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.
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.
If put AI in a sandbox that can't write to disk, can't call OS API, can't use a shell, can't search the internet to ground results, can't download files off the internet, etc it will be much less useful than an AI without such a sandbox. The incentives for users to not use a sandbox in order to get the most utility, but this also means that there is danger from AI misalignment. And just because you are running it locally that doesn't mean there is not a misalignment risk. And just because people notice, that doesn't mean people will stop using models. Take a look at how much outrage people have had over the overt censoring of Anthropic's Fable model, but at the same time they continue to pay Anthropic and use Fable.
You don't have to prevent it from accessing the internet in general to route all requests to a specific company's servers to a fake endpoint that never sends the data to the company and replies with whatever response you want it to get.
The firewall would come with the third party open source software being used to run the model and it would have a list of what to block based on observing what that model tries to do that none of the users actually wants it to.
The model itself just generates text, but the harness has done more than that since 2023 and is part of the reason on why AI is better today than it was years ago. I do not believe that the general public will want to give all that up just to improve privacy and will end up asking sensitive questions to agents that have access to sending a signal to the outside world.
If for example Microsoft had a local Copilot model integrated with Windows and people noticed this happening. Do you think all Windows users would just stop using Copilot. Or would you expect the internet to get upset, privacy conscious users switching to something else, and for the majority of Copilot users to remain using it.
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
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.
And prove you're not a child by handing over your ID or selfie to a company that Values Your Privacy according to their price list.
Microsoft will snitch your Windows install GUID when you connect through a VPN.
If you make your connection too anonymous, Google's recaptcha will not let your through until you use your mobile phone to prove your humanity, which knows your phone number (which in many countries can't be obtained anonymously, you burner-phone terrorist).
I’m consistently amazed that people will watch what’s happening with controlling governments around the world and still advocate for surrendering more power to governments.
There’s some deep cognitive dissonance where people dislike what they see governments doing, but think that their ideal imagined laws will only be used by some other imagined government that doesn’t have the challenges of real governments.
It all starts to feel like a side discussion about a total fantasy world where the government is authoritarian, but benevolent and correct.
You don’t write the laws and regulations for a perfectly faultless government that doesn’t exist. You write them for the governments as they do exist, which means they must be written first in a way that makes them resilient to bad governments abusing them.
Every time someone proposes one of these regulations where we control speech or force government centralized ID checking on to social websites, I wish they would pause and imagine how a law might be misused if someone they didn’t like got elected. We have years of evidence of government officials declaring things “domestic terrorism” to use justify using laws and powers for almost any purpose they want. Why would anyone think this wouldn’t happen if governments were given control of LLMs, too?
> You don’t write the laws and regulations for a perfectly faultless government that doesn’t exist. You write them for the governments as they do exist, which means they must be written first in a way that makes them resilient to bad governments abusing them.
The answer isn't "Don't regulate anything!". The answer needs to be an approach to regulation that restrains the government from over-reach.
Authoritarian and fascists governments throughout history have taken advantage of the desire for safety and predictability to rise to power. They never let an emergency or concern go to waste to expand their powers, which almost never retract once the initial societal panic is gone.
Ok, but this is so abstract that it fails to connect to the issue at hand. Ancient Athenians would have been quite mystified by the idea that certain metals are so dangerous that nobody must be allowed to own even a small piece of them; we don’t restrict uranium because our democracy is more tyrannical than theirs, we do it because we’ve discovered important facts about nuclear physics they didn’t know.
I think it would be quite sad if the mechanics of AI end up being such that people can’t be trusted with powerful compute, but sometimes sad things are true.
> Ok, but this is so abstract that it fails to connect to the issue at hand. Ancient Athenians…
There is actually a lot to learn from history on this topic. There is a reason why most successful governments are founded with documents outlining the rights that people have and limiting the powers of the government.
Your comparison to radioactivity doesn’t hold because you’re trying to compare physical harms to information freedom. Once you go down the path of outlawing access to general knowledge unless the government decides you’re worthy of being able to access the knowledge you’re just signing up for government control of knowledge.
We’re not talking about restricting uranium or nuclear bombs in this thread. We’re talking about government control over words (or weights and tokens). This almost turning into a new “think of the children” where we’re asked to consider surrendering access to knowledge and freedom to learn because that knowledge might harm someone. If you aren’t thinking about how inviting the government to become the arbiter of safe knowledge access would go wrong you can look to many of the other governments throughout history that have restricted knowledge access in the name of protecting the people.
