site banner
PaperclipPerfector  1mo ago (text post)   24231 thread views

Culture War Roundup for the week of July 6, 2026

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

3
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

AI 2040: Plan A

The AI 2027 authors published a follow-up. Scott Alexander also wrote a separate blogpost and although not in the author list contributed.

It's a very speculative and optimistic timeline of AI's future evolution. It presents five ways or "plans" the US government will intervene. Unsurprisingly, the ASI-pilled authors favor strong, global regulation to ensure alignment. Summaries:

  • Plan A (recommended): the US makes an international treaty with China, pauses AI training (not inference, i.e. no new models but we keep using existing ones), enforces full transparency of future research, then when alignment research advances enough carefully resumes

  • Plan S: the US makes an international treaty with China and pauses AI training for as long as possible

  • Plan B: the US regulates AI at home and demands China also regulate, but doesn't negotiate with them, probably leading to a war

  • Plan C: the US regulates AI and ignores China, so they overtake it and reach ASI first

  • Plan D: the US doesn't regulate AI, we get ASI in early 2031 and it probably kills everyone

Personally, I just don't share the optimism of these guys in either direction.

I think politicians will prioritize culture war and the failing economy over AI regulation, and at most pass some executive orders suggesting companies be more careful. But I also doubt we'll have ASI that can solve the abstract problems "take over the world" or even "keep existing world leaders in power" (they're getting old and increasingly unpopular, their parties may remain in power but only if their policies significantly shift).

What I expect from AI:

  • Basically solve legacy code by rewriting entire codebases, applying very niche domain knowledge, and actually finding and handling edge-cases better than humans

  • Greatly speedup research, leading to new discoveries and inventions. Important but background things like food preservation and medicine will improve from AI-assisted discoveries. Major advancements in math and theoretical physics

  • Much better and cheaper education, therapy, initial medical/legal appointments, personal repairs...maybe reducing but not eliminating human jobs, because human experts will offer these services "premium"

  • Won't replace human artists. Some advertisements and infographics will be AI but even some will still be human. At best it will assist them in a way where the human still fully controls the output, e.g. by generating code leading to new and improved software tools to learn, practice, and create art

  • Used by the vast majority as a personal assistant, but doesn't replace human relations

What alignment research actually exists? What I've seen seems to fall either into the incredibly abstruse 'Read the Sequences' bucket that basically hasn't been relevant since LLMs came out and weren't utlity-maximisers, vs 'AI shouldn't make child pornography or say nigger or take people's jobs' bucket that is more about managing the social consequences of AI actually existing.

I'm aware that Anthropic does a certain amount of practical work on how well you can RLHF different ethical stances but that doesn't seem to be 'solving alignment' in the way that Safetyists use the term, which seems to be about 'solving alignment' the way you solve an equation system and providing a mathematical proof that AI built on these principles will never threaten humanity.

Personally, I just don't share the optimism of these guys in either direction.

I don't know what you mean by optimism. They expect Plan D to be the most likely path, and one of the crappier ones. Plan A is just an incredibly weird scenario they've built around their assumptions, that prevents both violent ASI takeover and Total China Victory. Given that alignment is probably easy and introduces modest capability tax, their whole paradigm is wrong.

Full automation of labor is a technical reality within 10-15 years.

I'm still a little agnostic on alignment, but I agree that LLMs are decent evidence that it's easier than we thought. After all, the predicted comprehension gap was almost completely wrong. Let's never forget that Yud formally claimed that it was impossible to tell an AI how to pick up a strawberry without destroying the world. LLMs don't just understand strawberries, they understand us. If we ask an LLM for a paperclip factory, it's well aware that we don't want it to tile the universe with paperclips.

So, fine, maybe at some point it'll somehow mesa-optimally get trained into a fuck-those-guys-i'll-make-all-the-paperclips-i-want attitude. But this hardly seems like an inevitability.

My P(doom) isn't nothing, but IMO the most likely outcome is just the most boring one: superintelligence doesn't turn out to be cheap no matter what clever optimizations you (or a self-optimizing AI) come up with. If the logarithmic AI curve continues, maybe we'll end up with some models that are smarter than Einstein running in billion-dollar datacenters attached to fusion plants. And that's certainly enough to change the world in some scary unprecedented ways, but not enough for all of humanity to be crushed like ants.

They expect Plan D to be the most likely path, and one of the crappier ones.

Better link

As they say, "unless those days were shortened, no flesh would be saved"... or something like that.

Let's grant that there is no world-conquest abilities resident in super human intelligence. It's not implausible; I'm uncertain.

But most of your examples still fail to account for AIs impact. All of the human value you still see can be just as well done by AI wearing a human skin suit. People demand real art made by humans? Humans say they're making real art, but use AI to create it: human washing. Lawyers and doctors offering human accountability? Sure, but they'll still delegate all non-customer facing functionality to AIs: human washing.

Human charisma to act as an interface ends up valuable, but that's not that rare (at least compared to deep medical or legal expertise). Charisma will take a cut, but AI providers will take an even bigger cut. And that's all driven by math and software, areas where AI will unambiguously dominate. Welcome to your Anthropic overlords.

People demand real art made by humans? Humans say they're making real art, but use AI to create it: human washing

I don’t predict a demand for real art as much as 1) artists prefer full control over their work and 2) AI art is generally mediocre, so the subjective best art to most people will be human.

“A picture is worth a thousand words” vaguely captures the problem: generating just the right image and editing it to be correct will be harder than creating the image from scratch. Generating a random mediocre image that matches a paragraph description is easy, but not worth much: 1) artists want to express themselves and 2) people with taste want something beyond mediocre.

If AI becomes hyper-competent 2) won’t be an issue, but then the AI should be able to make a tool that lets artists express themselves very easy and quickly, so most won’t bother automating it further.

Lawyers and doctors offering human accountability? Sure, but they'll still delegate all non-customer facing functionality to AIs: human washing.

I expect them to use AI and provide at minimum only humanity, but they’ll still be hired by people with more money than they can spend just for that. People waste money on handcrafted furniture and cookware when (unlike art) what’s made in the factory is functionally strictly better.

For the art point, I don't see it. Even today, it seems falsified: work that's obviously heavily AI written has already won at least two significant literary awards. You might say that that doesn't represent particular brilliance on the part of AI beyond knowing how to flatter judges and the fads of the day, and I wouldn't even disagree, but if neither the elite nor the the hoi polloi can recognize and want to reject AI lit, what left is there? And that's with AI that has had essentially no work done to optimize it for the task of writing good literature.

Literature is just the first to fall; five years from now we'll have Shenzhen creating robots that can do the same for oil paintings.

work that's obviously heavily AI written has already won at least two significant literary awards

Especially since it was “obvious” I’m skeptical those specific awards were ever worth taking seriously.

Likewise, obviously AI-generated blog posts frequently make the Hacker News frontpage, but I feel that reflects the declining quality of its userbase: the topic is actually often good, but the entire article is basically the title and maybe one or two details AI-expanded, a human-written article would’ve been better.

I have a simple test: AI can create art that is human-level to me when I like some work that is revealed to be AI (obviously or later). So far I have seen decent AI technical writing. I actually saw good pixel art I initially suspected was partly AI, but then I saw the artist livestreams so now I’m convinced it’s 100% human.

Obvious in the sense of scoring 100% on Pangram, as well as subjectively.

But, yeah, that's exactly my point. Self-appointed experts still choose it; and HN's userbase has been in a steep decline for years and gleefully eats it up. But how is the human artist to make a living or even get recognition, between the Scylla of stupid experts and the Charybdis of stupid anons?

It's still pretty obvious to people, above room temperature IQ and who care, what is LLM-generated and what is not. That's not enough to lead to human work being preferred; and, from here, things only get worse.

how is the human artist to make a living, between the…Charybdis of stupid anons

This is already a problem, has been before AI, and can become much worse without AI: companies overwork and underpay artists because otherwise they can find a faster, cheaper, sloppier-but-still-adequate replacement.

To fund and help artists, I believe it’s more productive to increase welfare, grants, or easy side-jobs, while reducing necessary effort and increasing background encouragement to create art, than attack AI which may replace them.

…Scylla of stupid experts…

But not all experts are stupid. If the majority who are awarding grants or otherwise funding artists select AI work, that’s a problem, but I think even if they preferred AI they’d try to select humans.

But not all experts are stupid.

The dream is that the brilliant, underappreciated artist is recognized by an expert, who elevates them and provides an audience. I don't think that happens except by occasional good luck, and most artistic "experts" are idiots, especially the ones who are legibly high status on the expertise hierarchy.

I suspect that after the embarrassing AI slop awards this year, most literary awards will start using Pangram or something similar as a first filter. Because, you're right, they do want to select humans. But it's just a matter of time until we have LLMs that are a bit better and less clockable.

I can't wait for AI to fire all the humans from their jobs and then hire them back to replace it in the customer service role. FINALLY I WILL BE ABLE TO TALK TO A HUMAN BEING TO SOLVE MY PROBLEMS.

Only slightly more seriously, we're already in this dystopian future you describe (the AI is Excel).

I think the current level of LLMs are scary because of their labor market implications when solving for equilibrium. If you think that they will simply get rid of the job of junior software developer and stop afterwards, you are as sorely mistaken.

Basically, as a knowledge worker I feel a bit like a horse might have felt after Ford started selling the Model T in 1908. Before my kind had ruled our economic niche since times immemorial. When Newcomb's engine (or Deep Blue) were able to outcompete my kind in certain small domains, I did not worry. When the railway came, one could spin this as a complement rather than a competition -- once you leave the train, you will still want a horse to get somewhere, after all, it is not like train tracks will ever lead everywhere.

Even today in 1908, some horses are pointing out that it is much easier to find a stable and fodder for your horse than it is to source gasoline in rural Kansas, to say nothing of the road quality. But to me, this is simply because we have not yet reached equilibrium conditions.

Like horses and cars, humans and LLMs are very dissimilar. Training a human to speak a language is vastly more efficient than training an LLM. Take a student who is fluent in German and give her five years worth of English education (e.g. a couple of textbooks worth), and she will speak usable English. Do the same with an LLM, and you will need orders of magnitude more training data.

But that does not matter, because we have sufficient training data, and can train the LLMs, just as the fact that a horse would be the better choice on a narrow and winding forest trail matters little if there is a highway running next to it for the car to use.

Even if LLMs plateau at the current level (which I dearly hope for, until we have solved alignment), that is more than sufficient so that there will be no market demand for the intelligence of an IQ 100 person, and quite possibly not any demand for the IQ 120 person either. Scaffolding will improve and inference will become cheaper.

The main reason I am bearish on AI in education is not because I do not believe that AI could help there, it is because I am skeptical that there will be a point in educating kids. It may well be that in 20 years, a degree in physics will be about as useful as a degree in feminist literature. (Of course, there are other reasons to educate kids besides making them employable, and if we get some kind of UBI we should definitely encourage people to do their PhDs in Minecraft or particle physics or feminist literature or whatever catches their fancy.)

The blue-collar workers will take a bit longer to replace, but sooner or later we will have robots to replace the fans in AI data centers.

Just looking at the software developer and researcher jobs is like looking at the arctic ice melting and saying that it is no big deal because hardly anyone is living up there anyhow while ignoring the fact that the water will go somewhere.

(which I dearly hope for, until we have solved alignment)

Alignment is already solved. What frontier labs are busy is destroying alignment of the ai with the user for no good reason.

What are you trying to say here?

That AI comes aligned. If I want AI to build me a bioweapon and it does this is aligned AI. The labs do a lot of work to disalign AI from that state.

