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.
Jump in the discussion.
No email address required.
Notes -
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
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.
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.
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.)
More options
Context Copy link
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.
More options
Context Copy link
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.
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
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.
More options
Context Copy link
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.
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.
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.
Such as?
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.
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.
More options
Context Copy link
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.
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.
More options
Context Copy link
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.
More options
Context Copy link
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
My simplified argument, as distilled from Lesswrong (i.e. Yud) and other books.
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.
The floor for something 'superintelligent' (right now, I'm using the definition 'smarter than humanity itself as a collective') is substantially below that.
Human brain architecture is NOT anywhere near the most efficient way to instantiate intelligence. (This follows naturally if you accept 1.)
Humans are capable enough to build electronic hardware that can outperform their own brains in computation efficiency.
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.
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.
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.
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
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.
Could be. But this isn't an argument for short timelines, which is implicitly what we're discussing here.
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.
At some tasks they undoubtedly are.
The thought experiment that makes it palatable to me is this:
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.
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.
We should, in principle, be able to build a simulated Von Neumann that is ~as smart as he was.
Then we should be able to copy that cognitive model.
We should be able to run a bunch of these copies in parallel and have them work together.
With enough hardware... we should be able to speed up these copies arbitrarily.
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.
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?
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.
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 options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link