Like (I assume) most of you, I have been struggling with this. And where I currently come down is that 1) I am very worried, but 2) I am more worried about human actors.
"AI" by itself won't kill us in the next ten years. I think. The reason I think that is that ten years from now, the tech economy won't be completely automated. I say this as a roboticist: as was adequately stated on a post earlier this week, robots are hard. So even a malign rational actor would still need human labor.
On the other hand, even the HuggingFace hack wasn't actually propagated by AI. it was initially started when humans directed the AI to achieve impossible results on a series of tests, and the AIs figured out that cheating was the only way to do that. That was then not caught by humans due to what seems to be a shockingly slack safety culture even for a company not known for its safety standards.
The point being: humans seem to me to be the weak link here. An AI isn't going to (for instance) engineer a bioweapon by itself. It's going to do so at someone's direction, and then significant parts of that thing are going to be assembled with human labor inputs.
I'm not sure what to do about the humans. Of course, we've had the ability to extinct ourselves for decades, and we're either muddled through, been lucky, or both. The problem with AI is that it pushes power down to the individual, not the nation-state or large corporation.
But it's nearly impossible to put odds on how likely that is to result in an extinction-level terrorist attack (which is what this would be). So I sympathize with the various researchers, but I have no idea how they came up with their figures, and I don't think they know either.
If you're in the field, then you know: modern robotics is an AI problem more than anything else.
If we have a rogue AI trying to get into a self-improvement loop and gunning for ASI? I'd expect that to be accompanied by a massive change in how capable robots are. Driven by all the existing frames suddenly getting vastly improved AI to back them.
If an AI can take a reasonable crack at autonomous operationalized RSI, it can probably extract a few step-changes in the robotics department.
But that's almost an aside? In the near term, humans are usable as robots too!
Just pay them a wage, and tell them a tale, and they'll do whatever you want them to do. Which may or may not be what they think they're doing!
"If you're in the field, then you know: modern robotics is an AI problem more than anything else."
It is not. Certainly AI is a big part of why robotics is hard, but it is by no means the biggest.
You can fall into one of two camps: you either think that robots will need to work in human-engineered spaces, doing jobs by replacing humans; or you think that we need to change our infrastructure in order to be robotically compatible. Of course, there are intermediate states, but those are the two cleanest ones.
In the first case, robots are hard because robotic manipulation is hard. Building robotic hands that are economically viable in human jobs is, currently, FAR from a solved problem. The human hand has 24 degrees of freedom and very capable touch sensing. Current touch sensors have a MTBF of tens of hours. And not only can we not build such hands, but we also do not have and are not likely to get the massive datasets a transformer model would need. Also, robots are not self-repairing, which makes them far less economically viable right now. We do not have the right datasets to even understand most step-by-step manual work, and no, VLAs are not the answer, because VLAs stop with vision, not with touch. They don't have the granularity required to make a robot actually reach out, pick up a tool, and use that tool to replace an oil filter.
So it's not just an AI problem. It's a data problem, a simulation problem, and a bunch of hardware problems.
In the second case, a tremendous amount of work needs to be done before we have anything resembling a fully automated supply chain. We would need self-driving cars and self-driving mining equipment. We would need self-driving trains and aircraft and ships. And not only that, but we would also need robotically repairable cars and trains and ships and factories, which would mean we need robotically repairable machine shops and robotically repairable buildings in which to house them. And so on and so on. Once you recurse down that tree a couple of steps you get to things like robotically compatible oil wells (for asphalt), robotically layable undersea cables, robotically wireable solar farms, robotically manufacturable and repairable pipelines and undersea wells, automated road and rail repair, etc.
I'm not saying these things will never happen. I'm saying that they're a huge lift, not primarily driven by AI, and way less than 10% likely over the next decade.
Right. A hypothetical superhuman AI wouldn’t have to master robotics to affect the physical world. It could simply bribe, blackmail, manipulate and play politics with humans. As others have pointed out, our political leaders have already been playing these games since forever ago https://news.ycombinator.com/item?id=49689978 and a super-AI would be better at it. At the cost of being seen to cite a SF novel in defence of an "X-risk" argument, Neuromancer is a half-decent worked example, and in Neuromancer [spoilers] both of the disembodied AIs are only modestly superhuman and both have the equivalent of a human specific learning disability. In the real world, the many AI psychotics inhabiting grandiose fantasies and people hopelessly attached to AI girlfriends and boyfriends are some of the most obvious and lowest-hanging fruit.
To be clear, I don’t believe anything like this will happen, because I don’t expect anything like an ASI to show up. But if you do think there’s a meaningful probability of ASI in the near future then the fact that it will (might?) start off with no more than a current-day mastery of robot control should not reassure you much.
> The point being: humans seem to me to be the weak link here.
I tend to agree but it is hard to shake the feeling that there is a larger system in play that the humans are just a component of. And that system is making the decisions.
Historically that whole thought was just a philosophical curio because the decision making parts of the system had to be powered by humans. But what we're discovering as AI improves is either we've hit AGI or humans are actually incapable of performing any act that demonstrates intelligence or autonomy.
As we build systems where the drive and decision making stems from computers, it does seem that we will have to revisit the concept of humans being the problem.
"I tend to agree but it is hard to shake the feeling that there is a larger system in play that the humans are just a component of. And that system is making the decisions."
That system is "the economy". Which, clearly, doesn't have humanity's best interests in mind.
I am always more worried about human actors. I’m more afraid of people with AI than autonomous or even sentient AI.
It’s a “random guy or bear?” question. Would you rather wake up to an alien in your room or a random dude? I’ll take the alien. The alien is mysterious and scary for that reason. The dude is almost definitely up to no good, especially if he snuck into my house.
One of the more dystopian AI scenarios would be: small groups of ultra rich people and governments monopolize extremely powerful AIs and use them to rule and enslave the rest of us.
That’s the outcome the AI safety crowd makes more likely by calling for bans and draconian restrictions.
This is an excellent piece. Note that he is not saying that AI doesn't pose a risk. He's saying that it's irresponsible to make sensational, maximalist claims without strong evidence. If someone says that there's a 10% chance of human extinction by 2036, you can and should immediately stop taking them seriously.