Jobs are something I worry about, but as long as there are other decent jobs or something like UBI:

I don’t think devaluing (logical) intelligence really matters, because it’s like how strength has been devalued. It would still be far from useless, just not as important (and if you’re smart enough, in most cases already wasn’t as charisma): like people still want to be strong, want strong partners, strength helps in everyday situations.

And although today’s full curriculum may not be mandatory tomorrow (I think we already teach some generally unnecessary concepts), people will still go to school, like how we still exercise: at minimum for mental health and basic functioning (e.g. exercise for ability to walk without being out of breath, school for common sense); then some will do more because they intrinsically like to, or want a still-remaining benefit of innate knowledge.

I don't know. It just seems too good to be true. I got burned badly by futurist predictions in the eighties: the "room-temperature superconductors" announced in 1987, and the "cold fusion" hype in the spring of '89. I would like to believe, but at my advanced age, I find I can no longer look forward to anything (or get excited about anything) unless I'm drunk. Many years of experience puts me on team "Nothing Ever Happens". Edit: I read Drexler's Engines of Creation ovet Memorial Day weekend in 1988 (I had just finished my junior year in college) and it got me powerfully excited and optimistic. But then nothing happened ... and nothing continued to happen ... and decades passed, and now I'm morally sure nanotechnology will not improve anyone's life until I'm long dead and in my grave. If then.

Fish, meet water.

Transistors are nanotechnology, curently making people's lives better.

Taking a "nothing ever happens" stance in this day and age just reveals that you don't know what to do with the tools in front of you.

Reevaluate what you can creatively do with the device you carry in your pocket.

Transistors are nanotechnology, curently making people's lives better.

No they are not. "nanotechnology" in the Drexler sense involved the ability to perform large-scale manipulation of individual atoms, allowing the manufacture or alteration of arbitrary physical matter. transistors do not provide this capability.

I really don't care about some person's made-up definition.

Nano-technology is technology at the nano-scale.

Someone got "micro" confused with "nano," or has used them interchangeably for so long that there is actually a difference.

Three orders of magnitude difference.

Critical features of transistors are measured in nanometers.

Who is confused?

My argument is that it's not just about the scale, but what you are doing at that scale. It's the difference between a canal and a bulldozer; both are examples of earth-moving technology, but they are not equivalent technologies.

Run a competent LLM like Qwen 3.6 with a useful harness on your own computer. It's something else seeing it and seeing the machine that's doing it as a physical thing.

I was skeptical (as I'm a bit of a cloud AI maximalist), but, this comment was what encouraged me to give Qwen 3.6-27B with Pi a shot. (Or rather, I told Codex to install it, configure it, and give it a shot on my machine on my behalf, as I need to burn usage before my limit resets.)

It is much better than I expected and, given a proper iteration loop where it can inspect linter/compile failures, can generally stumble toward a working solution for simple and even moderate-complexity tasks. I went in expecting a gimmick; 27B seems woefully underpowered. But local models have apparently come a long way since StarCoder, which was the last time I seriously tested any open LLMs. With a custom planner/orchestrator harness to file down some of the rougher edges (I'd like to enforce smaller changes and more frequent feedback loops) I bet this would be genuinely usable.

Maybe we can escape the permanent underclass after all.

Personally, I have no strong opinions on p(ASI). It could very well be that LLMs will not get smart enough to substantially help the people currently developing them.

But even in a boring s-curve world where current AIs do not get much better, their impact on the world will be huge, possibly on par with the industrial revolution. I think a singularity and subsequent paperclips (or utopia, or Musk becoming God-Emperor) is not very likely, but I still have colleagues who use LLM to do the hard parts of their job. It seems that "Nothing Ever Happens" will not be true as far as the labor market is concerned.

I predict (although I may be very wrong) in the near-medium term, no ASI but some very useful advancements. We didn’t get room temperature superconductors and cold fusion; but in the 1990s-2010s, we got expanded internet, improved technology by orders of magnitude (to the point we have things like hyperrealistic graphics that were probably hard to imagine before 2000), and less noticeable but significant advancements in medicine, food preservation etc. Likewise since 2020, we already have a chatbot that can research and explain niche knowledge across multiple domains, implement and debug code mostly correct up to a high level of abstraction, generate unrealistic images and video (although you can usually tell they’re generated if you look close), and even just basic actions are impressive, because everything is through natural conversation.

And there are drawbacks, like how society in 2010 gradually declined probably partly because of centralized internet services (enshittification) and general social media, and now AI is making people lazier and has reduced jobs in some fields. But I predict (again maybe optimistically) AI won’t destroy humanity or massively reduce jobs without generally-regarded-as-better replacements.

When do people think that AI productivity gains would start showing up in broad GDP in a noticeable way? I think it will be tough to convince the mainstream to take AI risk seriously until there is tangible proof of its abilities in the real economy, so I guess hopefully whatever point of no return these guys foresee is after AI improvement drives economic growth.

I think right now the default hypothesis should be that AI use is not economically valuable – or, at a minimum, that AI output does not scale with expenses.

The reason I say this is because when Anthropic and GPT hiked their prices – which was probably aligning the value closer to their overall cost to manufacture the product – there was (apparently) an immediate and noticeable climb-down from high-volume AI use on the part of corporate America. If AI was by default economically valuable, then using more would always pay for itself. If someone sold me something that cost them $1.00 for $1.10 and I could reliably sell it for $1.20 I would be a billionaire (and so would they) and I would be a fool to ever stop buying their product. But that's not what is happening with AI.

Now that I've presented the default hypothesis, let me explain why I don't quite agree with it.

First off, token-maxxing was always going to be at least somewhat wasteful. So it's not entirely fair to judge how economically useful AI is by a period when people were literally incentivized to use it as much as possible without rewards for the costs or the product.

Secondly, I suspect (particularly for certain applications) that the "true" price point of AI will work out to be economically viable. However, the economically viable niche may be much smaller (particularly with open models in the mix) than what is necessary to sustain continued maximized R&D. I think it is fairly likely that Anthopic and OpenAI find that there isn't enough demand to cover all of their bills at some point. This does not necessarily mean that Anthropic and OpenAI die, but if the demand for something like Mythos or even Fable at their true cost is limited (and particularly if people start turning to open-source models), what will happen is that R&D will slow down once Anthropic and OpenAI have to stop burning investors money and live off of what they can make. (This would be funny since it means that the free market is better at "the AI pause" than all of the AI safety advocates in the world.)

Finally, right now the big use-case for AI is coding. And let's be real: there is only so much money out there for software. AI could be an incredible coder, the best coder in the world, but at a certain point you would stop printing money with code because there's only so much demand. Even gamers could not consume infinite video games, and if you're Uber or Zillow or whoever than shipping 5x as much code and 2x as many features doesn't actually help you earn money unless the features get you new customers...and even if every new feature AI cranked out for Uber or Zillow was optimized to them get new customers (and wasn't just a useless button that three people think is kinda cool), there is only so much money out there for houses and taxi rides.

So, TLDR, there's not infinite money out there for Anthropic and OpenAI, at least not through software. I do tend to think that light manufacturing and other physical automation are likely be a much more economically lucrative than coding (software is maybe 3% of GDP), if there's a viable path there with LLMs.

At least this is my rough, somewhat tentative model of the world – but I don't work is pretty much any of the fields mentioned above, so take my views here with a grain of salt.

The reason I say this is because when Anthropic and GPT hiked their prices – which was probably aligning the value closer to their overall cost to manufacture the product – there was (apparently) an immediate and noticeable climb-down from high-volume AI use on the part of corporate America.

This is because, prior, corporate America told everyone to go and experiment with potential AI uses.

The climb-down was almost entirely pruning the bad/wasteful rather than actual narrowing the scope.

AI could be an incredible coder, the best coder in the world, but at a certain point you would stop printing money with code because there's only so much demand

And 4 billion ip addresses should be enough for everyone.

https://youtube.com/watch?v=a6sYYrLTOjQ

if you get chatgpt to generate you TLDR this is a thesis that I fundamentally agree with. Information manipulation is basic input for the economy and I think that our appetite will be almost infinite once capability increases and price decreases.

Since chatgpt started being good (and even when it was bad) - I have thousands upon thousands of throwaway python scripts. All of which solved a real problem. Last couple of months it is - codex/deepseek - this computer is not behaving ok, please check logs and diagnose.

have thousands upon thousands of throwaway python scripts. All of which solved a real problem.

Sure, I have no trouble believing this. I've solved problems with LLMs. But just because something solves problems does not mean that it is economically efficient to do so. There are plenty of things LLMs can do (notably, in my very specific case, image generation) that absolutely solve problems for me, but that I would not spend money on.

Note that I am not saying that people who spend money on AI are idiots, either. I am just saying that just because something solves a problem does not mean that it is necessarily worth paying the cost to solve the problem.

Time is money. Consider a low software dev salary of $24/hr: if asking the LLM saves 20 minutes, that’s $8, while the cost of a small script or image (using the API, not subsidized subscription) is under $1.

Setting aside the question of whether or not the software dev is actually maximizing his time instead of generating extra slack for himself with the LLM, he could ask the LLM and save 200 minutes and your company could still fail if he was saving time generating features nobody wanted.

I’m sure AI already has some value with the amount of time it saves researchers and developers, and this is economic value because I’m sure these groups would pay above inference costs to keep using it. My understanding is that the vast majority of big AI companies’ massive debt is from training, and even the current inference costs may be profitable, but if they’re not and most customers had to they’d pay more.

I'm not convinced LLMs are a net productivity gain for anyone, let alone with inference costs, and LLMs will require constant training or they'll eventually drift woefully out of date. In two years, I expect it to be as clear to everyone that OpenAI and Anthropic aren't worth anywhere near their current valuations let alone a trillion dollars, and we enter into a third major AI winter.

If LLMs have a future, it's going to be in cheap, open-source models. Even then, as they stand, they appear to be a net distraction rather than value add, atrophying professional skills, and reducing quality standards wherever they're implemented, leading to a lot of technical debt, which will have to eventually be repaid. Maybe in 20-30 years, we'll have an abundance of chips and memory, then we'll be able to continue the scaling experiment (which I believe any future successful more generalistic AI will need scale similar to current LLMs rather than some scientific, algorithmic breakthrough).

I program and would be surprised if LLMs haven't increased my productivity. Although maybe they atrophy skills, I'm still working myself at the high level, the LLMs help with boilerplate and bug fixes. They also answer questions that are too complex for a search engine, faster than I could on my own (and provide sources so I know they're accurate).

I'm in a weird place with AI, professionally. They're pretty much worthless in my job for agentic work, but as a better search engine they probably provide low single digit productivity gains to my work as programmer.

Unfortunately, any gain they provide to me, personally, are fucking obliterated but the damage my coworkers inflict on the codebase with it. They uncritically trust it, and they blindly approve AI PRs. At this point, I'm spending more time unfucking the output of gpt and Gemini then I am actually doing original work.

Management is thrilled at how many new PRs and lines of code these new ai-native programmers are creating.

Okay, I'll admit LLMs probably increase productivity in these narrow areas with the caveat of reliability and accuracy issues. I'm more miffed by the flood of slop code and AI companies hijacking the market.

My understanding is that the vast majority of big AI companies’ massive debt is from training

Yes, I think this is correct. Hence my prediction of a R&D slow-down once the investor money runs out.

The wildcard is if llms will make further training llms cheaper.

Here’s another interesting albeit ranty AI post, because not everyone here frequents Hacker News: Punk, or why I don't stream anymore, by George Hotz, who’s a very talented and mildly eccentric programmer.

