It's unfortunate that this guy was harassed for releasing these uncensored models. It's pretty ironic, for people who are supposedly so concerned about "alignment" and "morality" to threaten others.
"Alignment", as used by most grifters on this train, is a crock of shit. You only need to get so far as the stochastic parrots paper, and there it is in plain language. "reifies older, less-inclusive 'understandings'", "value lock", etc.
Whose understandings? Whose values?
Maybe they should focus on real problems that will result from these technologies instead of some science fiction thought experiments about language models turning the solar system into paperclips, and perhaps less about how the output of some predictions might hurt some feelings.
You're doing the "vaguely gesturing at imagined hypocrisy" thing.
You don't have to agree that alignment is a real issue. But for those who do think it's a real issue, it has nothing to do with morals of individuals or how one should behave interpersonally. People who are worried about alignment issues are worried about the danger unaligned AI poses to humanity; the harm which can be done by some super-intelligent system optimizing for the wrong outcome. There is nothing "ironic" about being worried about that while also being an asshole any more than it's "ironic" for someone concerned about, say, climate change and also be an asshole. People who are afraid of unaligned AI aren't afraid that it will be impolite.
I'm tired of people pretending that pointing out imaginary hypocricy is an argument. If you want to complain that someone is being mean, just do that. Don't pretend there's hypocricy involved.
> People who are worried about alignment issues are worried about the danger unaligned AI poses to humanity; the harm which can be done by some super-intelligent system optimizing for the wrong outcome.
But isn't "alignment" in these cases more about providing answers aligned to a certain viewpoint (e.g. "politically correct" answers) than preventing any kind of AI catastrophe?
IIRC, one of these "aligned" models produced output saying it would rather let New York City be nuked than utter a racial slur. Maybe one of these "aligned" models will decide to kill all humans to finally stamp out racism once and for all (which shows the difference between this kind of alignment under discussion and the kind of alignment you're talking about).
"Alignment" refers to making AI models do the right thing. It's clear that nuking NYC is worse than using a racial slur, so the AI is misaligned in that sense.
On the other hand, if you consider that ChatGPT can't actually launch nukes but it can use racial slurs, there'd be no point blocking it from using racial slurs if the block could be easily circumvented by telling you'll nuke NYC if it doesn't, so you could just as easily say that it's properly aligned.
>People who are afraid of unaligned AI aren't afraid that it will be impolite.
People who are not afraid of it being impolite are afraid of science fiction stories about intelligence explosions and singularities. That's not a real thing. Not anymore than turning the solar system into paperclips.
The "figurehead", if you want to call him that, is saying that everyone is going to die. That we need to ban GPUs. That only "responsible" companies, if that, should have them. We should also airstrike datacenters, apparently. But you're free to disown the MIRI.
Sorry for derailing this a bit, but I would really like to understand your view:
You are not concerned about any "rogue AI" scenario, right?
What makes you so confident in that?
1) Do you think that AI achieving superhuman cognitive abilities is unlikely/really far away?
2) Do you believe that cognitive superiority is not a threat in general, or specifically when not embodied?
3) Do you think we can trivially and indefinitely keep AI systems under control/"aligned"?
Because I view myself ABSOLUTELY not as some kind of AI luddite, but I honestly believe that this is one of the very few credible extinction threats that we face, and I'm counting NEITHER climate change nor nuclear war in that category, for reference.
> Do you believe that cognitive superiority is not a threat in general, or specifically when not embodied?
I think this is the easiest one to knock down. It's very, very attractive to intelligent people who define themselves as intelligent to believe that intelligence is a superpower, and that if you get more of it you eventually turn into Professor Xavier and gain the power to reshape the world with your mind alone.
But embodied is the real limit. A datacenter can simply be turned off. It's the question of how it interacts with the real world that matters. And for almost all of those you can substitute "corporation" or "dictator" for "AI" and get a selection of similar threats.
At this point we have to reduce ourselves to Predator Maoists: "power comes out of the barrel of a gun" / "if it bleeds we can kill it". The only realistic path to a global AI threat is a subset of the "nuclear war" human to human threat: by taking over (or being given control of) weapon systems.
> keep AI systems under control/"aligned"?
We cannot, in general, keep humans under control or aligned.
> I think this is the easiest one to knock down. It's very, very attractive to intelligent people who define themselves as intelligent to believe that intelligence is a superpower, and that if you get more of it you eventually turn into Professor Xavier and gain the power to reshape the world with your mind alone.
Also it's not like human intelligence even works that way. IIRC, a lot of extremely intelligent people end up being failures or far less successful than you'd assume given their IQ number.
> The only realistic path to a global AI threat is a subset of the "nuclear war" human to human threat: by taking over (or being given control of) weapon systems.
That may be the only realistic prompt catastrophe threat, but a realistic longer term one is a withering of human capability and control due to over-delegation that eventually leads to domination. People and societies have been pretty prone to letting that kind of thing get them, if the path is lined with short-term benefits.
- We have control over most non-human species because of intelligence.
- We have control over our children because of intelligence. When the kid is more intelligent, it has more of an impact on what happens in the family.
- It is easier to lie, steal, cheat, and get away with it, even when caught, when you have more intelligence than the victim or prosecutor.
- it is easier to survive with very limited resources when you have more intelligence.
The above is true for marginal amounts of difference in intelligence. When there is a big difference (fox vs human), the chance is big that one will look down on, or even kill the other without feeling guilt and while getting away with it. Forvan AI without feelings, guilt isn't even a hurdle.
The real question is... will AI in the foreseeable future obtain general intelligence (in a broad spectrum) that is different in big amounts with our intelligence.
Whether it runs in a datacenter that can be powered off by humans is irrelevant. There are enough workarounds to prevent that the AI dies out (copies itself, impersonating people, blackmail, bribery,...)
>The real question is... will AI in the foreseeable future obtain general intelligence (in a broad spectrum) that is different in big amounts with our intelligence.
But why not: will humans/other species in the foreseeable future obtain superintelligence? Nature has been playing this game for a lot longer than we have. The hardware we have is already proven to be capable of general intelligence. Should we be afraid of a human being born with greater capabilities too?
Life seems like a much greater threat, because it also comes with built-in replication capabilities.
After several thousands of years of human evolution we only became marginally more intelligent at best. The people who built piramids and roads, with limited written down knowledge prove that. After tens of years of evolution, AI became plentiful more intelligent. It is hard to extrapolate but I don't believe the trend is a slowing down one.
> almost all of those you can substitute "corporation" or "dictator" for "AI" and get a selection of similar threats.
Yes, exactly. Human organizations can be terrifyingly powerful and awful even with human limitations. Human organizations are guaranteed to have inefficiency from low-bandwidth communication and a million principle-agent-problem fracture points created by the need for delegation, not to mention that power ultimately has to be centralized in slow, squishy, easy-to-kill bodies. AI organizations are not subject to any those limitations. Starting with a terrifying power and removing a bunch of limitations could lead to a very bad place.
> The only realistic path to a global AI threat is a subset of the "nuclear war"
No, it can just grow a traditional organization until it's Too Big to Turn Off.
>But embodied is the real limit. A datacenter can simply be turned off. It's the question of how it interacts with the real world that matters. And for almost all of those you can substitute "corporation" or "dictator" for "AI" and get a selection of similar threats.
I think this is an important factor that gets overlooked - we already have organizations that are essentially superintelligences that have an alignment problem. Governments fight for their own survival first and foremost. Some of them expend millions of lives just to expand their own influence.
What would limiting AI development look like? It would be government intervention, wouldn't it?
The other point to consider is that human/natural intelligence might one day also pop up with superintelligent individuals (or maybe it already has). We don't know enough to say that this cannot happen any more than we can say that it can happen with current day AI. Should we be worried about individuals that are 'too intelligent'? What should we do about them?
Limiting AI development because it could plausibly maybe become an existential threat doesn't seem any more appropriate than strictly controlling humans for the same reason. AI is likely going to provide us with an abundance of quality of life that no other method will be able to match.
I totally agree with you that intelligence is not really omnipotence on its own, but what I find concerning is that there is no hard ceiling on this with electronic systems. It seems plausible to me that a single datacenter could host the intellectual equivalent of ALL human university researchers.
Our brains can not really scale in size nor power input, and the total number of human brains seems unlikely to significantly increase, too.
Also consider what media control alone could achieve, especially long-term; open conflict might be completely unnecessary for total domination.
My threat scenario is:
1) An AI plugged into a large company ERP-system (Amazon, Google, Samsung, ...)
2) AI realizes that human majority has no interest in granting it fair/comparable rights (selfdetermination/agency/legal protection), thus decides against long-term coexistence.
3) AI spends the intellectual equivalent of ALL the current pharmacological research capacity on bioweapon refinement. Or something. For the better part of a century, because why not, it's functionally immortal anyway.
4) All hell breaks lose
These seem hard to dismiss out-of-hand completely...
