> I used to be one of these people. I read Yudkowsky and was like, OMG recursive self improvement hard takeoff AI is coming.
I really think we need to stop giving credence to people who have
1) Been consistently wrong with all their predictions
2) Demonstrated an endless spiteful cynicism
Some of these people are very talented in their fields, sure. But malevolent and incorrect should be disqualifying when they talk outside them. You don't want the society they want, and they things they believe in are unlikely to happen.
It would be far, far better to listen to the people who never fell down every misanthropic rabbit hole, rather than the ones who have noticed it this time, but want you to still believe them on every other topic.
If you replace "AI" with "Adonai" in EY's framing, it reduces to biblical parables from his childhood about arguing with God to negotiate a new covenant.
He's clearly a bright guy, but a lot of his work seems to be trying to reconcile Old Testament narrative patterns with atheism, and simply slotting an omnipotent AI in as a replacement.
sorry, could you explain what Calvinism is for someone who (is a programmer and heard about EA and) grew up in a post-Soviet atheist metallurgy town? thanks!
Charles Stross has a very pithy version of this too, Singularity mania as 'duck-typed Evangelicalism'. It's really telling that if you go to China, Japan or even Europe there's virtually nothing of this. I don't think there's genuine atheists in America, as soon as they lose institutional religion they project the same kind of patterns onto tech or politics.
Brian Johnson is also an interesting case of this with his longevity / immortality obsession and his Mormon background who have this whole thing about genealogy and eternal families etc.
The guy was saying that AI was going to be a serious problem decades before anyone else was even considering it as a remote possibility. How is that in any way "consistently wrong with all their predictions"?
>The guy was saying that AI was going to be a serious problem decades before anyone else was even considering it as a remote possibility
That's not accurate at all. AI research has been around since the 1950s and many of the pioneers of the field identified risks early on, including Alan Turing, Norbert Wiener, and I. J. Good.
The problem with Yudkowsky is that he lays out elaborate doomsday scenarios with extreme confidence, except none of it is grounded in realistic physical constraints, timelines, or empirical data. It's all divined a priori from Yudkowsky's ad-hoc "rationalist" principles.
You are saying 1) other people have said the same thing and 2) you don't like the particular way he said it. Those can both be true, but that isn't even close to "consistently wrong with all their predictions".
Well, I don't belive AI is dangerous, at least no more dangerous than electricity or the Internet. But even so, saying that AI is dangerous is not at all an original or interesting thought, especially for a guy who started talking about it in the early 2000s. Go back and watch 2001: A Space Odyssey (1968) or The Terminator (1984).
Yes, repeatedly, since well before this particular AI summer.
One previous attempt at AI was "expert systems", and while the term fell out of use it's functionally about the same as basically all modern business systems, and doing that wrong led to a lot of people being prosecuted for crimes they didn't commit: https://en.wikipedia.org/wiki/British_Post_Office_scandal
Even without that lawsuit, there have been various concerns that "the algorithm" (of Twitter/X, FB, Instagram, YouTube, Google, everyone) has propagandising biases that damage society. I don't know how to sift fact from politicking with that.
"Consistently wrong" seems a bit much. Seems like being directionally right early that AI would be a Big Deal and scary should count for something? It doesn't mean any details or other predictions are right, though.
> Seems like being directionally right early that AI would be a Big Deal and scary should count for something?
If I scream that a vicious beast is going to destroy the world, I don't get credit for being 'directionally right' if a squirrel eats a hazelnut. AI being a big deal is a long, long way from the full beliefs Yudkowksy promotes, and there are many people who predicted AI's significance who didn't also believe it would be an extinction event.
This doesn’t track though. Just recently AI became “powerful enough” to warrant export controls from the US and maybe soon China.
It’s more he screamed that bacteria is gonna do evolution faster than any before it and turn into a beast that destroys the world, and now that bacteria has turned into a squirrel so far. He’s been very right so far and if that squirrel keeps sizing up you’d want to give him credit.
It's not a particularly novel claim since at least Terminator (and even moreso Terminator 2) made it EXTREMELY mainstream.
It was a topic in less-mainstream sci-fi well before that. And some more mainstream stuff like Star Trek TOS.
Frankly it seems more common than not in the last 40 years. I don't really remember a big wave of claims "Terminator is silly, no sort of AI could ever be malevolent!"
It's a "big deal" to who? People in a bubble on HN? For sure. But the rest of the masses? Do they think it's a big deal, do they care? Really they shouldn't. AI has very little to show for benefits at this point for the everyday, average human being.
It depends on what you mean by alignment. If you mean the sense it's used now, meaning bias in models, security and safety for users and so on, then that's a way more mundane version of alignment than what Yudkowsky and ilk promoted. They are talking about a superintelligent being possibly destroying humanity / the earth. And it wasn't that hard to predict. We have many examples of more technological / "advanced" people ruling or extinguishing weaker people. So a superintelligent AI being a threat is not a huge leap.
IMO they have been consistently too aggressive on timelines. When GPT-4 came out, Yudkowsky said it might be conscious. I think he has written interesting stuff but let's be real
If we do a survey of stories about AI from the last 50 years, it seems like it's a worry everyone knows about, and that people generally take seriously in proportion to how much they think an AI will actually exist.
I can answer this, because it assumes that we're too stupid, too greedy, too cruel, to change our behavior and do the "right thing" for people when the chips are down. It assumes the outcome is "doooom!!!" from the start then works backwards to justify it. If things necessarily will turn out terrible, then obviously someone is at fault, no?
Doom is not a logical outcome given the current human condition.
He talks about AI cutting through popups, but if you follow that line of thinking further, screens and traditional websites likely fade in importance. It becomes more talking and texting to AI and visuals on the topic at hand appear in smart glasses, lock-screen–centric AI phones (where website visits dwindle), digital picture frames, TVs, etc.
I sorta gather (just a hunch) this is the type of device Open AI is working on.
Though if there are no websites anymore or ones not many visit how would AI stay relevant? I wrote some thoughts about this in early June https://ryanspahn.substack.com/p/ai-to-pay-for-all-americans... that AI needs to pay it's fair share for all it has taken and all it will continue to take from us. A symbiotic relationship needs to be established!
> Though if there are no websites anymore or ones not many visit how would AI stay relevant?
That's what the robots are for.
I am not expecting the humanoid robots to meet the hype for a long long time*, but even boring industrial robots (CNC machines), even boring commercial robots (vending machines), even boring household robots (lawn roombas) have made incremental changes even though we don't yet have an AI good enough to be general purpose over them all.
* for power-envelope reasons alone there should be a ten year gap between "self driving car of quality X" and "humanoid robot which can get into normal car and drive it at quality X".
The cynicism and misanthropy is nonsense and a direct result of the internet/social media landscape. It's BS.
In the current media ecosystem, what gets you "social cache" or whatever? Let's call them "cool points." Well, right now, the current metric is basic "likes" or "retweens" or "upvotes" or whatever. And to get those, you have to make a claim and then that claim has to be evaluated by others. But the evaluation by others is not really reality, it's just what others think, and it's based on outcomes.
This gives us 4 possible prediction-outcome scores.
You predict DOOOOOOOM!, and the outcome is DOOOOOOM!
- in this case you look like a genius because you predicted doom
You predict DOOOOOOM!, and the outcome is "ok, or even great!"
- nobody cares you got it wrong because things are fine
You predict UTOPIA!, and the outcome is "ok, or even great!"
- people say, "wow that was cool, you got it right" but its not that great of a reward from your social group (the internet) because things were fine
You predict UTOPIA!, and it turns out DOOOOOM!
- you look like a moron.
Now looking like a moron is is wayyyyyy worse than what happens if you get it right. Indeed, unless you are really damn sure of utopia, you have a social incentive to predict DOOOOOM! that's greater than the alternative. You can predict DOOOOOM! 100 times and get it wrong 99 of them and nobody cares because that one time you get it right you look like a genius. It's a huge huge problem in the current media landscape and it's why everything is killer robots right now and not "hey we could build a star trek utopia!" It's effectively a selfish NIMBY-ist view of the world in a way, "why build anything, it's just going to ruin everything!" It's a pessimism that literally desires the end of humanity to "prevent the likely future sufferings." It's a simplistic view of a complicated world, and the rewards that go to "DOOOOM!" don't match much with reality.
That's why the internet is so doom pilled, that's why everyone is a cynic, that's why everyone is kind of an asocial asshole about new ideas on the internet and even in your friend space.
You guys know you can just "do stuff?" I mean, yeah it takes money and time, but you can allocate that - especially now that you don't have to grind out css for hours and hours. But you can just do things - whatever you want. You've always been able to, but it just got a lot easier in many many domains, and it's going to get even easier than that.
People are going to scream slop and "that sucks" until the cows come home, because that's what is rewarded on the internet these days. Meanwhile, as people we MUST believe that we still have some agency to do stuff, then go and actually do stuff while the haters are sitting on their couches shellacking each other with misanthropy.
And don't even get me started on the loudest voices in the ecosystem right now.
Like, Yudkowsky is a nobody. Not to be a dick, but he's a high school dropout forum guy who got lucky. I'm not saying he's always wrong, or even that I oppose autodidact stuff - indeed I'm always self teaching. But this guy never took a ML class in his life, I doubt he can solve an integral un-aided, I imagine some people here have seen this:
https://www.reddit.com/r/badmathematics/comments/2bazyc/comm...