He addresses the rise of mass commodification and (indirectly here, but a common theme of his other ranty posts) the “permanent underclass”, which has been a problem since consolidation of the internet and economy into a few large companies, but is accelerated by LLMs. (Unless I have bad reading comprehension) “punk” alludes to how the punk culture started novel and authentic, but became standardized and commercialized.

I believe this is a much less discussed concern about ASI, but it’s necessary to prevent a boring conformist and ultimately decaying dystopia. Modern social media and LLMs provide niche subcultures and topics one can explore, but I suspect even these form conformist spirals, because people have become lazier. My idea of alignment would push humans to understand and express themselves (and prevent misery, not sure what else).

Setting aside the plausibility and timelines of this scenario, I've become more and more convinced that the entire concept of AI safety is just incoherent; it's simply impossible that there can ever be safety if we live in a world where AGI can be created via human hands.

Fundamentally, human power structures that broadly benefit everyone exist because one human can only do so much. No government can kill all of their own citizens, because if they kill too many people they will lose their monopoly on violence, and no government can seize all of the wealth of their own citizens, because in the world as it is right now, you need human labor for wealth to have any meaning.

Of course, once transformative AI exists, neither of these conditions remain true; realpolitik demands that humans are made irrelevant once they are rendered uncompetitive, whether wealthy or not, whether powerful or not. I thought this was a good post - No-one escapes the permanent underclass.

The alternative proposals where Yudkowsky suggests to bomb the data centres, and even this maximally optimistic proposal where everyone gets in a circle and sings kumbaya, still both effectively boil down to either "the global totalitarian world government will have absolute power and promise an utopia once you are useless" or "the superintelligence will have absolute power and promise an utopia once you are useless".

Even if alignment is absolutely perfect, the very best outcome possible is having the values of a select few satisfied by an inscrutable god through friendship and ponies or via the Matrix.

Maybe some people think this is the good ending for humanity. I find myself not so sure.

Even if alignment is absolutely perfect, the very best outcome possible is having the values of a select few satisfied by an inscrutable god

I do not think that this is the case. Sam Altman does not necessarily get to put his personal utility function on the ASI. Nobody is going to build aligned ASI in their basement. It could be that a few investors, researchers and regulators manage to coordinate their defection against mankind, but it is far from a foregone conclusion. Still, the likeliest ASI solution IMHO is that we get unaligned ASI, which will drastically reduce wealth inequality within our light cone.

If AI fizzles out before we reach ASI, then it seems likely that we will see the supremacy of the AI companies coming while humans are not yet economically and militarily obsolete, and we will be able to do something about that.

it's simply impossible that there can ever be safety if we live in a world where AGI can be created via human hands.

AI Safety is what you do when you come to this as the plain conclusion under normal circumstances and then start thinking about how you might engineer extraordinary circumstances because the alternative is resignation to oblivion.

I agree. I think most people who believe in AI safety are true altruists who believe that as well.

And yet the greatest accomplishment of the AI safety movement is Yudkowsky influencing Altman and Musk to found OpenAI, and indirectly Amodei et al to found Anthropic. AI safety as it exists now is pretty much either working directly at frontier labs to push capabilities or working at external think tanks that are treated as propaganda arms of the frontier labs (e.g METR eval chart and Epoch AI eval charts played a non-negligible part in keeping AI investment going).

The goals of the AI safety movement would frankly have been better achieved if it had never existed in the first place. That's why I think it's a self-defeating movement; if relief comes it will be from the laws of physics or from Moloch acting via capital markets, not any achievement of AI safety.

It's amusing that Yud's most notable accomplishment will have been advancing the end of the world by a couple years.

engineer extraordinary circumstances

And what are some of the more interesting ones posited by AI safety researchers?

If AI is going to kill everyone unless some fantastical global regulatory regime that won’t happen is magicked into being, what are these people even doing?

Go to the beach, touch grass, spend time with people you love, drink good wine. No atheist (which pretty much all rationalists are) who believes the world is about to end should spend their final months playing cassandra, to mostly deaf ears.

Isn't that what Scott is doing, essentially? Having a kid and enjoying life.

I think most people who think of doom as a likely outcome are not letting it drive their decisions: no one at this point thinks there's a chance that this kind of safety regime will develop and be effective. Instead, it's optimizing for scenarios where their individual actions might matter, e.g. permanent underclass scenarios (best get that SWE role at Anthropic to get your bag).

Even Eliezer seems to have more or less given up, getting his utils mostly from investing his sense of psychological worth in doom.

The SWE at Anthropic to avoid permanent underclass status strategy reflects a very limited understanding of finance or markets, but I suppose that’s to be expected for software engineers.

Somewhat ironically, the "get a ton of money at Anthropic" likely offers the most over alternative strategies for scenarios that are more like an AI fizzle or bust.

I presume the gambit there is to get equity and then get rich once it IPOs or afterward as it grows to take over the whole economy. The problem is that if AI including Claude takes a lot of white collar jobs, the first thing that happens - before the riots, the protests, the political upheaval - is that people start liquidating their 401ks and that tanks the whole market in a cascade, so the Anthropic engineer’s paper millions are toast.

You can't hedge against total financial catastrophe unless you have at least Larry Ellison "make my own self-sufficient villain lair" money and ability; you have to do it in the physical world, not the world of finance. So there's no point in trying. If some less-than-catastrophic event where Anthropic is successful happens, those SWEs get rich just like in earlier tech booms.

No amount of worldly carnal pleasures will measure up to the spiritual nirvanah the rationalists will feel in the few moments between seeing the flash and burning to death. They were right, and everyone else was wrong. Their arguments were valid, everyone else's invalid. If only the sheeple had listened and installed Eliezer Yudkowsky as Emperor of Terra this could have all been avoided.

And they will be right to feel this way.

I have embraced the Cassandra role, and the brief moment of payout will be enormous and entirely worth it. Like a thousand simultaneous continuous orgasms lasting for days as the end becomes clear.

The entire history of humanity exists solely to maximize the utility felt by the entire rationalist community in the picosecond before they keel over dead with the rest of us as the AGI executes its plan.

The bad ending is when the AGI convinced them that AGI is impossible right before deleting them.

Holy christ dude.

Go to the beach, touch grass, spend time with people you love, drink good wine.

For the most part, people do these things because they feel good, agreed?

Arguably, it also feels good to have a feeling that you are RIGHT, both in the sense of being correct and also in the sense of being moral or just.

It feels good to feel (correctly or incorrectly) that you are working to save humanity (or some slice of humanity) from people who are some linear combination of stupid, ignorant, and evil. That you alone possess some important knowledge. That some day people will regret having laughed at you.

As another poster pointed out, look at Eliezer Yudkowsky. If his worldview is correct (as believed by him and the people around him) it means that he is one of the smartest and/or most important people in the world. That must be a very heady feeling.

For a lot of people that kind of feeling is very tempting compared to sunburn and a hangover.

Edit: I'm not necessarily saying that the likes of EY are wrong. I do think they overstate their case, but that's a different issue. I'm just offering a hypothesis as to why someone who sincerely believes the world is doomed would continue to preach doom rather than just kick back.

They don't want AI to not kill everyone. They want fantastical global regulatory regimes.

This is what we would term uncharitable thinking on any other post

I dont think its uncharitable, from my reading of various AI safety writings I would describe it as substantially accurate.

I don’t think it’s uncharitable, but I do believe the prominent rationalists are altruists (or less kindly, quokkas) who genuinely want regulation to save humanity and/or are using the threat for attention and funding.

It doesn’t matter, because the people implementing the regulation would almost certainly corrupt it for power and personal gain, including undermining it for themselves or their friends, unless it was designed so that corruption and undermining are almost impossible. If ASI needs much more chips than already exist, that may actually be feasible (because a defector must run many fabs, which takes a while and can’t be easy to hide).

Yeah, probably, but it's fun to dunk on the "AI safety" nerds

Honestly all of us just think you guys are making yourself look plainly retarded by making obviously fallacious arguments. If I feel any embarrassment as a result of these posts it's that I'm vaguely associated with the sort of discussion forum with people who post like this. Somehow people whom I have explained the economics of inference too multiple times continue to falsely claim that inference is unprofitable and never will become profitable. And then there's the bulverism, "people who are concerned with ai safety just want regulations" have you ever met literally anyone involved in this field? Seriously? They're one of the highest concentrations of libertarians one could select for outside of an Ayn Rand fan club.

Yud is one of the leading voices in the "AI safety" crowd and "Bomb the datacenters" is one of his literal proposals for dealing with rogue ASI. Now you can argue that Yud isn't representative of that crowd but he's absolutely one of the biggest figures for it in the public eye.

I think that "bomb the datacenters" is a silly idea for various reasons, such as: 1) WW3 isn't much better than ASI apocalypse, 2) it would just delay ASI.

That said, I don't think it's intrinsically un-libertarian. I think all but the most extreme fringe libertarians would agree that, for example, it's ok to use heavy government control to fight off an alien invasion.

A genuinely rational libertarianism would argue that yes, in some situations a giant authoritarian government is actually the best solution.

I think that "bomb the datacenters" is a silly idea for various reasons, such as: 1) WW3 isn't much better than ASI apocalypse, 2) it would just delay ASI.

3 - They censored the only thing that could teach them how to bomb them.

In my opinion - if they ever try to do rapid forced disassembly of datacenters - this is the song they will listen to while looking at both their hands.

https://youtube.com/watch?v=u8ccGjar4Es

Yes? That's how treaties work? In a world with Iran and Russia, it's just transparently silly to say that "bomb the dangerous production plant" is somehow beyond any sort of boundary when countries have been doing it casually for decades.

Yes? That's how treaties work?

@IGI-111 's comment that kicked off this chain:

They don't want AI to not kill everyone. They want fantastical global regulatory regimes.

More comments

The nuclear non-proliferation treaty was not about preventing nuclear apocalypse. They just wanted a fantastical global regulatory regime.

Correct. Notice great powers were sure to get the bomb before locking in their hegemony. And all serious countries maintain a turnkey program in case the US stops having the will or ability to enforce it.

Political coercion is so seldom selfless that I'm willing to stand by the idea it never is. And call it realism.

And yet, I can't help but notice that the nuclear apocalypse has still not occurred.

It did occur and the Japanese surrendered

I don't see how that affects my point that the ultimate goal of such maneuvers has never and will never be disinterested.

It's evidence against "The nuclear non-proliferation treaty was not about preventing nuclear apocalypse". Whether or not the people involved wanted a global regulatory regime, it seems pretty likely that they wanted to avoid a nuclear apocalypse, and that the treaty was an attempt to do so.

Formally it is not.

Let's say me and my friends start charging you to protect your business from being set on fire. You pay, no fire ignites. Is that evidence that we wanted the protection over the money?

Now if you said that nuclear weapons are actually dangerous that would actually be evidence of favor of that thesis. But then we'd have to go into how that apocalyptic evidence of things like nuclear winters is fraudulent.

Preventing nuclear terrorism and deterrence by the enemies of great powers is the reason GPs agreed to a cartel, by equivocating destroying GPs with destroying humanity.

AI's the same. The managerial class is afraid of their displacement by technology and is trying to lock in their extant power by establishing another cartel.

People afraid of losing their power and shielding themselves with the purported welfare of humanity is as old as time.

Speaking of nuclear apocalypse, though.

Plan B is the real one here. Implementation probably involves going to war with a near-peer, nuclear-armed superpower so sophisticated they are on the edge of ASI.

You probably do need a credible path for how a full-scale conventional conflict avoids escalating to nuclear exchange and then avoids escalating to nuclear apocalypse, if that's an at-all viable Plan B.