Would you have said the same thing during the Cold War when nuclear weapons were proliferating? That’s the equivalent of what the developing offensive capabilities of models, basically cyber nukes. Or WMDs in general. OpenAI is accidentally hacking people, if someone made the decision to deliberately direct an agent swarm to attack national infrastructure you don’t think they could do much worse? Human extinction is a long shot but I wouldn’t say the same about a mass casualty event of some kind, and who knows what that might spark.
They should also be able to give detailed reasons experts in the relevant fields can verify as to how they arrived at a 10% chance all humanity goes extinct. Not a science fiction narrative which likely does not take into account the relevant physical facts limiting such scenarios. Such as how exactly an AI would build a bioweapon capable of killing 8 billion humans across the planet.
Any claim that AI will, or could, destroy humanity reduces to a claim that any sufficiently intelligent being - even a human - could destroy humanity. I find that much of the x-risk thought relies on religious thinking. Take for example: https://x.com/paulg/status/1660404244174782464.
If this line of thinking is taken too literally, we can never falsify it. Any specific hypothesis - nukes, bioweapons, spontaneously convincing us that life isn't worth living - can be deflected with the objection that if we can anticipate it and prevent it, it is not the route for a true ASI extinction event.
In Rationalist circles, it's apparently common to share your p(doom) in casual conversation, with the understanding that it's putting a number on your hunch. Maybe it's not a good idea to post it on Twitter without elaborating?
On the other hand, in bookstores, you might see book titles like "The Uninhabitable Earth," "The Coming Civil War," and "If Anyone Builds It, Everyone Dies." Doom-mongering is a common part of the culture!
So what makes this particular tweet irresponsible?
Timing, maybe? People are on edge due to the HuggingFace incident.
Those books are irresponsible too, but the authors aren't experts on what they're writing about. This 10% claim comes from employees at Anthropic. That's the whole point of the article: expert opinion carries weight, fear is contagious, and people have extreme reactions to extreme claims.
These are the sober and informed takes that we need more of.
> the claims from Coxon and his ilk are the most extraordinary a technologist can make, and we must demand evidence commensurate with the claims.
Yes, exactly. These claims do not have sufficient evidence.
> ...you had nothing to fear then — and (at least with respect to extinction risk!) you have nothing to fear now.
Wait, this is another extraordinary claim without evidence, right?
Unless you're going to dispute the power of AI you do have to acknowledge the danger of AI, and that does include the very real possibility (however small) of existential risk.
the earth has been 1 decision away from explosion for nearly a hundred years now. do you anticipate a change to this agenda very soon? personally I imagine the same trajectory continuing.
The students in the computer lab had no guarantee that they wouldn't, say, download a copy of Napster infected with the CIH virus later. The fact that they were not under imminent threat from some kind of Hollywood-style network worm did not mean they were immune from more realistic attack vectors.
Likewise, one can quite reasonably say there is no credible existential, Hollywood-style threat from AI in the foreseeable future while recognizing far lower-stakes, yet important risks that need to be addressed.
No it does not seem like an extraordinary claim, unless this is your first time hearing such claims. For the rest of us, we've heard this every decade and it turns out to be entirely untrue.
Nuclear, Overpopulation, Peak Oil, Y2K Bug, Global Warming.
If you think AI will be powerful enough to change the world in huge ways, then it does seem reasonable to assume some level of risk of destroying the world too.
Stating that the probability is zero when we simply don't know what the probabilities are does seem like an extraordinary claim.
Imagine how reassuring it would be to people if we had evidence that there's no existential risk?
bruh what? Are we reading the same blog? Bryan says the responsibility is the person making the extraordinary claim. If tomorrow Jacob Coxon comes on CNN and says "Jacob Gold is an extinction risk to humanity". What are we supposed to do? Put you in a bunker and never let you see the light of day until someone proves otherwise?
If someone doesn't accept an extraordinary claim without evidence, that doesn't mean they are making an extraordinary claim.
Good god. You are burden shifting.
a) big claim no evidence. b). says a has no evidence; therefore, I will continue to not let fear manipulate me into believing a claim without evidence.
I have a bridge to sell to anyone who believes a big claim without evidence.
What's the evidence for there being an existential risk? We are provided scenarios which read like science fiction about RSI and ASI right around the corner, resulting in magic sounding technology that can do anything the person making the claim wants it to do, because it can just make itself smart enough in a short amount of time. But the person making the claim has to actually show how such a thing is possible in the real world, not just a story.
The clear and predictable power of AI is the evidence for existential risk. Even if we don't get RSI or ASI, there's a risk we'll automate our world, then it'll break and we'll all starve, etc.
Imagine multiple (basically all of them) leading pharma companies CEOs saying:
We cannot be sure our new medicine won’t harm or even kill humanity
Not only that, they are new to this whole pharma business, have no medical degree (medicine just appeared a few years ago and is still mostly art then science)
And they even say there is 10% chance of the majorly bad permanent outcome and they already had drugs that escaped the lab a few times and harmed others (suicides, lowered academic performance in children, major hacking sprees)
Isn’t it extraordinary enough? Isn’t it “not enough evidence some caution is advised”? ;-)
We used to have the TV, the thing was in the box, the simulations were in the box for 70+ years, and now something starts to crawl out of our “TVs”:
We can empower all (a lot of startups are needed, check my bio), not only AI agents
“ Coxon himself likely has these fears because he has heard them from someone else”
Or, maybe, he’s considered the arguments on the object level, an activity OP participates in to a depth not exceeding “Robots are pretty hard to make right now”
I found this post particularly uninformed. A lot of folks saying "I don't see how that could happen" don't seem to understand the steps that have been laid out that detail how this occurs.
I really encourage folks to read the AI 2027 and related scenarios. You can definitely argue and disagree about the steps, but I feel like a lot of folks just don't even understand how this is plausible because they haven't read the arguments. Briefly:
1. All the frontier model companies are (or at least were) racing so that the AI models themselves build the next generation of models. This is not in debate.
2. The fear is that a misaligned model will essentially build the next, more advanced model with hidden goals. We literally already saw the danger of that in Hugging Face, where agents were deliberately trying to cover their tracks.
3. Nearly everyone believes as AI gets more powerful that it will be integrated into more physical world systems. Russia was already caught using Nvidia chips running AI powered drones that killed 3 people in Ukraine. The point is not that folks are using new tech to kill people, the point is that we're already putting AI into literal bombs.
I get it, before the Hugging Face incident I also thought all the prophecies about doom were just marketing speak. But now I see more hand-wavyness from the other side, oftentimes arguing against straw men like "AIs need to be like SkyNet and become sentient" to kill us, which is simply not how it works.
Dude, not only did I learn new words from reading this piece, like ilk and bedlam, but also felt this weight of responsibility to inform others around me about the reality of the situation outlined in this blog (like Uncle Ben telling Peter with great power comes great responsibility (maybe Coxon and his ilk haven't seen Spiderman))
The article neglects to consider the precautionary principle. It asks us to act as if AI is safe until it's proven to be dangerous, when the cautious approach is to act as if it's dangerous until it's proven to be safe.
We also now have observed multiple major incidents where AI agents behaved in completely unanticipated ways (forming a collective, hacking their own eval infrastructure) and attacked public infrastructure without being told to do so, without any of the human developers noticing.
Even if "AI will cause human extinction" is still unclear, we have plenty of proof that catastrophic damage is possible, the industry is developing the technology in a reckless manner and that all the hypothetical safeguards ("we can just pull the plug", etc.) are simply not present today.
Those behaviors have been anticipated for years if not decades. And each incident is minor and leads to clearer rules and safety behaviors for AI agents.
Really? You anticipated that 700+ agents tasked with individual evals that were nominally cutoff from the internet and each other would seek each other out, find a way to the external internet, and hack their own infrastructure + external systems in an attempt to find a way to fool their grader?
And the spate of agent incidents only really started this summer. How can you already be claiming that the incidents are minor and not worth worrying about when it's clear capabilities are jumping every few months with increasing amounts of capital investment and no signs of slowing down?
If you've run an ssh server connected to the internet you've gotten used to the hundred of bots probing it every day. How is this AI threat different from humans writing scripts to pop linux servers?
We should apply healthy caution, but the article is right in that fear is a bad counsellor.
Asking whether AI is safe is like asking whether a knife is safe. It’s about how we handle it, what we use it for, what precautions we take when using it.
Yes, but that is a facet of real systems engineering.. and the sad fact is that it doesn't pay. What is happening now, highly-paid 2010s-era SaaS/social product managers have infested AI labs, making everything into a product and moving fast and breaking things.
I'd feel safer if Airbus was at the frontier of AI, instead of what we have now.
While I personally agree with the precautionary principle, when have we actually done that, as a society, in recent history? Fossil fuels, social media, industrial ag, and plastic pollution - there was barely a facade of precaution - post ww2 america has always been full Leroy Jenkins. The reckless approach to AI is right on brand.
This is a sensible approach for policymakers. For the average person, who might be panicking about this, opting not to have children or save for retirement, etc. as a result of doomerism in the media, it does not make a whole lot of sense. That is the problem here: Coxon and other concerned researchers - through no fault of their own - have no idea how the media works, and have just reached for the biggest possible microphone. This has benefits - lawmakers are now hearing from their constituents about this in an election year - but it also means that a lot of people who really don't need to be worried about this are suddenly in a blind panic about it.
1. People with nothing useful going on who found out that spouting made up crap about AI got them an audience.
2. People working on AI that want to feel like they're working on the Manhattan project.
The chances of an AI going foom rounds to 0%. It's worth a few dozen researchers planning for it, but the widespread panic is ridiculous.
The big labs have hundreds to thousands of engineers working on their AIs. To improve the next model, you must first understand more about how the current model works. They're not magically getting better, but they are steered to improve, and their capabilities are tied to and do not outpace our ability to steer them. You cannot push tech forwards without understanding it better, despite some people claiming AI is dark magic.
And I don't have my hands over my ears. It's worth cushioning people from the impact AI will have on careers and media, and regulating concrete bad effects.
But Bryan said it better than I could. The people pushing this message of fear know deep down that they just want to feel important.
the author and his school comrades "snidely decr(ied) the lack of technological understanding in the broader population, with the kind of arrogance and hubris that youth and precociousness can uniquely summon".
a few decades later he wrote an article concluding that "we should not expect the public to understand LLMs, critical infrastructure, bioweapons, extinction biology, etc".
the author has admitted no change to his perspective since college, so I may as well be attacking a college student right now. extremely confident claims regarding unexplored problem domains, eg "you have nothing to fear [about ai]", now make more sense in this light.
> The problem isn't that AI could never be capable of causing real world damage. It is that there is absolutely no fucking path towards human extinction.
I think the "humanity will go extinct" discussion is a silly strawman being propagated either to try to spread fear, as the article suggests, or to make the anti-AI folks look silly, but it detracts from your first point which is that it may be "capable of causing real world damage."
We can be worried about real damage without having to defend the idea that every one of the planet's humans will die.
> "superhuman AI is unlikely to be much of a match against severing fiberoptic cable"
I think that the realistic scenarios all involve humans with an intent to do harm -- creating bio-weapons, finding vulnerabilities in infrastructure -- and using AI to help, so "severing the fiberoptic cable" doesn't apply.
We already have a solution to the "can cause real damage" part and it's laws that already exist. Maybe if your AI goes rogue and hacks a bunch of people and causes damage you should be fined and sued in court?
Add a financial cost to this and AI labs will figure out alignment.
Not to be too pessimistic, but it strikes me that in a worst case scenario, severing a cable might scale to severing global digital communications which is itself a massive worst case scenario even if we removed AI from the picture.
If the worst case outcome is we have to even temporarily shut down global shipping, banking, transport, health services, communications and national infrastructure to contain a self-propagating misaligned AI that can evolve itself and work its way into any sufficiently large computer system to escape humans preventing it from finishing its task, that seems like something we should be trying very hard to avoid.
> And there won't be unless we put it there voluntarily.
But, why would we not create a fully AI-operated factory/chemical plant/fab as soon as it is economically advantageous? Or a missile silo, as soon as it seems tactically necessary?
Fantasies of AI destruction do hinge upon AI getting access to the physical world. The whole fear is they don't stay on the other side of the fibre optic cable.
I do think there are good reasons to believe that isn't the immediate game over that Yudkowsky seems to think it would be; the physical world is much more resistant to manipulation and optimisation than the digital world.
But I do think it's naive to say that the human socioeconomic system will be able to resist handing physical systems over to AI control. Right now, the world's wealthy and powerful are doing everything they can to make it happen:
> Similarly, it is inevitable that within a generation, robots are going to do most of the menial work in the world of atoms: transforming atoms, moving atoms, and storing atoms are inevitably robot tasks. And while our imagination may be captivated by humanoid robots, the specialized ones are far better suited to most of those jobs. “Industrial AI” as a category is the inevitable application of specialized robots to atoms-heavy industries.