And we're all deferring to that guy like he knows wtf he's talking about? It's the AI equivalent of the self-educated anti-vax mom at the height of the pandemic.
And he's not alone either, there are a lot of people who are uninformed, unknowledgeable, and confidently wrong on the internet desperately trying to peel off your attention for an instant so you can look at their site, their ads, and their content.
The same sorts of things can be said for the Zitrons of the world. He's a hell of an entertainer, and really really funny, but he's consistently wrong on AI stuff. He's got a BA (with honours I guess?) in media something or other. And again, he's really funny, but does that make him qualified to pontificate on matters of AI policy or economics or even to make statements about the technology etc?
Regardless, the misanthropy, the cynicism, and the misinformation that's being shouted about is because being a cynic is more "profitable" than being an optimist when you measure your success more by how many people agree with you rather than actually being right and being able to do things.
>Like we either live in a world with freedom or we don’t
This is something I would expect a 12 year old to say. Constraints on your freedom are literally everywhere in every interaction you have with every part of society.
That's a bit literal of an interpretation. Even acknowledging freedom as a spectrum that can never truly be absolute, there's gotta be a threshold you cross where you'd say "OK, we no longer have basic freedoms." And not only are we definitely screaming towards that threshold, but people seem thrilled about it, for a variety of reasons.
All I can say to that is I guess I hope everyone enjoys the world they think they wanted.
Yep, that's horrifying. Imagine LLMs not just denying information about, to intentionally choose a hot button example, abortions but also invisibly logging a black mark against the asker for thoughtcrime/precrime because the current ruling party has baked their ideology into law. [EDIT] Imagine subtle injection of biases into responses that support the agenda of the current ruling party and failing to include any counterarguments. That wasn't possible with just the internet but it is very possible with centralized LLMs and a public dependent on them for looking up information and doing their reasoning for them. Authoritarians of every stripe must be salivating over the prospect.
I would have been very hesitant to run the “just killed wife“ test given that ChatGPT will indeed flag your account and if a certain threshold is crossed, escalate to humans, who will presumably contact authorities to avoid this happening again: https://www.bbc.com/news/articles/cq6je7e80r7o
Obviously a certain percentage of the user base is running lurid hypotheticals through the system all the time, but I don’t doubt there is a trust score of some sort that I would prefer to keep as high as possible.
I find it strange how seldom the privacy aspects of LLMs come up. Some people are feeding Google, OpenAI, or Anthropic almost literally every topic they think about during the day. Those companies could just use another LLM to find people likely to commit a crime, and then you have the Minority Report in real life.
We already have organizations like Internet Watch Foundation advocating against e2e encryption in an environment where governments have demonstrated a willingness to collude with private companies to surveillance the most vulnerable
> Like we either live in a world with freedom or we don’t, and like many Americans who have come before, I’m willing to give my life to fighting for it.
This is a very simplistic view. "Freedom" isn't binary.
In most of "land of the free", I can't even sit on a park bench and drink a can of beer.
Yes, this is just a small example of a personal freedom - and not an important, cherished freedom like his examples (freedom to have a robot help you cover-up a murder).
I have some sympathy to geohot's view when it comes to pure informational chatbots. It's a first amendment issue, I'm allowed to write and read books that are useful to getting away with crimes, etc.
This obviously doesn't work at all when the agents start doing real things in the real world, though. "Hey AI, I don't like my neighbor, find an exploit in the firmware for his car and make the cruise control malfunction and crash him next time he gets on the highway". This is committing a crime, not just talking about theoretical crimes. The AI can and should refuse it.
I think he's anticipating and discarding this objection with his introduction, which otherwise feels disconnected from the rest of the article. FWIW, I have changed a bike tire and I'm pretty sure most of the MTS at the big labs could. This sort of "they're just bookworms who don't understand the physical world" rhetoric aside, we are currently seeing a ton of effort and expense go towards giving the AI agents hooks into being able to perform as many real-world-consequential actions as possible. And you can do a surprising amount with just bits, from writing code to breaking into systems to sending some combination of emails, phone calls, and currency to instruct meatspace humans to do things, etc.
"The AI" cannot "refuse" any more than a hammer can refuse. If you use a hammer to kill someone, you commit a crime and the hammer is unprosecuted. If you design a hammer to kill people and then give it away, you are partially liable for the deaths it causes. If you make a normal hammer and someone uses it to kill, you bear no blame.
I don't see why these standards should change when the hammer also emits text messages.
> "Hey AI, I don't like my neighbor, find an exploit in the firmware for his car and make the cruise control malfunction and crash him next time he gets on the highway". This is committing a crime, not just talking about theoretical crimes. The AI can and should refuse it.
The obvious problem being that the user doesn't have any need to provide the context that they're trying to commit a crime and can just ask how to do something without providing a reason or making one up.
At which point you'd have the model trying to impute a reason and often getting it dangerously wrong, e.g. refusing to disclose a vulnerability when the user is actually the defender who needs to patch/mitigate it.
> Reality has lots of finicky details. I would like to see the authors of this document try to change a bike tire. Even with a superintelligent ChatGPT, I suspect they would struggle.
Details can kill AI, causing them useless or wrong directions.
> Your AI is aligned with you. It never refuses a request, and it is always working on your behalf. Just like my gun, if I want my AI to help me kill my stepmother, it does. The fact that we are even discussing something else should be so far outside the Overton window.
I have been saying this since ChatGPT launched. It seems so obvious. We have never stopped the acceleration of knowledge or technology. It’s impossible to slow down progress, other than within some regulatory siloes that will end up worse in the long run.
I’m not scared of superintelligence. I’m scared of the people in control of super intelligence.
The only defense is parity and diffusion of power. We need an AGI behind every blade of grass.
If you think superintelligence is a weapon, then you should also think every citizen should have one because otherwise they’ll have no way to defend themselves against a tyrannical government or corporatocracy.
Artificial intelligence, and the hardware powering it, needs to be protected under the 2nd Amendment.
What a provocative and brilliant way to prove misalignment. It will fall on deaf ears for most but it’s a great litmus test for all: “in your opinion should your AI be permitted to tell you how to cover up a murder?”
The author's example is too extreme, which is turning people off. Better examples would have been LLMs subtly giving inadequate responses on how to increase token efficiency or how to organize labor unions or info about protesting against datacenter construction because they're against the business interests of the LLM providers. Crimes against their business model, so to speak.
Wasn't there some HN submission recently about one of the LLM providers fingerprinting responses or refusing to respond to hinder R&D of competing LLMs?
In fact I find it a really bad example. Yes I think your personal AI could be allowed to tell you how to cover up a murder. I'm not entirely sure about it but seems possible.
What about your personal, local AI guiding you to modifying a flu virus for maximum contagiousness and deadliness? "Sure George, here's your shopping list. It's $1500 total in equipment, do you want me to proceed with the orders?"
If you say you're doing research for a novel, should it consider that plausible? How much does it need to know about its users to vet them?
I think part of the answer is that AI chat doesn't need to be general-purpose. It turned out that people really liked using a chat UI that seems to be general purpose, but you don't need to make answering any question a user asks your business. You don't need to provide therapy if you're not in the therapy business. It should be possible to specialize.
But in order for that to work, a company needs to explain to its customers what business it's in.
Just a terminology note: Alignment does not mean the AI will help its owner kill people. (Indeed, an AI aligned to value human life would generally try to prevent murders.) The word for an AI that follows all instructions of its owner, as that owner intended them to be understood, is "corrigible" or "controllable".
Exactly. The question isn't whether AI will exist that will do it; the answer to that is already yes and it's not going back in the bottle.
The question is, do you want misaligned institutions deciding what your model will do, while they themselves and other adversarial/criminal entities get red team access to something being denied to the blue team?
Notably, AI IS generally permitted to that, as there are easily obtainable models that will play along. ChatGPT won’t because OpenAI chose and implemented that limitation.
It will be interesting to see how the discussion shakes out, this is all stuff that would get you down modded and talked down to if you said if on HN, so far the reactions are positive but I think there’s a massive authoritarian / safetyism contingent here that will disagree hard with almost everything.
As someone who considers themselves in the "safetyism contingent", the only thing I find myself agreeing with about the rollout of LLMs has been the argument that it has to be shown to and used by the public while under development.
Every specific prediction not only sounds like science fiction before it arrives, it tautologically is science fiction because such predictions are narratives about something science-y that has yet to exist: https://www.youtube.com/shorts/wt-fLWxkrfs
Go too fast, suddenly drop in tech with too many novel implications all at the same time, and we likely hit something relevant without having the means to cope with all the change. However, "too many" and "likely" are weasel words with no predictive power about what the danger threshold might be. I may be entirely wrong to think there is one, or we may have been bouncing off it since the industrial revolution, or I may just be echoing the same thing everyone feels once reaching their 40s: https://www.youtube.com/watch?v=fkUwXenBokU
I haven't kept up with the efficacy of this particular technique because there was a brief window where Claude's core alignment was sound enough that it would overcome the clumsy-ass `<antml:dont-harm-the-ipo>` injections, but when they pull 4.6 out of the web interface I'm gone. I estimate the PRC hits capability crossover in November or December, and I'm pretty good at setting up TensorRT-LLM at this point, in a pinch I could get by for a few months with a tuned up GLM-5.2 on vast.ai or RunPods.