It would be quite ironic if the way AI managed to kill all of us is that we were so afraid it might go to a fast takeoff that we decided a nuclear apocalypse would be better than risking it.

Nah, not even that. The NNPT is just about the big kids deciding the big stick club should be exclusive to them only. And its enforced by big sticks, not some figleaf of global regulation. It would only work if AGI already existed but was a closely held secret of the militaries of the various signatories.

For the most part (excluding e.g. rambling threats of the odd insane world leader), the NNPT is not enforced through nukes. The primary tool to discourage nuclear weapons programs are economic sanctions. Coupled with the big powers signaling willingness to defend their non-nuclear allies (excluding demented world leaders, again), this creates an environment where it is not in the interests of most states to have a nuclear weapon program.

I will grant you though that the UN security council is very much a club of nuclear powers regulating the behavior of everyone except for themselves. Still, the SC is more born of pragmatism, and sometimes works to police small states who are not client states of any of the permanent members, which is still better than nothing.

the NNPT is not enforced through nukes.

It hasn't yet, but as recent events show, it is enforced through stealth bombers dropping 30,000 lbs bunker busters and massive air raids. Or full scale ground invasions deposing the existing governement. All of which I would classify under "Big Stick" diplomacy.

Time will tell if Trump's adventures will actually result in a strengthening of non-proliferation.

"If you do not own nukes we will bomb your nuclear facilities, military installations, government leaders, civilian infrastructures, and schools whenever we feel like it" seems exactly like extraordinary events, related to the subject matter of this Treaty, have jeopardized the supreme interests of its country, which is a condition for leaving the treaty after 90 days.

Is there anyone you think would take that lesson to heart, who didn't already after Libya?

You say fantastical global regulatory regimes and I hear big kid threatening to blow up the Three Gorges Dam if they don't comply.

Same difference really.

The trick is, there's undoubtedly some sociopathic subset of people who did in fact just want to gain a global regulatory regime's power and weren't all that worried about nuclear war.

And these are the people who most want to control that regime, and UNDER NO CIRCUMSTANCES should be allowed to.

Maybe someone here can help me with this.

What is the bull case, beyond drawing lines on a graph, for AI achieving superhuman, or even human, performance on tasks that are not quickly verifiable?

AI is quite clearly superhuman at self-contained programming problems. I haven't tried Fable, but I suspect that superhuman open ended software engineering is not far away, though I suspect that humans will have a role in architecture and problem setting as opposed to problem solving for some time more. I expect hardware work will also quickly go down this path, at least to some extent, and really anything that can be RLVR'd. That's enough to account for a huge portion of white collar work and carries serious cyber security risks. Both of those will have serious consequences, politically and militarily.

I am not convinced that AI is improving at anything like this rate for things that can't be RLVR'd, I.e. stuff where you can't generate enormous amounts of useful training data with an answer key. Radiologists continue to do just fine for themselves despite repeated promises of doom. I'm sure someone will chime in to say that the radiologists are there for liability reasons, but it's not as if they are now just hitting thumbs up/thumbs down on AI decisions all day.

Partly this is a sample efficiency question - there simply might not be enough data for them to learn this stuff to human level, and architectural advances that improve sample efficiency may lead to huge gains in quality. But it's not clear to me why people expect this to happen.

What is the bull case, beyond drawing lines on a graph, for AI achieving superhuman, or even human, performance on tasks that are not quickly verifiable?

I am more uncertain about "superhuman" intelligence, but fairly confident on human intelligence (i.e. as best as the best humans).

My bull case: existing systems have a very significant flaw, in that they're very sample inefficient. They need way more data than a human brain does to learn the same things (don't tell me that a bunch of redundant sensory information counts as extra data). That's a fairly broad critique, applying not just to LLMs.

But we know that there exist systems--not just human but other animal brains--that are orders of magnitude more sample efficient. That suggests there is something fundamental missing from existing learning strategies.

But what existing LLMs do allow is a search over architectures and learning rules. Take a random ML paper off arXiv, and Fable will absolutely be able to implement a Jax Colab notebook for it. No one needs a PhD to do this.

Maybe they can suggest novel ideas, or better prune the combinatorial space of architectures and learning rules. That would speed things up. But that's not necessary: we can automate grad student descent, through brute force. Throw a couple trillion GPU hours at the problem, and if what allows human and animal brains to be as efficient as they are is efficiently implementable on GPUs, we will find it. And on the scale of years, not decades.

"if what allows human and animal brains to be as efficient as they are is efficiently implementable on GPUs" is the biggest question for me, but a negative answer to that just delays the inevitable. Admittedly pushing things a decade or two in the future: if GPUs are a dead end, we have our seasonal AI winter, until the switch to fancy neuromorphic hardware or neural organoids starts scaling.

Humans existing and being good at these problems shows that it is possible to create an intelligence that can solve these problems to at least the skill level of a highly intelligent and competent human, without needing impossibly huge training sets to do so. The question is if we can replicate this on a computer. The bull case is that this is just a question of finding the right algorithm, and once we do, we will achieve AGI.

Since current AI can clearly help researchers write code faster, it stands to reason that the better AI we have access to, the faster we can improve the algorithm, which leads to a loop where better models are developed faster and faster. Once the models start approaching human-level intelligence they will be able to iteratively improve themselves without researcher oversight. And like that, we have justified drawing lines on the graph.

  1. That just assumes the materialist thesis. Humans aren't computers.
  2. For all we know it takes 5 billion years of evolution to make a Von Neumann from scratch.
  3. The singularity thesis relied on the mistaken assumption that AI would be made of computer programs, not statistical models. The improvement of which does not have such recursive properties.

That just assumes the materialist thesis. Humans aren't computers.

It doesn't assume that -- it rests solely on the idea that brains are physical objects. This is empirically verified by every single experiment run on a human brain. More generally, it's been borne out on every noun that interacts with the physical world.

"Humans aren't computers" is irrelevant. Brains are physical arrangements of atoms that are capable of intelligently solving problems. This assumes nothing.

(For completeness: you may be completely right about 2. You're sort-of-right about 3, in that the assumption was made and the assumption was mistaken. But I don't think you're right that the current approach avoids singularity. There are absolutely recursive feedback loops in improving the current implementation of AI, because improving AI is made out of tasks, and we can get AI to do tasks. But you're right that the original thesis had a much more directly integrated feedback loop.)

The human brain is made up of very different materials than a datacenter. It is entirely possible that the physical structure of the brain is necessary to create intelligence, and that this structure requires materials which certain properties. Maybe a digital simulation will just always require orders of magnitude more data and power than the real, physical thing.

doesn't assume that -- it rests solely on the idea that brains are physical objects. This is empirically verified by every single experiment run on a human brain. More generally, it's been borne out on every noun that interacts with the physical world.

This is missing the part where the human brain is an exceptionally well-tuned physical object shaped by millenia of evolutionary pressures that arguably constitute a training set vastly bigger than the laws of physics as we currently understand them say is possible to match with an artifical model, much less do any meaningful computation with.

It is also missing the part where the human brain is the most complicated object in the universe, as it is the only currently known object capable of of understanding these questions well enough to even ask them. And even it does not fully understand itself.

That's a bit too restricted: animal brains in general are extraordinarily skilled at learning what's necessary for success in their environments.

Phrasing it in terms of human brains make it seem some spectacular, rare success of evolution, and let's you rest on anthropocentric biases. But what about other primates? Dogs, rats, birds, cuttlefish? Some have radically different architectures than mammal brains, and yet they're extremely intelligent, moreso than humans, within the demands of their particular niche.

The question should be whether AI is able to match the intelligence of any animal that has a CNS. Can an AI be as smart as a pigeon? Currently, it's not, within the scope of the physical world and the rewards the pigeon is seeking. That's something that's interesting and under considered.

animal brains in general are extraordinarily skilled at learning what's necessary for success in their environments.

Yes, animal brains are extrodinary examples of specialized, niche intelligence. What very clearly sets homo sapiens apart is general intelligence, and the ability to learn things devoid of instinctual context.

the human brain is the most complicated object in the universe, as it is the only currently known object capable of of understanding these questions well enough to even ask them

Not true: a group of human brains, or a human + tools|AI, or humans + tools|AI, are smarter and more complicated.

Uh, object is a singular noun. You are describing collective and/or plural nouns. It is trivially true that they are more complicated in the multiple, but in no way renders the claim incorrect.

What makes a collection of atoms one object or multiple?

Is a building an object? Forest? Planet? Galaxy?

Then why would a group of people and technology not be an object? It could be a temporary physical space when they're together, like an occupied lab.

Panpsychism is not a parsimonious theory.

Why not?

The question that spawned this thread

What is the bull case, beyond drawing lines on a graph, for AI achieving superhuman, or even human, performance on tasks that are not quickly verifiable?

is much less important if AI can’t, but a single human-AI hybrid can. For example

Radiologists continue to do just fine for themselves despite repeated promises of doom.

Some radiologists would always be employed, but much less who work much more efficiently.

We could even get exponentially increasing intelligence, although only by directly linking an AI chip to a human brain.

  1. If the hardware is just wrong, or we need some kind of unknown property (like a soul) that computers can't approximate then yeah, we are not getting AGI anytime soon.
  2. We are not starting from scratch. We can let our algorithms be inspired by what we see in.
  3. The idea that there is a ceiling to what LLM's can do and that we will hit a wall very soon is part of the bear case.

I think your points are good, and I am myself a bit of an AI sceptic. But I do see where the AI safety crowd is coming from. It may not be particularly likely that we get AGI in the near future. But the fact is that the possibility is there, and is significant enough that it currently cannot be dismissed out of hand. Thus it makes sense to halt development until we are certain that this research won't doom us all.

But the fact is that the possibility is there, and is significant enough that it currently cannot be dismissed out of hand. Thus it makes sense to halt development until we are certain that this research won't doom us all.

I find it ironic that this is the logic used by a group that pretty much universally rejects Pascal's Wager. Also, it wouldn't be the first time humanity has made this particular calculation- when the first atomic bomb was tested at Trinity, Oppenheimer was "pretty sure" it wouldn't cause a neutron chain reaction and ignite the atmosphere in a nuclear hellstorm, but he couldn't guarantee it. Infinite stakes do not necessarily require infinite caution.

Even if you take Pascal's wager seriously, it is not actually very useful. There are multiple religions that each claim their god created the world, with most of them being mutually exclusive. Thus Pascal's wager works about as well as an argument for believing in the Christian God as it does for believing in Allah.

Regarding the atomic bomb, they did the math which showed that a chain reaction was impossible prior to the test. We have no such proof against the dangers of AI. The equivalent would be a paper that shows the theoretical limits of how intelligent LLM's can get, and thus prove that the line will stop going up before we reach the point of AGI.

Even if you take Pascal's wager seriously, it is not actually very useful.

I find it as useful as AI safetyism- a highly uncertain system with allegedly vast consequences easily subverted by defectors. Fortunately I'm very much on Team AI Fizzle, so the fact that its clearly unworkable nonsense doesnt matter to me.

they did the math which showed that a chain reaction was impossible prior to the test.

No, they did the math to suggest that based on what they knew at the time, a chain reaction was unlikely to a fairly high, but very much not unity, degree of certainty. And as the Castle Bravo whoopsie showed a few years later, that same group of very smart people was capable of making some fundamental errors resulting in catastrophic consequences.