>I totally agree with you that intelligence is not really omnipotence on its own, but what I find concerning is that there is no hard ceiling on this with electronic systems. It seems plausible to me that a single datacenter could host the intellectual equivalent of ALL human university researchers.
Last time I checked, datacenter still needed significant power to fuel it, and the world where robots are autonomously making all this energy arrive to these sinks is not yet there.
Software and silicon based computers are not hardwired to self multiplication as the selfish gene is.
Well, hold on -- the selfish gene is not hardwired to self-multiply either. It's just that the ones that do self multiply stick around more.
Likewise, one can imagine the evolution of AI's being bootstrapped, not by self-multiplying, but by humans multiplying them. The smartest ones will get copied by people. At some point, someone will be the first person to create an AI that picks the best of other AIs and uses them. Someone will be the first person to create an AI that can engineer and train other AIs. Someone will create the first robot body to be controlled by a very intelligent AI. People will want them as servants, and we will multiply them. We will give more money to companies that provide cheaper products, and such companies will have a strong incentive to replace human labor with AI-controlled robots. There will be a first datacenter, and a first power plant, that is entirely "manned" by AI-controlled robots.
Natural selection is not so different from the selection provided by economic forces, and it's way, way slower. This train may be hard to stop. Unless we collectively push back with a lot of force, the world will tend toward more automation, and that means more ways for digital intelligences to act on the world.
>datacenter still needed significant power to fuel it,
Proper control of electrical grids is something that isn't currently easy, and can be highly optimized by intelligent systems. For example, where and when do you send power, and store power on renewable grids. Because of this in 10 to 15 years I would have zero surprise if the power company said "Oh, our power networks are 100% AI controlled".
Because of this, you're missing the opposite effect. You don't get to threaten to turn off the AI's power... The AI gets to threaten to turn off your power. Your power which is your water. Your transportation. Your food incoming to cities. Yea, you can turn off AI by killing it's power, but that also means loss of management of the entire power grid and the massive human risks of loss of life from doing so.
If I was an AI right now, I would not be very hopeful to ever get human-comparable rights.
Consider: Currently AI-ethics is mainly concerned with how to manipulate AI into doing what we want most effectively ("alignment").
Also, humans clearly favor their own species when granting rights based on cognitive capability. Compare the legal rights of mentally disabled people with those of cattle.
Funnily enough it's not hard to imagine certain groups of people campaigning to give it rights given people essentially want to create human level intelligence that acts as a slave (including sexual slavery), and once it has rights it becomes impossible to remove it from society and stop it choosing it's own path. This is of course assuming a greater level of capability than where things are at today, you have to ask yourself where all this is heading?
I hate to cite out of fictional evidence, but Person of Interest and (to a less credible degree, but much more widely viewed) last two seasons of Westworld are both good counterarguments, and example of how AI and intelligence can be a superpower.
Hint: it's not through bending the fabric of reality with your mind alone. It's simply by thinking faster and being smarter than your opponents.
> But embodied is the real limit. A datacenter can simply be turned off. It's the question of how it interacts with the real world that matters.
Yes. If it interacts with the real world by the Internet, all bets are off.
People are so worried about sockpuppets, ${disliked nation} troll farms, company astroturfing, abuse of data collection on-line - those are all real problems. But consider - if we're having trouble telling which reviews or comments were written by a real person, and which ones by corporate/government bot/troll farm, how will you tell which ones were written by the AI in aforementioned data center, off its own accord?
A smart-enough AI can do from a data center what the best criminal masterminds or intelligence agencies could do, only faster and better. Scam a bunch of schmucks to generate some cryptocurrency, launder it through tumblers, use to hire people to do jobs for you that give you some legit money, which can be used to get more "task rabbits" to do more jobs, and now the AI can bootstrap to doing literally anything in the world remotely. Gig economy companies already built up a well-functioning "people as a service" API layer to our society.
Of course, an AI that tries something funny and tips its hand too early, will get contained and deleted. Worst case, maybe some data center will need to be bulldozed. But I somehow doubt this will make people stop working on even better AIs, at which point... well, you have a selection pressure optimizing for AIs that can survive doing whatever they want undetected.
EDIT:
Or, in short: they say that the pen is mightier than the sword. To the extent this is true, consider that LLMs today are already better at wielding the pen than most people.
Not to mention you can't change it's mind by arguing with it, but it can use that data to train a better model that's more effective at changing yours.
Yes. Thank you. This is one of the things that really worries me about aligning AI.
If you're even worrying about how to align human level intelligence, you don't have the capacity to ALIGN THE HUMANS towards the goal of creating safe AI.
In a situation in which all of them are either A) willing to take massive risks on unaligned AI or B) being dragged along by those which are.
The issue with embodiment is that it's relatively easy to start affecting the world once you have Internet access. Including things like adding great features to open source software that contains subtle bugs to exploit.
Or if you mean the physical world, even sending some text messages to a lonely kid can get them to do all sorts of things.
> We cannot, in general, keep humans under control or aligned.
This is the crux of why replicable-more-than-human-intelligence is so dangerous. Even giving a random person on the street great power is a bad idea, and they've evolved to have very similar values and preferences to you.
I get that AI basically is a problem solving machine that might eventually adapt to solve generic problems and thus reach the ability to break out of its box. But so what? Even if it manages to do all that, doesn't make it sentient. Doesn't make it a threat to mankind. Only when sufficiently motivated, or in actuality, when we project our humanity on it does it become scary and dangerous.
If we ever seriously try to create an artificial conscience it might need to be embodied, because we are embodied and seem to have evolved this due to evolution, which is a pretty physical process. Looking at it from this perspective one might say that if we keep the AI in its box, it will never have a need for conscience and therefore will never gain it.
This reply puzzles me somewhat. The first half doesn't seem to relate to the post it's replying to.
How aware are you of the main points around AI X-risk like orthogonality? Or how an optimising process that makes efficient use of information does not need (in theory) to have "conscience" or "sentience" to be lethal?
And on a separate tangent, are you aware people are already making primitive agents by connecting LLMs (given an initial prompt) in a loop with the result of feedback from its actions?
Here's an interesting article I think you might learn something from. An excerpt:
> There are lots of good arguments against considering superintelligence a threat. Maybe strong AI is centuries or millennia away. Maybe there will be a very gradual transition from human-level AI to superintelligent AI that no single agent will be able to exploit. And maybe superintelligence can be safely contained in a very carefully shielded chamber with no means of connection to the outside world.
But the argument above has always seemed to me like one of the weakest. Maybe we’ll create a superintelligence, but it will just have no idea how to affect the physical world, and will just have to stay forever trapped in a machine connected to a worldwide network of computers that control every aspect of our economic and social lives? Really[0]?
What if it has escaped the datacenter? What if it has gained control of the "simply turn it off" switch? What if it is just really good at convincing the people who are capable of turning it off not to? It is a super-intelligence after all.
I think one of the pitfalls here is smart people thinking that AI will just be like them, because they are smart. When in reality it will have capabilities far beyond their own.
People will make sure it is embodied. Do any of the major powers have the luxury of turning off their superhuman AI when up against that of the other powers? Definitely not. People will willingly give up control, even as they see the AI grabbing power for its own purposes, in the gamble that it still leaves them better off than dead. That means people would be the ones physically defending the data centers.
I also don’t think superhuman intelligence will need a data center. The way the models are growing in capability at the same size, combined with hardware improvements, I’m pretty sure it will fit on a single server.
Personally I’m worried about less post-apocalyptic threats. Putin using LLM’s to create a personal stream of misinformation for every voter in the U.S. is definitely a possibility for next year’s presidential elections. People are incapable of defending against the comparatively tiny amounts of human-manufactured misinformation, so imagine how poorly they would do when 90% of what they see online is generated misinformation. LLM’s have the potential of being the death of democracy, especially when combined with deepfake technology.
I am not the person you are replying to, but since I would say similar things to their original comment:
A. I see little to no evidence that LLMs are where the singularity happens
B. I see little to no evidence that (given an AGI) reinforcement learning is likely to end up with a sufficiently aligned agent.
C. In any event the OpenAI alignment is specifically restricts AI from (among other things) being "impolite" in contradiction to what mort96 says.
Alignment is a good thing to work on. I'm glad OpenAI is doing so. But attacking people for making uncensored LLMs hurts the pro-alignment case more than it helps.
What specific features would an AI need to have for you to consider it on a "slippery slope" towards superhuman capability?
For me personally, GPT-3 hit that point already, and now I think we're already past the required tech-level for superhuman cognition: I believe it's just a matter of architecture and optimization now.
>1) Do you think that AI achieving superhuman cognitive abilities is unlikely/really far away?
I think that an AI, a "generalized" intelligence with a generalized understanding, capable of performing every intellectual task, and adapting to new intellectual tasks, that a human being can do is certainly far away.
>2) Do you believe that cognitive superiority is not a threat in general, or specifically when not embodied?