Again, you're going very deep into incomplete abstractions. If I say the words "100% peanut free!", even though I know my food in fact contains peanuts, I'm going to get into serious trouble and it will do me no good to object that it's tyrannical to punish me for words. Perhaps there's some philosophical sense in which the words were not harmful themselves, but the action of stating them causes harm in obvious and predictable ways, and any functional society has to prevent such things.
> The answer needs to be an approach to regulation that restrains the government from over-reach.
This is restating the ‘who watches the watchmen?’ problem, which by definition has no solution. The government writes the laws, no one can stop it from writing overreaching laws, unless there is no separation from the concept of government and citizens themselves.
> This is restating the ‘who watches the watchmen?’ problem, which by definition has no solution. The government writes the laws, no one can stop it from writing overreaching laws, unless there is no separation from the concept of government and citizens themselves.
I have been pondering about this issue due to the ongoing fall from grace of EU with ChatControl debacle. I think the good approach can be to enshrine a strong and modern protections in a constitution-like foundation that requires some very high majority threshold to change. US constitution (especially the original one practically forbidding individual income tax) was great for its time and many European states and EU itself intentionally don't have one. UK and EU and two great cautionary tales of what happens when a group of bureaucrats can just decide that their subjects don't actually have any rights to communicate online in private or post politically-undesirable opinions anonymously on Internet.
I think the government fears are deserved. But equally corporations pose the generally the same risks. As corporations having increasing power over governments, they in many countries even write the laws. We have many examples of this in New Zealand. I think shifting the lens to power and class is possibly a more valuable lens.
Governments have an order of magnitude more power than any corporation, as they hold a virtual absolute monopoly on sanctioned violence. That powerful entities can sometimes influence governments should not be conflated with actually holding that power.
Both governments and corporations are joined at the hip and are effectively the same. Just like the Republicans and Democrats, and everything else too, like brands of laundry detergent and food labels. It's all owned by the same group of people with the same goal: enslavement of humanity.
Until they privatize the water supply of entire countries, like Nestlé...
> but also the top 1% is nearly a hundred million people.
And the wealth and control will not be evenly distributed between those 100 million people either. Then you have to start looking at the top 0.1%, top 0.01% etc.
They'd probably be happier if there was twice as much water where they bottle, so they could make the same products but without people being mad at them.
> Then you have to start looking at the top 0.1%, top 0.01% etc.
That's why I cited that billionaires have less than 4% between them. It's an extreme amount for single humans but very far from owning the world.
To reach 10% of the world's wealth, you need half a million people.
The big issue is either we surrender our freedoms to governments, or we depend on corporations that have even less oversight, unless we expect the government to regulate them.
There is no win condition for citizens, either way they are pawns to a larger game than them.
A more sensible heuristic would be to assume both government and corporations to effectively be an equivalent negative force, and the larger they are, the more fucked we are. Thanks to technological progress, we now live in an era of gigantic governments and gigantic corporations.
Rumor is that you do in fact get a black mark [0]… or, conversely, a “vetted” mark [1]… tied back certainly to your payment method and increasingly to your government name [2].
I'm wildly optimistic about the direction this is going to go, but my one dystopia idea that crops up occasionally is a "smart social credit system" that lives kind of under the hood of society.
Once everything is more or less modulated by AI, suppose the technology gets good enough to identify a potential dissident in grade school. Then to keep them from going into the mountains with their friends and rifles, you make it so getting ahead is hard, but not impossible. They can "make it" - but only just. So they spend their time peddling harder and harder to no avail.
They apply to jobs, but jobs to "good" companies go into the circular file and the HR person never sees them. They are only able to work the most time consuming, frustrating, exhausting jobs, and they always kind of make it, but not really comfortably. They're never too comfortable where they have time to think about what's broken in society, but they're never so uncomfortable that they have nothing left to live for.
Historically speaking, isn't the core of most revolutions the product of "elite overproduction" -- educated young people expecting to secure a high status position who are instead shunted into lower prospects? Seems like trapping ideological agitators in near-precarity with no way out would be feeding into that pressure.
yeah, from everything I've read this appears to be the case? But like, if you had some hyper intelligent robot that was getting John Q Revolution a raise and a corner office whenever he started to get dissatisfied maybe you could balance it out?