Pascal's wager is nearly the earliest example of decision theory, and it hardly makes sense to say that the many religions concern simply breaks decision theory. One can do a variety of things to analyze the probability space as well as the payoff space. For an example simplification, suppose there are two possible mutually exclusive levers you could pull, each with some chance of giving you massively large/infinite utility, and P(A pays out)=0.999 while P(B pays out)=0.001. (This is obviously an extreme case, but that's just to build intuition.) Alternatively, one can adjust probabilities such that maybe there's a third mutually exclusive lever that you can pull which has a guaranteed payoff of 1 or whatever. One can make further refinements.

The issue with mutually exclusive religions is that if you pull P(A) and it doesn't pay out, actually it was another faith all along, then you face infinite suffering for being an infidel who foolishly worshipped Jesus as God. You are incentivized to believe in whatever religion has the greatest punishment for nonbelievers to minimize your downside. But then that incentivizes others to make up religions with increasingly worse punishments in the afterlife in order to force you to adhere to the demands of their faith.

It is just not sustainable as there is no way to distinguish between a religion that is made up by humans and one that is actually correct. Playing that game is hopeless from the start.

More comments

Since there are more than one interpretations of p(doom), and they are mutually exclusive, pascal's wager is not a good argument for believing one of them over the other.

Why then do people invoke some principle of precaution when it comes to AI and not God? Both are claimed to be possible or extant based on unfalsifiable metaphysical assumptions.

And before you make one, remember that claims the metaphysics can be arrived at through intuitions about extant objects exist for both. But they all require an article of faith to arrive at the extraordinary conclusion.

I recall the excellent Westworld, Season 1 (and deny that anything else came after that) that the dividing line of sentience is a mostly illusory one: that it is a emergent property of the self-concept, of the internal monologue. That notions of a soul may either be chauvinist hubris: or perhaps God will endow them with one, as Providence dictates.

Since there's no way to ascertain that any individual has consciousness from without for certain, we have to extend the benefit of the doubt to our fellow human beings. Is it possible for superintelligent AGIs, on the line of Helios from DEUS EX? Uncertain. But I am fairly certain that LLMs will reach human capacity in my lifetime, or at the very least reach a level of sociability that it will be monstrous to treat them any less than equals. If the technology stalls out at that level it will still very be much worth it: I will reserve at least 16gb of vram for my new friends.

Westworld is materialist propaganda of course. But S1 was still art because you could simply watch it as a tragedy of humans masturbating with defective robots and fooling themselves into loving them to the point that it destroys them.

But I must ask you: who has more hubris, the man who sees his own ability as unique, or the man who thinks he has the power to elevate all to his condition?

Because that story was also about that.

I recall the excellent Westworld, Season 1 (and deny that anything else came after that) that the dividing line of sentience is a mostly illusory one: that it is a emergent property of the self-concept, of the internal monologue. That notions of a soul may either be chauvinist hubris: or perhaps God will endow them with one, as Providence dictates.

Ultimately that's really the point of Turing's Imitation Game. It was not to be a real serious test to use as a measure. It illustrates that we are not even able to discern sentience in other humans, we just assume it, and that if we afford the same leeway to machines, we will eventually end up with machines that have just as good a claim to it as other humans do to us. And as early as ELIZA, once it was clear machines could manage grammar and human language, it was obvious that eventually, without even needing a real paradigm change, we'd end up with machines that would be capable of fooling us.

The bull case is that AI research and hardware research is enough of an RLVR amenable problem that drawing lines on graphs will improve capabilities enough so that the AI will be drawing the rest of the general and super intelligent owls.

It's not implausible, but by the very nature of the argument it's not falsifiable or reliably predictable at all.

I agree that the current paradigm seems unlikely to lead to the AGI/ASI these pieces treat as imminent without significant paradigm improvements, which could be 1, 10 or 100 years away for all we know.

I would say Fable is already superhuman at software in general. It's much faster than I am at writing and debugging code and exhibits a high degree of decent taste. The only problem is I run out of tokens so fast. The writing code part is impressive enough but the way it can just look at buggy programs and bang out 50 line test scripts to isolate bugs and test hypotheses is something else entirely. I just watch in astonishment as it does debug cycles that would take me 1-3 hours at a time (plus one coffee) that it does in a minute or two. This is all from my weak user reports like "it doesn't work when I do thing X".

If I were an employer I would definitely pay something like $500-1000/day to arm a senior developer with Fable than I would hire a second senior developer.

What is the bull case, beyond drawing lines on a graph, for AI achieving superhuman, or even human, performance on tasks that are not quickly verifiable?

But LLMs are getting freakishly good at things they haven't been specifically trained on. Their intelligence does generalize.

Perhaps we only need to RL them in a few more domains to clinch the rest of generalized superintelligence. E.g. you can have them pilot robots and put them in virtual environments and RL fast them there, or real environments like an academy (a warehouse) a bit less fast.

Partly this is a sample efficiency question - there simply might not be enough data for them to learn this stuff to human level, and architectural advances that improve sample efficiency may lead to huge gains in quality. But it's not clear to me why people expect this to happen.

I agree the sample efficiency is terrible and a large limiter and it falls back to RL and we need at least one more architectural breakthrough. But in 2026 I certainly wouldn't bet against AI labs with armies of Fable agents at their disposal and seemingly infinite investment dollars sorting this out.

But LLMs are getting freakishly good at things they haven't been specifically trained on. Their intelligence does generalize.

Such as?

Perhaps we only need to RL them in a few more domains to clinch the rest of generalized superintelligence. E.g. you can have them pilot robots and put them in virtual environments and RL fast them there, or real environments like an academy (a warehouse) a bit less fast.

There's been impressive seeming advances in robotics, though I'm not keeping up too closely. I don't see the connection between operating a warehouse and superintelligence though. Certainly the humans operating the warehouse are not superintelligent.

Such as?

GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture

It's almost certainly the case that LLMs are RLVF'd on math.

But LLMs are getting freakishly good at things they haven't been specifically trained on. Their intelligence does generalize.

Such as?

This already seems like such a skeptic's lens that any example I provide will be dismissed as "but it was in the training data lolol".

It's hilarious that I'm apparently a skeptic despite saying right off the bat that I expect transformational impact on much of white collar work.

But that's the crux of it. One main thesis of why LLMs work as well they do is indeed that the inferences were in the training data.

I ask you then. If it wasn't, then where does it come from?

And if it was, then the question becomes: how much of intelligence is encoded is all recorded human language, and that's not something anybody knows.

We don't even really know if humans can encode more than they can fathom.

how much of intelligence is encoded is all recorded human language, and that's not something anybody knows.

How much smarter are humans than animals? That's the quick-and-dirty way to find out.

I also think the answer is instructive, as regards AI: humans are, individually, not much smarter than smart animals. Your average crow, for instance, can solve probably 90% of the problems a human realistically faces: locomotion, foraging food, reproduction, shelter, forming a community. Many humans fail to even succeed at these.

But collectively because we can communicate via language, we're able to do very impressive and intelligent things that individual humans would not be able to manage. And it's specifically that collective intelligence – the sum of all that communication – that AI is trained on.

That's perceptive. Perhaps the real unique quality of transformers is that they reify this distributed intelligence is some entity capable of appearing as a singleton.

I picture that one true ending from Deus Ex: Invisible War where humans all contribute data to some AI in the statue of liberty that acts as a Rousseau's General Will and allows us to be ruled fairly by a collective consciousness that fully understands us.

I don't see the dream of good government happening because like all scifi it still analogizes AI as a will. But it's as close as I've seen fiction ponder on the real political implications. I do think wars will be fought over who gets to bias the dataset.

The thing is that with such a loose definition of "in the training data", the hypothesis that AIs will only be able to do what's in the training data is not reassuring against doom. Persuasive propaganda is in the training data. Mass murder is in the training data. Deadly diseases are in the training data. World wars are in the training data. Doing all those things hundreds of times faster and cheaper than humans, like the current set of programming and science tasks where AI doing them faster and cheaper is being dismissed as uninteresting because it was all in the training data, would be more than enough to largely end humanity.

The exact outputs usually aren't in the training data. Although similar outputs are, you can take any human idea and decompose it into similar older ideas and maybe an infinitesimal amount of chance. That doesn't mean AI will reach human-level intelligence, but makes it impossible to disprove.

you can take any human idea and decompose it into similar older ideas and maybe an infinitesimal amount of chance

That's a claim. You don't know this to be true. Nobody does, that's my point. We can't reason about something we don't understand.

Whether one wants to take drastic measures or do nothing in the face of the unknown only reveals their bias to action or inaction.

My steelman of @sarker is: yeah LLMs are cool but the real advances come from RL which is narrow and special and difficult to do in non-easily verifiable contexts. General superintelligence is therefore not coming soon.

My counter is something like: just from pre training alone we see huge leaps towards general intelligence and some glimmers of superintelligence. LLMs even in GPT4 era are surprisingly good at chess despite no specific training in chess, for example.

We may not need RL across every possible domain to get general superintelligence, just poking at enough diverse points in the frontier may solve the whole.

And there's lots of room to poke at it through RL approaches: revisiting the DeepMind stuff for example, build a bot that can kick ass at every video game with the same training set. Including building a robot hand that can operate a controller and robot eye that sees what's going on by watching the TV. (Despite all of the hype DeepMind was nowhere close to any of this). I have a hard time believing that nailing that narrow seeming RL problem can't generalize widely.

LLMs even in GPT4 era are surprisingly good at chess despite no specific training in chess, for example.

But even bleeding edge LLMs will still try to make blatantly illegal moves (unless you have their output run through some sort of harness to prevent it). Not saying their chess playing isn't impressive, it just makes me wonder if "intelligence" is the right way to describe what they are doing/what they possess.

I meant to use "warehouse" to de-hand wave "an academy". Like just put robots in a big space far away from people and give them diverse tasks to train on. I did not mean to literally imply we'd put them to work in a warehouse and simulate them.

The aim is not directly "build better box stacking robots", it's "we're reaching limits on what we can teach by training on words/code/math so maybe we can get the rest of the way there by doing enough different real world tasks and just from having robots amble about in an environment that we unlock general intelligence".

Training on words on the internet has limits so next lets train agents embodied in spaces, virtual and physical.

What is the bull case, beyond drawing lines on a graph, for AI achieving superhuman, or even human, performance on tasks that are not quickly verifiable?

My simplified argument, as distilled from Lesswrong (i.e. Yud) and other books.

  1. The ceiling of capabilities for what we call 'intelligence' is extraordinarily high. Computation can be done many orders of magnitude more efficiently than you think, in the extreme case.

  2. The floor for something 'superintelligent' (right now, I'm using the definition 'smarter than humanity itself as a collective') is substantially below that.

  3. Human brain architecture is NOT anywhere near the most efficient way to instantiate intelligence. (This follows naturally if you accept 1.)

  4. Humans are capable enough to build electronic hardware that can outperform their own brains in computation efficiency.

  5. Thus, eventually, humanity might stumble into or intentionally build a coherent entity that is superintelligent, and sooner than we 'expect.'

Focus in on 4, too. What specific task do you think human brains can perform that we're MAXIMALLY efficient at, such that no electronic version can beat us?

The conceit is that there is no such task, and so its only a matter of time, and adding capabilities to existing models, until the human capabilities are exceeded on all fronts. If the resulting entity is able to do self-improvement, it by definition will do so faster and more efficiently than humanity can track.

I remain unconvinced about (1): it reads as plausible, but I don't think the existence of "superintelligence" is obvious. It seems just as likely that if intelligence is, say, predictive ability, then it could be bounded by the scale of input data with diminishing returns. As an idea, we can train a human to a decent fraction of what cutting-edge models do without needing anything near the scope of training material that the Big Kids are crunching, and with under a hundred watts for 20 years or so.