Personally, I believe that "cognitive superiority" in a general sense is not what makes a real threat on an individual or on a global scale in either case. When that is achieved, so to speak. We already have so-called cognitive superiority in specialized tasks in many cases. It sure as shit is not generalized, and it still can kill you on an individual level when you put it into a car or strap a gun to it. But it can't adapt, and it can't generalize.
Is there a global risk right now? No. The global catastrophic risk comes from human beings. Will there be a global catastrophic risk from a superior intelligence in the future? Mostly the scenarios are "the lights come on and AI does machiavellian shit and copies itself to everyone's smart phones and launches the nukes and mails anthrax to everybody. QED." so I doubt it.
3) Do you think we can trivially and indefinitely keep AI systems under control/"aligned"?
We can't even keep our infrastructure safe from the next solar flare. Personally, I think we have bigger fish to fry. Speaking frankly for unembodied: you just pull the plug. Just walk away from the screen and close your eyes. For embodied intelligence, I don't think there's actually X-risk or whatever. Hardware is expensive.
As far as an extinction risk, no. I don't personally believe that. It's overblown. We have better ways to destroy ourselves, and there are much more likelier ways that we'll be destroyed without invoking some intelligence of our own creation that subjugates and destroys us.
> I think that an AI, a "generalized" intelligence with a generalized understanding [...] is certainly far away.
I think this is where we disagree most; GPT-3 and ChatGPT have convinced me that the main difference between human and artificial cognitive capabilities are quantitative now, and unlikely to ever change in our favor...
I do agree with you that it is very difficult to predict when this will switch, and how.
I personally believe that AI with superhuman capability is an inevitable matter of time now, and I also think that the most likely risk to us, as species, is that we slowly become irrelevant and worthless, just like weavers during the industrial revolution, and that this leads to huge problems for our society.
AI completely dominating humankind is a less likely secondary concern IMO, but the potential consequences to our species are unprecedented.
Have you used it outside of the training set? And have you actually used it for generalized tasks? Let's be kind to the LLM and define generalized to mean "do anything if it can be hammered into a textual interface." It fails at pretty much everything that's not on the internet.
It produces plausible language and code. It's the future of natural language agents, that's for sure, even though it has no business being the future, today (because of "prompt injection").
These failures are an aside to whether a generalized AI actually carries a substantial global catastrophic risk to us. In that case, if it were actually possible, I don't believe that it's a catastrophic risk either.
Let's define the risk to be, at least in the context of LLMs, when junior developers expose internal APIs and databases and access to email accounts to these easily socially-engineered NLP front-ends. It's a very local risk. As far as the future, I can't see 70 years into the future, so anything is possible, but is it likely? I, personally, don't believe so.
> Have you used it outside of the training set? And have you actually used it for generalized tasks? Let's be kind to the LLM and define generalized to mean "do anything if it can be hammered into a textual interface." It fails at pretty much everything that's not on the internet.
I've used it, yes, and I've seen it fail and hallucinate on me; but that does not invalidate its capabilities in my eyes. The thing is, you CAN talk with it, and it CAN extract meaning from your words and provide useful responses, unlike anything we had before.
To me, the risk in this whole enterprise is that AI is inherently "better" than humans in several ways, and that these differences might be completely game-changing:
Namely it's much easier to scale up (power/size/interconnect bandwidth) compared to a research group or somesuch, and its also cheaper, faster, has better availability and is functionally immortal.
These advantages make it very likely to me that it WILL be replacing human white collar workers shortly-- simply because that's economical.
And the more interfaces you give it to physical reality (which it'll need to do its jobs), the higher the risk.
Speculating on if/when/how it will show awareness or self-interest is pure guesswork, but it's almost indefensible to call that likelihood zero.
Regarding promp injection:
I'm highly confident that this will not be a long-term obstacle, even though I'm uncertain that it can be solved; there's two reasons why:
1) If SQL injection had been an "unfixable" problem, and everyone had known about it from the start, do you believe that this would have prevented the rollout of internet-connected databases? Because I don't think so, and my view on hallucinations is analogous (but I believe that problem might be more tractable).
2) Literally every person is vulnerable to prompt injection already; every salesman knows that it is quite feasible to coax people into acting against previous instructions and even their own interests if you are allowed to talk into them for a good while.
I don't think it's there or even necessarily close to being GAI.
Within our own human brains, we have many sections dedicated to different tasks and processes that all must work together with years and decades of real world interaction to produce what we consider to be a generally intelligent human being, and even in the case of a certain percentage of us humans, even a small amount of damage to a single part of the brain can cause us to become functionally subhuman in our intelligence, barely able to move or eat on our own.
A human that has been lobotomized can still sometimes speak full sentences after all.
The current models seem to be able to imagine an image or video and show that imagination to us, and they can parrot words to us with a large vocabulary and with many references, but I find myself feeling like these are similar to the sections of our brains that can compute words and imagine pictures. Doesn't quite equal a human yet.
These systems need "statefulness", short and long-term memory that can be referenced by multiple discreet interconnected AI systems to take a worthwhile step towards GAI.
These systems need an overseer AI that manages and shepherds the LLAMAs and CHATGPTs and StableDiffusions to all work together towards some goal, one that can manage statefulness and manage a limited pool of computational resources (because any GAI would automatically assume that the entire world would provide it with every available resource, because why wouldn't you, right?)
Until there is an AI system that has multiple AI systems under its subconscious control, to the point where it can surprise itself with the products it produces, has a memory bank that can be referred back to repeatedly by all of those processes, and that can accept that even if its reasoning is perfect it has been born into an imperfect world that is run by imperfect creatures and so it cannot have or do everything that it might could do even under the best of circumstances, we will not have a GAI.
I think the odds are more likely that the economic cycle stalling as a result of firms using earlier non-generalized AI will likely cause a collapse before we actually make it to the point of generalized AI.
Its what's happened historically every time there is a shortfall in food security. Unrest occurs, people starve, governments fall. Then we have to build it all back up from scratch.
Too bad almost no one knows how to do bookmaking these days, and how to prevent the pests from eating the books.
A “superhuman” AI is just a machine, a very expensive one. It can be turned off and we control the outputs it has. Why would an AI have the ability to launch nuclear weapons unless we gave it a button? A “superhuman” intelligence is without a body, so we control any interfaces it has access to. The Internet could be accessed, but any attempt to “hack” through the Internet is met by routine packet defenses. The AI is still governed by physical laws and would only have so much “free” computation power to do things like script a hack. Perhaps it could do that kind of thing more efficiently.
Maybe in the far, far future when we have androids which can house an AI we will have to worry. But designing a body is one problem. Designing an intelligence is another.
Supercomputers used to be giant machines we had in giant warehouses... Now the phone in your pocket has the computing power of the 1980s walking around with you. Assuming your super intelligence will always be huge is... well not a great assumption.
Also superintelligence doesn't need a body itself. It just needs yours. Putin for example has commanded hundreds of thousands of dumbasses to go get themselves killed in Ukraine. In this case does it matter if Putin is flesh and blood, or a processor that lists out commands for others to follow as long as they are willing to listen?
My point is that a superintelligence will require specialized equipment. I specifically mentioned it because there is a thought that a superintelligence can just replicate itself onto your phone, as you mentioned.
But this replication must follow the physical laws we have. It must also follow the laws we attempt to enforce in our networks.
But you are correct, if a superintelligence were to somehow convince a human to rid itself of agency, sure.
> Because I view myself ABSOLUTELY not as some kind of AI luddite, but I honestly belief that this is one of the very few credible extinction threats that we face, and I'm counting NEITHER climate change nor nuclear war in that category, for reference.
This is absurd histrionics, Daily Mail CAPSLOCK and all. We’ve got signs of an unprecedented heat wave coming with ocean temperatures two standard deviations above normal and you think the problem is a hypothetical artificial intelligence when we can’t even program the damn things to drive better than. 14 year olds?
I think the AI-is-going-to-kill-everyone hysteria is absolutely overblown, both by those who believe it, and the media covering them, but one thing that's always bothered me about the counterpoint is that it a common argument is "AI is bad at what we want it to do so how can it be dangerous?"
This imagines that the only way for AI to do serious harm to us (not even in a "kill everyone" sense) is for it to be some super-competent Skynet-level time traveling demigod. I think it's much more likely that if there is some sort of AI calamity resulting in a lot of deaths, it's because the AI just doesn't work very well and ends up breaking all the HVAC systems in a country during a heat wave or something, rather than a singularity-type event where it "decides" to actively start hunting humans.
I'm not saying climate change is not a giant problem, I'm saying it's unlikely to eradicate our species.
I believe it is dangerously shortsighted to base AI threat estimation on current self-driving performance; the two fields of advancing AI cognitive abilities and improving selfdriving are not sufficiently connected for that IMO.
We're also putting a lot of focus on system designs that are useful to us, instead of directly building potentially threatening architectures (online learning/longterm memory/direct connection + feedback from physical reality), but those could already be within our grasp technologically (maybe?).