Personally, I think it's likely that we end up with some post-material scarcity sort of thing and the need to fly off the handle and rebel declines a bit? Hard to justify overturning the apple cart when everything seems pretty great for most people? But yeah, I could imagine a smart enough set of tools figuring out how to walk the knife's edge beteen elite over production and a nation of serfs.
well, so, you wouldn't want to do that because a lot of people have opposing views, and if they're just shut out of the economy, you couldn't use their labor. A truly efficient and nefarious robot overlord would be fine with everyone not liking it, but use it's superior predicting powers to just eliminate the people who are likely to cause a problem and dissuade them specifically.
Like, everyone else applying to Mego-Corp gets their resume reviewed, but John Q Revolution's resume doesn't attach to the application or whatever.
They don’t have to be authoritarians. Equally bad would be a group of people getting control of them thinking that they were doing all this for our own good.
> Of all tyrannies, a tyranny sincerely exercised for the good of its victims may be the most oppressive. It would be better to live under robber barons than under omnipotent moral busybodies. The robber baron's cruelty may sometimes sleep, his cupidity may at some point be satiated; but those who torment us for our own good will torment us without end for they do so with the approval of their own conscience.
We have trackers everywhere now. If you gave access to that data, it seems very possible identify every person who watches certain categories of YouTube videos or follows certain groups of people on social media and mark them according to political criteria.
It's a bit ironic that everyone writes about "surveillance capitalism", but there is still no connection made to actual surveillance...
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...
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.
How will the government prevent you from owning it realistically? They could put some barriers in front of buying pre-made ones, but gamer CPUs and GPUs can run LLMs just fine, even if trailing data-center models in performance. It is inevitable that whichever SOTA models currently require $$$ data-center hardware will be runnable on mid-range gamer box in 10 years.
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.
> 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.
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.
>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.
Or maybe the people who do so aren't actually atheists. There are plenty of atheists out there who don't do 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.
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!"
One thing I find weird is that despite the stock photo of the Terminator's shiny endoskeleton holding a gun being used to illustrate so many AI news stories, now we have ChatGPT people keep saying "how can AI possibly hurt us?"
We've even had the dichotomy of some Doctor Who episodes where the self-driving car crashes itself to kill the occupant to silence them, against repeated real-life news stories about self-driving cars killing people, and yet the connection isn't getting made that software controls hardware.
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.
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!
I'm curious how the value of your website would be determined in your model of the future and why an AI (at least one of the cloud provider models) wouldn't just persist your data after accessing it once. In other words, the way I am understanding it is that your data has a package value. An AI, if it wanted to access the entire contents, pays the whole package value. And every time it does this, you get paid. So, to really benefit, your data needs some sort of minimum value that is accessed at an appropriate frequency.
Also, in this model Cloudflare basically becomes the centralized gate keeper of you getting paid. And to me, this kind of sucks. It's not a very productive line of thought on my end -- I could see something like this happening as the internet trends toward centralization over time -- but it just feels bad.
Thank you for reading and just one idea from one human. But our content keeps Facebook, Google and the digital world spinning. Further in our daily lives we all create content which parts of it or a lot of it is content that Facebook, Google, etc make multiples of billions off of. Why are we not making that money off it ourselves is my thinking. It might not be a lot of money but it would be one way of many that humans get paid for keeping AI relevant.
Though this human would be fine if the universe just unplugged AI and I could go back to working as a well paid UX Engineer/Designer where jobs were plentiful and it was easy to get a job (not have to go through 15 interviews). Thanks AI, but as a startup person I am deep into vibe coding now all my crazy ideas and without having to lean on others (so far). That's pretty cool!
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 if 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".
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.
Disagree, words drift but they still mean things. Nobody building accounting software in the 90s called it AI (or even an expert systems, barring the most inventive salescritters).
Besides what kind of slippery slope argument is this? All software is AI, so if you do harm with software, AI is a harmful technology? Seems like a not very helpful model of the world.
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 ...
Your theory is pretty silly. If the outcome is doom, nobody cares whether you were right, because humans are dead or irrelevant. The asymmetry in the reward function isn't there.
> they things they believe in are unlikely to happen.
Sorry, but most times I've talked to someone who says this (about AI completely replacing humanity), they don't have anything to say about why it's unlikely to happen other than:
- "well, it's just such an extreme outcome, it must be improbable"
- "humanity has survived near scrapes with extinction before"/"all the previous doomsday predictions have been wrong"
There is a cliche that all teenagers deep down believe that they are invincible. It seems to me that humanity is still a teenager in this respect: We don't take seriously the possibility of our own extinction. While one might think that the invention of nuclear weapons would serve as a wake-up call, if anything it has done the opposite.
I'm willing to hear arguments besides the two above, if you have them. (And to be clear, being replaced by AI doesn't necessarily mean being replaced by LLMs in particular. They are a relatively new development.)
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. ;-)
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".
No, training data has nothing to do with this. My point was that "you can't take over the world with tokens" is not a valid counterpoint in the doomer's position. Because their view is that singularity can happen. And when that happens things become murky. And 20 years of progress can go in 2 months "real time". Or two weeks. You get the point.
(I don't share their view, btw. I was just trying to say that the argument of tokens can't affect the real world is not a good one, if you steelman their maximalist position)
> 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.
By most mutually agreed upon metrics, USA hasn't been at the tippity top for a while.
He speaks of an AI that cuts through the popups which personally if im going to follow in his futuristic thinking ... i don't think we will be dealing with screens as much and or even going to websites. Rather just texting and talking to AI and it on demand showing us visual representations per the topic at hand. Whether its showing the visuals on your smart glasses, an AI phone where all the action happens on the lock screen, a digital picture frame (those Internet connected ones took almost 20 years to finally become pervasive in households), TVs and any other place visual information can appear.
I sorta gather (just a hunch) this is the type of device Open AI is working on.
Though if there are no websites anymore 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. Without us it's irrelevant... humanity and AI should both thrive together!
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.
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.
IMHO, freedom is not one thing or even a value. It's a concept that applies to a lot of stuff, speech, thought, association, and tool use included. In practice, one has some freedoms but not some others. I fail to see any need to have them all to be able to discuss about one in particular.
blurring lines of definitions, I see. you can call a potato a freedom, if you wish. this is just counter-productive. arguing about definitions is dangeours slippery slope.
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 crime?”
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.
> 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.
I was under the impression that the lack of specialization is an aspect of the models themselves, not merely the UI or harnesses.
For example, the recent OpenAI success in a mathematics proof was accomplished with a general purpose model [0].
> The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular.
Once the model is trained on general purpose data, specialization is just another kind of guardrail, as vulnerable to “jailbreaking” or prompt injection as any other content-based restriction.
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?
The extremeness is the point. He’s not trying to make an argument over what’s acceptable and what’s not, which just would devolve into pedantic bickering anyway. He’s saying there’s nothing so heinous that an LLM aligned with its user shouldn’t do it.
It can seem unpleasant, it’s meant to, but likewise with other tools, putting controls on them because of some hypothetical bad thing a person could do is stupid and not aligned with the user or compatible with free society.
This is a fundamental problem in politics in my opinion. Everybody has different values and points of view so they may see the same thing completely differently. Its what makes humans great but also frustrating. geohots example will work for some, others will be convinced of the opposite opinion, and others will just let is pass them by and not understand the implications. The same could be applied to your examples to a different degree.
Politicians try to use different examples for the audience they are speaking to.
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.
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".
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?"
Why not? There are perfectly pro-social reasons a person might ask such a question. And a bio-terrorist would have many other avenues to answer these questions.
The concept of a bioweapon attack is commonly cited but in fact there are many other defenses against actually developing and deploying such a weapon. It's commonly portrayed as easily achievable but is in fact very, very far from it. Moteover, although there is no evidence that AI makes it easier in any way, somehow this is constantly invoked as an example of a thought crime that must be forestalled at all costs.
In the AI era people seem to want to redefine the criminal act from "doing the thing" to "asking or investigating how to do the thing". Not a mindset I share.
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.
> 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"
> 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.
Am I missing something here? Bike tires are pretty much the easiest tires to change. You don’t even really need irons and it involves a few very basic steps.
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.
Not sure if you are talking about extremist actions or just rhetoric but if so, the numbers aren't there. Violence isn't the answer as the government has gotten better at stomping down any resistance with each passing generation. Some say Kent State killed the resistance mentality among boomers. Their kids started with that watered down mentality and a more emboldened government went further to erode it further. I don't really know about this for sure but it seems this is the case. I hope to see how it all plays out in the coming decades though.
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
>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.