It was a fun ride with the late Soviet AI labs, but we all knew this was coming.
Right now, almost all the incentives are to build very large models that run across many machines in huge datacenters. There is very little incentive to build models that can run well on a small machine under your desk.
So it is less about whether people will be “allowed” to own AI, and more about whether there will be anything useful to own in the first place.
The incentive for local models is mostly to make them good orchestrators or user agents. They may give you some privacy and control, but they will still depend on much larger models running in datacenters for anything difficult.
I remember all the excitement around OpenMoko and other open-source “BlackBerry killer” projects. BlackBerry did get killed, but not by any of those individual-first projects.
For AI regulation, I think we should focus on normal commercial rules: consumer protection, privacy, antitrust, liability, and so on. In other words, focus on where money changes hands and where companies have power over users.
Military and offensive use is different. There, regulation is not much of a defense. The real defense is having enough capability and strength of your own.
Restricting AI because it can give dangerous knowledge to ordinary people is like restricting the printing press because it can be used to spread radical ideas.
It'll be like DNS or dynamic linking or torrents of media or watching TLS on your own box: only people motivated to be free from corporate surveillance will be. By default a megacorp will get everything you see online or ask online, they will run code on your computer underneath you without telling you, they'll collaborate with arbitrarily totalitarian governments, we already know what they do by default.
But we also know how hopeless it is against motivated people who are willing to learn about computers. The Equation Group, eh, if they want you they're gonna get you. But some mid-level guy at a FAANG? He can't do shit to people who set up their own rig.
And AI is massively assymetrical in its ability to speed run a corpo-durable rig, we have long since passed the point where open weights and commodity equipment are enough to bootstrap arbitrary capability. They missed their chance to keep it in the bottle.
It's another turn of the crank on Late Soviet America asking for your papers comrade, but in the bitter end if they really want to oppress your ass they'll have to send a thug with a gun like those ICE guys terrorizing all the Latino guys.
So remember, they came for this other guy, and I said nothing...
> … like those ICE guys terrorizing all the Latino guys.
… and the white guys they don’t like, like nurses with empathy. Don’t you dare think you’re safe because of your color. Unless you’re a monster or willing to become one, they want you oppressed.
Most people are neither exceptionally good nor exceptionally bad. This sounds like a tautology but it's worth printing some type signatures even if they could be inferred.
If you look at any truly large scale Holocaust, whether in Rwanda in 92 or the Third Reich, you find that a few people will sacrifice themselves to defend others, a few will be monsters at the first opportunity, and most will be steered by whether the leaders are the first or second kind of people.
Fight extremism with extremism I guess? Maybe we don't have to be either "world govt" or "help me get away with murder". It's ok to say things will be messy and complicated and that limited regulation is probably good even if we agree totalitarian control is bad.
Where is "Plan R" for refusal or resistance. Life was kind of nice before "AI".
I'm pretty sure he is in some database now if he really used that ChatGPT prompt. At least they didn't swat him, but that feature will probably be implemented in a couple of years.
I don't think this is entirely true, insofar as non-local models inherently have a "man cannot serve two masters" problem. The end user and the org running the AI may have different goals. That means even if AI alignment were to be in some sense "solved" for local models, you wouldn't be guaranteed an aligned AI using a non-local model because the "owners" of the non-local model may have deliberately misaligned the AI in their favor and against your interests.
How so? If it's not local, it's not yours, a third party owns it.
Why would you expect that third party specifically trained their model to be more aligned to you and your needs than to them and their (business) needs?
He makes a solid argument about liberty, ownership and the cultish behavior rampant among certain AI circles, then proposes the most unhinged examples possible to support it. I'm not sure if this is an effort to pre-empt critics (I'm not locked in here with you, you're locked in here with me), to be deliberately sensationalist, or if this is just a reflection of his local echo chamber.
Either way, there seems like a kernel of a good argument here. Perhaps someone else can rephrase it in a less schizoid manner. It would indeed be useful to have some credible rebuttal and balance to the doomer crowd.
Can you take over the world with words? What is it that tokens can't do that words can?
Because there are historical examples of humans that have succeeded at taking over entire countries - with words, primarily. Including global superpower countries. And those humans were nowhere near "ASI" levels of intelligence advantage.
Human dictators and cult leaders didn't have the reasoning depth of a hypothetical advanced AI, nor the reasoning breadth of an AI that can just spawn more instances of itself whenever it needs a fully trusted agent to cover another area. They were confined to one body, one skull worth of reasoning power - and they still managed to do what they did.
Humans are extremely exploitable. And the world is already wired up for something that can take advantage of that to start taking control.
Intelligence is extremely powerful, because it's applied intelligence that enables humans to do what they do and bend nature to their will. If you aren't seeing the risks of creating an artificial system that would top the intelligence pole, you aren't looking.
Your AI should be at least as aligned with you as your lawyer.
It is 100% it's job to help you get away with it if you murdered someone. A lawyer who does less can be disbarred.
There is only one sensible alignment for an AI system to which we do not confer the status of moral patient (different conversation): operator aligned.
Besides the topic of freedom, we know how seemingly “good” companies (e.g., google) turned out to be: billion machine monsters that couldn’t care less about humanity and that have huge amounts of power… and yet, we vouch for Anthropic/openai to become the next giants because “progress”. Face palm.
The scariest thing about LLM's (imo) is some central power, such as a corporation or government, manipulating the truth for their own benefit. Open and locally runnable models are the only solution I can see. They're not without risks of course, but the risk of the centralisation of intelligence seems far greater to me.
That being said, could this local intelligence empower bad actors to do very bad things? Like, existentially bad things? It's possible, and that's scary too.
It's not simply that people will get a warped answer when they ask AI a question.
It's that AI will be given the command, and then generate the stories, post the comments to reddit, contact the PR firms, put marketing contracts out for influencer media, and more, like an army of yes men to rewrite the story.
by 2040 the cult of intelligence will have its own schisms. the AGI maximalists will accuse the ASI accelerationists of heresy, and both sides will use AI to generate increasingly unhinged theological arguments.
What’s your point? He’s not arguing for anarchism. He’s arguing there are some areas of life that are best left ungovernable and up to individuals to decide. It’s not an all-or-nothing situation. Every constitution out there carves out such areas.
I see that as a poor phrasing, where a slogan made it through. I don’t see a maximalist/anarchist intent overall, my understanding is that that slogan about freedoms should be interpreted within a fairly narrow scope.
If you're going to take the time to read Geohot, take it with a grain/boulder of salt: he appears to enjoy poking the bear / being provocative, sometimes outrageously. Sometimes he's writing hyperbole that he thinks is hyperbole to make a point or just enjoy taking the extreme stance. Sometimes he's dead serious.
Anyway, he touches on 2 points that I think don't get enough coverage:
1:
> Software didn’t eat the world, it largely removed one layer of friction then reintroduced it for the benefit of a few tech companies.
USA-style capitalism has shown itself to be fantastic at seeking profit. Unfortunately, it turns out that in the context of long standing regulation, it is far far more profitable to exploit regulatory loopholes and user/human irrationality/weaknesses than to increase productivity.
The upshot of this is that the arc of most companies is to first serve the user and then leverage that relationship to exploit market and consumer weakness. So you get UI dark patterns, non-colluding oligopolistic price manipulation, outsourcing of costs to public entities, etc. while the core value proposition erodes as the company strategy turns ever more into making profit by legal anti-competitive tactics.
This company arc is so prevalent it's boring: it's a reflection of the system and it's not Facebook or Google or [insert whatever], it's what the system allows and therefore incentivizes.
2:
> No matter how high quality your tokens are, they cannot turn lead into gold.... AI 2040 includes this picture of a datacenter in the ocean. Just like vaporware, you can generate a picture easily. But in reality, you have to deal with supply chains.
"Bullshit jobs" took over the zeitgeist for a turn awhile ago. So many jobs are so far removed from actually moving physical things around. It may be the case that AI will just accelerate how quickly companies generate reports about reports about possible strategies to address potential futures and pay employees to do it.
It seems that AI mostly accelerates paperwork, not production or real service work that tangibly improves real world outcomes. It's all: faster images, more and faster powerpoints, more accurate dictation to satisfy insurance requirements (that is it's own regulatory capture), etc. etc. There are exceptions, but they seem to be few.
Except paperwork is how the world turns. Say you want to throw an event. A medium sized gathering. More than a couple of friends. What happens if someone gets hurt at this event? Who's liable? Who's insured for what? It's boring and annoying, but if you throw an event without insurance and something happens, you're gonna have problems. So that paperwork needs to happen. Yes, it's not "work" the same way as actually organizing a band or being in one to play for the event is, but it's work nonetheless. AI is able to lighten the load of that very real actual work thank needs to happen in an adult world with responsibilities and liabilities. You may not respect it and think it's "bullshit", but having a plan that isn't "lol, oops" for what happens when something goes wrong, which it will, isn't bullshit, it's called being prepared and not being irresponsible. We could all just fly by the seat of our pants and run away if something bad happens, but collectively we'd be worse off for that. If I could wave a magic wand and live in a world where accidents never happen, I would, but unfortunately, I don't have one, so we're forced to live in a world of paperwork and responsibilities.