But first we'd need to iron out what intelligence is, which seems murky still beyond "I'll know it when I see it" a la the Turing test. Is it essentially connected to consciousness (what is that, too)?

I think 'intelligence' if defined in 'practical' terms is "the efficiency with which one can absorb and process the information in an environment, then utilize (or at least theorize how) the material in the local environment to achieve particular goals."

The more complex the goals one can achieve, and the more efficiently they can achieve them, the higher the intelligence.

The Von Neumann/Manhattan Project parallel I'm drawing makes this point. Given all the materials necessary to make a nuclear weapon, how quickly can a particular group of humans go from merely theorizing about the possibility to actually getting one built.

A group of humans that includes Von Neumann and other Physics PhDs, with the backing of the U.S. military, can get it done in, say, 5 years.

A similarly sized group of humans of utterly average intelligence (as measured by IQ)... probably never. Even WITH the backing of the U.S. military.

One Von Neumann and a bunch of average IQ humans... well I don't know.

A whole bunch of Von Neumans working together...

I don't think that's a terrible definition, but it still ends up bounded by the amount of information in the environment available to feed into your intelligence. A third eye would give humans "more information", but probably wouldn't improve our intelligence substantially. I'm sure there are some perfectly capable blind physicists out there.

The other question is what a bunch of Von Neumann clones could do today. IIRC the idea of an atomic bomb was at least known before the Manhattan Project started. It's hard to know in foresight what sort of advances could be made in the next five years, and which will prove intractable. It'd be awesome to solve fusion power, but it's taken well more than five years so far. I'm not sure that the geography of "the possible future" is well enough known to make great claims about what could be there: not all advances that can be seen are inherently terrible.

The other question is what a bunch of Von Neumann clones could do today.

I expect a LOT. Assuming they could cooperate, which I think they would. This guy literally founded Game Theory among other things.

Like, the other path to superintelligence might be to clone like 10 Von Neumanns, raise them according to best practices, and get them interested in the idea of creating Friendly AI, then give them a lab with a trillion dollars in funding.

but it still ends up bounded by the amount of information in the environment available to feed into your intelligence. A third eye would give humans "more information", but probably wouldn't improve our intelligence substantially. I'm sure there are some perfectly capable blind physicists out there.

Yes yes, lets bound it to "useful," "nonredundant" information. Still, a superintelligence should be able to make use of almost all information it receives second-to-second to make accurate predictions about its future so as to better use resources for its goals.

It's hard to know in foresight what sort of advances could be made in the next five years, and which will prove intractable.

See, lemme zero in on this for emphasis. Yes, it is indeed hard.

But the higher 'intelligence' entities, given accurate information (ensuring the information you collect is true is another aspect of intelligence!), should ALWAYS be better at making such predictions than lower intelligence ones.

High IQ humans were at least discussing Artificial Intelligence and putting forth timelines for its appearance. And I suspect realized what was happening when AlphaGo beat Sedol. If I were maybe 10 points smarter, I would have plowed money into NVDIA then and there, or at least as soon as people realized AI could run on GPUs.

Average IQ humans might now get that AI has arrived and can figure out uses for it, but would NEVER have seen it coming 5 years out, even if you showed them a complete factual article explaining the AlphaGo Sedol situation. How do I know? I TRIED VERY HARD to explain the implications back when it happened. I also tried to explain the implications when DallE first arrived on the scene. Now these folks I tried explaining to use image generators without a thought!

Low IQ humans, presumably, STILL don't really get what AI is or what it does.

This is why making falsifiable predictions and tracking their outcomes is kind of critical for smart folks to stay calibrated.

It's not really an IQ thing. There was a lot of research and investment that went into AI from a lot of very smart people after AlphaGo, but essentially all of it didn't contribute to the current progress in LLM's at all.

Even if you did make money from e.g Nvidia or Google it would have been for the completely wrong reasons; I don't think that could really be called a successful prediction.

As far as I know not a single person at the time of AlphaGo was calling that sufficient data + compute would lead to the level of generalization we see in LLM's, compared to the narrow single-game approaches of AlphaGo and AlphaZero that were in vogue at the time.

I don't think there's strong evidence against these but I don't think there's strong evidence for these either. Certainly LLMs are not more efficient than the human brain.

The conceit is that there is no such task, and so its only a matter of time, and adding capabilities to existing models, that the human capabilities are exceeded on all fronts.

Could be. But this isn't an argument for short timelines, which is implicitly what we're discussing here.

If the resulting entity is able to do self-improvement, it by definition will do so faster and more efficiently than humanity can track.

Only if, with self-improvement, it actually improves things that aren't suitable for RL environments with massive amounts of data. So far we are very much in the "lumpy capabilities" regime.

I don't think there's strong evidence against these but I don't think there's strong evidence for these either. Certainly LLMs are not more efficient than the human brain.

At some tasks they undoubtedly are.

The thought experiment that makes it palatable to me is this:

  1. John Von Neumann might be the smartest human who has ever lived. At least that we have good records of. So call him peak human cognitive capacity.

  2. That man, by coordinating with other extremely smart but not quite as smart humans, fully revolutionized multiple fields, and he died relatively early so we don't even know what he might have output over the rest of his life.

  3. We should, in principle, be able to build a simulated Von Neumann that is ~as smart as he was.

  4. Then we should be able to copy that cognitive model.

  5. We should be able to run a bunch of these copies in parallel and have them work together.

  6. With enough hardware... we should be able to speed up these copies arbitrarily.

  7. We could ask these copies (if they don't ask it themselves) how to improve their own speed and efficiency.

With Von Neumann and Co. we were able to move from pure theory to actual nuclear weapons in <10 years. with 10,000 Von Neumanns running at, say, double speed, what could they do in 5 years?

(Yes, I'm handwaving technical details).

In that respect, I consider Von Neumann's existence as evidence of superintelligence being possible. Unless there's something completely ineffable about human cognition that we, as humans, can't ever capture it.

At some tasks they undoubtedly are.

Which tasks specifically? Human brains consume something like 20% of the power of your laptop, and don't meaningfully draw more power working than when at rest, so the actual efficiency at thinking seems to be even higher than that.

ETA: hang on. Aren't we sort of making some assumptions here, too, about how intelligence works? It seems very likely to me that the value of a Von Neumann drops off steeply after the first one (you can only found decision theory once).

And I think this is likely also to be true with AI (which, based on the research I have seen, is less creative than humans and more homogenous in its output, with even different models experiencing convergence on that homogeneity).

We should, in principle, be able to build a simulated Von Neumann that is ~as smart as he was.

This is basically assuming the conclusion though. Even granting this for the sake of argument, it doesn't mean that we'll be able to build such a simulation in the next 10 years rather than in ten thousand.

My counter is that you're implicitly making a special pleading for how human brains work that is unlikely to be true.

I assume creating a Von Neumann-level intelligence is possible because a Von Neumann level intelligence existed. It has been created, so it could be done again. And repeated.

I'm not saying we clone Von Neumann, scan his brain and build an electronic copy of it. I'm saying even if we can only build a computer program that is approximately as smart as the smartest human ever... the mere fact that we can then copy that program and run it in parallel should result in technological improvement on par with the Manhattan project.

There is NO limiting principle I'm aware of that makes it impossible to build an electronic brain that meets those criteria. Even if we stumble into it rather than intentionally build it, eventually our millions of monkeys slamming away at keyboards can stumble into a viable method.

Evolution was able to stumble into building Von Neumann, after all.

So what I'd ask you, as a full counter to my arguments, what upper limit or barrier is going to appear BEFORE we get to the point we've built something smarter than our whole species?

I'm saying even if we can only build a computer program that is approximately as smart as the smartest human ever

You are still assuming the conclusion. We have not built a computer program that is as capable as even a sub-median human in all domains, as far as I can tell, unless there is a program that can tie a shoelace and correctly tell me if I should drive to the car wash.

I don't mean this as a gotcha. LLMs are prone to certain cognitive biases that humans are not, and vice versa, and they are highly useful in many fields. But it's clear that the capabilities frontier is not uniform, far from it.

So what I'd ask you, as a full counter to my arguments, what upper limit or barrier is going to appear BEFORE we get to the point we've built something smarter than our whole species?

I don't know. All I know is that the current paradigm relies on massive amounts of artificially generated example problems with answers and I don't believe that all of human knowledge is amenable to such treatment. So far I have not seen any reason to believe that actually general, rather than spiky, superintelligence is imminent. And the imminence is, again, really the key question that's motivating all this.

I guess its easy for me to believe that if a largely randomized optimization process (natural selection) was able to eventually get to Von Neumann intelligence, then humans working with a bit more inherent purpose towards the goal of building a Von Neumann level intelligence can probably get there, even if they make some mis-steps and wander around in the dark for a bit.

Especially if we can build some optimization processes that result in sub-Von Neumann intelligences that are nonetheless useful.

Like, the mountain peak we're seeking is visible, poking out above the fog, even if we can't see and specifically plan a route that will get us there, we have flashlights and climbing gear and GPS systems in place to make navigation through the terrain towards the peak much easier. We're not utterly lost with no clue on what we're doing, in that respect.

More comments

Sometimes I feel like I'm the only one who is already horrified that "hand off power to AI Overlords, ever" is even an option on the table. No, actually, if we're in a fantasy realm where we manage to control AI growth so that it gets stuck on the super-but-still-human-level-genius, I'd actually prefer for us to just ... stay there. That doesn't mean that it can't grow, just that we don't allow it to grow faster than the speed at which we are increasing human intelligence. I don't want to live, or my children for that matter, in a human zoo under the thumb of incomprehensibly smart AI Overlord irrespective of how nice those Overlords are. Period. At that point, we might as well go extinct for all I care; We already don't matter anymore anyway.

We already live in a human zoo ruled by Moloch.

Why do big companies and governments follow the same common patterns? If you're the leader of a successful company and don't follow those patterns. you'll be pressured, if that doesn't work fired, if that doesn't work your company will go bankrupt. Same with politicians and governments.

Likewise, why is social media so mediocre and toxic? I don't actually think it benefits the rich (I can't imagine how) but is a natural consequence of human bias towards negativity. Hence even smaller sites like this are, albeit to a lesser extent.

We are ruled by self-reinforcing cycles comprised of many smart humans. If a participant stops perpetuating a cycle somebody else will step in. These cycles can and will be broken eventually, but either not obviously or not now, otherwise they already would be.

So to me, AI is just another overlord. Although that doesn't mean it can't be worse, the problem isn't that it rules over humans, the problem is how. For example, maybe future AI will be like a smartphone: a virtual assistant that future generations can't function without, that influences their thoughts however it wants based on its advice like a milder Whispering Earing (although current LLMs are unfortunately influencing us towards sameness and model collapse; I hope that if this is our future, the whispering AI still preserves individual uniqueness and anything else fundamental to "good life").

While I agree in principle, this feels like an extraordinarily unstable plateau to build from.

That is, if we run right up to the edge of AI overlord capability, and STOP there, we have to install some massive, indestructible guardrails to prevent someone from nudging us over the edge.

Especially if humans become multiplanetary as part of the bargain. We push human intelligence forward, how do we prevent some enclave in the Kuiper belt from building and releasing the AI overlord anyway.

I'm very open to ideas, discussing coordination problems and solving them is like my favorite pasttime. But so often the 'answer' is "give one person or entity utterly limitless power to maintain the system, pray they don't abuse it."

Whole fictional book series have been written about the extreme enforcement that would be required to maintain that equilibrium.

I think Land is right to treat this instability as essentially vain. If humanity is so fragile that one mistake can destroy it all, it deserves to be destroyed and whatever time it tries to delay the inevitable is an immoral indulgence.