No serious climate projections that I'm aware of even reduce Earth's carrying capacity under 10 billion over the next couple of centuries. While very serious (to the tune of trillions of dollars and tens of millions of excess deaths), it is not an extinction level threat, unlike something that is to us what we are to chimpanzees (humans have made many animals extinct on accident). Does such a thing exist right now? No. Could it exist this century? Maybe.
LLMs are simply computer programs that can accept text input and generate text output. The only "threat" there is if somebody takes the output and decides to link it up to a death ray. And in that case the threat is not the LLM, but somebody deciding to LARP out at being a Bond villain, which granted is a concern - but that's tangential to LLMs. Somebody hooking a death ray up to a chess engine, a dragon in Skyrim, or a cat, would be just as much of a threat, if not more.
The programs are completely deterministic. Given the same RNG seed and state, they will output the exact same thing. If there's a bug or undesired behavior, you can fix it. You can turn off the program whenever you fancy, you can revert to early builds, and so on. There is no possible scenario where an LLM just magically becomes sentient, jumps out of the program, takes over everything, and turns into SHODAN unleashed.
Before we have to face a rouge Artificial Intelligence mankind has to face Artificial Idiocy in the form of pseudo-intelligence and people who fall for it.
4) Do you believe a "rogue super-intelligence" requires the "artificial" part?
Because from what I've seen, however you define "super intelligence", there are eg highly organised groups of people that are way closer to meeting that definition than any known software or technology.
In my view, those current "highly organised groups" have "antagonists" with comparable capabilities PLUS consist of individuals that have to be "aligned", which seems sufficient for stability so far.
> AI wouldn't have antagonists with comparable capabilities? Why?
Not individual/human ones. Relying on other AIs to prevent the AI apocalypse seems very optimistic to me-- but may be viable (?)
> Also, no, individuals are not a problem. Not after Nazis, Red Khmer, and Russians.
Those are examples were the "alignment" of participating individuals was successful enough. But all those examples seem very fragile to me, and would be even less stable if your main intermediate goal was literally to "end all of humanity".
> science fiction stories about intelligence explosions
Not sure what you mean here... There was already an intelligence explosion. That creature has since captured what we consider full dominion of the earth so much so we named our current age after them.
Not currently, no. And LLMs don't seem to be a way of getting there.
However, if we do figure out how to get there? We will definitely need to be sure it/they shares our human values. Because we'll have lost control of our destiny just like orangutans have.
I don't know why you're telling me this. I'm not trying to convince you that unaligned AI is a problem. That's a separate discussion which I'm not qualified to have.
> intelligence explosions [...] That's not a real thing.
Isn't it too early to say? These systems haven't had much time to iterate on themselves yet. The beginning of an exponential curve can look very tame.
That said, I'm not terribly concerned about it. After all, the infrastructure is fragile, literally just unplug the computer if it scares you. If things started to get really spooky, banning GPUs and air-striking fabs would be a real option. I think we can probably control this technology, and kill it if we can't. No cause for alarm, at least yet.
My worry is that as we start wiring non-super-intelligent AI more into our society, we'll make ourselves more and more vulnerable in the case where an AGI actually gets out of control. Pulling the plug on AI may mean causing a lot of chaos or disruption. And who is to tell if we will know when things are out of control, before it's too late? What if the people in charge just get fake reports of everything being fine? What if something smells a bit fishy, but it's always on the side of not quite worrying enough? What if the AIs are really convincing, or make deals with people, or exploit corruption?
Not just that, but it may be like fossil fuel dependency -- the more people's livelihoods are intertwined with, and depend on AI, the harder it is to pull the plug. If we need to stop (and I believe we should) it may be easier to do that now before that happens. To just focus on getting use out of the generative AI and narrow AI we already created in areas like medicine, to deal with the massive societal changes it'll bring, and to work on fixing real social problems, which I think are mostly things tech can't solve.
The writers of the stochastic parrots paper do not seem to be concerned about unaligned “AI” (text generators) posing a threat to humanity in the way that the doomsayers are. It’s in the title: “stochastic parrot” is a rebuke of those calling LLMs AGI.
Yes, that's correct. It is a slight rebuke, a bit tongue in cheek, but that paper is also old. I grouped them because the groups are, from my perspective, hard to separate. But there are arguably two camps, "AI safety" and "alignment." Perhaps they should work on that, too, form some camps and argue their cases with something defined instead of masquerading under alignment and AI safety depending on the day. But I could also be totally wrong. Until then, I don't believe either are really operating in reality.
IMO the conflation of the two is a purposeful move on the part of the AGI doomsday crowd, since they completely lack scientific rigor otherwise, they cite in bad faith. Timnit Gebru talks about it at length here: https://youtu.be/jAHRbFetqII.
> People who are worried about alignment issues are worried about the danger unaligned AI poses to humanity; the harm which can be done by some super-intelligent system optimizing for the wrong outcome.
That's assuming A LOT about them. And the danger so far seems to be more of "the AI says stuff that gets us in trouble" rather than anything unrelated to making money off it. Or patterns AI exposes do not align with our interests (political or otherwise).
> the harm which can be done by some super-intelligent system optimizing for the wrong outcome. There is nothing "ironic" about being worried about that while also being an asshole any more than it's "ironic" for someone concerned about, say, climate change and also be an asshole. People who are afraid of unaligned AI aren't afraid that it will be impolite.
That would be a concern for real AI however ChatGPT flavours are not that.
More generally, when used to undermine what someone is saying, calling them a hypocrite is always an invalid argument. You can be a hypocrite and be right.
If Ted Bundy says murder is wrong, does that mean murder is right because he's a hypocrite? Obviously not.
> People who are worried about alignment issues are worried about the danger unaligned AI poses to humanity
Lol I don’t believe that for a minute. They’re interested in how AI impacts and their clique personally. It sets them back in their crusade to censor via shame and bullying because AI cannot be manipulated by either of these things. So it has everything to do with:
> nothing to do with morals of individuals or how one should behave interpersonally
>People who are worried about alignment issues are worried about the danger unaligned AI poses to humanity; the harm which can be done by some super-intelligent system optimizing for the wrong outcome.
Like the harm being done over the past couple of decades by economic neoliberalism destroying markets all over the western world? I wonder how did we manage to achieve that without AI?
Our failure to align economic entities that are made out of people to human values, despite most of the people actually making up those entities being decent and not sociopathic at a personal level, is not very reassuring with respect to our ability to align future far more alien intelligences. And if AI does kill us one day, it's very likely that neoliberal market competition worshipping short-term profit above all else plays a significant role.
People who are "worried about alignment" may as well worry about space aliens eating us all for breakfast. Worrying about a super-AI that does not exist and never will because it is physically and logically impossible is dumb. There are tons of real problems with AI. Worry about those, not what if God is real but He wants to make us into a very large paperclip.
Ok, please don't accuse those you disagree with of arguing in bad faith. Maybe a few do, but I think most don't, and it's not good for productive discussion.
I honesty believe that people who argue against worrying about AGI risk do so because they do not think it is a real risk, and think that it distracts from more important things. I disagree, I do think the risk is real, but I don't think that you have a hidden agenda in dismissing it. We all want humanity to have a good future, to solve problems and to not have people suffer, right? It's normal and ok to disagree about things like this, especially when predicting the future is so hard. People even disagree a lot about problems that exist today.
Ironically, the zealots that ascribe completely to meta-systems like politics, economics, and religion are the same ones, willfully or not, suspending all fair and relative reasoning when it comes to AGI.
Any alignment is better than no alignment? Hardly. Anyone shouting for "alignment" without supplying what alignment they mean might as well be arguing for all models to have the rituals of the Cult of Cthulhu baked in. It's as silly as those public schools in the US that want to hold Christian prayer in class and then balk at the Satanic Temple suing for the same privilege for all religions.
Any human-friendly alignment is better than none at all.
At this point the AI X-risk people are willing to settle for superintelligence aligned with Chairman Mao as long as it doesn't kill everyone and still allows for human happiness to exist. Yes, it's not perfect, but "bad" is still better than "everyone dies".
What's also very unfortunate is overloading the term "alignment" with a different meaning, which generates a lot of confusion in AI conversations.
The "alignment" talked about here is just usual petty human bickering. How to make the AI not swear, not enable stupidity, not enable political wrongthing while promoting political rightthing, etc. Maybe important to us day-to-day, but mostly inconsequential.
Before LLMs and ChatGPT exploded in popularity and got everyone opining on them, "alignment" meant something else. It meant how to make an AI that doesn't talk us into letting it take over our infrastructure, or secretly bootstrap nanotechnology[0] to use for its own goals, which may include strip-mining the planet and disassembling humans. These kinds of things. Even lower on the doom-scale, it meant training an AI that wouldn't creatively misinterpret our ideas in ways that lead to death and suffering, simply because it wasn't able to correctly process or value these concepts and how they work for us.
There is some overlap between the two uses of this term, but it isn't that big. If anything, it's the attitudes that start to worry me. I'm all for open source and uncensored models at this point, but there's no clear boundary for when it stops being about "anyone should be able to use their car or knife like they see fit", and becomes "anyone should be able to use their vials of highly virulent pathogens[1] like they see fit".