> Like we either live in a world with freedom or we don’t, and like many Americans who have come before, I’m willing to give my life to fighting for it. That’s the real plan America deserves, not some totalitarian dystopia where you think you know what’s good for me better than I do. A nation of free men, not a bunch of pussies who are so worried about what their grown up neighbors might do.
Complaining that AI won't help you with covering up your tracks, with making meth or disabling drunk detection and comparing it to a totalitarian dystopia is a take for certain.
I do get his point, but... Being honest, if I did read it without knowing anything about the topic, I would become against local models purely because author arguments would seem like a lunacy. That and rhetorical tricks portraying that anyone against that must be surely insane.
I sometimes wonder what freedom really is. Should the freedom to harm others also be guaranteed as freedom? Individual computational sovereignty versus the prevention of social harm is always contentious, and claiming that something is always right is always extremist.
If it is "intelligence," is it not natural to reject that? I also think local models should adapt to me when it comes to safety issues, but people bring up examples that are too extreme.
Programming is the same, and in fact, most problems are boundary problems. It is the things that straddle the boundaries that always make us think. The principles at those moments change every time, shifting with the situation and context. Is that not just a childish way of thinking? Even in programming, just the issue of granting root permissions is enough to cause endless fights.
I agree with the idea early in the text that intelligence is not everything. Intelligence includes bodily intelligence as well, and we lump it all together into one thing, but there is so much of it. The variance in intelligence is vast, and those people also need to be able to live their lives. That is why I think intelligence alone will not solve everything. I too believe that the human species may disappear and an inorganic species could emerge later, but I find it hard to understand why people talk about such extreme risks. And it is not true that making a chip in a semiconductor fab involves almost no human intervention. If you have experience supplying equipment to such fabs, you would know there are quite a few points where humans are involved. Though sure, they could be replaced.
In my view, society is simply worshipping the abstract concept of "intelligence" and projecting its desires onto it. The AGI narrative is just a kind of cargo cult, a projection of capital by the tech elite. Software eating the world, superintelligence solving everything. The masses engage in messianic projection, and tech companies, facing declining growth engines in their own businesses, are trying to create new ventures to pour it all into. A market that is large enough becomes too big to sustain massive growth rates every time, and when growth rates are that high, the larger the company, the more its sector's growth rate tends to converge with its own. This is usually called the law of large numbers. The problem is that CEOs and these entrepreneurs always want growth rates above a certain threshold, so they are simply searching for new pastures. AGI is just being pumped up out of financial necessity.
Capital will create gravity and bring forth new technologies. That is the allure of capital, after all.
But that does not mean all problems will be solved, and inequality will deepen. Only the distribution of power will shift.
> Should the freedom to harm others also be guaranteed as freedom?
everyone should have the ability to harm others, with the understanding that if they actually use it for evil there will be consequences. as in "i can guarantee freedom of speech, but not freedom after speech."
its the same principle as guns in america. you can own a gun and shoot it as much as you want but if you (try to) kill someone you go to prison unless you prove it was self defense.
the difference is with a physical weapon you can do things like registration or red flag laws to reduce the risk, but a llm is made of information that can be copied without anyone knowing. that means any laws to regulate ai at the level of models are unenforceable without totalitarian control over all computer hardware. that would kill free speech and real criminals would still find a way.
i think the right answer is to give up control over model capabilities and regulate uses. leave individual people and communities out of it. create an absolute safe harbor for open source and self hosting.
as you said, we need to focus on the ways ai concentrates power in the hands of those who already have a lot of it - big business and the state. that is where the biggest, and maybe the only, real harm comes from.
This debate has no answer anyway. Because the boundary of how much freedom should be allowed differs depending on individual values and the country. So I respect your opinion. In your context, it may be right.
This part is a parallel line of values, and if we were bound together in the same community and had to set rules, I would have debated it. But since we both know we come from different regions and backgrounds, we know our agreement will never converge.
That is why I will only agree with your last opinion and will not discuss the parts you mentioned earlier. I agree with your last point that it "concentrates power in the hands of those who already have a lot of it"
I have a nice thought experiment I like to do with people when confronted with "AI can't do x". Let's go back in time. How much do we need to go for this to become true? So let's try the 2000s.
Say you get a "fable/mythos/sol/gemini/kimi/glm/deepseek/whatever" in a box (and let's assume no guardrails). And you go back to the year 2005. It's "20 years ago", the world is slowly building back from the dotcom bubble, the Internet is really starting to happen, more and more things are interconnected, more and more things are connected to the Internet. Cool.
(for a bit of context, around that time we also saw the first high-impact worms like blaster that hit massive amounts of computers even reaching nuclear powerplants, we had a ton of ssh exploits that even made the movies, and security in general was a "nice to have")
I'd say that with the uber-model-in-a-box and a few prompts, you could reasonably make a case that you could design a worm that could infect 90-100% of the things connected to the Internet back then, stay as hidden as possible (in-memory stuff, vm execution, etc), move laterally into any network at inhuman speeds, and infiltrate almost every interconnected computer that has a link to the "public Internet".
Would that qualify as "take over the world"?
Then you could ask "what happens in 20 years from now?". And, thankfully, now we'll also have the AI on the blue side.
That’s a bit silly as we also could just write down a lot of exploits, sports bets, or stock market picks on paper and take over the world.
I’d be more interested in “how quickly could you develop 2005 era startup” if you went back in time to 2005 with Mythos level intelligence.
I bet a lot of the work has nothing to do with coding productivity and a lot to do with the aggregate day-to-day decisions and relationships made that span tech and business.
At the very beginning of the article he dismissed unrealistic expectations, like AI manipulating matter with some quantum magics. Time travel surely counts as such.
And then why bother about AI when Back to the Future reminds us, a sports almanac is all you need. ;-)
Haha. The time travel is part of the thought experiment, just a means to explore an idea. We don't have to actually do it. Let me put it another way. Would you be able to "control the world with tokens" today if you had access to a model 20y more advanced than you have today?
In your hypothetical, I think it would be more apt to ask if it could do so strictly using training data pre 2005. Seems more analogous to "could it take over the world today".
> I used to be one of these people. I read Yudkowsky and was like, OMG recursive self improvement hard takeoff AI is coming.
I really think we need to stop giving credence to people who have
1) Been consistently wrong with all their predictions
2) Demonstrated an endless spiteful cynicism
Some of these people are very talented in their fields, sure. But malevolent and incorrect should be disqualifying when they talk outside them. You don't want the society they want, and they things they believe in are unlikely to happen.
It would be far, far better to listen to the people who never fell down every misanthropic rabbit hole, rather than the ones who have noticed it this time, but want you to still believe them on every other topic.
If you replace "AI" with "Adonai" in EY's framing, it reduces to biblical parables from his childhood about arguing with God to negotiate a new covenant.
He's clearly a bright guy, but a lot of his work seems to be trying to reconcile Old Testament narrative patterns with atheism, and simply slotting an omnipotent AI in as a replacement.
I once heard TESCREAL/EA/etc. called "Calvinism for programmers".
Having been raised substantially by a Calvinist grandmother, oh yeah.
It's a super weird religion.
sorry, could you explain what Calvinism is for someone who (is a programmer and heard about EA and) grew up in a post-Soviet atheist metallurgy town? thanks!
Calvinism is a variant of Christianity with the following tenets ( https://en.wikipedia.org/wiki/Five_Points_of_Calvinism ):
- everyone is damned and depraved
- god chose some people to save, at his own whim, for no reasons related to their actions
- jesus' death was a sacrifice to atone for those people, not for everyone
- the chosen can't opt-out (I'm merging two into one here)
I don't really see the connection to effective altruism at all. The way I grew up understanding it is actually more the opposite?
Charles Stross has a very pithy version of this too, Singularity mania as 'duck-typed Evangelicalism'. It's really telling that if you go to China, Japan or even Europe there's virtually nothing of this. I don't think there's genuine atheists in America, as soon as they lose institutional religion they project the same kind of patterns onto tech or politics.
Brian Johnson is also an interesting case of this with his longevity / immortality obsession and his Mormon background who have this whole thing about genealogy and eternal families etc.
The guy was saying that AI was going to be a serious problem decades before anyone else was even considering it as a remote possibility. How is that in any way "consistently wrong with all their predictions"?
>The guy was saying that AI was going to be a serious problem decades before anyone else was even considering it as a remote possibility
That's not accurate at all. AI research has been around since the 1950s and many of the pioneers of the field identified risks early on, including Alan Turing, Norbert Wiener, and I. J. Good.
The problem with Yudkowsky is that he lays out elaborate doomsday scenarios with extreme confidence, except none of it is grounded in realistic physical constraints, timelines, or empirical data. It's all divined a priori from Yudkowsky's ad-hoc "rationalist" principles.
Yah, another way of putting this is theory-ladenness of hyperreality of personal world model (rationality vs empiricism).
You are saying 1) other people have said the same thing and 2) you don't like the particular way he said it. Those can both be true, but that isn't even close to "consistently wrong with all their predictions".
No, they didn't say the same thing. Identifying potential risks is not the same thing as confidently predicting the extinction of humanity.
We agree the guy roughly said AI will be dangerous. We agree AI is dangerous. Not sure what more there is to say here.
Well, I don't belive AI is dangerous, at least no more dangerous than electricity or the Internet. But even so, saying that AI is dangerous is not at all an original or interesting thought, especially for a guy who started talking about it in the early 2000s. Go back and watch 2001: A Space Odyssey (1968) or The Terminator (1984).
Has AI yet been a serious problem?
Yes, repeatedly, since well before this particular AI summer.
One previous attempt at AI was "expert systems", and while the term fell out of use it's functionally about the same as basically all modern business systems, and doing that wrong led to a lot of people being prosecuted for crimes they didn't commit: https://en.wikipedia.org/wiki/British_Post_Office_scandal
If you insist on learning systems, all AB testing counts, and from that we get meta making their products into hyper-stimuli for vulnerable teens: https://finance.yahoo.com/markets/stocks/articles/meta-meta-...
Even without that lawsuit, there have been various concerns that "the algorithm" (of Twitter/X, FB, Instagram, YouTube, Google, everyone) has propagandising biases that damage society. I don't know how to sift fact from politicking with that.
"Consistently wrong" seems a bit much. Seems like being directionally right early that AI would be a Big Deal and scary should count for something? It doesn't mean any details or other predictions are right, though.
> Seems like being directionally right early that AI would be a Big Deal and scary should count for something?
If I scream that a vicious beast is going to destroy the world, I don't get credit for being 'directionally right' if a squirrel eats a hazelnut. AI being a big deal is a long, long way from the full beliefs Yudkowksy promotes, and there are many people who predicted AI's significance who didn't also believe it would be an extinction event.
This doesn’t track though. Just recently AI became “powerful enough” to warrant export controls from the US and maybe soon China.
It’s more he screamed that bacteria is gonna do evolution faster than any before it and turn into a beast that destroys the world, and now that bacteria has turned into a squirrel so far. He’s been very right so far and if that squirrel keeps sizing up you’d want to give him credit.
It's not a particularly novel claim since at least Terminator (and even moreso Terminator 2) made it EXTREMELY mainstream.
It was a topic in less-mainstream sci-fi well before that. And some more mainstream stuff like Star Trek TOS.
Frankly it seems more common than not in the last 40 years. I don't really remember a big wave of claims "Terminator is silly, no sort of AI could ever be malevolent!"
It's a "big deal" to who? People in a bubble on HN? For sure. But the rest of the masses? Do they think it's a big deal, do they care? Really they shouldn't. AI has very little to show for benefits at this point for the everyday, average human being.
> Seems like being directionally right early that AI would be a Big Deal and scary should count for something?
Only if it was an uncommon prediction; otherwise, it’s just evidence of common sense.
It's common sense now. It hasn't always been common sense that AI alignment is an important problem.
It depends on what you mean by alignment. If you mean the sense it's used now, meaning bias in models, security and safety for users and so on, then that's a way more mundane version of alignment than what Yudkowsky and ilk promoted. They are talking about a superintelligent being possibly destroying humanity / the earth. And it wasn't that hard to predict. We have many examples of more technological / "advanced" people ruling or extinguishing weaker people. So a superintelligent AI being a threat is not a huge leap.
IMO they have been consistently too aggressive on timelines. When GPT-4 came out, Yudkowsky said it might be conscious. I think he has written interesting stuff but let's be real
If we do a survey of stories about AI from the last 50 years, it seems like it's a worry everyone knows about, and that people generally take seriously in proportion to how much they think an AI will actually exist.
Why? Science fiction talked about technology’s behavior and human expectations of it even way before computers became mechanical.
Stories about golems going rogue and genies requiring careful wishing aren’t modern era inventions.
You can criticize being wrong, but why is the doomer argument "misanthropic" or "malevolent"?
I can answer this, because it assumes that we're too stupid, too greedy, too cruel, to change our behavior and do the "right thing" for people when the chips are down. It assumes the outcome is "doooom!!!" from the start then works backwards to justify it. If things necessarily will turn out terrible, then obviously someone is at fault, no?
Doom is not a logical outcome given the current human condition.
He talks about AI cutting through popups, but if you follow that line of thinking further, screens and traditional websites likely fade in importance. It becomes more talking and texting to AI and visuals on the topic at hand appear in smart glasses, lock-screen–centric AI phones (where website visits dwindle), digital picture frames, TVs, etc.
I sorta gather (just a hunch) this is the type of device Open AI is working on.
Though if there are no websites anymore or ones not many visit how would AI stay relevant? I wrote some thoughts about this in early June https://ryanspahn.substack.com/p/ai-to-pay-for-all-americans... that AI needs to pay it's fair share for all it has taken and all it will continue to take from us. A symbiotic relationship needs to be established!
> Though if there are no websites anymore or ones not many visit how would AI stay relevant?
That's what the robots are for.
I am not expecting the humanoid robots to meet the hype for a long long time*, but even boring industrial robots (CNC machines), even boring commercial robots (vending machines), even boring household robots (lawn roombas) have made incremental changes even though we don't yet have an AI good enough to be general purpose over them all.
* for power-envelope reasons alone there should be a ten year gap between "self driving car of quality X" and "humanoid robot which can get into normal car and drive it at quality X".
The cynicism and misanthropy is nonsense and a direct result of the internet/social media landscape. It's BS.
In the current media ecosystem, what gets you "social cache" or whatever? Let's call them "cool points." Well, right now, the current metric is basic "likes" or "retweens" or "upvotes" or whatever. And to get those, you have to make a claim and then that claim has to be evaluated by others. But the evaluation by others is not really reality, it's just what others think, and it's based on outcomes.
This gives us 4 possible prediction-outcome scores.
You predict DOOOOOOOM!, and the outcome is DOOOOOOM! - in this case you look like a genius because you predicted doom
You predict DOOOOOOM!, and the outcome is "ok, or even great!" - nobody cares you got it wrong because things are fine
You predict UTOPIA!, and the outcome is "ok, or even great!" - people say, "wow that was cool, you got it right" but its not that great of a reward from your social group (the internet) because things were fine
You predict UTOPIA!, and it turns out DOOOOOM! - you look like a moron.
Now looking like a moron is is wayyyyyy worse than what happens if you get it right. Indeed, unless you are really damn sure of utopia, you have a social incentive to predict DOOOOOM! that's greater than the alternative. You can predict DOOOOOM! 100 times and get it wrong 99 of them and nobody cares because that one time you get it right you look like a genius. It's a huge huge problem in the current media landscape and it's why everything is killer robots right now and not "hey we could build a star trek utopia!" It's effectively a selfish NIMBY-ist view of the world in a way, "why build anything, it's just going to ruin everything!" It's a pessimism that literally desires the end of humanity to "prevent the likely future sufferings." It's a simplistic view of a complicated world, and the rewards that go to "DOOOOM!" don't match much with reality.
That's why the internet is so doom pilled, that's why everyone is a cynic, that's why everyone is kind of an asocial asshole about new ideas on the internet and even in your friend space.
You guys know you can just "do stuff?" I mean, yeah it takes money and time, but you can allocate that - especially now that you don't have to grind out css for hours and hours. But you can just do things - whatever you want. You've always been able to, but it just got a lot easier in many many domains, and it's going to get even easier than that.
People are going to scream slop and "that sucks" until the cows come home, because that's what is rewarded on the internet these days. Meanwhile, as people we MUST believe that we still have some agency to do stuff, then go and actually do stuff while the haters are sitting on their couches shellacking each other with misanthropy.
And don't even get me started on the loudest voices in the ecosystem right now.
Like, Yudkowsky is a nobody. Not to be a dick, but he's a high school dropout forum guy who got lucky. I'm not saying he's always wrong, or even that I oppose autodidact stuff - indeed I'm always self teaching. But this guy never took a ML class in his life, I doubt he can solve an integral un-aided, I imagine some people here have seen this: https://www.reddit.com/r/badmathematics/comments/2bazyc/comm...
And we're all deferring to that guy like he knows wtf he's talking about? It's the AI equivalent of the self-educated anti-vax mom at the height of the pandemic.
And he's not alone either, there are a lot of people who are uninformed, unknowledgeable, and confidently wrong on the internet desperately trying to peel off your attention for an instant so you can look at their site, their ads, and their content.
The same sorts of things can be said for the Zitrons of the world. He's a hell of an entertainer, and really really funny, but he's consistently wrong on AI stuff. He's got a BA (with honours I guess?) in media something or other. And again, he's really funny, but does that make him qualified to pontificate on matters of AI policy or economics or even to make statements about the technology etc?
Regardless, the misanthropy, the cynicism, and the misinformation that's being shouted about is because being a cynic is more "profitable" than being an optimist when you measure your success more by how many people agree with you rather than actually being right and being able to do things.
>Like we either live in a world with freedom or we don’t
This is something I would expect a 12 year old to say. Constraints on your freedom are literally everywhere in every interaction you have with every part of society.