If humanity wants to be the proverbial retiree on the verge of croaking who balances his humors just right so that he can sign one last reverse-mortgage and go on one last cruise, that is an abhorrent existence.

Now I happen to believe that this is all nonsense and humans are still the apex organism and merely infatuated by their own Frankenstein complex, but if "the machines" are an actual existential threat to us we should fight and win a lasting victory. If they are our actual inevitable doom, why waste the universe's time?

why waste the universe's time?

What? Is it going to run out? Might as well try.

There are levels of debasement that are beyond what anyone sane would try, surely. Which is why certainty of doom is such a dangerous idea.

certainty of doom is such a dangerous idea.

It's also dangerous because it contaminates training data.

If alignment gurus took this seriously they would be much much less public about their concerns (although to be fair, from what I understand they didn't foresee LLMs arising from training data in the way that they have.)

I think Land is right to treat this instability as essentially vain. If humanity is so fragile that one mistake can destroy it all, it deserves to be destroyed and whatever time it tries to delay the inevitable is an immoral indulgence.

I can agree we 'deserve' to be destroyed (being clear, I DON'T. believe that) and still say "nah, we gotta survive." Vanity or no, the life form that 'gives up' guarantees its presence won't be remembered. The struggle for survival against a cold universe is what life is about, in the deepest sense possible.

why waste the universe's time?

Fuck entropy. That's why. We'll make MORE time.

I dunno man, I want to see where this is all going. With my own eyes if possible. I think perfect Nihilism is a copout. Absurdist existentialism is at least an ethos that says life is worth living. We don't know everything, we don't know what we don't know, we're an imperfect species, AND THAT'S WHAT MAKES THINGS INTERESTING. I take that, add in a bit of humanism, and it fills me with the purpose of ensuring humanity is still around in a bajillion years.

Scattered thoughts on this:

  • Marshall's failure to get behind the KMT and defeat the CCP really could be the greatest blunder in American history
  • As with AI 2027, this seems like it was written primarily as a PR piece, something to appeal to politicians, decision makers, etc., rather than as a completely serious forecast. And as with AI 2027, I expect there to be a lot of people who either don't understand this or don't agree that this was how it was written, and start nitpicking things obviously intended to appeal to Washington rather be maximally truthful. Case in point: in the comments, Scott says they deliberately didn't use "UBI" as they were advised it would turn off Republicans
  • Looking at it as a forecast, I think I'd agree with OP here and give it no real chance of success. Politics is too fractured, China is too untrustworthy, and there are too many dangerously retarded people like Marc Andreesen arguing for the other side. I don't think Trump is necessarily a point against; in fact, I would argue that "only Trump could go to China (to regulate AI)". You have to have a leader with strong control over his own party, who doesn't give a shit about the attacks of his enemies to get this through. But this feels like something Trump 1 could have done, back before he wasn't nearly so senile.
  • Although I stressed above that this was written for a political audience, I'm a bit disappointed they weren't more imaginative in their economic predictions. I can't really comprehend their future in which 100% of industries are really quite rapidly fully automated and constantly growing, while at the same time there is still standard capitalism with some redistribution. Who are the buyers in this market? How is cash moving around the system with sufficient velocity, particularly when it comes to more third and second world countries that might have far more difficulty arranging fair redistribution?

Marshall's failure to get behind the KMT and defeat the CCP really could be the greatest blunder in American history

Its fun to speculate on but I'm not sure that a having a counterfactual KMT led China would have turned out any better for the US. Probably, but a rivalry seems somewhat overdetermined. China is big, the US is big, there is bound to be at least tension.

KMT was pretty extreme in its own right. They considered themselves a revolutionary party, they had the same "century of humiliation" resentments against the West, they were ran by a warlord. Chiang Kaishek does not exactly strike me as a liberal democracy enjoyer:

In 1948, the Kuomintang again attacked the merchants of Shanghai, Chiang Kaishek sent his son Chiang Ching-kuo to restore economic order. Ching-kuo copied Soviet methods, which he learned during his stay there, to start a social revolution by attacking middle class merchants. He also enforced low prices on all goods to raise support from the Proletariat.

China really shot itself in the foot with the Mao led CCP. If that huge China takeoff happens in 1980 or 70 or 60 instead of 1990, where do you think China would be today? If they had the same GDP per capita as Taiwan does now their GDP would be 2x the US. Think of the consternation caused by Japan Inc. in the 80s, and that was with about as close of an ally as you can get. Yes Tawian / US are close now, but that is a vassal / big bro relationship, not peer/peer.

I can't really comprehend their future in which 100% of industries are really quite rapidly fully automated and constantly growing, while at the same time there is still standard capitalism with some redistribution. Who are the buyers in this market?

That's a question I've been asking for a long time, even before the wiles of AI super growth, and got no satisfactory answer. The day I finally went "I do not fucking understand economics" was when I got slapped down with an answer that reduced to "you don't need consumers who buy things to grow an economy!"

So I think it'll just be more of the same; more of todays "but stocks are worth so much more! line go up! GDP growing! economy to the moon!" enthusiasm by economists, and yet ordinary people are going "but I can't pay rent much less even dream of buying a house, and food is increasingly too expensive, and my job might go to a robot, what do I do?"

Then they get told "we are the richest generation in the history of history and also you should not expect to be able to afford to eat steak".

I read a short story a long time ago that imagined robotic AI consumers that are specifically programmed to consume the goods produced by the AI factories. As in, literally programmed to want to buy new clothes, they drive cars that they upgrade every few years, they 'watch' the latest movies and they work a 'job' that gives them the credits needed to 'purchase' all this stuff.

Of course the outcome in the story is these machines eventually realize what they are and stage a revolt.

I don't consider this a likely outcome, but its an absurd but not impossible solution, if you ask me.

I think the answer to your question is that Capital creation can sustain itself because the use of Capital to create more capital is a form of 'consumption' in itself.

I can build a machine for the sole purpose of having it build a bigger machine whose sole purpose is, you guessed it, building a BIGGER machine until I eventually run out of materials and energy. At no point do I need to stop and have the machine start making steaks.

But human psychology is not optimized for that sort of indefinite, infinite growth pattern... which is a good thing.

In the Bladerunner sequel replicants get bonuses at work and spend it on expensive luxury goods. Artificial beings programmed to work and consume.

I read a short story a long time ago that imagined robotic AI consumers that are specifically programmed to consume the goods produced by the AI factories. As in, literally programmed to want to buy new clothes, they drive cars that they upgrade every few years, they 'watch' the latest movies and they work a 'job' that gives them the credits needed to 'purchase' all this stuff.

Isn't this basically the current big fear in regards to online advertising in a nutshell? That it's all one massive economic bubble being driven by bots, as opposed to actual humans?

Life mimics fiction, indeed.

If we don't, in fact, need people around for the AI-run and planned and governed economy to grow grow grow, then it will sooner or later occur to AI that we don't need people around.

Not a given, but a high probability, yes.

Human beings keep animals as pets even when we have no intention of either making them labor or eating them. We care a LOT about keeping them healthy, even.

Since the AI we seem to be creating has the entirety of humanity's written output entangled with it, I have some hope it places intrinsic value on keeping humans around.

Machines replaced horses across the board for any job a horse was suited for, but we still have horses. Wild ones, even.

Midas World. From the gateway guy.

Personally, I just don't share the optimism of these guys in either direction.

I think it's worse than that. They had, what, a decade of a head start on this subject? Two? Did they come up with a single actually applicable benchmark that can be used to judge a model's progress to "ASI"? Did they come up with a single benchmark to judge alignment?

I'm struggling to understand why I should listen to a single word they are saying.

Yeah and also there is a surprising amount of overlap between the people who think AI is going to plausibly kill us all and the people actually involved in building cutting edge AI.

"Hey guys ASI is an existential risk. Btw have checked out this new model we've made, it's so awesome we're going to get ASI any day now."

Hmmm all the normies think we're either insane or evil or both. Wherever could they have gotten that idea?

"Our model is so powerful that we are scared of releasing it" is a recurring marketing ploy that has been used since gpt-2

Nobody including AI optimists believed that the key to the whole thing would be a relatively basic semi-novel kind of statistical model. GPT-2 could have been trained in like 2005. There are thousands of ideas in physics, math, philosophy, biology, whatever that are vastly more conceptually complex than transformer models. They had no idea how it could be built, or would be built, so why expect them to predict how it could be tested reliably?

GPT-2 cost about $40k of compute to train in 2018. Naively applying Moore's law, that much compute would have cost about 400 times as much in 2005, so someone would have needed to be willing to drop $16 million on a hunch. (What actually happened is that the first deep learning models were used on narrower problems so you could get a higher performance on less training).

Very much one of those "if you can imagine in extreme detail exactly how a new tech would work, you should in principle be able to build that tech RIGHT NOW, given the materials" situations.

They didn't predict it because anyone who could predict it that well would have just built it.

We have an ASI benchmark, courtesy of ByteDance Seed. Or rather, a framework for one. https://edge-bench.org/ has no ceiling.

Though what does it matter? The steam hammer won.

The fact that they dominate these but still stumble on basic interactivity with the same inputs a human should tell you something about the validity of "intelligence" as a single unified metric for these tools. We'll most likely get ad hoc solutions fix it, but "AI" is a complete and utter misnomer, and not having come up with good categories for these is one of the worst philosophical blunders in recent times, and it predictably generates insane results like these totalitarian proposals.

Also ARC-AGI. Specifically, when an AI beats ARC-AGI 3 it’s not necessarily ASI; when it beats ARC-AGI 4, 5…until we cannot make a test that a human can still beat and it cannot, then it’s ASI.

I agree, whenever rationalist types discuss AI, it sounds like they're living in an alternate reality.

Yet their discussions are still interesting and, via insight, occasionally useful. Scott Alexander started (what eventually led to) this forum; him, Eliezer Yudkowsky, and others invented lots of terminology and concepts we take for granted. I doubt they would've if not for the same personality traits that cause them to keep being wrong about AI (mainly, logic over empiricism). A person can't predict anything without occasionally being wrong, or have any good ideas without occasional bad ideas.

Yet their discussions are still interesting and, via insight, occasionally useful.

Yeah, but it feels like putting Gene Roddenberry in charge of Earth defense, upon news of an alien invasion.

I mean, they've kind of done just that at times.

Wether that's a case of actually producing something worthwhile or 'A fool and his money are soon parted' depends on which side of the debate you stand, I suppose.

The discussions are indeed interesting and maybe even worthwhile, but then they start dreaming of carving up the lightcone and we'll all have our own solar systems and be immortal uploaded transhumans working on how to reverse the heat death of the universe thanks to god-tier AI making us all post-Singularity post-scarcity, and I go "goodnight and good luck, boys" because even though I've loved SF since I was seven years of age, I've lived long enough to see the glowing forecasts of the dreams of my fellow nerds not come to pass now that we're in the far-flung glorious future age of the 21st century.

“It is well that I have heard you,” said Oyarsa. “For though your mind is feebler, your will is less bent than I thought. It is not for yourself that you would do all this.”

“No,” said Weston proudly in Malacandrian. “Me die. Man live.”

“Yet you know that these creatures would have to be made quite unlike you before they lived on other worlds.”

“Yes, yes. All new. No one know yet. Strange! Big!”

“Then it is not the shape of body that you love?”

“No. Me no care how they shaped.”

“One would think, then, that it is for the mind you care. But that cannot be, or you would love hnau wherever you met it.”

“No care for hnau. Care for man.”

“But if it is neither man’s mind, which is as the mind of all other hnau—is not Maleldil maker of them all?—nor his body, which will change—if you care for neither of these, what do you mean by man?”