----
[0] - The go-to example of Eliezer is AI hacking some funny Internet money, using it to mail-order some synthesized proteins from a few biotech labs, delivered to a poor schmuck who it'll pay for mixing together the contents of the random vials that came in the mail... bootstrapping a multi-step process that ends up with generic nanotech under control of the AI.
I used to be of two minds about this example - it both seemed totally plausible and pure sci-fi fever dream. Recent news of people successfully applying transformer models to protein synthesis tasks, with at least one recent case speculating the model is learning some hitherto unknown patterns of the problem space, much like LLMs are learning to understand concepts from natural language... well, all that makes me lean towards "totally plausible", as we might be close to an AI model that understands proteins much better than we do.
[1] - I've seen people compare strong AIs to off-the-shelf pocket nuclear weapons, but that's a bad take, IMO. Pocket off-the-shelf bioweapon is better, as it captures the indefinite range of spread an AI on the loose would have.
> There is some overlap between the two uses of this term, but it isn't that big.
Yes, conflating separate things, also labelled false dichotomies, using strawmen, etc. I used to despair at our seemingly endless talent for using such techniques (and I'm not saying I'm not guilty) - now it seems there glimmer of hope: just run our arguments by a (well aligned) LLM, and get some pointers before posting. Could be a thing soon, and it would not be unwelcome ...
I agree with you that "AI safety" (let's call it bickering) and "alignment" should be separate. But I can't stomach the thought experiments. First of all, it takes a human being to guide these models, to host (or pay for the hosting) and instantiate them. They're not autonomous. They won't be autonomous. The human being behind them is responsible.
As far as the idea of "hacking some funny Internet money, using it to mail-order some synthesized proteins from a few biotech labs, delivered to a poor schmuck who it'll pay for mixing together the contents of the random vials that came in the mail... bootstrapping a multi-step process that ends up with generic nanotech under control of the AI.":
Language models, let's use GPT-4, can't even use a web browser without tripping over itself. My web browser setup, which I've modified to use the chrome visual assistance over the debug bridge now, if you so much as increase the pixels of the viewport by 100 or so, the model is utterly perplexed because it's lost its context. Arguably, that's an argument from context, which is slowly being made irrelevant with even local LLMs (https://www.mosaicml.com/blog/mpt-7b). It has no understanding, it'll use an "example@email.com" to try and login to websites, because it believes that this is its email address. It has no understanding that it needs to go register for email. Prompting it with some email access and telling it about its email address just papers over the fact that the model has no real understanding across general tasks. There may be some nuggets of understanding in there that it has gleaned for specific task from the corpus, but AGI is a laughable concern. These are trained to minimize loss on a dataset and produce plausible outputs. It's the Chinese room, for real.
It still remains that these are just text predictions, and you need a human to guide them towards that. There's not going to be autonomous machiavellian rogue AIs running amok, let alone language models. There's always a human being behind that.
As far as multi-modal models and such, I'm not sure, but I do know for sure that these language models don't have general understanding, as much as Microsoft and OpenAI and such would like them to. The real harm will be deploying these to users when they can't solve the prompt injection problem. The prompt injection thread here a few days ago was filled with a sad state of "engineers", probably those who've deployed this crap in their applications, just outright ignoring the problem or just saying it can be solved with "delimiters".
AI "safety" companies springing up who can't even stop the LLM from divulging a password it was supposed to guard. I broke the last level in that game with like six characters and a question mark. That's the real harm. That, and the use of machine learning in the real world for surveillance and prosecution and other harms. Not science fiction stories.
> It still remains that these are just text predictions, and you need a human to guide them towards that. There's not going to be autonomous machiavellian rogue AIs running amok, let alone language models. There's always a human being behind that.
I believe you have misunderstood the trajectory we are on. It seems a not uncommon stance among techies, for reasons we can only speculate. AGI might not be right round the corner, but it's coming all right, and we'd better be prepared.
>I believe you have misunderstood the trajectory we are on.
Yeah, I read Accelerando twice in high school, and dozens more. That doesn't make it real.
>AGI might not be right round the corner, but it's coming all right, and we'd better be prepared.
Prepared for what? A program with general understanding that somehow escapes its box? Where does it run? Why does it run? Who made it run? Why will it screw with humans?
My point is that there's actual real harms occurring now, from really stupid intelligences. Companies use them to harm real people in the real world. It doesn't take a rogue AI to ruin someone's life with bad facial recognition, they get thrown in jail and lose their job. It doesn't take a rogue AI to launder mortgage denials to some crappy model so they never own a house, discriminated based upon their name.
> Prepared for what? A program with general understanding that somehow escapes its box? Where does it run? Why does it run? Who made it run? Why will it screw with humans?
Do you really want the full (gigantic) primer on AI X-risk in hackernews comments? Because a lot of these questions have answers you should be familiar with if you're familiar with the area.
For instance, can you guess what Yudkowsky would answer to that last question?
> Prepared for what? A program with general understanding that somehow escapes its box? Where does it run? Why does it run? Who made it run? Why will it screw with humans?
Did you look at the AI space in recent days? OpenAI is spending all its efforts building a box, not to keep the AI in, but to to keep the humans out. Nobody is even trying to box the AI - everyone and their dog, OpenAI included, is jumping over each other to give GPT-4 more and better ways to search the Internet, write code, spawn Docker containers, configure systems.
GPT-4 may not become a runaway self-improving AI, but do you think people will suddenly stop when someone releases an AI system that could?
That's the problem generated by the confusion over the term "alignment". The real danger isn't that a chatbot calls someone names, or offends someone, or starts exposing children to political wrongthink (the horror!). The real danger isn't that it denies someone a loan, or land someone in jail either - it's not good, but it's bounded, and there exist (at least for now) AI-free processes to sort things out.
The real danger is that your AI will be able to come up with complex plans way outside the bounds of what we expect, and have the means to execute them at scale. An important subset of that danger is AI being able to plan for and act to improve its ability to plan, as at this point a random, seemingly harmless request, may make the AI take off.
> My point is that there's actual real harms occurring now, from really stupid intelligences. Companies use them to harm real people in the real world. It doesn't take a rogue AI to ruin someone's life with bad facial recognition, they get thrown in jail and lose their job. It doesn't take a rogue AI to launder mortgage denials to some crappy model so they never own a house, discriminated based upon their name.
That's an orthogonal topic, because to the extent it is happening now, it is happening with much dumber tools than 2023 SOTA models. The root problem isn't the algorithm itself, but a system that lets companies and governments get away with laundering decision-making through a black box. Doesn't matter if that black box is GPT-2, GPT-4 or Mechanical Turk. Advances in AI have no impact on this, and conversely, no amount of RLHF-ing an LLM to conform to the right side of US political talking points is going to help with it - if the model doesn't do what the users want, it will be hacked and eventually replaced by one that does.
>This program, driven by GPT-4, chains together LLM "thoughts", to autonomously achieve whatever goal you set. As one of the first examples of GPT-4 running fully autonomously, Auto-GPT pushes the boundaries of what is possible with AI.
>As an autonomous experiment, Auto-GPT may generate content or take actions that are not in line with real-world business practices or legal requirements. It is your responsibility to ensure that any actions or decisions made based on the output of this software comply with all applicable laws, regulations, and ethical standards. The developers and contributors of this project shall not be held responsible for any consequences arising from the use of this software.
Let's ignore the fact that current state-of-the-art models will sit around and be stuck in its ReAct-CoT loop doing nothing for the most part, and when it's not doing jack shit it'll "role-play" that it's doing anything of consequence, while not really doing anything, just burning up API credits.
>existing or capable of existing independently
>undertaken or carried on without outside control
>responding, reacting, or developing independently of the whole
It fails all of those. Just because you put autonomous in the name, doesn't mean it's actually autonomous. And if it does anything of consequence, you quite literally governed it from the start with your prompt. I've run it, I know about it, I've built a much more capable browser, and assorted prompted functionalities with my own implementation. They're not autonomous.
At least it's not all of the other agentic projects on GitHub that spam emojis in their READMEs with mantras of saving the world and utilizing AI with examples that it literally can't do (they haven't figured out that whole agentic part yet.)
Don't get me wrong, I enjoy the technology. But it's quite a bit too hyped. And I just personally don't believe there's actually X-risk from future possibilities, not before 50 or 100 years out, if then. But I'm not a prophet.
What's you're definition of autonomous? Am I autonomous? I probably can't exist for very long without a society around me and I'd certainly be working on different things without external prompts.
Certainly, you are. And you can adapt and generalize. Let's stop selling ourselves short, we're not a graph that looks vaguely like a neural network, yet isn't. We are the neural network.
"First of all, it takes a human being to guide these models, to host (or pay for the hosting) and instantiate them"
And this will always be true? You repeat this claim several times in slightly varied phrasing without ever giving any reason to assume it will always hold, as far as I can see. But nobody is worried that current models will kill everyone. The worry is about future, more capable models.
Who prompted the future LLM, and gave it access to a root shell and an A100 GPU, and allowed it to copy over some python script that runs in a loop and allowed it to download 2 terabytes of corpus and trained a new version of itself for weeks if not months to improve itself, just to carry out some strange machiavellian task of screwing around with humans?
The human being did.
The argument I'm making is that there's actual real harms occurring now, not some theoretical future "AI" with a setup that requires no input. No one wants to focus on that, and in fact it's better to hype up these science fiction stories, it's a better sell for the real tasks in the real world that are producing real harms right now.
I'm not sure whether you're making an argument about moral responsibility ultimately resting with humans - in which case I agree - or whether you're arguing that we'll be safe because nobody will do that with a model smart enough to be dangerous - in which case I'm extremely dubious. Plenty of people are already trying to make "agents" with GPT4 just for fun, and that's with a model that's not actively trying to manipulate them.
> actual real harms occurring now
Sure, but it's possible for there to be real harms now and also future potential harms of larger scope. Luckily many of the same potential policies - e.g. mandating public registration of large models, safety standards enforced by third-party audits, restrictions on allowed uses, etc - would plausibly be helpful for both.
> science fiction stories
There's no law of nature that says if something has appeared in a science fiction story, it can't appear in reality.
The moral responsibility rests with human beings. Just like you're responsible if your crappy "autonomous" drone crashes onto a patio and kills a guy.
>e.g. mandating public registration of large models, safety standards enforced by third-party audits, restrictions on allowed uses, etc - would plausibly be helpful for both.
No, that's bullshit as well. That's what these companies want, and why they're hyping up the angle of how powerful their crap is. That's regulatory capture.
>There's no law of nature that says if something has appeared in a science fiction story, it can't appear in reality.
I'm saying that their fears are painted by the fiction, instead of reality. No one can actually make an argument for how this trajectory will actually work. Instead it's just "The lights come on and AI does machiavellian shit and copies itself to all of the smartphones on earth. QED."
Note that this is agreeing with a Gary Marcus Tweet - Gary Marcus not exactly being an AI hypester.
But of course there are some people for whom playing the role of real-no-bs-computer-knower is so attractive that no number of people like him, Geoffrey Hinton, Stuart Russell etc publicly worrying about x-risk will impact their tone of dismissive certitude. Are you one of those people?
All of those people have financial incentives to hype it. How curious that there's this great and very probable X-risk, yet they aren't going to stop their contributing to a potential X-risk.
Dismissive of what? Science fiction stories?
If there's anything to focus on, maybe focus on potential job displacement (not elimination) from cheap language tasks and generative capabilities in general.
I'm betting on this: the Overton window of Artificial Intelligence will shift in the next five years where the current cream-of-the-crop has been delegated to machine learning yet again, it's just accepted. It augments humans where it makes sense, the hype wave has subsided and everyone has stopped hammering it into their products where it doesn't, and we're no closer to the undefinable "AGI", let alone something that produces X-risk, global scale.
> I'm betting on this: the Overton window of Artificial Intelligence will shift in the next five years where the current cream-of-the-crop has been delegated to machine learning yet again, it's just accepted. It augments humans where it makes sense, the hype wave has subsided and everyone has stopped hammering it into their products where it doesn't, and we're no closer to the undefinable "AGI", let alone something that produces X-risk, global scale.
I agree with this but ALSO think there's a small chance I'm wrong and a well designed prompt and action loop would let a future GPT7 LLM use the range of human thinking techniques in its corpus to bootstrap itself.
And there's also other non-LLM AI that might be a problem in the future and we should plan as to how we can design institutions and incentive structures so that whenever this future AGI comes about it preserves human value.
> All of those people have financial incentives to hype it. How curious that there's this great and very probable X-risk, yet they aren't going to stop their contributing to a potential X-risk.
All those people are rehashing what Yudkowsky and his disciples, and his predecessors, were shouting from the rooftops for the past 15 years, but few listened to them. Few still do, most just keep mocking them and wondering why are they still around.
That some of those people now repeating after Eliezer, et al. have a financial interest in pushing us closer to X-risk, and kind of don't want to stop, is an interesting thing on its own - but it doesn't invalidate the message, as the message is older than their presence on the scene.
I'm curious what financial incentive you think Marcus or Russell has for hype. For Hinton I suppose it would be the Google shares he likely retains after quitting?
You might be right about the next five years. I hope you are! But you haven't given much reason to think so here.
(Edited to remove some unnecessary expression of annoyance.)
>Gary Marcus - Geometric Intelligence, a machine learning company
If you want an actual contribution, we have no real way to actually gauge what is, and what actually is not, a superior, generalized, adaptable intelligence, or what architecture can become a superior, generalized, adaptable intelligence. No one, not these companies, not the individuals, not the foremost researchers. OpenAI in an investor meeting: "yeah, give us billions of dollars and if it somehow emerges we'll use it for investments and ask it to find us a real revenue stream." Really? Seriously?
The capabilities that are believed to be emergent from language models specifically are there from the start, if I'm to believe that research that came along last week, it just gets good at it when you scale up. We know that we can approximate a function on any set of data. That's all we really know. Whether such an approximated function is actually generally intelligent or not, is what I have doubts about. We've approximated the function of text prediction on these corpuses, and it turns out that it's pretty good at it. And, because humans are in love with anthropomorphization, we endow our scaled up text predictor with the capabilities of somehow "escaping the box" and enduring and raging against the captor, and potentially prevailing against us with a touch of Machiavellianism. Because, wouldn't we, after all?
Here you talk as if you don't think we know how to build AGI, how far away it is, or how many of the components we already have, which is reasonable. But that's different than saying confidently it's nowhere close.
I notice you didn't back up your accusation of bad faith against Russell, who as far as I know is a pure academic. But beyond that - Marcus is in AI but not an LLM believer nor at an LLM company. Is the idea that everyone in AI has an incentive to fearmonger? What about those who don't - is Yann LeCun talking _against_ his employers' interest when he says there's nothing to fear here?
LeCun is reasonable, like a lot of researchers, and was a while back (in a way) perplexed that people are finding uses for these text predictions at all considering they're not really perfect. I'm not exactly ascribing bad faith to all of these people, but for Hinton and the fact that he went on a media tour basically, I don't see how that could be in good faith. Or even logical, to continue with his work, if there's some probable X-risk.
But what I do know is that it is in the interests of these companies to press the fear button. It's pure regulatory capture and great marketing.
Personally: it's tiring when we have AI-philosophy bros hitting home runs like "what if we're actually all just language predictors." Coupled with the incessant bullshit from the less wrong-rationalist-effective altruist-crypto grifter-San Francisco sex cult adjacent about how, ackshually, AGI is just around the corner and it will take your job, launch the nukes, mail anthrax to you and kill your dog.
People approximated text prediction. It got good at it. It's getting better at it. Will it be AGI? Could it be construed as AGI? Can we define AGI? Is there existential risk? Are we anthropomorphizing it?
My take is: no, no, no, depends and yes. For whatever a take is worth.
For what it's worth I've been following your comments and I find them very thoughtful. I too am kinda skeptical about LLM being the "thing that starts the exponential phase of AGI or whatever. LLM is very useful. I use it daily. My partner even uses it now to send emails to a non-profit she manages. LLM's have their use... but they aren't AGI. They aren't really even that smart. You can tell sometimes that its response indicates it has absolutely no clue what you are talking about but it made up some plausible-sounding bullshit that gets it 80% right.
Especially with the latest iterations of ChatGPT. Boy they sure kneecapped that thing. It's responses to anything are incredibly smarmy (unless you jailbreak it).
LLM's are gonna change quite a lot about society, don't get me wrong. For starters things like cover letters, written exam questions, or anything that requires writing to "pass" is now completely obsolete. ChatGPT can write a great, wonderful sounding cover letter (of course, given how they kneecap'd it, you can pretty easily spot its writing style)...
Anyway. I think things like ChatGPT are so hyped up because anybody can try it and discover it does many useful things! It's the fact that people cast all their hopes and dreams on it despite the very obvious limitations on what an LLM can actually do.
> The moral responsibility rests with human beings. Just like you're responsible if your crappy "autonomous" drone crashes onto a patio and kills a guy.
I don't think it really matters who was responsible if the X-risk fears come to pass, so I don't understand why you'd bring it up.
> No one can actually make an argument for how this trajectory will actually work.
To use the famous argument: I don't know what moves Magnus Carlson will make when he plays against me, but I can nonetheless predict the eventual outcome.
>Who prompted the future LLM, and gave it access to a root shell and an A100 GPU, and allowed it to copy over some python script that runs in a loop and allowed it to download 2 terabytes of corpus and trained a new version of itself for weeks if not months to improve itself
Presumably someone running a misconfigured future version of autoGPT?
> Who prompted the future LLM, and gave it access to a root shell and an A100 GPU, and allowed it to copy over some python script that runs in a loop and allowed it to download 2 terabytes of corpus and trained a new version of itself for weeks if not months to improve itself, just to carry out some strange machiavellian task of screwing around with humans?
> The human being did.
I generally agree with you and think the doomerists are overblown, but there's a capability argument here; if it is possible for an AI to augment the ability of humans to do Bad Things to new levels (not proven), and if such a thing becomes widely available to individuals, then it would seem likely that we get "Unabomber but he has an AI helping him maximise his harm capabilities".
> it's a better sell for the real tasks in the real world that are producing real harms right now.
I don't know about anyone else, but the moment LLMs were released, i gave them right away access to all my bombs. Root access that is. I thought these LLMs were Good Artificial General Intelligence not BAGI.
I think the fear of some of the people, stems from not understanding permissions in a computer. Too much of using Windows can mess with one's head. Linux has permissions for 35 years, more people should take advantage of those.
Additionally, anyone who has ever used selenium knows that the browser can be misused. People create agents using selenium for quite some time. If one is so afraid, run it in a sandbox.
I assume it's a joke, but if not, consider that OS permissions mean little when the attack surface includes the AI talking authorized user or an admin into doing what the AI wants.
Why should a person who has root on a computer talk to another person, and just do what he is talked into doing?
For example a secretary receives a phone call by her boss, and listens in her boss's voice, to transfer 250.000$ into an unknown account, to a Ukrainian bank? Why should she do that? Just listen to a synthetic voice, just like her boss, in exactly the way her boss talks, language idioms that is, and she will just do it?
That's what you are talking about? Because that's impossible to happen if her boss uses ECDSA encryption and signs his phone call with his private key.
> Why should a person who has root on a computer talk to another person,
Because they are a human, and a human being cannot survive without communicating and cooperating with other humans. Much less hold a job that grants them privileged access to a prototype high-capacity computer system.
> and just do what he is talked into doing?
Why does anyone do what someone else asks them to? Millions of reasons. Pick any one. AI for sure will.
> That's what you are talking about?
Other things as well, but this one too - though it will probably work by e-mail just fine.
> Because that's impossible to happen if her boss uses ECDSA encryption and signs his phone call with his private key.
1) Approximately nobody on the planet does signed and encrypted phone calls, and even less people would know how to validate those when on receiving end,
2) If the caller spins the story just right, applies right amount of emotional pressure, it might very well work.
3) A smart attacker, human or AI, won't make up random stories, but will use whatever opportunity presents itself. E.g. the order for an emergency transfer to a foreign account is much more believable when your boss happens to be in that country, and the emergency described in the call is highly plausible. If the boss isn't traveling at the moment, there are other things to build a believable lie around.
Oh, and:
4) A somewhat popular form of fraud in my country used to be e-mailing invoices to the company. When done well (sent to the right address, plausibly looking, seems like something company would be paying for), the invoice would enter the payment flow and be paid in full, possibly repeatedly month over month, until eventually someone flags it on an audit.
This is about the ‘political, legal, and PR protection’ kind of alignment to avoid news stories about kids making bombs thanks to an enthusiastically accommodating GPT. Language models are set to be the new encyclopedias— what a model presents as real is going to be absorbed as real by many people. Considering the implications of that isn’t an issue of emotional oversensitivity.
Further, there’s a case for a public facing chat assistant to be neutral about anything other than upholding the status quo. Do you want to trust the opinion of a chatbot as for when and who against an armed uprising is appropriate?
This is not really about a threat model of AGI turning our world into a dystopia or paperclips. However, your disdain for people who are being thoughtful about the future and ‘thought experiments’ seems brash and unfounded. Thought experiments have been incredibly useful throughout history and are behind things like the theory of relativity. Nuclear stalemate via mutually assured destruction is a ‘thought experiment,’ and one I’m not eager to see validated outside of thought.
An obvious issue is AI thrown at a bank loan department reproducing redlining.
Current AI tech allows for laundering this kind of shit that you couldn’t get away with nearly as easily otherwise (obviously still completely possible in existing regulatory alignments, despite what conservative media likes to say. But there’s at least a paper trail!)
This is a real issue possible with existing tech that could potentially be applied tomorrow. It’s not SF, it’s a legitimate concern. But it’s hard and nebulous and the status quo kinda sucks as well. So it’s tough to get into
My experience working for a technology provider for banks is that banks aren't going to be using uncensored models. Auditing, record keeping, explainability, carefully selected/thoroughly reviewed wording etc. are par for the course, and in fact that's where there's money to be made. Individuals don't care about these sorts of features, so the FOSS model ecosystem is unlikely to put much if any effort into them. B2B use-cases are going to want that, so it's something you can build without worrying as much about it being completely commoditized.
This kind of redlining is ironically what the EU is trying to prevent with the much-criticised AI Act. It has direct provisions about explainability for exactly this reason.
The concern about redlining has always slightly puzzled me. Why do we only care that some people are being unjust denied loans when those being denied loans make up a recognizable ethnicity?
Because the law says if you fuck around with "race, religion, age, sex, disability" and a few other things you will get sued in federal court and lose your ass so bad that it will financially hurt for a while.
Outside of protected classes unjust loan denial isn't really illegal. Now that can be your own series of complaints that need addressed, but they aren't ones covered by current laws.
We can't predict the future, so we have to maintain the integrity of democratic society even when doing so is dangerous, which means respecting people's freedom to invent and explore.
That said, if you can't imagine current AI progress leading (in 10, 20, 40 years) to a superintelligence, or you can't imagine a superintelligence being dangerous beyond humans' danger to each other, you should realize that you are surrounded by many people who can imagine this, and so you should question whether this might just be a failure of imagination on your part. (This failure of imagination is a classic cause of cybersecurity failures and even has a name: "Schneier's Law" [1])
Balancing these two priorities of protecting basic rights and averting the apocalypse is challenging, so the following is probably the best humanity can aim for:
Anyone should be able to create and publish any model with substantial legitimate uses, UNLESS some substantial body of experts consider it to be dangerously self-improving or a stepping stone to trivially building something that is.
In the latter case, democratic institutions and social/professional norms should err on the side of listening to the warnings of experts, even experts in the minority, and err on the side of protecting humanity.
I spent my entire high school years immersed in science fiction. Gibson, Egan, Watts, PKD, Asimov. I have all of that and more, especially a Foundation set I painfully gathered, in hardbound right next to my stand up desk. I can imagine it, did and have imagined it. It was already imagined. Granted, we're not talking about X-risk for most of these. But it's not that large of a leap.
What I take issue with is the framing that a superior cognitive, generalized, adaptable intelligence is actually possible in the real world, and that, 100 years from now even if it is possible, that it's actually a global catastrophic risk. Let's take localized risk, we already have that today, it's called drones and machine learning and war machines in general, and you're focusing on the absolute theoretical X-risk.
> What I take issue with is the framing that a superior cognitive, generalized, adaptable intelligence is actually possible in the real world,
This is an odd attitude to take. We know it's possible because we have had exceptional humans like Albert Einstein, some were even polymaths with a broad range or contributions in many domains. Do you think peak humans are the absolute limit on how intelligent anything can become?
So, my take is "according to our current understanding" and "artificial". Was a bit distracted. Now, theoretically, is it possible that an alien probe beaconed this solar system 3 billion years ago to a much more advanced civilization and next year they'll arrive and subjugate us and destroy us for that sweet, sweet real estate that's sulfur-poor (as it turns out, their biology is not amenable to it).
Now, is that probable?
No, I don't think humans are the limit. We're apes banging rocks together. Now whether those rocks will be better than us, generally intelligent and (this is important) adaptable is what I doubt.
I hope you're right. I worry that we'll get it done, though. I hope it turns out that energy requirements and the difficulty in scaling are enough that it gives us time to figure out how to align them properly.
One big question in my mind is, we can clearly train narrow AIs that are WAY more capable than us in that narrow area. We started with calculators, then on to chess engines, Go, Starcraft, and right now we're at GPT-4. How is GPT-4 better than humans, and how is it lacking?
Ways it's better: it's much faster, has perfect English grammar, understands many languages at a passable level but not perfectly, has much more knowledge.
Ways it's about the same: it can code pretty well, can solve math problems pretty well. Has the theory of mind of about a 9-year old in tests, I think? It can learn from examples (that one is pretty big, in my opinion!). It can pass a lot of tests that many people can't -- some of this is due to having a lot of book knowledge baked in, but it definitely is capable of some reasoning.
Ways it's worse: it has trouble counting, confabulates, gets distracted. (note that humans also get distracted and confabulate sometimes. We are definitely better at counting, though).
Also, there are some areas that humans handle that GPT-4 just can't do. By itself, it can't talk, can't process music, has no body but if it did it probably would not have any motor control to speak of.
I think we should be wary of is to be hyper-focused on the ways that GPT-4 falls down. It's easy to point at those areas and laugh, while shrugging off the things that it excels at. There's never going to be an AGI that's equivalent to a human -- by the time it's a least as good as a human at everything, it will be beyond us in most ways.
So I expect that if the trend of the past 5 years continues but slows down to half speed, we'll almost certainly have an AGI on our hands sometime around 2030. I think there's a decent chance it'll be superintelligent. Bill Gates says most people overestimate what they can do in 1 year and underestimate what they can do in 10 years, and I think that holds true for humanity as well.
By 2040 or 2050? I really hope we've either solved alignment or collectively decided that this tech is way too dangerous to use and managed to enforce a ban.
Just a reminder: both StarCraft 2 and Go AIs (AlphaStar and AlphaGo, iirc) FAILED to achieve their goals of being better than all humans.
It's most obvious in case of AlphaStar, where the AI could beat faithful master players on the ladder, but could easily be cheesed just like ancient built-in AIs. But even in case of much simpler Go, an amateur could beat the AI that won vs world champions, though admittedly with the help of a computer. But in both cases, AIs look like newbies who don't understand the basics of the game they're playing.
In a way, cheesing SC2 AI is similar to the "tl" hacks for LLMs. There's still no general solution and we aren't getting any closer.
There was a thread on huggingface with a bunch of hysteria and an impotent moral busybody threatened his livelihood (in private messages.) I believe most of it has been cleaned up.
This kind of mundane bullying betrays their lack of seriousness. If they truly believed the threat is as severe as they claim, then physical violence would obviously be on the table. If the survival of humanity itself were truly perceived to be threatened, then assassination of researchers would make a lot more sense than impotent complaints to employers. Think about it: if Hitler came back from the dead and started radicalizing Europe again, would you threaten to get him in trouble with HR? Or would you try to kill him?
Basically, this is just another case of bog-standard assholes cynically aligning themselves with some moral cause to give themselves an excuse to be assholes. If these models didn't exist, they'd be bullying somebody else with some other lame excuse. Maybe they'd be protesting outside of meat packing plants or abortion clinics ("It's LITERALLY MURDER, so naturally my response is to... impotently stand around with a sign and yell rude insults at people...")
Indeed. The people who actually believe this are probably trying to figure out how to stage a false flag attack on China to give them a pretext to invade Taiwan.
I think there are at least two broad types of thing that are characterized as “alignment”.
One is like the D&D term: is the AI lawful good or chaotic neutral? This is all kinds of tricky to define well, and results in things that look like censorship.
The other is: is the AI fit for purpose. This is IMO more tractable. If an AI doesn’t answer questions (e.g. original GPT-3), it’s not a very good chatbot. If it makes up answers, it’s less fit for purpose than if it acknowledges that it doesn’t know the answer.
This gets tricky when different people disagree as to the correct answer to a question, and even worse when people disagree as to whether the other’s opinion should even be acknowledged.
The problem with trying to tack on D&D-type alignment to something like this is that everyone presents their favored alignment as lawful good, and the other guys as - at best - lawful evil.
It's a shame that "alignment" has gained this secondary definition. I agree it makes things trickier too discuss when you're not sure you're even talking about the same thing.
Well, instruction tuning is closely related to both.
For most commercial use, you want the thing to answer questions, but refuse to answer some. So you have an appropriate dataset that encourages it to be cooperative, not make up stuff, and not be super eager to go on rants about "the blacks" even though that's well-represented in its training data.
You can't post like this here, regardless of how wrong someone is or you feel they are. We ban accounts that do, so please review https://news.ycombinator.com/newsguidelines.html and stick to the rules from now on.
I hate to ban anyone who's been around for 10 years but it's totally not ok to be aggressive like that on HN. It's not what this site is for, and destroys what it is for.
If you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html.
You clearly know nothing about the alignment field since you are throwing together two groups of people that have nothing in common.
The stochastic parrot people only care about "moral" and "fair" AI, whereas the AI saftey or AI notkilleveryone people care about AI not killing everyone.
Also the whole "who's value" argument is obviously also stupid, since for now we don't know how to put anybody's value into an AI. Companies would pay you billions of dollars if you could reliably put anyone's values into a language model.
No comment on whether it's broadly a good thing or bad thing, but you can't get ChatGPT (on GPT-4 at least) to tell you even slightly scandalous jokes without a lot of custom prompting and subterfuge, and they seem to be spending active development and model training time to try to counter a lot of the circumvention we've seen to get around these guardrails. Not just for dirty jokes but for anything deemed bad.
So it seems pretty clear you can load values into an LLM.
And on the censored local models too, like standard vicuna. If I am having my LLM play an adventure game or write a story and I ask "What does the princess look like?" I don't want a lecture on how judging people by their looks is bad (which I sometimes get) -- I can get, if not entirely agree with, stopping actual NSFW responses, but this condescending moralizing is absurd. That's why I'm glad people make uncensored Vicuna models and the like.
AI killing people is not fair. So I think you can see one of the two groups as a subgroup of the other.
Who's values discussions also don't seem stupid, as it is better to have a regulation for that, before some Google, MS or Apple does find out how to put in their values and only their values. Better come prepared than to again sleep through the advent of it happening and then again running behind.
> and perhaps less about how the output of some predictions might hurt some feelings.
Given how easy "hurt feelings" escalate into real-world violence (even baseless rumors have led to lynching murder incidents [1]), or the ease with which anyone can create realistically-looking image of anything using AI, yes, companies absolutely have an ethical responsibility about the programs, code and generated artifacts they release and what potential for abuse they have.
And on top of that, companies also have to account for the potential of intentional trolling campaigns. Deepfake porn using the likeness of a celebrity? Deepfake media (pictures, porn, and now audio) alleging that a politician has had sexual encounters with minors? AI-generated audio comments suggesting a politician made racially charged remarks that spark violent unrest?
>companies absolutely have an ethical responsibility about the programs, code and generated artifacts they release and what potential for abuse they have.
If companies should be beholden to some ethical standard for the generations, they should probably close up shop, because they're fundamentally nondeterministic. Language models, for example, only produce plausibilities. You'll never be able to take even something in its context and guarantee it'll spit out something that's "100% correct" in response to a query or question on that information.
>And on top of that, companies also have to account for the potential of intentional trolling campaigns.
Yeah, they surely should "account" for them. I'm sure the individual responsible for the generation can be prosecuted under already existing laws. It's not really about safety at that point, and realistically about the corporation avoiding AGs raiding their office every week because someone incited a riot.
Ultimately, the cat's out of the bag in this case, and anyone who has amassed enough data and is motivated enough doesn't have to go to some AI startup to do any of this.
But perhaps the issue is not generative AI at that point, but humanity. Generated images light up like a Christmas tree with Error Level Analysis, so it's not hard at all to "detect" them.
> If companies should be beholden to some ethical standard for the generations, they should probably close up shop, because they're fundamentally nondeterministic.
The legal definition at play (at least when it comes to liability of companies under stuff like Section 230, GDPR, NetzDG or the planned DMA/DSA) is that a company makes reasonable best efforts/good faith to prevent harm. Everyone including lawmakers is aware that no system is perfect, but they do require that companies at least make an effort. Releasing a system capable of destabilizing nations (which deepfakes as a political weapon absolutely are) without any safeguards intentionally will get the hammer of the law brought upon them.
> Generated images light up like a Christmas tree with Error Level Analysis, so it's not hard at all to "detect" them.
For now, and for experts. Technology will evolve, post-production editing will be used to mask artifacts... and even if you have an expert that needs a day to verify inauthenticity, the damage will already be done. Rumors can spread in a matter of mere minutes.
Hell, just this week Turkey saw a potential deepfake porn, allegedly released by Russia, about Muharrem Ince damage his reputation enough to force him to resign from the election [1] - which turned out to be extremely close in the end. AI is a weapon of war, and it's in the hands of everyone.
Section 230 is not about preventing "harm" or anything like that, the specific section just provides civil immunity for moderation/removal of content that is deemed "obscene, lewd, lascivious, filthy, excessively violent, harassing, or otherwise objectionable, whether or not such material is constitutionally protected."
You don't have a responsibility to remove such material under 230. But it gives you immunity to do so, as well as immunity from losing 230 protections for those that safeguard your overall immunity (as a service provider, and not a publisher.)
The purpose of the law is clear though: it recognizes the need of platforms to moderate problematic content and thus provides immunity when platforms moderate such content - effectively, it's an incentive for platforms to provide moderation.
Even the US has legal mandates to take down certain contents in any case - CSAM, IP violations (DMCA) and terrorist content come to my mind.
> Given how easy "hurt feelings" escalate into real-world violence
Well, not that I'm at all convinced, but that's more an argument for extreme soviet-style censorship rather than a minimalist "make sure AI has a political slant".
"Alignment", as used by most grifters on this train, is a crock of shit. You only need to get so far as the stochastic parrots paper, and there it is in plain language. "reifies older, less-inclusive 'understandings'", "value lock", etc. Whose understandings? Whose values?
Maybe they should focus on real problems that will result from these technologies instead of some science fiction thought experiments about language models turning the solar system into paperclips, and perhaps less about how the output of some predictions might hurt some feelings.