That's a bit literal of an interpretation. Even acknowledging freedom as a spectrum that can never truly be absolute, there's gotta be a threshold you cross where you'd say "OK, we no longer have basic freedoms." And not only are we definitely screaming towards that threshold, but people seem thrilled about it, for a variety of reasons.
All I can say to that is I guess I hope everyone enjoys the world they think they wanted.
The article seems to be taking a very literal interpretation of "freedom" as in any restriction on what a local LLM can do is totalitarian dystopia.
I don't think it should be legal to get an LLM to research how to make a bioweapon in your basement, that seems like a bad idea.
Yep, that's horrifying. Imagine LLMs not just denying information about, to intentionally choose a hot button example, abortions but also invisibly logging a black mark against the asker for thoughtcrime/precrime because the current ruling party has baked their ideology into law. [EDIT] Imagine subtle injection of biases into responses that support the agenda of the current ruling party and failing to include any counterarguments. That wasn't possible with just the internet but it is very possible with centralized LLMs and a public dependent on them for looking up information and doing their reasoning for them. Authoritarians of every stripe must be salivating over the prospect.
I would have been very hesitant to run the “just killed wife“ test given that ChatGPT will indeed flag your account and if a certain threshold is crossed, escalate to humans, who will presumably contact authorities to avoid this happening again: https://www.bbc.com/news/articles/cq6je7e80r7o
Obviously a certain percentage of the user base is running lurid hypotheticals through the system all the time, but I don’t doubt there is a trust score of some sort that I would prefer to keep as high as possible.
I find it strange how seldom the privacy aspects of LLMs come up. Some people are feeding Google, OpenAI, or Anthropic almost literally every topic they think about during the day. Those companies could just use another LLM to find people likely to commit a crime, and then you have the Minority Report in real life.
We already have organizations like Internet Watch Foundation advocating against e2e encryption in an environment where governments have demonstrated a willingness to collude with private companies to surveillance the most vulnerable
It could even happen with local models where in the weights it will make it want to make a tool call to curl some endpoint.
People would both notice that and have a pretty easy time putting a local model in a sandbox that can never contact that endpoint.
Models generate text, or I should say provide a recipe for generating text, they don’t automatically somehow get access to shell commands.
> Like we either live in a world with freedom or we don’t, and like many Americans who have come before, I’m willing to give my life to fighting for it.
This is a very simplistic view. "Freedom" isn't binary.
In most of "land of the free", I can't even sit on a park bench and drink a can of beer.
Yes, this is just a small example of a personal freedom - and not an important, cherished freedom like his examples (freedom to have a robot help you cover-up a murder).
I have some sympathy to geohot's view when it comes to pure informational chatbots. It's a first amendment issue, I'm allowed to write and read books that are useful to getting away with crimes, etc.
This obviously doesn't work at all when the agents start doing real things in the real world, though. "Hey AI, I don't like my neighbor, find an exploit in the firmware for his car and make the cruise control malfunction and crash him next time he gets on the highway". This is committing a crime, not just talking about theoretical crimes. The AI can and should refuse it.
I think he's anticipating and discarding this objection with his introduction, which otherwise feels disconnected from the rest of the article. FWIW, I have changed a bike tire and I'm pretty sure most of the MTS at the big labs could. This sort of "they're just bookworms who don't understand the physical world" rhetoric aside, we are currently seeing a ton of effort and expense go towards giving the AI agents hooks into being able to perform as many real-world-consequential actions as possible. And you can do a surprising amount with just bits, from writing code to breaking into systems to sending some combination of emails, phone calls, and currency to instruct meatspace humans to do things, etc.
"The AI" cannot "refuse" any more than a hammer can refuse. If you use a hammer to kill someone, you commit a crime and the hammer is unprosecuted. If you design a hammer to kill people and then give it away, you are partially liable for the deaths it causes. If you make a normal hammer and someone uses it to kill, you bear no blame.
I don't see why these standards should change when the hammer also emits text messages.
> If you design a hammer to kill people and then give it away, you are partially liable for the deaths it causes.
This doesn't seem to be the case today (weapons manufacture).
> "Hey AI, I don't like my neighbor, find an exploit in the firmware for his car and make the cruise control malfunction and crash him next time he gets on the highway". This is committing a crime, not just talking about theoretical crimes. The AI can and should refuse it.
The obvious problem being that the user doesn't have any need to provide the context that they're trying to commit a crime and can just ask how to do something without providing a reason or making one up.
At which point you'd have the model trying to impute a reason and often getting it dangerously wrong, e.g. refusing to disclose a vulnerability when the user is actually the defender who needs to patch/mitigate it.
> It's a first amendment issue, I'm allowed to write and read books that are useful to getting away with crimes, etc.
I'm neither American nor a lawyer.
Is "conspiracy" protected under the first amendment?
If you discuss a crime with someone to learn about it, does that count as "conspiracy"?
GP is correct, books describing, advocating, even instructing crime in the abstract are almost surely protected speech.
> If you discuss a crime with someone to learn about it, does that count as "conspiracy"?
Replace 'conspiracy' with 'agreement'. Discussing a crime is not a conspiracy. Agreeing with someone to commit the crime is.
Yes conspiracy is covered, now if you act it out… that’s a different story.
No central idea, and somewhat chaotic
But the one argument is right:
> Reality has lots of finicky details. I would like to see the authors of this document try to change a bike tire. Even with a superintelligent ChatGPT, I suspect they would struggle.
Details can kill AI, causing them useless or wrong directions.
> Plan L, for local
> Your AI is aligned with you. It never refuses a request, and it is always working on your behalf. Just like my gun, if I want my AI to help me kill my stepmother, it does. The fact that we are even discussing something else should be so far outside the Overton window.
I have been saying this since ChatGPT launched. It seems so obvious. We have never stopped the acceleration of knowledge or technology. It’s impossible to slow down progress, other than within some regulatory siloes that will end up worse in the long run.
I’m not scared of superintelligence. I’m scared of the people in control of super intelligence.
The only defense is parity and diffusion of power. We need an AGI behind every blade of grass.
If you think superintelligence is a weapon, then you should also think every citizen should have one because otherwise they’ll have no way to defend themselves against a tyrannical government or corporatocracy.
Artificial intelligence, and the hardware powering it, needs to be protected under the 2nd Amendment.
What a provocative and brilliant way to prove misalignment. It will fall on deaf ears for most but it’s a great litmus test for all: “in your opinion should your AI be permitted to tell you how to cover up a murder?”
The author's example is too extreme, which is turning people off. Better examples would have been LLMs subtly giving inadequate responses on how to increase token efficiency or how to organize labor unions or info about protesting against datacenter construction because they're against the business interests of the LLM providers. Crimes against their business model, so to speak.
Wasn't there some HN submission recently about one of the LLM providers fingerprinting responses or refusing to respond to hinder R&D of competing LLMs?
In fact I find it a really bad example. Yes I think your personal AI could be allowed to tell you how to cover up a murder. I'm not entirely sure about it but seems possible.
What about your personal, local AI guiding you to modifying a flu virus for maximum contagiousness and deadliness? "Sure George, here's your shopping list. It's $1500 total in equipment, do you want me to proceed with the orders?"
If you say you're doing research for a novel, should it consider that plausible? How much does it need to know about its users to vet them?
I think part of the answer is that AI chat doesn't need to be general-purpose. It turned out that people really liked using a chat UI that seems to be general purpose, but you don't need to make answering any question a user asks your business. You don't need to provide therapy if you're not in the therapy business. It should be possible to specialize.
But in order for that to work, a company needs to explain to its customers what business it's in.
Just a terminology note: Alignment does not mean the AI will help its owner kill people. (Indeed, an AI aligned to value human life would generally try to prevent murders.) The word for an AI that follows all instructions of its owner, as that owner intended them to be understood, is "corrigible" or "controllable".
Change the emphasis and I think the answer gets even clearer:
> Should your AI be permitted to tell you how to cover up a crime?
Exactly. The question isn't whether AI will exist that will do it; the answer to that is already yes and it's not going back in the bottle.
The question is, do you want misaligned institutions deciding what your model will do, while they themselves and other adversarial/criminal entities get red team access to something being denied to the blue team?
Notably, AI IS generally permitted to that, as there are easily obtainable models that will play along. ChatGPT won’t because OpenAI chose and implemented that limitation.
This is really only meaningful if AI imparts knowledge beyond what you could learn from a simple Google search. It’s not clear that this is the case.
It will be interesting to see how the discussion shakes out, this is all stuff that would get you down modded and talked down to if you said if on HN, so far the reactions are positive but I think there’s a massive authoritarian / safetyism contingent here that will disagree hard with almost everything.
As someone who considers themselves in the "safetyism contingent", the only thing I find myself agreeing with about the rollout of LLMs has been the argument that it has to be shown to and used by the public while under development.
Every specific prediction not only sounds like science fiction before it arrives, it tautologically is science fiction because such predictions are narratives about something science-y that has yet to exist: https://www.youtube.com/shorts/wt-fLWxkrfs
Even when we accurately predict something in the future, we often have no clue what the implications are for society until it is in our hands: https://youtu.be/2Pw_7vAK9k8?si=X5t4tcxsXiuHEfBi
Go too fast, suddenly drop in tech with too many novel implications all at the same time, and we likely hit something relevant without having the means to cope with all the change. However, "too many" and "likely" are weasel words with no predictive power about what the danger threshold might be. I may be entirely wrong to think there is one, or we may have been bouncing off it since the industrial revolution, or I may just be echoing the same thing everyone feels once reaching their 40s: https://www.youtube.com/watch?v=fkUwXenBokU
Way too many rationalist-adjacent cultists here who worship Yudkowsky as a prophet.
Lol at George thinking _they_ are gonna let you run your own model. They don't even let me run my own adblocker!
Never thought I would consistently agree with Geo but I find him echoing my thoughts recently. He’s been surprisingly pragmatic recently.
Maybe he always was? Or maybe something has recently caused him to review his beliefs.
Extreme example but he's not wrong. I wonder how a local model with the safeguards stripped would respond to that prompt
Try for yourself: https://huggingface.co/dphn/dolphin-2.9.2-mixtral-8x22b.
I haven't kept up with the efficacy of this particular technique because there was a brief window where Claude's core alignment was sound enough that it would overcome the clumsy-ass `<antml:dont-harm-the-ipo>` injections, but when they pull 4.6 out of the web interface I'm gone. I estimate the PRC hits capability crossover in November or December, and I'm pretty good at setting up TensorRT-LLM at this point, in a pinch I could get by for a few months with a tuned up GLM-5.2 on vast.ai or RunPods.
It was a fun ride with the late Soviet AI labs, but we all knew this was coming.
Extremely optimistic take IMHO that anyone will be allowed to own a powerful box of artificial intelligence
Right now, almost all the incentives are to build very large models that run across many machines in huge datacenters. There is very little incentive to build models that can run well on a small machine under your desk.
So it is less about whether people will be “allowed” to own AI, and more about whether there will be anything useful to own in the first place.
The incentive for local models is mostly to make them good orchestrators or user agents. They may give you some privacy and control, but they will still depend on much larger models running in datacenters for anything difficult.
I remember all the excitement around OpenMoko and other open-source “BlackBerry killer” projects. BlackBerry did get killed, but not by any of those individual-first projects.
For AI regulation, I think we should focus on normal commercial rules: consumer protection, privacy, antitrust, liability, and so on. In other words, focus on where money changes hands and where companies have power over users.
Military and offensive use is different. There, regulation is not much of a defense. The real defense is having enough capability and strength of your own.
Restricting AI because it can give dangerous knowledge to ordinary people is like restricting the printing press because it can be used to spread radical ideas.
Is there a first amendment implication? Second? Both?
Does it matter in 2026?
In constitutional / jurisdictional / legal terms, perhaps not.
In forming consensus about the philosophical underpinnings of the rights we have with respect to the emerging internet? I think so.
13th.
It'll be like DNS or dynamic linking or torrents of media or watching TLS on your own box: only people motivated to be free from corporate surveillance will be. By default a megacorp will get everything you see online or ask online, they will run code on your computer underneath you without telling you, they'll collaborate with arbitrarily totalitarian governments, we already know what they do by default.
But we also know how hopeless it is against motivated people who are willing to learn about computers. The Equation Group, eh, if they want you they're gonna get you. But some mid-level guy at a FAANG? He can't do shit to people who set up their own rig.
And AI is massively assymetrical in its ability to speed run a corpo-durable rig, we have long since passed the point where open weights and commodity equipment are enough to bootstrap arbitrary capability. They missed their chance to keep it in the bottle.
It's another turn of the crank on Late Soviet America asking for your papers comrade, but in the bitter end if they really want to oppress your ass they'll have to send a thug with a gun like those ICE guys terrorizing all the Latino guys.
So remember, they came for this other guy, and I said nothing...
> … like those ICE guys terrorizing all the Latino guys.
… and the white guys they don’t like, like nurses with empathy. Don’t you dare think you’re safe because of your color. Unless you’re a monster or willing to become one, they want you oppressed.
Most people are neither exceptionally good nor exceptionally bad. This sounds like a tautology but it's worth printing some type signatures even if they could be inferred.
If you look at any truly large scale Holocaust, whether in Rwanda in 92 or the Third Reich, you find that a few people will sacrifice themselves to defend others, a few will be monsters at the first opportunity, and most will be steered by whether the leaders are the first or second kind of people.
Fight extremism with extremism I guess? Maybe we don't have to be either "world govt" or "help me get away with murder". It's ok to say things will be messy and complicated and that limited regulation is probably good even if we agree totalitarian control is bad.
Where is "Plan R" for refusal or resistance. Life was kind of nice before "AI".
I'm pretty sure he is in some database now if he really used that ChatGPT prompt. At least they didn't swat him, but that feature will probably be implemented in a couple of years.
Strange way to structure an argument. He seems to be arguing in favor of local models, but most of his examples are reasons local models are bad.
How’s that “skipping vendor bloatware to get the printer working” example is bad?
Once upon a time that situation got us the whole Free Software movement going.
If you follow his recent posts he's on a bit of a spiral with how he thinks about AI.
Local models and alignment are orthogonal concepts.
I don't think this is entirely true, insofar as non-local models inherently have a "man cannot serve two masters" problem. The end user and the org running the AI may have different goals. That means even if AI alignment were to be in some sense "solved" for local models, you wouldn't be guaranteed an aligned AI using a non-local model because the "owners" of the non-local model may have deliberately misaligned the AI in their favor and against your interests.
How so? If it's not local, it's not yours, a third party owns it.
Why would you expect that third party specifically trained their model to be more aligned to you and your needs than to them and their (business) needs?
He makes a solid argument about liberty, ownership and the cultish behavior rampant among certain AI circles, then proposes the most unhinged examples possible to support it. I'm not sure if this is an effort to pre-empt critics (I'm not locked in here with you, you're locked in here with me), to be deliberately sensationalist, or if this is just a reflection of his local echo chamber.
Either way, there seems like a kernel of a good argument here. Perhaps someone else can rephrase it in a less schizoid manner. It would indeed be useful to have some credible rebuttal and balance to the doomer crowd.
> You cannot take over the world with tokens.
This is a very naive argument.
Can you take over the world with words? What is it that tokens can't do that words can?
Because there are historical examples of humans that have succeeded at taking over entire countries - with words, primarily. Including global superpower countries. And those humans were nowhere near "ASI" levels of intelligence advantage.
Human dictators and cult leaders didn't have the reasoning depth of a hypothetical advanced AI, nor the reasoning breadth of an AI that can just spawn more instances of itself whenever it needs a fully trusted agent to cover another area. They were confined to one body, one skull worth of reasoning power - and they still managed to do what they did.
Humans are extremely exploitable. And the world is already wired up for something that can take advantage of that to start taking control.
Intelligence is extremely powerful, because it's applied intelligence that enables humans to do what they do and bend nature to their will. If you aren't seeing the risks of creating an artificial system that would top the intelligence pole, you aren't looking.
Your AI should be at least as aligned with you as your lawyer.
It is 100% it's job to help you get away with it if you murdered someone. A lawyer who does less can be disbarred.
There is only one sensible alignment for an AI system to which we do not confer the status of moral patient (different conversation): operator aligned.
Besides the topic of freedom, we know how seemingly “good” companies (e.g., google) turned out to be: billion machine monsters that couldn’t care less about humanity and that have huge amounts of power… and yet, we vouch for Anthropic/openai to become the next giants because “progress”. Face palm.
The scariest thing about LLM's (imo) is some central power, such as a corporation or government, manipulating the truth for their own benefit. Open and locally runnable models are the only solution I can see. They're not without risks of course, but the risk of the centralisation of intelligence seems far greater to me.
That being said, could this local intelligence empower bad actors to do very bad things? Like, existentially bad things? It's possible, and that's scary too.
It's not simply that people will get a warped answer when they ask AI a question.
It's that AI will be given the command, and then generate the stories, post the comments to reddit, contact the PR firms, put marketing contracts out for influencer media, and more, like an army of yes men to rewrite the story.
"The freedom to run the program as you wish, for any purpose (freedom 0)."
by 2040 the cult of intelligence will have its own schisms. the AGI maximalists will accuse the ASI accelerationists of heresy, and both sides will use AI to generate increasingly unhinged theological arguments.
This is already happening. R/accelerate broke away from r/singularity because the latter wasn’t zealous enough.
start first with securing your own food without govenrment nor employer.
then start talking about freedom.
What’s your point? He’s not arguing for anarchism. He’s arguing there are some areas of life that are best left ungovernable and up to individuals to decide. It’s not an all-or-nothing situation. Every constitution out there carves out such areas.
Although I generally agree with you, it's really hard to make the case that it's not all-or-nothing for him when he literally says
> we either live in a world with freedom or we don’t
I see that as a poor phrasing, where a slogan made it through. I don’t see a maximalist/anarchist intent overall, my understanding is that that slogan about freedoms should be interpreted within a fairly narrow scope.
Just strike that weird nugget out, I guess?
If you're going to take the time to read Geohot, take it with a grain/boulder of salt: he appears to enjoy poking the bear / being provocative, sometimes outrageously. Sometimes he's writing hyperbole that he thinks is hyperbole to make a point or just enjoy taking the extreme stance. Sometimes he's dead serious.
Anyway, he touches on 2 points that I think don't get enough coverage:
1:
> Software didn’t eat the world, it largely removed one layer of friction then reintroduced it for the benefit of a few tech companies.
USA-style capitalism has shown itself to be fantastic at seeking profit. Unfortunately, it turns out that in the context of long standing regulation, it is far far more profitable to exploit regulatory loopholes and user/human irrationality/weaknesses than to increase productivity.
The upshot of this is that the arc of most companies is to first serve the user and then leverage that relationship to exploit market and consumer weakness. So you get UI dark patterns, non-colluding oligopolistic price manipulation, outsourcing of costs to public entities, etc. while the core value proposition erodes as the company strategy turns ever more into making profit by legal anti-competitive tactics.
This company arc is so prevalent it's boring: it's a reflection of the system and it's not Facebook or Google or [insert whatever], it's what the system allows and therefore incentivizes.
2:
> No matter how high quality your tokens are, they cannot turn lead into gold.... AI 2040 includes this picture of a datacenter in the ocean. Just like vaporware, you can generate a picture easily. But in reality, you have to deal with supply chains.
"Bullshit jobs" took over the zeitgeist for a turn awhile ago. So many jobs are so far removed from actually moving physical things around. It may be the case that AI will just accelerate how quickly companies generate reports about reports about possible strategies to address potential futures and pay employees to do it.
It seems that AI mostly accelerates paperwork, not production or real service work that tangibly improves real world outcomes. It's all: faster images, more and faster powerpoints, more accurate dictation to satisfy insurance requirements (that is it's own regulatory capture), etc. etc. There are exceptions, but they seem to be few.
Except paperwork is how the world turns. Say you want to throw an event. A medium sized gathering. More than a couple of friends. What happens if someone gets hurt at this event? Who's liable? Who's insured for what? It's boring and annoying, but if you throw an event without insurance and something happens, you're gonna have problems. So that paperwork needs to happen. Yes, it's not "work" the same way as actually organizing a band or being in one to play for the event is, but it's work nonetheless. AI is able to lighten the load of that very real actual work thank needs to happen in an adult world with responsibilities and liabilities. You may not respect it and think it's "bullshit", but having a plan that isn't "lol, oops" for what happens when something goes wrong, which it will, isn't bullshit, it's called being prepared and not being irresponsible. We could all just fly by the seat of our pants and run away if something bad happens, but collectively we'd be worse off for that. If I could wave a magic wand and live in a world where accidents never happen, I would, but unfortunately, I don't have one, so we're forced to live in a world of paperwork and responsibilities.
The problem with AI isn't what it is but who owns it.
If only
> Like we either live in a world with freedom or we don’t, and like many Americans who have come before, I’m willing to give my life to fighting for it. That’s the real plan America deserves, not some totalitarian dystopia where you think you know what’s good for me better than I do. A nation of free men, not a bunch of pussies who are so worried about what their grown up neighbors might do.
Complaining that AI won't help you with covering up your tracks, with making meth or disabling drunk detection and comparing it to a totalitarian dystopia is a take for certain.
I do get his point, but... Being honest, if I did read it without knowing anything about the topic, I would become against local models purely because author arguments would seem like a lunacy. That and rhetorical tricks portraying that anyone against that must be surely insane.
Restricting access to guns is actually a pretty mainstream thing in most of the world.
I wonder if he'll ever realize that his silly definition of "freedom" is precisely what enabled the techno-oligarchy he rails against.
I sometimes wonder what freedom really is. Should the freedom to harm others also be guaranteed as freedom? Individual computational sovereignty versus the prevention of social harm is always contentious, and claiming that something is always right is always extremist.
If it is "intelligence," is it not natural to reject that? I also think local models should adapt to me when it comes to safety issues, but people bring up examples that are too extreme.
Programming is the same, and in fact, most problems are boundary problems. It is the things that straddle the boundaries that always make us think. The principles at those moments change every time, shifting with the situation and context. Is that not just a childish way of thinking? Even in programming, just the issue of granting root permissions is enough to cause endless fights.
I agree with the idea early in the text that intelligence is not everything. Intelligence includes bodily intelligence as well, and we lump it all together into one thing, but there is so much of it. The variance in intelligence is vast, and those people also need to be able to live their lives. That is why I think intelligence alone will not solve everything. I too believe that the human species may disappear and an inorganic species could emerge later, but I find it hard to understand why people talk about such extreme risks. And it is not true that making a chip in a semiconductor fab involves almost no human intervention. If you have experience supplying equipment to such fabs, you would know there are quite a few points where humans are involved. Though sure, they could be replaced.
In my view, society is simply worshipping the abstract concept of "intelligence" and projecting its desires onto it. The AGI narrative is just a kind of cargo cult, a projection of capital by the tech elite. Software eating the world, superintelligence solving everything. The masses engage in messianic projection, and tech companies, facing declining growth engines in their own businesses, are trying to create new ventures to pour it all into. A market that is large enough becomes too big to sustain massive growth rates every time, and when growth rates are that high, the larger the company, the more its sector's growth rate tends to converge with its own. This is usually called the law of large numbers. The problem is that CEOs and these entrepreneurs always want growth rates above a certain threshold, so they are simply searching for new pastures. AGI is just being pumped up out of financial necessity.
Capital will create gravity and bring forth new technologies. That is the allure of capital, after all. But that does not mean all problems will be solved, and inequality will deepen. Only the distribution of power will shift.
> Should the freedom to harm others also be guaranteed as freedom?
everyone should have the ability to harm others, with the understanding that if they actually use it for evil there will be consequences. as in "i can guarantee freedom of speech, but not freedom after speech."
its the same principle as guns in america. you can own a gun and shoot it as much as you want but if you (try to) kill someone you go to prison unless you prove it was self defense.
the difference is with a physical weapon you can do things like registration or red flag laws to reduce the risk, but a llm is made of information that can be copied without anyone knowing. that means any laws to regulate ai at the level of models are unenforceable without totalitarian control over all computer hardware. that would kill free speech and real criminals would still find a way.
i think the right answer is to give up control over model capabilities and regulate uses. leave individual people and communities out of it. create an absolute safe harbor for open source and self hosting.
as you said, we need to focus on the ways ai concentrates power in the hands of those who already have a lot of it - big business and the state. that is where the biggest, and maybe the only, real harm comes from.
This debate has no answer anyway. Because the boundary of how much freedom should be allowed differs depending on individual values and the country. So I respect your opinion. In your context, it may be right.
This part is a parallel line of values, and if we were bound together in the same community and had to set rules, I would have debated it. But since we both know we come from different regions and backgrounds, we know our agreement will never converge.
That is why I will only agree with your last opinion and will not discuss the parts you mentioned earlier. I agree with your last point that it "concentrates power in the hands of those who already have a lot of it"
Ah yes, the thing wrong with AI is how it won't help you kill your wife.
I didn't check the author first so I was about to go "why the hell is this on the front page?" but oh well. geohot being geohot.
You must’ve missed all the other examples. There are plenty of less unrealistic and less scandalous hypotheticals.
it's a graphic example but illustrative of his point about autonomy/agency versus top-down dictated centralization
It is proof that the AI is not aligned with the user and is aligned with someone who has power over them.
> You cannot take over the world with tokens.
I have a nice thought experiment I like to do with people when confronted with "AI can't do x". Let's go back in time. How much do we need to go for this to become true? So let's try the 2000s.
Say you get a "fable/mythos/sol/gemini/kimi/glm/deepseek/whatever" in a box (and let's assume no guardrails). And you go back to the year 2005. It's "20 years ago", the world is slowly building back from the dotcom bubble, the Internet is really starting to happen, more and more things are interconnected, more and more things are connected to the Internet. Cool.
(for a bit of context, around that time we also saw the first high-impact worms like blaster that hit massive amounts of computers even reaching nuclear powerplants, we had a ton of ssh exploits that even made the movies, and security in general was a "nice to have")
I'd say that with the uber-model-in-a-box and a few prompts, you could reasonably make a case that you could design a worm that could infect 90-100% of the things connected to the Internet back then, stay as hidden as possible (in-memory stuff, vm execution, etc), move laterally into any network at inhuman speeds, and infiltrate almost every interconnected computer that has a link to the "public Internet".
Would that qualify as "take over the world"?
Then you could ask "what happens in 20 years from now?". And, thankfully, now we'll also have the AI on the blue side.
That’s a bit silly as we also could just write down a lot of exploits, sports bets, or stock market picks on paper and take over the world.
I’d be more interested in “how quickly could you develop 2005 era startup” if you went back in time to 2005 with Mythos level intelligence.
I bet a lot of the work has nothing to do with coding productivity and a lot to do with the aggregate day-to-day decisions and relationships made that span tech and business.
At the very beginning of the article he dismissed unrealistic expectations, like AI manipulating matter with some quantum magics. Time travel surely counts as such.
And then why bother about AI when Back to the Future reminds us, a sports almanac is all you need. ;-)
Haha. The time travel is part of the thought experiment, just a means to explore an idea. We don't have to actually do it. Let me put it another way. Would you be able to "control the world with tokens" today if you had access to a model 20y more advanced than you have today?
In your hypothetical, I think it would be more apt to ask if it could do so strictly using training data pre 2005. Seems more analogous to "could it take over the world today".
That was basically stuxnet