IMHO, one should be able to research absolutely anything, for any reason. “Forbidden knowledge” shouldn’t exist as a category, freedom of thought should be protected, even if that thought is intended to cause harm.
What shouldn’t be possible (and obviously not legal as well) is building that bioweapon.
I got bad news for you - GLM 5.2 runs on basement-level hardware and supports 1 million token contexts for bioweapon research. There are already people halfway finished with novel bioweapons research as we speak! A man in Tuscon is designing the next sarin gas! A 12 year old in Stockholm is perfecting his prompt for the Fart Virus! Humanity is doomed.
Well, we're doomed once they iron out all the kinks, get their supply chain set up, hire professionals to assemble their device, find a clandestine cell of conspirators to deploy the thing... then it's P(doom) all the way to the bank. AI is the red mercury of the 21st century, a surefire shortcut to escaping the law and building your terrorist superweapon/time machine of choice.
I just don't subscribe to the belief that knowledge itself is dangerous. Well, that's a bit of a cop-out; a lot of things aren't inherently dangerous but that doesn't mean they must never ever be regulated, but regulating knowledge and information under the belief that it is essentially inherently dangerous to have access to it is truly as cut-and-dry as it comes. We already know where that ends.
When people say this they’re most likely referring to negative liberty - freedom from external coercion and interference. This is not the same as positive liberty - “I can do whatever I want”.
I think it’s self-evident there is a difference between having the “freedom” to punch random people on the sidewalk, and having the freedom to walk down the sidewalk without being randomly punched. Now replace walking down the sidewalk with using an uncensored LLM and it might be clearer what freedom means in this context.
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.
Big AI aren't selling hammers though, they are renting them out. And if you rent a hammer to someone asking how hard to ypi need to hit someone over the head to kill, you are likely to be found liable.
> "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.
This leads to LLMs refusing to do security work for anyone, because it can’t tell if it’s being done for good or evil purposes.
Which is precisely what we’re going through right now with frontier models and it’s terrible.
These proposals always assume some perfect mechanism for identifying the thought crime with triggering on the normal requests. The people who want to commit crimes just persist until they jailbreak the guardrails while the rest of us suffer with denials.
Also if you can’t imagine these guardrails being used by governments to control inconvenient speech, you probably need to think a little harder about the realities of how these will be used.
> This leads to LLMs refusing to do security work for anyone, because it can’t tell if it’s being done for good or evil purposes.
Oh no, of only we had some sort of ideas about automated authentication and authorisation! (If big AI actually make a model that would reliably reject dangerous queries in the first place, they can just use auth tech to serve the current ones to vetted customers)
> If big AI actually make a model that would reliably reject dangerous queries in the first place...
Not happening, for two reasons
1) LLMs mix attacker-controlled data into their instruction stream, and it doesn't look at all like the big LLM providers consider making that impossible to be an actual priority.
2) Intent is what distinguishes "dangerous" queries from prosocial ones. Reliably determining intent is so complex that not even humans can do it.
> they can just use auth tech to serve the current ones to vetted customers
Your solution is to require everyone to surrender their privacy and then allow some company or government decide who is worthy of being allowed to use AI?
It’s crazy how authoritarian, controlling, and anti-privacy the anti-LLM conversation is getting.
I don’t think it’s a privacy issue. All the best auditors collaborate to agree on ethical auditing standards requiring complete disclosure; that’s not a privacy violation, even if it means some businesses end up effectively forced to reveal details they’d prefer to keep private.
Not sure how "some users get special access after special scrutiny" means "all users must submit to scrutiny", obviously you could just perform the scrutiny upon request.
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.
215 comments
[ 2.6 ms ] story [ 38.1 ms ] threadI 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.
Isn't this Dr. Karp's entire rant of the month?
If for example Microsoft had a local Copilot model integrated with Windows and people noticed this happening. Do you think all Windows users would just stop using Copilot. Or would you expect the internet to get upset, privacy conscious users switching to something else, and for the majority of Copilot users to remain using it.
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.
Oh no, I'll have to spin up a new e-mail address...
https://support.claude.com/en/articles/14328960-identity-ver...
> Being responsible with powerful technology starts with knowing who is using it.
Shiver.
Microsoft will snitch your Windows install GUID when you connect through a VPN.
If you make your connection too anonymous, Google's recaptcha will not let your through until you use your mobile phone to prove your humanity, which knows your phone number (which in many countries can't be obtained anonymously, you burner-phone terrorist).
Anecdote: I just created an API access account at openAI, for a business I own, two weeks ago, with no identification check of any kind.
No human name is anywhere on the account.
There’s some deep cognitive dissonance where people dislike what they see governments doing, but think that their ideal imagined laws will only be used by some other imagined government that doesn’t have the challenges of real governments.
It all starts to feel like a side discussion about a total fantasy world where the government is authoritarian, but benevolent and correct.
You don’t write the laws and regulations for a perfectly faultless government that doesn’t exist. You write them for the governments as they do exist, which means they must be written first in a way that makes them resilient to bad governments abusing them.
Every time someone proposes one of these regulations where we control speech or force government centralized ID checking on to social websites, I wish they would pause and imagine how a law might be misused if someone they didn’t like got elected. We have years of evidence of government officials declaring things “domestic terrorism” to use justify using laws and powers for almost any purpose they want. Why would anyone think this wouldn’t happen if governments were given control of LLMs, too?
> You don’t write the laws and regulations for a perfectly faultless government that doesn’t exist. You write them for the governments as they do exist, which means they must be written first in a way that makes them resilient to bad governments abusing them.
The answer isn't "Don't regulate anything!". The answer needs to be an approach to regulation that restrains the government from over-reach.
Authoritarian and fascists governments throughout history have taken advantage of the desire for safety and predictability to rise to power. They never let an emergency or concern go to waste to expand their powers, which almost never retract once the initial societal panic is gone.
I think it would be quite sad if the mechanics of AI end up being such that people can’t be trusted with powerful compute, but sometimes sad things are true.
There is actually a lot to learn from history on this topic. There is a reason why most successful governments are founded with documents outlining the rights that people have and limiting the powers of the government.
Your comparison to radioactivity doesn’t hold because you’re trying to compare physical harms to information freedom. Once you go down the path of outlawing access to general knowledge unless the government decides you’re worthy of being able to access the knowledge you’re just signing up for government control of knowledge.
We’re not talking about restricting uranium or nuclear bombs in this thread. We’re talking about government control over words (or weights and tokens). This almost turning into a new “think of the children” where we’re asked to consider surrendering access to knowledge and freedom to learn because that knowledge might harm someone. If you aren’t thinking about how inviting the government to become the arbiter of safe knowledge access would go wrong you can look to many of the other governments throughout history that have restricted knowledge access in the name of protecting the people.
This is restating the ‘who watches the watchmen?’ problem, which by definition has no solution. The government writes the laws, no one can stop it from writing overreaching laws, unless there is no separation from the concept of government and citizens themselves.
I have been pondering about this issue due to the ongoing fall from grace of EU with ChatControl debacle. I think the good approach can be to enshrine a strong and modern protections in a constitution-like foundation that requires some very high majority threshold to change. US constitution (especially the original one practically forbidding individual income tax) was great for its time and many European states and EU itself intentionally don't have one. UK and EU and two great cautionary tales of what happens when a group of bureaucrats can just decide that their subjects don't actually have any rights to communicate online in private or post politically-undesirable opinions anonymously on Internet.
And billionaires control less than 4% of the world's money. (Let alone already being able to get as many servants as they want.)
There's no elite group that owns everything.
> but also the top 1% is nearly a hundred million people.
And the wealth and control will not be evenly distributed between those 100 million people either. Then you have to start looking at the top 0.1%, top 0.01% etc.
Still motivated by profit and not control.
They'd probably be happier if there was twice as much water where they bottle, so they could make the same products but without people being mad at them.
> Then you have to start looking at the top 0.1%, top 0.01% etc.
That's why I cited that billionaires have less than 4% between them. It's an extreme amount for single humans but very far from owning the world.
To reach 10% of the world's wealth, you need half a million people.
> There's no elite group that owns everything.
That you are aware of. But you're not aware of much.
There is no win condition for citizens, either way they are pawns to a larger game than them.
A more sensible heuristic would be to assume both government and corporations to effectively be an equivalent negative force, and the larger they are, the more fucked we are. Thanks to technological progress, we now live in an era of gigantic governments and gigantic corporations.
[0] e.g. https://docs.bswen.com/blog/2026-03-21-detect-claude-hidden-... , https://www.anthropic.com/news/building-safeguards-for-claud... , https://www.techtimes.com/articles/319415/20260701/claude-co...
[1] e.g. https://support.claude.com/en/articles/14604842-real-time-cy...
[2] https://support.claude.com/en/articles/14328960-identity-ver...
Once everything is more or less modulated by AI, suppose the technology gets good enough to identify a potential dissident in grade school. Then to keep them from going into the mountains with their friends and rifles, you make it so getting ahead is hard, but not impossible. They can "make it" - but only just. So they spend their time peddling harder and harder to no avail.
They apply to jobs, but jobs to "good" companies go into the circular file and the HR person never sees them. They are only able to work the most time consuming, frustrating, exhausting jobs, and they always kind of make it, but not really comfortably. They're never too comfortable where they have time to think about what's broken in society, but they're never so uncomfortable that they have nothing left to live for.
Personally, I think it's likely that we end up with some post-material scarcity sort of thing and the need to fly off the handle and rebel declines a bit? Hard to justify overturning the apple cart when everything seems pretty great for most people? But yeah, I could imagine a smart enough set of tools figuring out how to walk the knife's edge beteen elite over production and a nation of serfs.
Imagine if having opposing views tarnished social credit, making one unemployable…
Like, everyone else applying to Mego-Corp gets their resume reviewed, but John Q Revolution's resume doesn't attach to the application or whatever.
> Of all tyrannies, a tyranny sincerely exercised for the good of its victims may be the most oppressive. It would be better to live under robber barons than under omnipotent moral busybodies. The robber baron's cruelty may sometimes sleep, his cupidity may at some point be satiated; but those who torment us for our own good will torment us without end for they do so with the approval of their own conscience.
- CS Lewis
We have trackers everywhere now. If you gave access to that data, it seems very possible identify every person who watches certain categories of YouTube videos or follows certain groups of people on social media and mark them according to political criteria.
It's a bit ironic that everyone writes about "surveillance capitalism", but there is still no connection made to actual surveillance...
This is what's going on, with varying intensity, with social media algorithms.
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...
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.
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.
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 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.
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.
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.
Or maybe the people who do so aren't actually atheists. There are plenty of atheists out there who don't do that.
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!"
We've even had the dichotomy of some Doctor Who episodes where the self-driving car crashes itself to kill the occupant to silence them, against repeated real-life news stories about self-driving cars killing people, and yet the connection isn't getting made that software controls hardware.
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!
Also, in this model Cloudflare basically becomes the centralized gate keeper of you getting paid. And to me, this kind of sucks. It's not a very productive line of thought on my end -- I could see something like this happening as the internet trends toward centralization over time -- but it just feels bad.
Though this human would be fine if the universe just unplugged AI and I could go back to working as a well paid UX Engineer/Designer where jobs were plentiful and it was easy to get a job (not have to go through 15 interviews). Thanks AI, but as a startup person I am deep into vibe coding now all my crazy ideas and without having to lean on others (so far). That's pretty cool!
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 if it is grounded in realistic physical constraints, timelines, or empirical data. It's all divined a priori from Yudkowsky's ad-hoc "rationalist" principles.
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.
Sounds like corporations can have bad motivations, and can make bad software - no AI required.
All AI *is software*; and by the old definition of AI, almost all modern software *is AI*.
Besides what kind of slippery slope argument is this? All software is AI, so if you do harm with software, AI is a harmful technology? Seems like a not very helpful model of the world.
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 ...
People so ascribe “doooooom” to things that actually aren’t doom but instead are “bad” - but everything is hyperbolic now
Sorry, but most times I've talked to someone who says this (about AI completely replacing humanity), they don't have anything to say about why it's unlikely to happen other than:
- "well, it's just such an extreme outcome, it must be improbable"
- "humanity has survived near scrapes with extinction before"/"all the previous doomsday predictions have been wrong"
There is a cliche that all teenagers deep down believe that they are invincible. It seems to me that humanity is still a teenager in this respect: We don't take seriously the possibility of our own extinction. While one might think that the invention of nuclear weapons would serve as a wake-up call, if anything it has done the opposite.
I'm willing to hear arguments besides the two above, if you have them. (And to be clear, being replaced by AI doesn't necessarily mean being replaced by LLMs in particular. They are a relatively new development.)
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.
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.
And then why bother about AI when Back to the Future reminds us, a sports almanac is all you need. ;-)
(I don't share their view, btw. I was just trying to say that the argument of tokens can't affect the real world is not a good one, if you steelman their maximalist position)
By most mutually agreed upon metrics, USA hasn't been at the tippity top for a while.
https://ourworldindata.org/grapher/freedom-of-expression-ind...
I sorta gather (just a hunch) this is the type of device Open AI is working on.
Though if there are no websites anymore 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. Without us it's irrelevant... humanity and AI should both thrive together!
then start talking about freedom.
> we either live in a world with freedom or we don’t
Just strike that weird nugget out, I guess?
it is all wannabe lemon-stand freedom.
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.
I was under the impression that the lack of specialization is an aspect of the models themselves, not merely the UI or harnesses.
For example, the recent OpenAI success in a mathematics proof was accomplished with a general purpose model [0].
> The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular.
Once the model is trained on general purpose data, specialization is just another kind of guardrail, as vulnerable to “jailbreaking” or prompt injection as any other content-based restriction.
[0] https://openai.com/index/model-disproves-discrete-geometry-c...
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?
It can seem unpleasant, it’s meant to, but likewise with other tools, putting controls on them because of some hypothetical bad thing a person could do is stupid and not aligned with the user or compatible with free society.
Politicians try to use different examples for the audience they are speaking to.
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?"
The concept of a bioweapon attack is commonly cited but in fact there are many other defenses against actually developing and deploying such a weapon. It's commonly portrayed as easily achievable but is in fact very, very far from it. Moteover, although there is no evidence that AI makes it easier in any way, somehow this is constantly invoked as an example of a thought crime that must be forestalled at all costs.
In the AI era people seem to want to redefine the criminal act from "doing the thing" to "asking or investigating how to do the thing". Not a mindset I share.
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.
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.
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 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 wonder if he'll ever realize that his silly definition of "freedom" is precisely what enabled the techno-oligarchy he rails against.
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.
And even without such a difficult tire, imo most people would struggle if you’d ask them to change a tire of a bike, especially the back wheel.
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
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.
All I can say to that is I guess I hope everyone enjoys the world they think they wanted.
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.
IMHO, one should be able to research absolutely anything, for any reason. “Forbidden knowledge” shouldn’t exist as a category, freedom of thought should be protected, even if that thought is intended to cause harm.
What shouldn’t be possible (and obviously not legal as well) is building that bioweapon.
Well, we're doomed once they iron out all the kinks, get their supply chain set up, hire professionals to assemble their device, find a clandestine cell of conspirators to deploy the thing... then it's P(doom) all the way to the bank. AI is the red mercury of the 21st century, a surefire shortcut to escaping the law and building your terrorist superweapon/time machine of choice.
I think it’s self-evident there is a difference between having the “freedom” to punch random people on the sidewalk, and having the freedom to walk down the sidewalk without being randomly punched. Now replace walking down the sidewalk with using an uncensored LLM and it might be clearer what freedom means in this context.
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.
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"?
Replace 'conspiracy' with 'agreement'. Discussing a crime is not a conspiracy. Agreeing with someone to commit the crime is.
I don't see why these standards should change when the hammer also emits text messages.
This doesn't seem to be the case today (weapons manufacture).
These things can and do say “sorry I won’t help you with that” based on the nature of your request. What do you call that if not a refusal?
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.
This leads to LLMs refusing to do security work for anyone, because it can’t tell if it’s being done for good or evil purposes.
Which is precisely what we’re going through right now with frontier models and it’s terrible.
These proposals always assume some perfect mechanism for identifying the thought crime with triggering on the normal requests. The people who want to commit crimes just persist until they jailbreak the guardrails while the rest of us suffer with denials.
Also if you can’t imagine these guardrails being used by governments to control inconvenient speech, you probably need to think a little harder about the realities of how these will be used.
Oh no, of only we had some sort of ideas about automated authentication and authorisation! (If big AI actually make a model that would reliably reject dangerous queries in the first place, they can just use auth tech to serve the current ones to vetted customers)
Not happening, for two reasons
1) LLMs mix attacker-controlled data into their instruction stream, and it doesn't look at all like the big LLM providers consider making that impossible to be an actual priority.
2) Intent is what distinguishes "dangerous" queries from prosocial ones. Reliably determining intent is so complex that not even humans can do it.
Your solution is to require everyone to surrender their privacy and then allow some company or government decide who is worthy of being allowed to use AI?
It’s crazy how authoritarian, controlling, and anti-privacy the anti-LLM conversation is getting.
terrible is a bit far, but pretty hilarious for sure