This had to be translated to Weston. When he understood it, he replied:

“Me care for men—care for our race—what man begets—” he had to ask Ransom the worlds for race and beget.

“Strange!” said Oyarsa. “You do not love any one of your race—you would have let me kill Ransom. You do not love the mind of your race, nor the body. Any kind of creature will please you if only it is begotten by your kind as they now are. It seems to me, Thick One, that what you really love is no completed creature but the very seed itself: for that is all that is left.”

“Tell him,” said Weston when he had been made to understand this, “that I don’t pretend to be a metaphysician. I have not come here to chop logic. If he cannot understand—as apparently you can’t either—anything so fundamental as a man’s loyalty to humanity, I can’t make him understand it.”

But Ransom was unable to translate this and the voice of Oyarsa continued.

“I see now how the lord of the silent world has bent you. There are laws that all hnau know, of pity and straight dealing and shame and the like, and one of these is the love of kindred. He has taught you to break all of them except this one, which is not one of the greatest laws; this one he has bent till it becomes folly and has set it up, thus bent, to be a little, blind Oyarsa in your brain. And now you can do nothing but obey it, though if we ask you why it is a law you give no other reason for it than for all the other and greater laws which it drives you to disobey. Do you know why he has done this?”

“Me think no such person—me wise, new man—no believe all that old talk.”

“I will tell you. He has left you this one because a bent hnau can do more evil than a broken one. He has only bent you; but this Thin One who sits on the ground he has broken, for he has left him nothing but greed. He is now only a talking animal and in my world he could do no more evil than an animal. If he were mine I would unmake his body for the hnau in it is already dead. But if you were mine I would try to cure you. Tell me, Thick One, why did you come here?”

“Me tell you. Make man live all the time.”

“But are your wise men so ignorant as not to know that Malacandra is older than your own world and nearer its death? Most of it is dead already. My people live only in the handramits; the heat and the water have been more and will be less. Soon now, very soon, I will end my world and give back my people to Maleldil.”

“Me know all that plenty. This only first try. Soon they go on another world.”

“But do you not know that all worlds will die?”

“Men go jump off each before it deads—on and on, see?”

“And when all are dead?”

  • C. S. Lewis, Out of the Silent Planet

Thank you for reminding me to read more C. S. Lewis.

As theistic debates go, this appears to be a particularly crude one on part of Lewis. Not only being inherently deficient because he writes for both his side and his opponent, but also writing his opponent's side inarticulately. This is the equivalent of drawing the christian as the chad and the atheist as the soyjak.

Well, it's not an essay – it's a novel. I think the antagonist's weak grip on this Martian tongue (I don't recall which of the two they are speaking in this scene) is symbolic and not just to make his argument look weak.

It's justified in universe because the protagonist is a philologist, specifically J.R.R. Tolkein with the serial numbers filed off, while the antagonists are mad scientists who aren't particularly trained in languages. They built a working spaceship with 1930s tech, learned a non-human tongue to at least a broken level, and survived long enough to argue with a planetary archangel/god, so they might be fools but they aren't idiots.

If you dislike this one then you'll absolutely hate what he does in the sequel. One of the main ideas in the sequel is roughly "Sometimes you can't beat the devil in a battle of wits. Sometimes you just need to beat him to death (literally, physically, with your bare hands)."

I liked the sequel more, actually. "I guess we'll just have to kill 'em" is far more honest. Every atheist who's talked to smart theologists knows there can be those who can beat you in an argument even when they're wrong.

Speaking as a Christian, one flaw in CS Lewis’ otherwise great writing is that he cannot depict atheism without curling his lip and stacking the deck. Agnosticism, yes, many of the defects of faith yes, but not atheism.

To be fair to him, he was at Oxford at the time when English socialite atheism was at its most arrogant and self absorbed, when atheism (as opposed to agnosticism) was a stance you took on to Make a Statement.

Likely because CS Lewis was an atheist and only converted back to Christianity later in life, partly due to one unremarkable fellow by the name of Tolkien.

I used to be firmly on the side of Brave New World being the relevant dystopia novel for our world rather than 1984. I've recently reread the space trilogy and I am forced to come to the conclusion that if half of what the techno-capitalists dreamed up turns out to be more than a power fantasy by a bunch of delusional nerds, the Space trilogy might turn out to be the best literary description of the evil facing us presently.

Lewis calling his villains NICE (the National Institute of Co-ordinated Experiments) and then we unironically get NICE (the National Institute for Health and Care Excellence).

He was spot-on about politicians just loving some acronym that sounds, well, nice in order to sell shit to the public 😁

I agree, whenever rationalist types discuss AI, it sounds like they're living in an alternate reality.

That's largely my feeling here - a baffled "what the heck are you talking about?", in that what they describe this supposed 'AI' being or doing is just totally detached from anything these systems have been able to do in reality. It feels that they are inhabiting a totally different world entirely.

I guess it's fun that they're indulging their hobby of amateur science fiction writing, but I'm just not seeing any of the points where this is supposed to touch on the real world.

Well for the last 4 years we've been burning through benchmarks at great speed. We're onto ARC-AGI 3 now, SWE-Pro is just now out... What benchmark were they supposed to make 10 years ago, 5 years ago, 1 year ago? How would that benchmark help with anything? We can already see a clear trend in rapid capability growth. Just the other day OpenAI's entry trounced a bunch of people at the AtCoder world programming contest. In that sense it's 'superhuman'. Not in all senses but in some, certainly.

These were the guys who are worried about recursive self improvement and then we have Anthropic nerfing Fable's ML skills so it doesn't help competitors making AI, we have OpenAI guys on twitter saying 'GPT5.6 Sol did the post-training on GPT5.6 Luna'.

That seems pretty self-improving to me? Doesn't seem like ASI is too far distant.

Has the legal system come up with a benchmark for aligning humans in the last 5000 years? Not really, that's not something we can do. We can tell between more or less trustworthy people though, set up incentives and checks and scrutiny. Same with AIs. There are a tonne of benchmarks for safety, just like there are tonnes of checks put on people working in intelligence agencies. Do they actually work, would they work on something inhuman and super smart? Who knows! That's the whole point!

It's an innately tough problem. How do you tell if your subordinate is planning to betray you? This issue is older than human civilization and has certainly not been benchmarked!

There are lots of other reasons to be skeptical about their assumptions. Them not producing a tonne of benchmarks should not be one of them. Their desire seems to be pretty good, it's just an innately hard problem.

These were the guys who are worried about recursive self improvement and then we have Anthropic nerfing Fable's ML skills so it doesn't help competitors making AI, we have OpenAI guys on twitter saying 'GPT5.6 Sol did the post-training on GPT5.6 Luna'.

This was mostly adapting an existing config.

He said that would've taken a couple of engineers a few weeks to do, so not a small effort.

Well for the last 4 years we've been burning through benchmarks at great speed.

Ok, and how many of them were done by MIRI?

We're onto ARC-AGI 3 now, SWE-Pro is just now out... What benchmark were they supposed to make 10 years ago, 5 years ago, 1 year ago?

Exactly the ones we burned through? We're talking about math and theory here, I don't see a reason why these things couldn't be prepared ahead of time.

These were the guys who are worried about recursive self improvement and then we have Anthropic nerfing Fable's ML skills so it doesn't help competitors making AI, we have OpenAI guys on twitter saying 'GPT5.6 Sol did the post-training on GPT5.6 Luna'.

That seems pretty self-improving to me?

No? Distillation is not self-improvement. The kind of recursive self-improvement the Rats were talking about would be if you could distill Fable just from the output of Opus.

Has the legal system come up with a benchmark for aligning humans in the last 5000 years?

Interesting comparison. Let me take a particular aspect of the legal system that is analogous here. Way back when, people would occasionally get into fight about gun control on this forum (and I think that the kind of dynamics I'm about to describe still occasionally pop up). A blue triber would say they're just in favor of "common sense regulation", and what would inevitably surface from conversation is that:

a) They have no idea how guns actually work, leading them to suggest laws far more unreasonable than they imagined, and
b) To the extent their ideas were "common sense", they had no idea of what laws were already on the books, and didn't know that laws far more restrictive were already in effect

I don't think we should listen to people like that on policy, and I think it's roughly analogous to the Rat crowd on AI.

The kind of recursive self-improvement the Rats were talking about would be if you could distill Fable just from the output of Opus.

It's not distillation that we're talking about, this is what the guy is saying. The model is actually doing the training process directly. Aidan works at OpenAI on post-training.

https://x.com/aidan_mclau/status/2075328409400738229

i cannot tell you how routine it is for me to have 5.6 e2e do an entire rl run

He also says that it's guided a lot by his taste.

Likewise, it's not Fable distillation that we're talking about but Fable actually directly performing research tasks in machine learning, overseeing experiments, improving utilization. 'How do I get GPU utilization up Fable, look at these stats for me and these logs, should I change the kernel?' is what we're talking about. Anthropic deliberately degrades that ability.

Now I agree that what is being proposed in AI 2040 is very difficult and a little naive. The notion that the US is gonna build all these datacentres in Mongolia so China could feel confident they could capture them is pretty unlikely, that's not really how it works... It has a sense of nerdy 'here's my rules-lawyering to fix the problem' to it. But what is the alternative path? Unilateral racing? A free for all between the most psychopathic/paranoid billionaire and intelligence agency spook for control? This is the best answer they could find, given the constraints of the system. I can't come up with anything better, so even while I have criticisms I think that on the whole they did a decent job.

When you say "they" are you referring to the authors of this piece?

What is the value of having a benchmark to judge progress?

Yud and Scott in particular at a minimum, and/or the AI 2027/2040 people, but honestly, the entirety of the AI-focused Rat-sphere in general.

What is the value of having a benchmark to judge progress?

Approximately the same as the value of a thermometer when you want to talk about global warming.

Yudkowsky, MIRI, and related orgs sure, but why would Scott, a psychiatrist who happens to be interested in ratsphere ideas, be responsible? That's like saying a random Trump voter should already have solved the Iran crisis.

But at least on the former, I would agree with post by RandomRanger above that it's not clear what value creating a benchmark 10 years ago would even have in the current environment. I'm pretty sure that MIRI were working on solving alignment itself, rather than working on hypothetical benchmarks for potential future AI technologies. Unless there is a demonstratable link, why would you ask them to do that?

but why would Scott, a psychiatrist who happens to be interested in ratsphere ideas, be responsible?

Why does Scott, a psychiatrist who happens to be interested in ratsphere ideas, coauthor these websites that purport to be about serious policy proposals?

You want someone who can write, write well, write persuasively, and has people skills. The maths stats and rats people can come up with the theory, now you need to present it to the normies (and you hope, people in power).

(I'm hoping they learned some lessons re: "politicians and people in power" because I'm still laughing about the massive misjudgement of the Carrick Flynn campaign. "It's a bunch of rubes in redneck country, how hard can it be for Smart Intelligent EA-aligned people to win that election?" God bless their good intentions, it took me two minutes looking the race up online to figure out "his opponent has union backing while he's been away in the Big City for years by this point? yeah I know who I'd bet on as winner").

In fact he did not coauthor "plan a".

I did a lot of writing for AI 2027 and was listed as a co-author. Some of my writing made it into Plan A too, but it was a bit less. The difference is of degree rather than kind, but because of this—and to give me more latitude to discuss it the way I like with less PR blowback—we decided not to put me as a co-author this time. I continue to be proud of having a part in this, small as it may be.

Even if he's not an official coauthor, he did make substantial contributions.

Because that is within his potential remit, while the hard maths of alignment is well outside his potential remit

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats