Geoffrey Hinton: it might turn out we were better off without AI
theguardian.com
theguardian.com
The carelessness is also clear in how AI is rolled out. It's being presented as a disruptive technology, as is everything else presented by modern tech companies. The goal is a very large payout in a very short period of time, it's not to make anything better, or worse, either is just a side effect.
Had LLMs first been presented as a "better IntelliSense", where rather than auto-completing function names, your editor can now suggest implementations, very few people would have been upset. Same for writing, it's just a better version of Grammarly, it's fine, or as a tool in drawing software. The issue would have been, for the AI bosses, that this takes time and doesn't create trillion dollar companies. It only creates companies that they'd actually have to run, long term, and that's boring.
If it weren’t OAI and Anthropic, it would be others.
If it weren’t US companies, it would be others.
It is far too useful now to stop large scale deployment, it isn’t a matter of specific people or countries not being cautious or prudent enough. The entirety of the Earth is at issue now and there will always be people who cause the tragedy of the commons (if indeed AI turns out to be harmful in the long run, which is not remotely certain).
Look at what happened with fossil fuels. Even if the entire global west totally ceased use, there are many countries that simply will not stop because of local utility. AI is the exact same way.
There is something deeply rotten about the US mindset though, them being American is definitely a problem.
> Look at what happened with fossil fuels.
The west had a 100 years head start, on top of that these countries are producing everything you consume, don't start pointing fingers. Chinese current co2 per capita is the same as the US in 1900
Releasing artificial intelligence for the use of all of humanity is better than it being controlled by one or more shadow groups. That's what the Americans did, and the Chinese followed in.
Had AI been invented in the 20th century, it would have been kept secret and used only by state actors to facilitate genocides at mass scales. But we live in different times.
Uh? In what universe are openai and anthropic not shadow groups?
Oh, how nice of them, such charitable souls...
How could modern AI have helped those groups you mention in their endeavour? Really fucking easy and effectively, and you don't need a lot of imagination to think of how. From surveillance and interrogation to pinpoint determination of ethnicity and much more. Not to mention translation between human languages.
They literally targeted people for their past opinions, how they looked, etc. which you can now determine and process at break neck speed, you can even remove humans from the loop and have the machine determine who's an undesirable or not.
How is this a serious question lol?
What is that “something”?
The world’s biggest philanthropists are American.
The US has for decades invested into charities and human development around the worlds .
> deregulation at all cost
The US has regulations. The Constitution of the United States is the oldest and longest-standing written and codified national constitution in force in the world.
> Christian brained politicians
We have separation of church and state.
The Sackler family of opioid infamy were huge philanthropists. Not saying that the AI companies are like Purdue Pharma, but clearly philanthropy does not compensate for societal damage.
Americans and the US as a whole are the most charitable in the world.
The data does not lie.
Americans can be charitable individually, but that does not contradict the reality that US corporate and institutional structures operate on a strict "profit above all else" mandate.
Yeah because it includes lobbies and turbo billionaire pastors from texas using "charity" money to buy rolls royces as well as other highly retarded """charities""" that aren't even legal in the rest of the developed world
You have senators and wannabe senators quoting the bible on TV as their guiding star as we speak right now, give me a break
On top of that you can separate the church and the state all you want, when people like Thiel show up it doesn't matter, these people are deeply troubled, if they were muslim and saying 1/10th of what he says you'd send them to guantanamo
Doing what? Rushing to make trillion dollar companies, rather then developing the tech and their companies in a sustainable way? I don't even know what in the original comment this would be a response to.
The particular form of the pursuit of power practiced by OAI and Antrophic - the creation of trillion dollar companies, is one of the most benign in recorded history. It's so unnaturally benign, that I doubt it's going to survive for much longer.
Political power is when people act in concert, when leaders have people behind them because those people agree with how they're being lead etc., it's not something that happens unilaterally. You can physically force people, but for what when you can have robots that you don't have to force?
So that kind of leaves wiping out nearly all humans, for a few hundred thousand remaining people to get served by robots and play golf. Which I would say is what we're kinda on track for. But that wouldn't be a race between nations, that race would just be pretend to get the tech online; it would rather be a race between all humans and a bunch of sick people.
AI doesn't have to be used offensively to give a nation power. The machines can already design new software at the drop of the hat, it will at some point become capable of designing new consumer and industrial products. Putting a tireless intelligence at the task of looking for new optimizations in your industrial processes and factories will eventually give you economic dominance, one way or another.
Offensive uses don't need to be as extreme as you imagine them. We're already starting to see the effects of having meticulous automated hackers. Instead of a blitzkrieg, imagine your foes suffer from a constant stream of industrial accidents, unfortunate events and information leaks that can never quite be traced to your nation.
Finally, political power itself doesn't need to come from people acting in concert. There's always been wheeling, dealing, intrigue and bureaucracy. Once AI is sufficiently versed in legal matters, it'd allow political interest groups to overwhelm governments and private enterprises with objections and appeals, paralyzing their ability to carry out their agenda until they make concessions.
The possibilities of using intelligence for acquiring more power are literally endless..
But that doesn't mean that everyone has to use it.
We're only starting to see the beginning of the negative impact of AI to the job market and society. If if turns out that allowing most jobs to be replaced by AI is bad for society and the economy (super high unemployment - nobody buying more then necessities), then wouldn't it be smarter to not do that?
If that's the case, then each country can decide for themselves.
Dario's "Machines of loving grace" UBI/foodstamp "utopia" sure doesn't sound like a recipe for success.
Value is simply the mutual understanding that something should cost a certain amount, and the system we've built emphasizes the need for profit above all else. When a set of companies comes along with a new, impressive technology claiming that they can automate away the white collar workforce, they're trying to capture as much value as they can.
It's not a moral sickness, it's market actors behaving rationally. Believing that people should be willing to leave money on the table because it's the right thing to do is a dead end, the entire system needs a rethink.
I don't agree that "I am just a market actor behaving rationally" somehow absolves you of any moral duty
I'm not an ethicist or anything but I have studied quite a lot of ethical frameworks and I've never encountered one that I think would agree with this
I didn’t expect a full throated defense of mugging and burglary on HN this morning, but here we are.
Or have you forgotten that Sam Altman was ousted by individuals that shared your concerns, and it was the OpenAI employees that rebelled against the board and wanted Altman back?
That HN post was infamous and nearly took down this site.
No matter what practical benefits this technology might bring, if this runaway marketing campaign isn't stopped, in just a couple decades we'll see what the complete collapse of civil society looks like, a world where the institutions no longer see value in maintaining an educated democratic populous, and can relax back into the comfort of feudalism.
The AI companies are as much being pushed along by investors as they are in the driving seat, but the push is in the direction of getting real products in front of people and experimenting with this miraculous tech everywhere it could possible work.
How many, approximately?
Ignoring the implementation details, if we treat the whole global technology world as one single AI system made up of AI companies, independent developers/researchers, and end users, what would be the training goal and evaluation?
I feel like you’re not acknowledging the inevitability of the technology being developed further. You’re also ignoring competition and the basic fact that people do research and push the boundaries of their field naturally.
If the major frontier AI labs had decided, as you said, to make a fancy autocomplete and a better Grammarly and then simply stop and coast, do you really believe that every AI researcher and entrepreneur in the entire world would look at the situation and decide that the status quo is where it should end?
Humans don’t work that way. People would get ideas and want to push the limit. If the frontier labs stopped developing their technology, other companies would be created to carry on. Academic researchers would push the limits further. It would be discovered that the AI could go further and we’d get back to this same eventuality with a different set of people leading it.
I also disagree with your characterization of this as a get rich quick scheme for AI leaders who want to make a lot of money and then be done. I think this characterization feels truthy if your only conception of startups is as a gamble to retire early, but the people who build and run these companies actually do like building and running companies. If the AI leaders and executive management you’re talking about were separated from their jobs today, even with billions of equity, most of them would be out starting new companies next month because that’s what people like that do. I’m not saying this to “glaze” them or whatever I’ll surely be accused of in the comments, I’m stating some facts about this space. You don’t need to look far to find examples of this already happening.
This looks a lot like the “if I didn’t do $badthing someone else would have” line of reasoning.
With AI, the point is who uses it for what.
Think about it like a magic wand that makes your wishes come true. Would you really want everybody to have one? Who should and who shouldn't? Who gets to decide?
Humanity's strategy so far has been to pull marbles out of the bag as fast as we can. And that's been working out, because most marbles are white. But we're insanely lucky we haven't had a nuclear war. The USSR's early warning systems picked up 5 incoming american missiles in 1983. They didn't fire back because one of the officers on shift didn't believe the reports (why only 5 missiles?), and he declared the reports were fake.[1] If a different person was on shift, there's a good chance nuclear war would have broken out because of a sensor glitch. And that's far from the only instance. The list is terrifying: [2]
And now we're improving AI capabilities as fast as we can, despite it being pretty clearly another grey ball. AI has a lot of amazing uses. But why are we rushing so much? Doesn't seem wise.
[1] https://en.wikipedia.org/wiki/1983_Soviet_nuclear_false_alar...
> But why are we rushing so much?
Talking about these things with elaborate analogies only obscuring the point you already stumbled into: Humanity advances because humans are curious, driven, like to discover things, and like to push the boundaries.
Talking about humanity like one entity that operates together misses the point: There are billions of people in the world and some of them are going to discover new things.
There is no general counsel where humanity gets together and decides what humanity works on. There is no singular bag with a singular “humanity” pulling marbles out one at a time.
The analogy is too broken to be useful for the discussion.
This is true, but not all the time, and transparently not the case for the current leaders in AI development. Their motivator is money.
There are plenty of examples of this, e.g. when we discovered how to harness nuclear energy, we quickly realized that extreme measures needed to be taken to ensure we don’t collectively destroy ourselves.
You’re speaking of “we” as if the whole world was operating in unison, while there’s a war going on in Iran over nuclear disagreements.
The nuclear situation only reached stability after multiple countries fully developed it. I don’t see the parallels to halting AI.
The nuclear stabilization that occurred was the result of a collective understanding of the existential risk it posed leading to anti-proliferation agreements and superpowers going to great lengths to prevent nations deemed as unstable from obtaining such technology.
Some portion of the global community may collectively discover/decide that the risk AI presents is worthy of anti-proliferation treaties and wars may be fought over it.
The main gap in the comparison is that we have a much clearer understanding of the risks of unchecked nuclear weapons use than we do about what current and future AI tech might or might not unleash.
The underlying point is that we have historically placed limits on our curiosity when we deem certain paths of exploration existentially dangerous. It has yet to be seen if AI will reach those levels of existential risk, but if they do, there is plenty of precedent for stepping on the brakes.
Nuclear Non-Proliferation Treaty (1968)
Biological Weapons Convention (1972)
Asilomar Conference on Recombinant DNA (1975)
We've done this sort of thing plenty of times around specific, dangerous technologies.
Also the bio/DNA ones are quickly falling apart as the technology to do these kinds of things on a small scale are readily available. To the point where we're finding illegal wet labs in the US.
Nuclear was more of an exception because of many of difficulties in purifying stuff for weapons and its detectability.
AI is a much, much, much harder problem is it uses the already existing infrastructure as a base. Unless we regulate, monitor, and ban all private usage of technology AI can be developed at a small scale.
If there is a room full of people, and one person gets a gun, who would you pick to have it?
Yourself, obviously.
Maybe america will end up the only world power holding the secrets of LLMs? ... Oh no...
If nuclear weapons were as easy to build as AI models are to run, we would have all been dead decades ago.
There might have been. But the notes from the Manhattan project were leaked almost immediately to the russians by informants.
How is any of that good?
---
Simply put humans are at risk of any power scaling. We exist safely in a very tiny temperature and energy environment and the rest of the processes in the universe operate widely out of that strict safe area.
It's also a catch-22. There are large rocks in space that will eventually end us unless we do something about that. But having the power to do something about that gives us the same power to also end everyone on the earth.
> Nazi Germany undertook several research programs relating to nuclear technology, including nuclear weapons and nuclear reactors, before and during World War II. [...] The first effort started in April 1939, just months after the discovery of nuclear fission in Berlin in December 1938, but ended shortly ahead of the September 1939 German invasion of Poland, for which many German physicists were drafted into the Wehrmacht. A second effort under the administrative purview of the Wehrmacht's Heereswaffenamt began on September 1, 1939, the day of the invasion of Poland. The program eventually expanded into three main efforts: Uranmaschine (nuclear reactor) development, uranium and heavy water production, and uranium isotope separation. Eventually, the German military determined that nuclear fission would not contribute significantly to the war, and in January 1942 the Heereswaffenamt turned the program over to the Reich Research Council (Reichsforschungsrat) while continuing to fund the activity.
> The program was split up among nine major institutes where the directors dominated research and set their own objectives. Subsequently, the number of scientists working on applied nuclear fission began to diminish as many researchers applied their talents to more pressing wartime demands.
> [...] Heisenberg said that in 1939, the physicists at this second meeting concluded that "in principle atomic bombs could be made, but they also emphasized that it would take a number of years",[24] certainly not less than five. He also said that he "didn't report it to the Führer until two weeks later and very casually because I did not want the Führer to get so interested that he would order great efforts immediately to make the atomic bomb. Speer felt it was better that the whole thing should be dropped and the Führer also reacted that way."
> [...] Speer states that the project to develop the atom bomb was scuttled in the autumn of 1942. Though the scientific solution was there, it would have taken all of Germany's production resources to produce a bomb, and then no sooner than 1947.[39] Development did continue with a "uranium motor" for the navy and development of a German cyclotron. However, by the summer of 1943, Speer released the remaining 1200 metric tons of uranium stock for the production of solid-core ammunition.
There are ways of calling this "weren't working on a nuclear bomb", but it seems to me like "they wanted to, were moving in that direction, and would have done it if they had the resources and weren't distracted by shorter-term considerations".
That sounds very much like "they weren't really working on it". They explored the options and decided not to.
Whatever you want to call it, that's what I was referring to.
Still seems a moral cop-out, and while true, the timeline would certainly have changed, if not the inevitability, if key figures such as himself had opted out.
I'm not arguing about the consequences or responsibility. I'm asking which humans would be capable of stopping the development of AI at this point in history.
If we're talking game theory, a better model is the Stag Hunt [1]. At some point the equilibrium strategy is to cooperate.
Another difference is that real life is an iterated game, and the iterated prisoner's dilemma rewards cooperation.
It’s difficult to understand where this all is coming from because it’s so strangely disconnected from history. It’s not even consistent with current world events. The US is in a war with Iran over nuclear proliferation while also worrying about North Korea’s nuclear program. China and the US are in a trade war and Taiwan is under threat due to semiconductor capabilities (among other things, obviously). Open weight models allow anyone in any country to do AI work on their laptop. The top story on HN yesterday was a small team of academics making some cool discoveries on improving LLMs without frontier lab level compute.
How anyone can see all of this, in addition to decades of history of attempted nuclear diplomacy and wars, and posit that the whole problem is easily solved by simple agreements is confusing to me. The confidence with which these alternatives are being proposed and the unwillingness to acknowledge second order effects (like AI researchers moving to a different country and continuing to work) is baffling.
But yeah, there are a lot of comments from genie-rebottlers that are not very grounded in reality. They're functionally equivalent to saying "hey, everyone in the world, let's not actually use electricity after all" in 1905.
Despite this, when discussing the egregious behavior of the frontier labs, there is vanishingly little acceptance of the idea that 1) yes, it probably was inevitable in some ways AND 2) those involved, particularly at the top, are still responsible for the consequences of their (bad) decisions.
1. Is it both inevitable and still they are responsible? 2. Or it was inevitable and they were just an instrument to make it happen. 3. Or it wasn't inevitable and if it were not for some specific people then it would not have happened?
I am leaning more towards 2. But if that is the case why are only specific people reaping the benefits of this?
If it is 3 or 1, then why aren't we holding them accountable?
So yes, they are 100% responsible for damage they caused and did not had to happen. And no, "I had no choice, because it was the only way to become billionaire" is not an excuse.
They were NOT "instrument". No one manipulated them. No one forced them. Nothing forced them. They themselves wanted to and did not cared
People are starting to panic, and few things bring out full blown idiocy and disregard for ground truth faster than widespread panic.
Other labs seem to be working the exact same angles as them, but just have the good sense to not be so open about it. It's not the "lab" people have a problem with, it's their leadership. These people cannot be trusted to make the right decisions for society with this technology.
You must be joking. Altman is a sociopath, Amodei is either too stupid or too evil to stop working on tech he claims will have a 10% chance of destroying humanity (and in either case he should be imprisoned to protect the rest of us from him), and Musk is a man-child who will lash out at anyone and everything who dares to speak ill of him. There's absolutely nothing "lucky" about having these jokers in charge of stuff.
So yeah, I'm negative against their development of AI along with the general public, and for good reason it would seem.
Also, I don't want super intelligence. I want humans to remain in control. Why are they trying to build something to replace humanity? Why don't we got a vote in this? All 8 billion of us not developing frontier models with the expressed goal of achieving AGI followed closely by ASI.
It's a pity because as computer scientists this should have been such a fun technological moment for us. What a cool thing to experiment and explore new open doors with. But instead it was immediately the driver of mass job loss, societal upheaval, and a double digit percent chance of ending humanity, espoused by the AI lords themselves.
The list of possible reasons is pretty short and mostly not really good for most of us, as far as I see it
I do. Quite a lot of people do. And that is because we believe it will lead to a better life for most humans.
The models we currently have aren't going away, and especially the ones small enough to run at home will continue to be used, but if these data center investments end up bankrupting Big Tech, I don't think we're ever going to see anything beyond that. Not for a long time, at least.
My guess would be that once the money runs out, things slow down, but incremental progress continues to be made. A focus on smaller models will improve their capabilities. We’ll get better at driving them. It’ll be used in places where it truly makes sense, instead of anywhere & everywhere just because it’s easy.
I don’t actually know what the financials are, but even without the effective subsidy from VC money, I have to believe there are things AI’s doing today that’s way cheaper than what it would cost to have a human doing equivalent work. And that dynamic will drive further innovation in the space for a long time.
What does this actually mean? In what sense are people using this term. In the most strictly literal sense then sure, nothing is inevitable, because all paths are open.
But in the practical sense, I am quite happy to use the term inevitable for many technologies. I think if ASI is possible (and I think it is), then it is also inevitable. I would happily work towards developing ASI technology, because I believe it will be ultimately good for the world. But it sounds like a lot of people want to lock me up or even kill me if I did that.
AI indeed goes far beyond "autocomplete" and other nonsense. But the question is, whether it's a good idea to just let the "market" run wild with it.
Short-term capital interests don't align with the public good. AI greatly scales capability, but without anything to restrain it, that implies those non-aligned interests overwhelming society's defenses, outright destroying it in the process.
The idea of rich people being some kind of benevolent caretakers, who independently and dependably foresee and manage the consequences of their actions in everybody's best interest, is laughably wrong already.
With AI, that reality gap will swallow whatever is left.
Without capitalism, you wouldn't be able to restrain AI adoption any more than the armored knights of the middle ages were able to restrain the adoption of the firearm.
I think the vast majority of the rich are greedy bastards, but IMO being rich does not imply greedy bastard 100% of the time.
This doesn't seem right: it doesn't match the history of reality, in which capitalism has repeatedly stymied rather than fostered competition.
The argument from nature is a fallacy. Just because something is doesn't mean it ought to be.
I operate within the logic of how the world actually works.
Which I find infinitely more useful for practical topics like the reality of AI development in a global economy.
Fantasizing about alternative realities where all the leaders of the world unite and halt technology development without any of them secretly defecting and developing it further isn’t very useful to discuss. These things aren’t going to happen.
it has kinda worked for nuclear weapons
That’s the inverse of what I said.
AI has no such hurdles.
Transitioning from the current "order of things" to one compatible with the existence of things like AI is a necessity, not a "fantasy".
When you imagine people as savages to be outdone in their savagery, you become one yourself. Spend more effort on finding ways not to.
Ignorance of what? All of the previous successful attempts where the entire world reorganized itself into a single world government?
> Transitioning from the current "order of things" to one compatible with the existence of things like AI is a necessity, not a "fantasy".
I think you’re putting the cart before the horse. You’re assuming it’s necessary and then making another leap to assume that being necessary means it must also be possible.
Additionally, you don't need all the leaders of the world to agree to anything, you only need US and China.
Additionally x2, if you agree with the common stance that China makes the vast majority of their progress by distilling US models, you don't even need them to agree to anything.
You're really suggesting the following bans. A ban on computer science and algorithms. A ban on any device that can play games that aren't encrypted or DRM locked from an approval authority. A ban on any kind of unregulated computing without government approval and licensing.
>China makes the vast majority of their progress by distilling US models
Eh, this is a budgetary decision, not a technological one. Why spend billions when the US can spend billions and you can spend millions. China more than has the power to do what we're doing if they chose to do so, their biggest limitation is speed of making new compute, not money or brains.
I'm not rooting for AI companies or hustlers that are out to ride the wave and get rich.
However your claim is just hyperbole. I've used it in cross-cutting projects to extend protocol on desktop, iot device, which involves holding premise long enough to co-develop and test. It is also fantastic in debugging multi-threaded systems.
Some times it comes from people who try out ChatGPT free tier or the Google AI summary and think it represents the state of the art.
More recently I suspect it’s from people who build their opinion based on being contrarian and cynical: They dislike the AI people and distrust everything they say. If the AI people say it’s changing the world, then the cynical and contrarian belief is to assume that the technology is actually very bad and they’re just lying to make a lot of money and retire.
These takes were more understandable in the earlier phases of AI when it really was very limited, but it’s getting harder to understand how someone can be thinking of AI as just a slightly better autocomplete or slightly better Grammarly at the end of 2026.
That wasn't the point. My point was that if you had rolled LLMs out like that, as embedded technology in existing products. Slowly, over time, you'd have a lot less cynical people. I think the hostility towards AI product in large parts comes from the disruption and glancing over the negative parts. What we're doing now is the equivalent of putting millions of cars on roads which five years ago had never seem a car, only horses. We're doing so without any understanding of how that might affect society. Social media was a similar story. It was slower, yet we still didn't catch the majority of the negative consequences.
If it turns out that the people behind social media, and smart phones, don't care about e.g. peoples mental well being, then why would I trust that the people behind AI wouldn't behave the same way?
1. enhanced autocomplete -- either line (e.g. JetBrains small language models) or longer predictions;
2. chat -- asking questions related to the task you are doing;
3. tool assisted/agentic -- letting models interact with tools to perform tasks.
I personally find 1 & 2 helpful for my development tasks. For 3, I'm cautious of using it for several reasons:
1. lack of control of what gets sent to the model (e.g. if you have confidential user data, proprietary code and/or data, access tokens/private keys);
2. letting the models run arbitrary commands (modifying the system, deleting files, etc.);
3. not having control over the design process, formatting of the files, etc.
Presenting AI as something it isn't doesn't change the reality which is like the Wait But Why thing https://waitbutwhy.com/2015/01/artificial-intelligence-revol...
That was written in 2015 before LLMs were much of a thing and has been fairly spot on so far. This is not just about LLMs any more than the industrial revolution was just about steam engines.
Aside from the AI dungeon app, this was my first real exposure to LLMs wrt github's copilot extension. It sucked, and did a lot of stuff wrong, but they got exactly 5 bucks from me to try it.
It's sorta crazy to think how many years ago that was now.
Then the Chinese came along and showed that you can take those model weights, repackage them as open weights, and undercut the business model. Suddenly Dario et al. start talking about how dangerous AI is and how it needs to be regulated right now, before someone's abliterated open-weight Chinese-provided business-undercutting model destroys the world. They want their walled gardens regulated into being the only game available.
Both stages of the game—raising capital and regulatory capture—require hyping up the disruptive aspect of AI.
They certainly don't care when a model hallucinates or an agent does something it shouldn't be doing. It's your fault and they charge you for the tokens.
The only leader I’m not aware of really taking a more active role in safety on the daily is Elon Musk, so I’d like to see evidence supporting what he is doing or not doing on that front if someone from SpaceX can share their experience. The evidence on the environmental front from SpaceX is very concerning and is disregarding the wellbeing of the world and local populations in my opinion.
That’s why the AI heads all have extremely dramatic narratives. They are the end product of a selection tournament for raising massive amounts of money pre-product.
You make it sound like this was a feasible option. This technology would exist regardless of how any one company decides to frame it. If OpenAI decided to go with a more conservative approach rather than just throwing ChatGPT out for the masses, then all it would take is for one person to realize they didn't take full advantage of this technology to just make their own ChatGPT.
>The issue would have been, for the AI bosses, that this takes time and doesn't create trillion dollar companies. It only creates companies that they'd actually have to run, long term, and that's boring.
You can criticize the current tech leaders all you want, but what you're describing is a fundamental dynamic of capitalism. You must choose the most profitable option, otherwise you have chosen the wrong option. If the current tech billionaires all vanished just as ChatGPT was released, they'd be quickly and easily replaced with roughly equivalent individuals who would effectively operate the same exact way. That's true because they're operating based on the mostly deterministic policy of profit maximization. They don't really get a say in it; they must follow the path towards maximizing expected future returns, otherwise they'll get canned by the investors who are seeking maximized expected future returns.
None of this relieves them of their personal ethical responsibilities either. I'm just pointing out that your comment might as well be describing an alternate reality where everyone lives peacefully with each other and nobody goes without what society can afford to offer them. It's equally as unrealistic.
Can someone who thinks this about Altman explain to me why he has repeatedly declined to take any equity in OpenAI? I genuinely want to understand your theory of mind here. Do you think he's actually in it for the power, and for some reason thought taking equity would undermine his effort at accumulating power? It just makes no sense to me.
It seems obvious that this same dynamic is in play with the frontier AI companies now. They know if they win the race to the they'll be forgiven their sins, or anyway they'll be beyond the reach of effective punishment.
I know this sounds unusual, but I do think this wouldn't be the worst thing that could happen. Sure, it would cause a bit of chaos, but it might also make us think twice about continuing down our current path of unfettered development, and possibly even make us much more cautious, which might in turn even help us with the climate crisis.
Because human beings in a large mass only respond to one thing: immediate difficulty. They don't respond to delayed difficulty like "the climate is getting worse" or "technology will make life more difficult in the distant future". They only respond to immediate pain.
So I think it would be a good thing if the financial system had a serious threat. Then people with money, who tend to offset future problems with technical solutions and financial padding, would start to become much more suspicious of technology and defund AI and wasteful things that are likely to threaten their safety.
I disagree that something like this would be a good thing though. Whether it's a power outage, financial crisis, or something else people will die. If that's what happens and society ends up adjusting course after then that's what happens, but I can't call people dying a good thing.
I don't think that's a good thing. But if we DON'T do anything, then even more poeple are likely to suffer. So while people dying is not a good thing, a thousand people dying is better than a million people dying because we don't do anything.
Same thing about burning your hand. It's not a good thing to burn your hand, but burning your hand once when you're young probably saves you from a more serious accident later on.
We both don't know all the possible futures and we don't know if a power outage or similar would kill 1,000 or 1,000,000. We also don't know if that loss would lead to the future we want.
This sort of thing was going to happen, with or without AI (state sponsored hackers, the US or China taking down systems they control, or whatever).
The later it happens, the more we will be down the path that lets it happen, and the worse the consequences. We are getting gradually worse consequences: AWS outages taking down websites, GCP outages taking down banks, attacks on fridges disrupting good supplies...
Its better for us to change path when something happens that causes dozens of people die rather than wait until its millions.
The OP didn't say we should make it happen so I'm not intending to put those words in their mouth. Its a small leap from "it might be a good thing if the power is out for days" to someone hearing or thinking that themselves and acting on it.
If it turns out we have a power outage, people die, and we learn a lesson and adjust course then so be it. That would be what happens in that future but it isn't necessary and it isn't a good thing even if today we think it leads to the best outcome - we can't know that without knowing all the outcomes and would be generalizing into only the ones we can imagine.
Blockchain despite all its flaws, is the canary in the coal mine. They have been dealing with an increasing amount of cyberattacks. Because it all happens publicly in the blockchain, there is no way to hide it.
However major bank gets hacked in the tune of over 100 millions and we get to know it happened because somebody who has no relationship to the hack turned his eyes to a bunch of suspicious transactions happening on Ethereum.
This is why I think AI is so stupid. It's just too powerful for anyone to control, especially with the inertial mass of capitalistic society powering it through the incremental, short-term advantages it gives.
Nuclear weapons are indiscriminate and make everyone worse off if used. And everybody has them. Your optimism is unfounded.
The difference is most people have not experienced any immediate effects of nuclear weapons. But people will experience the immediate effects of AI taking down the electricity grid, for instance.
If we're at the point where all that matters is scale and reinforcement training then the public facing aspects could entirely disappear and progress would still continue especially when given an infinite budget of taxpayer dollars.
Maybe the long term changes such blackouts would cause would be "worth it" in a utilitarian sense, but let's not hope for such damage.
And I'd go one step further to say that it probably can not be true. This has always been true and probably will always be true. All LLMs are doing is demonstrating that. And the LLMs of tomorrow will demonstrate that the most LLM hardened system of today is still just a vulnerable mess. So the conclusion is simple - things should not be connected to the internet unless the worst case scenario from hacking/exploiting is relatively low. If somebody wants to connect their coffee maker to the internet, for one can only imagine whatever reason, then the worst case scenario is a broken coffee pot - who cares, go for it. But connecting e.g. power, water, or other core infrastructure to the internet? No, never.
Just because something hasn't been proven to be a problem today does not mean it won't be a problem tomorrow. We shouldn't learned that yesterday, hopefully we can learn it today.
Are they?
Consider the old joke, "what do you call the medical student that graduated bottom of his class? 'Doctor'".
Same thing applies to sysadmins and IT. Sure there's plenty of skilled people working in facilities out of bumfuck, but also consider how few people want to work in the low-paid, low-respect IT jobs.
Those roles are often filled by the apathetic, mid-skilled, and close in proximity. They're connecting shit to the internet because their boss wants more data points.
What makes this particularly embarrassing is that power companies literally run thousands of miles of cables and could easily build a private network instead of using the Internet. The regulations should be amended to require it.
Again, STUXNET shows that bugs get into authorized devices really easily. One of the biggest issues is you have power guys that don't know dick about computer security using fleets of laptops that will need a huge amount of training to bring them up to the needed high levels of security.
More so there are thousands of these devices out in the field in power company trucks. All it takes is someone picking one of these up and infecting it for the problems to spread.
It is an insanely hard problem to solve. Breaches happen in military networks pretty often. As a measure of secrecy, they never tell you about it.
While you certainly make a good point about the difficulties that come with securing the laptops and other equipment, the military has the same problem. Most of the users of classified military systems are not themselves experts in computer security. The necessary training is not a graduate level course in security; it is training people to always match cable colors to port colors, to only use laptops provisioned by the organization, etc. The military disables, sometimes by physical means, USB ports on laptops that are used on classified networks to close one of the key security holes Stuxnet took advantage of.
Sure, it is challenging, but this is not impossible to solve. Utilities would have to be forced to do it and they would need time to overhaul their operations, but they certainly can do it, and there is an argument that the NSA's IA division should help (as the electrical grid is a critical for national security). It can be done, we just have to be willing to do it (and pay the cost of doing it).
Now supermarket fridges are openly accessible over the Internet
I’m not sure it’s AI that’s the issue, it’s just a scapegoat for decades of bad security choices that nation state bad actors could have always exploited
Now that’s a perceived threat, ‘AI will do bad things’ is a great excuse for awful national infrastructure and security management
If someone manages with the help of AI to hack into a vital system, then the system was already in a critically bad state beforehand. If developers can't use AI to check the code for security flaws, then security gets worse and not better.
Insecure is insecure. These systems never were secure.
You’re placing blame on the wrong spot. It’s algorithms and social media incentive systems not AI. Our digital media incentives addictive minimum effort chum and slop. AI just makes it easier to make.
But also no, the crap AI is causing goes waay beyond algorithms and social media. It is ok to blame it.
> Human slop is slow and easy to contain and build containments around it. Generative AI is so fast that it breaks every possible containment strategy
If you think society prior to ChatGPT was any less off-the-rails than it is now, then you're suffering from recency bias. I'm willing to give you that AI certainly accelerates the slop, but it doesn't fundamentally change anything about humans being unable to regulate their consumption of slop.
I think all these laws/theories etc needs an adjustment to today's reality. Human slop is slow and easy to contain and build containments around it. Generative AI is so fast that it breaks every possible containment strategy
A snowflake is not a problem.
A lot of snowflakes can be an issue as the pile up.
A lot of snowflakes rushing down a mountain at 100 miles per hour is a big damned problem.
Having AI in my life is a hill I will die on. I will not want to go back to being dependent on having other people do for me what AI lets me do for myself.
Let's flip your last point: if the tech gets good enough to replace people at any task (including physical ones), do you think others will need you for much?
And then there's the systems that the corporations are making more powerful as fast as they can. Yes, they promise to do things like cure cancer, but they're also capable of scarier things like hacking banks. And they're progressing faster than we can understand them. Would it be so bad to delay the cancer cures by a couple years to ensure we still understand what the models are doing?
With AI-driven speech recognition, I have far more agency than I did without, period. I can write, I can code, I can communicate with various entities, people, government organizations, corporations, chatbots. I can fill in online forms instead of leaving a bunch of spastic scrawls that generate more questions. I can use it to drive OpenSCAD and 3d print something.
And here's why I start out by saying I don't care if others need me.
I think I'm congenitally incapable of not building/designing/exploring things. If we had an AI-dominated future and we were all on UBI, I'd still find myself saying “hmmm, I wonder” and going down a rat hole to see what I can find. I'm doing it for myself. My curiosity and satisfaction in bringing something to life, even if it's only a variation on what somebody else did.
Rephrasing something I said in a comment somewhere that I would spend time gardening, bird watching, and building a closer relationship with my partner and my pets. I would have no want of things to do or explore.
To your last point, we don't need to understand how something works for it to be useful. Doctors don't understand why treatments work and why they fail. If doctors waited, if treatments waited until they understood how and why it worked, I'd probably be dead. But because they don't, I now have a VP shunt in my head (aka, medical-grade juice-box straw), and my life has drastically improved. Toddlers are something parents don't understand—the how or why they do what they do—but they keep having them anyway.
WTF moments are an emergent property of toddlers, goats, border collies, working dogs, hunting dogs, and AI agents. And like with a toddler, we need to deal with it like we're the adults in the room.
Too many people in my life have died of cancer, and the agony and pain as their body hollows out to a waxen shell is something I'll probably never forget. If I had a choice between understanding how AI works and a cancer cure, I would choose the cancer cure, hands down.
One thing we do need to do is slow down the development or the build-out of AI data centers and put a resource constraint on the model companies, forcing them to get better with the same resources. Sort of an economic “you can't have dessert until you've eaten your vegetables” moment.
I really got to sit down and write that blog post on why agents are like toddlers and that toddlers should have figurative machine guns.
If one doesn't provide incremental value, how do they get food, shelter, and compute?
Perhaps we'll be lucky like horses, and be lucky enough to become pets once our population collapses.
Without accessibility tools, there's not much incremental value I can provide. When the RSI showed up, I was advised by many people to spend the rest of my life on SSDI. I refuse to do that, but it was difficult to show I could provide incremental value. Yes, my mind still worked; the designer, the architect were still there. I had speech recognition so I could write, but a common reaction was, “How can you do anything technical? Your hands don't work.”
Now to the general case that you seem to be worried about: that AI will remove everybody's ability to provide incremental value. No worries there. The entire system will collapse. Labor has no value. Without labor, capital has no value. And we're all jolly well fucked. Now, Your value is how good you can become at asymmetrical class warfare.
At least that's where the current nihilist fantasies about AI leave you.
The AI nihilist view works when you have the wrong enemy. AI is merely the puppet, the distraction, just like the H-1B visa, offshoring, undocumented workers, etc. All these distractions are so that we won't notice the billionaires are enshittifying society and won't seek to dethrone them.
IMO, I argue the fix involves returning the tools of production to the hands of labor, and taking power away from capital. But as with many problems in life, the solution is left as an exercise for the reader.
I sympathize with "dethroning the billionaires" but to me it looks a lot like powerful, general AI systems are the most capital enhancing technology we've ever seen. Wouldn't the class struggle benefit from slowing down frontier AI development while diffusing local models to ordinary people? I feel both can happen at the same time.
I'm reading this in the context of the article currently at the top:
AI is forcing us to pay off 40 years of technical debt
https://news.ycombinator.com/item?id=50364336
Is it bad we're finding all the bugs now? Or is it good that we're finally being forced to make our critical systems actually robust?
I guess we could look at it as going through a world-wide Code Red...
My prediction is that some people will be able to use AI to create essentially infinite money inside their bank accounts (through hacking); the prices of assets will grow to ridiculous sums as the people with infinite money buy up assets, but yields on all assets would soon approach 0%; still, one asset is going to grow at a faster rate than all other assets (E.g. an AI company backed by blockchain/tokenized shares that are impossible to hack) and everyone is going to start piling their infinite money into that asset; every other asset will crash because of the rapid, exponential price acceleration of the top asset making every other asset unattractive by comparison. Then fiat money would cease to exist and shares (or crypto tokens) of the top company will be used as money instead. Then prices of assets would stabilize again; denominated in the new currency... Then after that, we will finally be able to start innovating with AI.
Not to say that they're unhackable or unstealable, but merely that they dont have a central weak point that if broken, destroys the whole system
Traditional finance has the benefit that they can more easily reverse malicious transactions, but if hacks become too common, it will be impossible to correct fast enough and in a reliable way. With traditional banking, a hacker only needs to hack a few nodes to compromise the integrity of the entire system.
So yeah I also think crypto will win in the long run. I would bet on a tokenized stock with AI attached.
I think the top asset will combine the AI which is best at hacking with the blockchain which is hardest to hack. Ability to corrupt others whilst maintaining perfect integrity yourself.
AI for offence, blockchain for defence.
Programming has never been this awesome.
Also the future where there is no sacrifice to achieve your lifegoals seems just unnatural to me.
I am building personal software for family (home automation where it is not possible, time capsule (future letters for kids), study materials for kids and much more.
Alternative is, it would take me hours of manual work and I'm stuck at computer.
Yes, I'm in AI heaven right now that allows me to spend way more time with family than I would have been able to and still be productive.
It's no different from someone seeing their car as just a tool to get from A to B. That might be an understandable mindset for people on HN, whereas people at a car meet might find something off about that mindset.
Not to mention, sacrifice can come in many forms. The simplest example is, the time required to learn something or develop a skill. As long as people want to pursue things just out of interest, that's not going away.
(Unless we find a way to download the information directly into our brains, I suppose.)
Thank you, brave HN commenter. You've truly made the site a better place with your astute rebuttal.
“I think [Mark] Zuckerberg and [Elon] Musk and Sam Altman, they understand the risks, and they don’t give a shit. They’d rather have the money. That’s my belief.”
Mortality.
I don't buy this. There isn't some kind of voodoo "rogue AI(s)" out there which could randomly decide to take over their vendor stack or brute force their way coming into possession of the most critical infrastructures on earth. These kind of predictions could only be realized if a tiny minority who owns a huge amount of compute and a particular agenda decides to do this kind of thing, which also will require insane amount of data center capability to cause, especially some kind of financial "collapse".
Or some other government setting an AI to intentionally hack and crash the US?
>possession of the most critical infrastructures on earth.
I think you may underestimate how poorly we protect these things. They have all the same security flaws as our internet connected systems. They aren't written in some super special language, it's all C, C++, JS and other non-memory protected trash.
I decided long ago that - when there are plausible arguments either way - it's better to be an optimist. Optimism does not mean ignorance or denial; rather, optimists should acknowledge risks and problems while having confidence in our ability to generate solutions.
One of the biggest challenges facing us is a critical mass of pessimists. Not only regarding AI, but also with respect to the economy, democracy, liberalism, etc. And while these negative views are plausibly correct, they also act as a self-fulfilling prophecy in that they rob people of the agency that is needed to solve problems.
That's a touching sentiment or whatever but the nature of the thing is that it will inevitably be opened and that it cannot be closed thereafter. So (I think) we'd all be better off putting our efforts towards predicting the likely outcomes and figuring out how to make the best of them.
And if all the likely outcomes are very bad, then what?
You might as well ask "but what if all the options are bad" while standing on a sinking ship contemplating your next course of action.
lifting people's standards of living -- is _not_ a technological problem. So I don't understand why people think that AI might solve this, even more so because AI is not primarily being developed or used to solve human or social problems, because there is little financial incentive to doing so (not much ROI in developing something that's cheap or free).
The Earth isn't lacking in resources, they're just distributed highly unevenly -- and correcting that distribution is a social and political problem, not a technological one. In fact, AI is more likely to exacerbate inequality than reduce it.
Let's say that AI allows us to crack the code to nuclear fusion. That would be amazing from a technological point of view. But the idea that this will lead to free, or even cheap, energy for all is rather naive -- all we have to do is look at whether the invention of efficient solar power (a limitless resource that is now relatively inexpensive to develop) has lowered energy costs for the average American [0]
[0] https://www.electricchoice.com/electricity-prices-by-state/h...
I still don't know how Karpathy still works with those folks. I really think the guy is great but he makes me wonder
Seriously? It seems to be causing more problems, giving us more boring work, and less free time.... sounds like absolute bullshit.
We need to build it as fast as possible because it could cure all disease. Ignore whether that's possible, how it might happen, whether all we're missing to do it is more intelligence, etc.
The only way I see us curing all disease is to throw out all the names we've given to collections of symptoms. What disease? I don't see no disease, you just have a sore throat, runny nose, and mild fever.
Hell no. We need to risk pumping out even more CO2, causing more environmental disaster, ruining entire human communities, just to cure diseases? You might as well say it's better to be in prison because you're guaranteed three meals a day. The logic of some pro-AI people is baffling.
The other day I hired an accountant and even he sent me an AI answer that sounded right but was actually bullshit.
In short, it takes a LOT more time just to find some sanity now.
AI is shit.
Only 14 years old and accelerationists think we can just push the capitalism pedal further and somehow get social changes going their way
How so?
Looking back now, was it really happier before AI existed? I'm not sure. In the past, when I wanted to find out whether I was right or wrong about some piece of knowledge or information, I'd contact several people and, far from getting an answer, I'd often be told not to bother them. But AI, in that respect, is the best teacher I can think of.
I think the times are continuously getting better. It's just that there's a problem of power concentration among the people who handle it. Still, if AI is a tool of that nature, then I suppose that can happen.
But considering what access to knowledge was like before AI came along, I think having AI is better.
Most people are not interesting in learning for learning's sake, they are motivated to learn something they can apply (see: kids complaining about how learning Math in school doesn't transfer to useful skills). Why learn about architecture and technical debt, if you can tell the machine to make it faster?
