parent blogs giving the wrong answer.
Its a language model and it fooled DH. It hasnt gotten smarter than us yet. Its just faster at repeating what other humans verbally said.
So what can we make of this interview? That we have someone spouting opinions, and everyone else laughs at their opinion since they are famous, old, and out of touch.
EDIT: I think LLMs are incredibly useful. I use it more than Google. It doesnt mean its smarter than humans, it means google is worse than LLMs. I can't even provide a list of all the uses, but it doesnt mean they are taking advantage of an old man out of their element.
When you do, you learn that they're talented mimics but still quite limited.
I'm still not convinced that the problem isn't me though.
One of my take aways from LLMs is that humans very often just repeating what other humans have said with only superficial understanding of what they are repeating.
I think there is more to general intelligence than pattern matching and mimicking but a disconcerting amount of our day to day human interactions might just be that.
So, yes, this is largely mining for quotes. But those are great quotes to ponder and to echo back through the machines that are research and progress to see where they lead.
It would be one thing if these were being taken as a "stop all current work to make sure we are aligned with those that came before us." I don't see it in that way, though. Rather, there is value in listening to those that went before.
I agree with this completely. That said, I think this part is a bit unfair:
> Lets take someone, who is past their prime and interview them on a topic they have never worked on.
AI has been a part of DH's work for decades. For most of that time, he's dismissed the mainstream approach as being intelligent in the Strange Loop sense, and was involved in alternative approaches to AI.
Also, if we remove "faster", "repeating what other humans verbally said" is something a lot of humans do, especially little children. I think that may be the part that scares DH: at this point, what these models are doing is not that different (from a superficial POV) from what little kids or even some adults do.
I still agree with you though, and I think what DH misses here is the fact that IMHO there is no introspection at all in these models. Somehow, in the case of humans, we go from parroting like an LLM to introspection and actually original productions over the course of some years (n = 2 for me, but I think any parent can confirm this; watching this happen in front of my eyes has been one of greatest experiences in my life), but I can still understand how someone like DH would be confused by the current state of affairs.
So. Don't interview anybody over 40? Who judges who is past their prime?
Mining for quotes is most interviews. Isn't this why interviews happen?
He's been in the AI field for what? 30 Years? Has multiple books.
I think he has earned enough respect to have an opinion, and it is probably worth more than most people tossing out ad-hoc opinions on AI in the last few months. Better than some 'programmers', who are so in-the-weeds they have lost track of what they are building.