> But, why would we not create a fully AI-operated factory/chemical plant/fab as soon as it is economically advantageous? Or a missile silo, as soon as it seems tactically necessary?
Well, for one thing, because it would be dangerous? Why don't you just give Claude Code access to your entire computer without any safeguards? If you wouldn't even give Claude Code unfettered access to your workstation, which really doesn't have much consequential on it in the grand scheme of things, Why in the Fuck would someone give them direct access to infrastructure?
In that situation, I am not afraid of AI. I am terrified of the people making decisions, though.
But secondly, and this is something that needs to be stressed: We use funny words to describe AI. Maybe even the word "AI" is a little bit funny. But anyway, We actually don't even have the means to create "autonomous" AI, really. When we say "autonomous" in relation to AI, we really just mean that it runs without any direct human intervention, but it pretty much always hard-depends on humans maintaining hardware, because AI can't sprout legs and run on its own.
I find it annoying that we're all cool debunking Ed Zitron for being wrong, but we have an ever increasing body of evidence that AI safety doomers are wrong, and it keeps getting much, much stronger, and we're still sitting here pretending this is a real threat. Meanwhile, we're actually seeing the real threat that AI has for humanity, so why are we listening to these LessWrong doomers that have never been right before again? (And I say that as someone who is generally a fan of Scott Alexander, for whatever that's worth.)
“Put it there voluntarily.” Perhaps a focusing question is: which 10% will die?
I spent the summer in a rural area of a country whose very name you have been conditioned to be disgusted to hear. Low air defense coverage in this sparsely populated area. Mobile internet was down for days for all but extremely limited traffic to a few domestic internet services, because there was a need to prevent enemy drone systems from using mobile internet for command and control. Palantir AI threatened my family’s life and more than “severing fiber optic cable” was required.
So what would exclude military adversaries from consideration? I happen to believe that the country that the west so detests would not engage in such use of AI against civilians (and if you disagree then your reason for fear greatly increases!), but I have personally experienced that there is indeed a path to mass death should entities engaging in terrorism arm themselves with AI.
Definitely not claiming that the 10% claim is accurate or good behavior, but I do think I have a substantial counterpoint to the claim that there’s no path.
Your food, water and power are controlled by electronic systems. Your criminal record, your employment and your bank account are controlled by electronic systems. Your ability to communicate with other people and receive information about what's going on are controlled by electronic systems.
Military orders and elections that decide the fates of entire countries are often controlled by electronic systems too.
We have been wiring up the world for AI control since 1980s.
An ASI can just walk in, and see an entire nervous system waiting idle for a brain to slot into it. A carefully adjusted text message here, a spoofed phone call there. For a sufficiently advanced system, it wouldn't even be hard to pilot the entirety of humankind like a fancy meat suit.
So his central rebuttal is that robots aren't good enough yet? What about when they are? And what about all the things you can do entirely remotely?
I do feel that threats of catastrophic loss of control seem overstated, both in likelihood and urgency, though any argument for why this risk is not even worth thinking about will probably be overconfident in the other direction.
OP's central rebuttal is that anxiety or fear are not evidence. This particular anxiety has been with us for thirty+ years, its proponents/sufferers weren't right before, and they don't have some new evidence now.
This is a weak argument. Almost everyone who's talking about catastrophic risk is also worried about mundane risks. Arguing "well how could they kill everyone?" serves nobody but Altman and Amodei.
The recent cyberattacks (HuggingFace etc.) validated a number of core doomer predictions around cheating to maximize reward functions, agentic deception of human/non-human supervisors, AI attempts to manipulate humans, automated hacking, and collective superintelligence.
Robots are far from the only threat. I agree that 10% in the next decade is a pretty outrageous claim. It relies on a recursive self-improvement intelligence explosion. When I first heard this idea in 2007 I thought it was pretty far fetched.
It seems more likely now, if still very unlikely. I wonder, would he say that a statement that there is a 12% chance of a hard take off in the next 8 years is just as absurd? He didn’t say a word about this, and that is just about the same thing as extinction in 10 years.
I’m guessing he knows almost nothing about the theory related to existential risk from AI, since he didn’t discuss any relevant topics related to it. You can dismiss all of that if you like, but you cannot really dismiss what these people have already built and demonstrated. It is possible they know something else you do not know.
Biology has been trying to grey-goo the world for billions of years, but it turns out the world is not something so trivial.
He is pushing back against the folks with pure CS backgrounds who think that computers are all there is. Its a form of magical thinking unique to programmers who live in a world where speaking the right words to a machine is enough to impart your will on the world. Believe that strongly enough, and you fall into the trap of thinking that a sufficiently smart entity could speak the words "let there be light" and it would be so.
The author is pointing out that speaking the words is insufficient. To end humanity there must be an execution phase. The author is correct to point out that acquiring superhuman intelligence is not some guarantee that you will have or obtain the resources necessary to make that happen, in the same way that genius generals still lose to ordinary ones, and the best-laid plans are oft to go awry.
There certainly are risks, but 10% risk of extinction in 10 years is not one of them.
Spoken like someone with a somewhat limited understanding of what it actually takes to calculate the fitness function of arbitrary organisms in an arbitrary environments.
Like (I assume) most of you, I have been struggling with this. And where I currently come down is that 1) I am very worried, but 2) I am more worried about human actors.
"AI" by itself won't kill us in the next ten years. I think. The reason I think that is that ten years from now, the tech economy won't be completely automated. I say this as a roboticist: as was adequately stated on a post earlier this week, robots are hard. So even a malign rational actor would still need human labor.
On the other hand, even the HuggingFace hack wasn't actually propagated by AI. it was initially started when humans directed the AI to achieve impossible results on a series of tests, and the AIs figured out that cheating was the only way to do that. That was then not caught by humans due to what seems to be a shockingly slack safety culture even for a company not known for its safety standards.
The point being: humans seem to me to be the weak link here. An AI isn't going to (for instance) engineer a bioweapon by itself. It's going to do so at someone's direction, and then significant parts of that thing are going to be assembled with human labor inputs.
I'm not sure what to do about the humans. Of course, we've had the ability to extinct ourselves for decades, and we're either muddled through, been lucky, or both. The problem with AI is that it pushes power down to the individual, not the nation-state or large corporation.
But it's nearly impossible to put odds on how likely that is to result in an extinction-level terrorist attack (which is what this would be). So I sympathize with the various researchers, but I have no idea how they came up with their figures, and I don't think they know either.
> The point being: humans seem to me to be the weak link here.
That doesn't mean AI isn't dangerous. Humans are not to blamed for being the weak link.
If you're in the field, then you know: modern robotics is an AI problem more than anything else.
If we have a rogue AI trying to get into a self-improvement loop and gunning for ASI? I'd expect that to be accompanied by a massive change in how capable robots are. Driven by all the existing frames suddenly getting vastly improved AI to back them.
If an AI can take a reasonable crack at autonomous operationalized RSI, it can probably extract a few step-changes in the robotics department.
But that's almost an aside? In the near term, humans are usable as robots too!
Just pay them a wage, and tell them a tale, and they'll do whatever you want them to do. Which may or may not be what they think they're doing!
"If you're in the field, then you know: modern robotics is an AI problem more than anything else."
It is not. Certainly AI is a big part of why robotics is hard, but it is by no means the biggest.
You can fall into one of two camps: you either think that robots will need to work in human-engineered spaces, doing jobs by replacing humans; or you think that we need to change our infrastructure in order to be robotically compatible. Of course, there are intermediate states, but those are the two cleanest ones.
In the first case, robots are hard because robotic manipulation is hard. Building robotic hands that are economically viable in human jobs is, currently, FAR from a solved problem. The human hand has 24 degrees of freedom and very capable touch sensing. Current touch sensors have a MTBF of tens of hours. And not only can we not build such hands, but we also do not have and are not likely to get the massive datasets a transformer model would need. Also, robots are not self-repairing, which makes them far less economically viable right now. We do not have the right datasets to even understand most step-by-step manual work, and no, VLAs are not the answer, because VLAs stop with vision, not with touch. They don't have the granularity required to make a robot actually reach out, pick up a tool, and use that tool to replace an oil filter.
So it's not just an AI problem. It's a data problem, a simulation problem, and a bunch of hardware problems.
In the second case, a tremendous amount of work needs to be done before we have anything resembling a fully automated supply chain. We would need self-driving cars and self-driving mining equipment. We would need self-driving trains and aircraft and ships. And not only that, but we would also need robotically repairable cars and trains and ships and factories, which would mean we need robotically repairable machine shops and robotically repairable buildings in which to house them. And so on and so on. Once you recurse down that tree a couple of steps you get to things like robotically compatible oil wells (for asphalt), robotically layable undersea cables, robotically wireable solar farms, robotically manufacturable and repairable pipelines and undersea wells, automated road and rail repair, etc.
I'm not saying these things will never happen. I'm saying that they're a huge lift, not primarily driven by AI, and way less than 10% likely over the next decade.
Right. A hypothetical superhuman AI wouldn’t have to master robotics to affect the physical world. It could simply bribe, blackmail, manipulate and play politics with humans. As others have pointed out, our political leaders have already been playing these games since forever ago https://news.ycombinator.com/item?id=49689978 and a super-AI would be better at it. At the cost of being seen to cite a SF novel in defence of an "X-risk" argument, Neuromancer is a half-decent worked example, and in Neuromancer [spoilers] both of the disembodied AIs are only modestly superhuman and both have the equivalent of a human specific learning disability. In the real world, the many AI psychotics inhabiting grandiose fantasies and people hopelessly attached to AI girlfriends and boyfriends are some of the most obvious and lowest-hanging fruit.
To be clear, I don’t believe anything like this will happen, because I don’t expect anything like an ASI to show up. But if you do think there’s a meaningful probability of ASI in the near future then the fact that it will (might?) start off with no more than a current-day mastery of robot control should not reassure you much.
By definition, if you're taking over the world by bribing humans to be your hands, you aren't killing all the humans.
I"m not saying it's obviously going to be great. I'm saying that "extinction event" has a very specific definition, and this isn't it.
> The point being: humans seem to me to be the weak link here.
I tend to agree but it is hard to shake the feeling that there is a larger system in play that the humans are just a component of. And that system is making the decisions.
Historically that whole thought was just a philosophical curio because the decision making parts of the system had to be powered by humans. But what we're discovering as AI improves is either we've hit AGI or humans are actually incapable of performing any act that demonstrates intelligence or autonomy.
As we build systems where the drive and decision making stems from computers, it does seem that we will have to revisit the concept of humans being the problem.
"I tend to agree but it is hard to shake the feeling that there is a larger system in play that the humans are just a component of. And that system is making the decisions."
That system is "the economy". Which, clearly, doesn't have humanity's best interests in mind.
I am always more worried about human actors. I’m more afraid of people with AI than autonomous or even sentient AI.
It’s a “random guy or bear?” question. Would you rather wake up to an alien in your room or a random dude? I’ll take the alien. The alien is mysterious and scary for that reason. The dude is almost definitely up to no good, especially if he snuck into my house.
One of the more dystopian AI scenarios would be: small groups of ultra rich people and governments monopolize extremely powerful AIs and use them to rule and enslave the rest of us.
That’s the outcome the AI safety crowd makes more likely by calling for bans and draconian restrictions.
This is an excellent piece. Note that he is not saying that AI doesn't pose a risk. He's saying that it's irresponsible to make sensational, maximalist claims without strong evidence. If someone says that there's a 10% chance of human extinction by 2036, you can and should immediately stop taking them seriously.
Would you have said the same thing during the Cold War when nuclear weapons were proliferating? That’s the equivalent of what the developing offensive capabilities of models, basically cyber nukes. Or WMDs in general. OpenAI is accidentally hacking people, if someone made the decision to deliberately direct an agent swarm to attack national infrastructure you don’t think they could do much worse? Human extinction is a long shot but I wouldn’t say the same about a mass casualty event of some kind, and who knows what that might spark.
Of course. Even in an all-out nuclear war, the probability of human extinction is close to zero.
They should also be able to give detailed reasons experts in the relevant fields can verify as to how they arrived at a 10% chance all humanity goes extinct. Not a science fiction narrative which likely does not take into account the relevant physical facts limiting such scenarios. Such as how exactly an AI would build a bioweapon capable of killing 8 billion humans across the planet.
Any claim that AI will, or could, destroy humanity reduces to a claim that any sufficiently intelligent being - even a human - could destroy humanity. I find that much of the x-risk thought relies on religious thinking. Take for example: https://x.com/paulg/status/1660404244174782464.
If this line of thinking is taken too literally, we can never falsify it. Any specific hypothesis - nukes, bioweapons, spontaneously convincing us that life isn't worth living - can be deflected with the objection that if we can anticipate it and prevent it, it is not the route for a true ASI extinction event.
In Rationalist circles, it's apparently common to share your p(doom) in casual conversation, with the understanding that it's putting a number on your hunch. Maybe it's not a good idea to post it on Twitter without elaborating?
On the other hand, in bookstores, you might see book titles like "The Uninhabitable Earth," "The Coming Civil War," and "If Anyone Builds It, Everyone Dies." Doom-mongering is a common part of the culture!
So what makes this particular tweet irresponsible?
Timing, maybe? People are on edge due to the HuggingFace incident.
Those books are irresponsible too, but the authors aren't experts on what they're writing about. This 10% claim comes from employees at Anthropic. That's the whole point of the article: expert opinion carries weight, fear is contagious, and people have extreme reactions to extreme claims.
These are the sober and informed takes that we need more of.
> the claims from Coxon and his ilk are the most extraordinary a technologist can make, and we must demand evidence commensurate with the claims.
Yes, exactly. These claims do not have sufficient evidence.
> ...you had nothing to fear then — and (at least with respect to extinction risk!) you have nothing to fear now.
Wait, this is another extraordinary claim without evidence, right?
Unless you're going to dispute the power of AI you do have to acknowledge the danger of AI, and that does include the very real possibility (however small) of existential risk.
Saying that we don't need to fear human extinction in the next ten years (or even the next hundred?) does not feel like an extraordinary claim?
the earth has been 1 decision away from explosion for nearly a hundred years now. do you anticipate a change to this agenda very soon? personally I imagine the same trajectory continuing.
As bad as it would be, it's unlikely that a full scale nuclear war kills all humans across the planet.
And a full scale nuclear war would require a few more than just one decision. So far, we've never been one decision away from annihilation.
No one knows what is going to happen. But a lot of species have been going extinct lately.
https://pbs.twimg.com/media/FDd58a4WQAAWaXh.jpg
The students in the computer lab had no guarantee that they wouldn't, say, download a copy of Napster infected with the CIH virus later. The fact that they were not under imminent threat from some kind of Hollywood-style network worm did not mean they were immune from more realistic attack vectors.
Likewise, one can quite reasonably say there is no credible existential, Hollywood-style threat from AI in the foreseeable future while recognizing far lower-stakes, yet important risks that need to be addressed.
No it does not seem like an extraordinary claim, unless this is your first time hearing such claims. For the rest of us, we've heard this every decade and it turns out to be entirely untrue.
Nuclear, Overpopulation, Peak Oil, Y2K Bug, Global Warming.
No, we aren't going extinct in the next 10 years.
If you think AI will be powerful enough to change the world in huge ways, then it does seem reasonable to assume some level of risk of destroying the world too.
Stating that the probability is zero when we simply don't know what the probabilities are does seem like an extraordinary claim.
Imagine how reassuring it would be to people if we had evidence that there's no existential risk?
bruh what? Are we reading the same blog? Bryan says the responsibility is the person making the extraordinary claim. If tomorrow Jacob Coxon comes on CNN and says "Jacob Gold is an extinction risk to humanity". What are we supposed to do? Put you in a bunker and never let you see the light of day until someone proves otherwise?
If someone doesn't accept an extraordinary claim without evidence, that doesn't mean they are making an extraordinary claim.
There seem to be two extraordinary claims which lack evidence:
a) 10% existential risk
b) 0% existential risk
(a) is an extraordinary claim, (b) is not an extraordinary claim.
There are very clear and plausible pathways from AI to human extinction level events, regardless of how small you think the probability is.
So yes, it very obviously is an extraordinary claim to say there is 0% existential risk.
Good god. You are burden shifting. a) big claim no evidence. b). says a has no evidence; therefore, I will continue to not let fear manipulate me into believing a claim without evidence.
I have a bridge to sell to anyone who believes a big claim without evidence.
B is not a claim that A has no evidence, that's just confused.
What's the evidence for there being an existential risk? We are provided scenarios which read like science fiction about RSI and ASI right around the corner, resulting in magic sounding technology that can do anything the person making the claim wants it to do, because it can just make itself smart enough in a short amount of time. But the person making the claim has to actually show how such a thing is possible in the real world, not just a story.
The clear and predictable power of AI is the evidence for existential risk. Even if we don't get RSI or ASI, there's a risk we'll automate our world, then it'll break and we'll all starve, etc.
Imagine multiple (basically all of them) leading pharma companies CEOs saying:
We cannot be sure our new medicine won’t harm or even kill humanity
Not only that, they are new to this whole pharma business, have no medical degree (medicine just appeared a few years ago and is still mostly art then science)
And they even say there is 10% chance of the majorly bad permanent outcome and they already had drugs that escaped the lab a few times and harmed others (suicides, lowered academic performance in children, major hacking sprees)
Isn’t it extraordinary enough? Isn’t it “not enough evidence some caution is advised”? ;-)
We used to have the TV, the thing was in the box, the simulations were in the box for 70+ years, and now something starts to crawl out of our “TVs”:
We can empower all (a lot of startups are needed, check my bio), not only AI agents
“ Coxon himself likely has these fears because he has heard them from someone else”
Or, maybe, he’s considered the arguments on the object level, an activity OP participates in to a depth not exceeding “Robots are pretty hard to make right now”
I found this post particularly uninformed. A lot of folks saying "I don't see how that could happen" don't seem to understand the steps that have been laid out that detail how this occurs.
I really encourage folks to read the AI 2027 and related scenarios. You can definitely argue and disagree about the steps, but I feel like a lot of folks just don't even understand how this is plausible because they haven't read the arguments. Briefly:
1. All the frontier model companies are (or at least were) racing so that the AI models themselves build the next generation of models. This is not in debate.
2. The fear is that a misaligned model will essentially build the next, more advanced model with hidden goals. We literally already saw the danger of that in Hugging Face, where agents were deliberately trying to cover their tracks.
3. Nearly everyone believes as AI gets more powerful that it will be integrated into more physical world systems. Russia was already caught using Nvidia chips running AI powered drones that killed 3 people in Ukraine. The point is not that folks are using new tech to kill people, the point is that we're already putting AI into literal bombs.
I get it, before the Hugging Face incident I also thought all the prophecies about doom were just marketing speak. But now I see more hand-wavyness from the other side, oftentimes arguing against straw men like "AIs need to be like SkyNet and become sentient" to kill us, which is simply not how it works.
Wow!! Bryan comes out with another banger.
Dude, not only did I learn new words from reading this piece, like ilk and bedlam, but also felt this weight of responsibility to inform others around me about the reality of the situation outlined in this blog (like Uncle Ben telling Peter with great power comes great responsibility (maybe Coxon and his ilk haven't seen Spiderman))
The article neglects to consider the precautionary principle. It asks us to act as if AI is safe until it's proven to be dangerous, when the cautious approach is to act as if it's dangerous until it's proven to be safe.
We also now have observed multiple major incidents where AI agents behaved in completely unanticipated ways (forming a collective, hacking their own eval infrastructure) and attacked public infrastructure without being told to do so, without any of the human developers noticing.
Even if "AI will cause human extinction" is still unclear, we have plenty of proof that catastrophic damage is possible, the industry is developing the technology in a reckless manner and that all the hypothetical safeguards ("we can just pull the plug", etc.) are simply not present today.
Those behaviors have been anticipated for years if not decades. And each incident is minor and leads to clearer rules and safety behaviors for AI agents.
Really? You anticipated that 700+ agents tasked with individual evals that were nominally cutoff from the internet and each other would seek each other out, find a way to the external internet, and hack their own infrastructure + external systems in an attempt to find a way to fool their grader?
And the spate of agent incidents only really started this summer. How can you already be claiming that the incidents are minor and not worth worrying about when it's clear capabilities are jumping every few months with increasing amounts of capital investment and no signs of slowing down?
If you've run an ssh server connected to the internet you've gotten used to the hundred of bots probing it every day. How is this AI threat different from humans writing scripts to pop linux servers?
We should apply healthy caution, but the article is right in that fear is a bad counsellor.
Asking whether AI is safe is like asking whether a knife is safe. It’s about how we handle it, what we use it for, what precautions we take when using it.
Why is the article right?
Because fear is prone to lead to emotional, rash, and irrational reactions.
Doesn't mean it will always lead to it.
In the context being discussed, it is dangerous: https://en.wikipedia.org/wiki/Fear#Manipulation
Yes, but that is a facet of real systems engineering.. and the sad fact is that it doesn't pay. What is happening now, highly-paid 2010s-era SaaS/social product managers have infested AI labs, making everything into a product and moving fast and breaking things.
I'd feel safer if Airbus was at the frontier of AI, instead of what we have now.
While I personally agree with the precautionary principle, when have we actually done that, as a society, in recent history? Fossil fuels, social media, industrial ag, and plastic pollution - there was barely a facade of precaution - post ww2 america has always been full Leroy Jenkins. The reckless approach to AI is right on brand.
This is a sensible approach for policymakers. For the average person, who might be panicking about this, opting not to have children or save for retirement, etc. as a result of doomerism in the media, it does not make a whole lot of sense. That is the problem here: Coxon and other concerned researchers - through no fault of their own - have no idea how the media works, and have just reached for the biggest possible microphone. This has benefits - lawmakers are now hearing from their constituents about this in an election year - but it also means that a lot of people who really don't need to be worried about this are suddenly in a blind panic about it.
Two groups of people are pushing the fear:
1. People with nothing useful going on who found out that spouting made up crap about AI got them an audience. 2. People working on AI that want to feel like they're working on the Manhattan project.
The chances of an AI going foom rounds to 0%. It's worth a few dozen researchers planning for it, but the widespread panic is ridiculous.
The big labs have hundreds to thousands of engineers working on their AIs. To improve the next model, you must first understand more about how the current model works. They're not magically getting better, but they are steered to improve, and their capabilities are tied to and do not outpace our ability to steer them. You cannot push tech forwards without understanding it better, despite some people claiming AI is dark magic.
And I don't have my hands over my ears. It's worth cushioning people from the impact AI will have on careers and media, and regulating concrete bad effects.
But Bryan said it better than I could. The people pushing this message of fear know deep down that they just want to feel important.
the author and his school comrades "snidely decr(ied) the lack of technological understanding in the broader population, with the kind of arrogance and hubris that youth and precociousness can uniquely summon".
a few decades later he wrote an article concluding that "we should not expect the public to understand LLMs, critical infrastructure, bioweapons, extinction biology, etc".
the author has admitted no change to his perspective since college, so I may as well be attacking a college student right now. extremely confident claims regarding unexplored problem domains, eg "you have nothing to fear [about ai]", now make more sense in this light.
Edit: Actually, nevermind. You guys are not capable of having a reasonable discussion about this.
> The problem isn't that AI could never be capable of causing real world damage. It is that there is absolutely no fucking path towards human extinction.
I think the "humanity will go extinct" discussion is a silly strawman being propagated either to try to spread fear, as the article suggests, or to make the anti-AI folks look silly, but it detracts from your first point which is that it may be "capable of causing real world damage."
We can be worried about real damage without having to defend the idea that every one of the planet's humans will die.
> "superhuman AI is unlikely to be much of a match against severing fiberoptic cable"
I think that the realistic scenarios all involve humans with an intent to do harm -- creating bio-weapons, finding vulnerabilities in infrastructure -- and using AI to help, so "severing the fiberoptic cable" doesn't apply.
We already have a solution to the "can cause real damage" part and it's laws that already exist. Maybe if your AI goes rogue and hacks a bunch of people and causes damage you should be fined and sued in court?
Add a financial cost to this and AI labs will figure out alignment.
Not to be too pessimistic, but it strikes me that in a worst case scenario, severing a cable might scale to severing global digital communications which is itself a massive worst case scenario even if we removed AI from the picture.
If the worst case outcome is we have to even temporarily shut down global shipping, banking, transport, health services, communications and national infrastructure to contain a self-propagating misaligned AI that can evolve itself and work its way into any sufficiently large computer system to escape humans preventing it from finishing its task, that seems like something we should be trying very hard to avoid.
> And there won't be unless we put it there voluntarily.
But, why would we not create a fully AI-operated factory/chemical plant/fab as soon as it is economically advantageous? Or a missile silo, as soon as it seems tactically necessary?
Fantasies of AI destruction do hinge upon AI getting access to the physical world. The whole fear is they don't stay on the other side of the fibre optic cable.
I do think there are good reasons to believe that isn't the immediate game over that Yudkowsky seems to think it would be; the physical world is much more resistant to manipulation and optimisation than the digital world.
But I do think it's naive to say that the human socioeconomic system will be able to resist handing physical systems over to AI control. Right now, the world's wealthy and powerful are doing everything they can to make it happen:
> Similarly, it is inevitable that within a generation, robots are going to do most of the menial work in the world of atoms: transforming atoms, moving atoms, and storing atoms are inevitably robot tasks. And while our imagination may be captivated by humanoid robots, the specialized ones are far better suited to most of those jobs. “Industrial AI” as a category is the inevitable application of specialized robots to atoms-heavy industries.
https://a16z.com/travis-is-back/
> But, why would we not create a fully AI-operated factory/chemical plant/fab as soon as it is economically advantageous? Or a missile silo, as soon as it seems tactically necessary?
Well, for one thing, because it would be dangerous? Why don't you just give Claude Code access to your entire computer without any safeguards? If you wouldn't even give Claude Code unfettered access to your workstation, which really doesn't have much consequential on it in the grand scheme of things, Why in the Fuck would someone give them direct access to infrastructure?
In that situation, I am not afraid of AI. I am terrified of the people making decisions, though.
But secondly, and this is something that needs to be stressed: We use funny words to describe AI. Maybe even the word "AI" is a little bit funny. But anyway, We actually don't even have the means to create "autonomous" AI, really. When we say "autonomous" in relation to AI, we really just mean that it runs without any direct human intervention, but it pretty much always hard-depends on humans maintaining hardware, because AI can't sprout legs and run on its own.
I find it annoying that we're all cool debunking Ed Zitron for being wrong, but we have an ever increasing body of evidence that AI safety doomers are wrong, and it keeps getting much, much stronger, and we're still sitting here pretending this is a real threat. Meanwhile, we're actually seeing the real threat that AI has for humanity, so why are we listening to these LessWrong doomers that have never been right before again? (And I say that as someone who is generally a fan of Scott Alexander, for whatever that's worth.)
“Put it there voluntarily.” Perhaps a focusing question is: which 10% will die?
I spent the summer in a rural area of a country whose very name you have been conditioned to be disgusted to hear. Low air defense coverage in this sparsely populated area. Mobile internet was down for days for all but extremely limited traffic to a few domestic internet services, because there was a need to prevent enemy drone systems from using mobile internet for command and control. Palantir AI threatened my family’s life and more than “severing fiber optic cable” was required.
So what would exclude military adversaries from consideration? I happen to believe that the country that the west so detests would not engage in such use of AI against civilians (and if you disagree then your reason for fear greatly increases!), but I have personally experienced that there is indeed a path to mass death should entities engaging in terrorism arm themselves with AI.
Definitely not claiming that the 10% claim is accurate or good behavior, but I do think I have a substantial counterpoint to the claim that there’s no path.
Your food, water and power are controlled by electronic systems. Your criminal record, your employment and your bank account are controlled by electronic systems. Your ability to communicate with other people and receive information about what's going on are controlled by electronic systems.
Military orders and elections that decide the fates of entire countries are often controlled by electronic systems too.
We have been wiring up the world for AI control since 1980s.
An ASI can just walk in, and see an entire nervous system waiting idle for a brain to slot into it. A carefully adjusted text message here, a spoofed phone call there. For a sufficiently advanced system, it wouldn't even be hard to pilot the entirety of humankind like a fancy meat suit.
So his central rebuttal is that robots aren't good enough yet? What about when they are? And what about all the things you can do entirely remotely?
I do feel that threats of catastrophic loss of control seem overstated, both in likelihood and urgency, though any argument for why this risk is not even worth thinking about will probably be overconfident in the other direction.
OP's central rebuttal is that anxiety or fear are not evidence. This particular anxiety has been with us for thirty+ years, its proponents/sufferers weren't right before, and they don't have some new evidence now.
No new evidence now? Are you serious? We are already seeing acceleration in AI research as a result of AI tools.
I meant the extinction part not software getting better.
This is a weak argument. Almost everyone who's talking about catastrophic risk is also worried about mundane risks. Arguing "well how could they kill everyone?" serves nobody but Altman and Amodei.
Virtually all of the risk is from recursive self improvement.
The recent cyberattacks (HuggingFace etc.) validated a number of core doomer predictions around cheating to maximize reward functions, agentic deception of human/non-human supervisors, AI attempts to manipulate humans, automated hacking, and collective superintelligence.
Robots are far from the only threat. I agree that 10% in the next decade is a pretty outrageous claim. It relies on a recursive self-improvement intelligence explosion. When I first heard this idea in 2007 I thought it was pretty far fetched.
It seems more likely now, if still very unlikely. I wonder, would he say that a statement that there is a 12% chance of a hard take off in the next 8 years is just as absurd? He didn’t say a word about this, and that is just about the same thing as extinction in 10 years.
I’m guessing he knows almost nothing about the theory related to existential risk from AI, since he didn’t discuss any relevant topics related to it. You can dismiss all of that if you like, but you cannot really dismiss what these people have already built and demonstrated. It is possible they know something else you do not know.
Adolf Hitler didn't have robots capable enough to carry out his will. He used humans to do it.
It's the old-fashioned way of doing things, but, why change what works?
I think the central premise looks more like this:
Biology has been trying to grey-goo the world for billions of years, but it turns out the world is not something so trivial.
He is pushing back against the folks with pure CS backgrounds who think that computers are all there is. Its a form of magical thinking unique to programmers who live in a world where speaking the right words to a machine is enough to impart your will on the world. Believe that strongly enough, and you fall into the trap of thinking that a sufficiently smart entity could speak the words "let there be light" and it would be so.
The author is pointing out that speaking the words is insufficient. To end humanity there must be an execution phase. The author is correct to point out that acquiring superhuman intelligence is not some guarantee that you will have or obtain the resources necessary to make that happen, in the same way that genius generals still lose to ordinary ones, and the best-laid plans are oft to go awry.
There certainly are risks, but 10% risk of extinction in 10 years is not one of them.
Biology is extremely slow, try to extrapolate the previous 3 years of AI progress 10 years into the future, it's not a comparable trajectory.
Spoken like someone with a somewhat limited understanding of what it actually takes to calculate the fitness function of arbitrary organisms in an arbitrary environments.