I basically don't think you're wrong. However, books and reading have their information fixed at the point of publication, whereas depending on where you live or your community, the information you get buried in differs. Because I was in an Asian country, it was actually very hard for me to synchronize my programming skills with the latest technology in the US. The frameworks commonly used in my country are stuck at roughly 15 years ago. The time lag between US technology and my mainstream technology is about 5 years. In my area, it's about 15 years.
Anyway, thinking about your opinion, I think there are both valid points and dark points. I think learning for its own sake is important. I agree that people should have an attitude of learning for its own sake, but I don't think it's mandatory. I think it should vary depending on the environment. Still... I do think having that attitude is a good thing. I like your idealism, senior.
However, that holds for people who are middle class or above and can survive. Because rich kids can picture the future and are given the right to love learning. But if you're poor, that's very hard. And when you grow up in such a poor family, there's an emphasis on it having to lead to practical skills.
And even if someone loves learning for its own sake, I think quite a lot of people love the honor that comes from learning. But that too can be called loving learning itself. It's purely a matter of individual values.
I don't think you're wrong, senior, but I also wonder whether the old ways are right in today's era of excessive information. It's just that because our lives and the environments surrounding us are different, we came to different conclusions. I upvoted because I agree with your thoughts. But there are points where I differ. Have a good day.
AI has the benefit of having trained on all that material, it's a snapshot in a sense. But as people increasingly stop creating content like that and we turn to AI instead of asking for help on forums we stop growing the knowledgebase that AI draws from.
I don't think forums are bad, but I've questioned whether experiences in such forums were always pleasant. There was a lot of wrong information, and there were also many people who picked fights for no reason simply because they didn't like something. But AI really does just tell you knowledge easily.
Of course, forum activity gave me the opportunity to come across knowledge whose keywords I didn't even know, and that got me studying, but I question whether such forum activity was open to everyone.
I also somewhat agree that the problem of the knowledge base shrinking is real. But is that problem really unsolvable? As far as I know, the results on the degradation problem of synthetic data say that if there is synthetic data, there isn't that much degradation. People, too, tend to idealize the past as they get older. Like treating the inconveniences of that time as nostalgia.
I don't know. Rather, once the current SLOP era passes, certain knowledgeable people may become rare and be treated well then. The future can't be known.
Anyway, this is purely a matter of values "Some people may regret the destruction of knowledge communities, but I benefited from that destruction, so I can say it's good. People always judge by their own experience, don't they? I respect your past experience. In the cases you went through, I think AI produced more bad effects than good.
I've also seen many things disappear because of AI, but it let me try things I wanted to try, and I think that part is good. So, honestly, I don't know. I have both affection and hatred for AI. But the affection seems a little bigger.
Uh... we're already at the latter prediction.
He would still be right and the field practioners would be wrong.
Like it would be fucking awesome if he could try find some more positive takes because it's tiring here a guy who basically got us into this mess, whine about how screwed we are.
Like good on him trying to raise awareness but I just find the guy so frustrating.
I do find capsules interesting though.
He doesn't search the internet to find an answer. "Search" results are nearly useless but the AI load does a good job.
He doesn't lean on AI to synthesize large information.
He doesn't use AI to build spreadsheets.
Etc.
Fortunately, he probably also doesn't use AI for therapy which two of the already crazy people in my family do.
Because a certification has to be fixed at a particular point in time. Korea's certification exams start with the 23 GoF patterns and go on to rote memorization of countless patterns and layered terminology.
The problem is that the way you learn for this certification is the old way, so someone who gets the certification ends up writing legacy code.
But if you always update the questions to the latest, then nobody can pass the certification. So, having seen how the certification system works in East Asia, I don't think it's all that good
No way to tell until after, though. So far, mutually assured destruction is working. It works until it doesn’t.
Same with AI. Thus far we’re better off with it than without it. That may change.
I mean, look what we did with fossil fuels. Some people are even still willingly burning them, even though the data is clear that our use of them has already started the total destruction of our civilization.
Is there a name for this, a technology that is so insanely useful that there is no way we won’t exploit it, but where the worst case scenario is total collapse, and impossible to determine well in advance?
(Fossil fuels I suppose don’t count as even the energy companies knew unambiguously decades ago that large scale use would eventually bake the entire agricultural system out of existence.)
The reason for that was more that the 3-4 great empires involved in the previous two were exhausted, Japan was defeated, China was in no shape, and US vs USSR had no reason to push for one, since the Cold War era gave them plenty of breathing room, roughly half the world each.
Both the USSR and US fought multiple wars directly against the proxies of the other. They weren't simply hanging back and letting others do the dirty work. In most any other era one of those would have become a full blown war between the two. Both sides pushed the line of not directly involved and then both sides looked the other way versus escalating. Nukes made the alternative too much of a risk for both sides.
Rather, after WW2 the idea of millions of people dying in a war became anathema -- countries would simply revolt against the idea.
Also, an estimated 20-40 million people have died in wars since WW2 (depending on how you do the counting), so it's not like wars have stopped.
The FDA analogy is structurally wrong for AI. Drugs have a fixed chemical composition and a testable biological mechanism; the FDA approves a specific molecule for a specific indication. A frontier model is a weight file that can be fine-tuned, quantized, distilled, or used with different prompts and toolchains after release. Pre-release testing cannot anticipate every post-release adaptation. The thalidomide comparison also misses that thalidomide's harm was a single, repeatable teratogenic effect. AI harms are emergent from use patterns that don't exist at test time. This is the core tension that makes any pre-release approval regime unworkable without also controlling how models are used after deployment.
[ ] Radiologists “replaced by AI”
[ ] The “end of the world”
Your opinion on these two topics has always brought so much value to the world
... no shit Sherlock
TFA could have been two sentences:
He warns that AI might cause serious problems: power failures, job losses, and greater social inequality.
He believes AI might improve our lives, but governments must make sure it is safe.
That's a relatively silly question, don't you think?
>So one reason he might lament the status quo is because it has people thinking we have achieved AI when we haven't.
I don't think the article allows us to infer this at all.
>But also, it wasn't apparent until recently that even if the tech did work as promised, it will be an awful thing because humans will use it to bad ends. I don't think it's particularly blame-worthy that he didn't foresee that 50 years ago.
No, really, that would be extremely naive, we can expect better from scientists.