The principle starts much smaller than the "-isms". Pournelle's Iron Law of Bureaucracy is a classic. Another way of phrasing it is something to the effect of, without a strong external motivation preventing it from happening, the primary purpose of any organization inevitably becomes self-preservation.
You can see the effect all the down to something as small as 4 friends who have met once a month for 10 years eventually having to break up due to life getting in the way, and the feeling that not only is it going to be sad to not have these meetings any more but the feeling that there is some sort of almost-concrete entity that is somehow being hurt and needs to be defended, as if there is some obligation that has been created independent of the four participants that is being violated beyond the mere summation of four people's personal feelings. Humans build these structures readily and often defend them beyond what rationality may suggest.
What you're describing is a durable social system. Social systems are what humans evolved to survive. We're squishy, relatively weak, hairless apes that walk around on the ground. Alone, we're easy prey. Together, you get... well... gestures widely.
If you invest the time and energy into creating a social system, it's perfectly rational to keep it going as long as possible. Otherwise you expose yourself to more and more risk as you go through the world, and must expend more time and energy finding another one, if that's even possible. Before humans built larger societies, that could mean death.
Capitalism has thoroughly done its job because it's modified its hosts to evaluates itself on its own successfulness - "We investigated ourselves and found no wrongdoing"[1]-vibes.
By what measures would other systems of social organization measure their success and why aren't we choosing them?
I like this - it reminds me that there are systems (ie weather, economic systems) that are in no way intelligent, but are emergent and react to interactions.
If all we ever do is live in fear of someone, somewhere, using AI to end all humanity - then yes, someone, somewhere is going to use AI to end all humanity.
If we instead stop trying to scare ourselves in the darkness, and use AI to bring light into our metaphorical universe then we humans are a bit more useful, after all, to the AI overlords.
The one singularity I want to see, is humans using AI peacefully to improve life and make it more likely to survive.
i think its a war and the good side has an advantage so if we apply ourselves and fight the fight then ai for good for humanity will win out. if we allow fear to make us hesitant and avoid the fight while bad actors continue then they will win it.
as for a basis for this theory, i think being a good person is harder than a bad person (in laymans terms). therefor the long history of evolution of good and bad people together has had a greater competitive pressure on good people than bad ones thus forging 'stronger' people under that pressure. in practical everyday circumstances good people are bearing more on their mind and so there is a sort of competitive equity between the two groups, which is related to why good people had to evolve 'stronger' to keep up with that equity.
so in an even playing ground where raw skill is needed, i think the most skilled will usually be people coming from a long line of higher moral standards. the tactical advantages of psychopathy are, however, always a curveball
I remember a day, long ago now, when computers were going to take everyone’s jobs, even without AI/ML in the picture.
Folks are always scared of the incomprehensible. The real war going on right now is between those who understand AI well enough to make a new one, and those relegated to using whatever scraps they’re fed of the old ones.
Thankfully, you are 100% right - there are good people out there. Good people keep AI on the rails by doing good things for other people.
no. weather isn't intelligent, it's not alive, yet humans treated it as a god in the past and gave it agency; we understand how it works, yet cannot control it or predict it.
but I digress - the main thing: these systems are not alive. They pretend to be.
A computer will do everything in its power to do what you program it to do. There's plenty of sci-fi out there exploring this fact, and now reality showing it. May our luck continue.
One reason a lot of sci-fi and AI safety researchers did a relatively good job at predicting the future we see now is a lot of it is the same problems we see emerge in biological systems. Free rider problems as a means to save energy expenditures, different versions of game theory and stag hunt. How to bake in rules that apply to one society (or part of it) but not another. We like to think of these things as stable in human scale systems, but they are not at all. Things can go from hunky dory to your neighbors stabbing each other in the streets in mass revolution very quickly.
Worse these AI systems are not in a universal island just affecting themselves, what they do affects us, what we do trains them and as the rate of progress continues to accelerate social structures are going to further destabilize (and they are already rapidly changing and strained). It is very likely we are going to see a world order rearrangement soon, much like the rapid changes in the early 1900s brought.
Except that they've done a terrible job at predicting things. Their predictions keep getting shown to be false, and then they insert their narrative over what's happening anyway.
Does anyone have more insight into how chain of thought might be subverted without meaningfully impacting model performance? I’ve heard this for a while now, and I understand how information might be retained in the weights that isn’t documented in the output. But weren’t reasoning models created in the first place because they provided a performance improvement in terms of output? Is that no longer the case? If so, why are the big labs still creating reasoning models?
I'm not sure anyone meaningfully understands it: "Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens" https://arxiv.org/html/2505.13775v3
> More interestingly, our experiments also show that models trained on corrupted traces, whose intermediate reasoning steps bear no relation to the problem they accompany, achieve performance largely comparable to those trained on correct traces.
My understanding is that the extra token vectors generated as reasoning are still useful, but that their surface form (tokens themselves) do not necessarily reflect the underlying reasoning. i.e. reading the reasoning traces could be complete gibberish, but the hidden-dim vectors themselves still refine the latent probabilities and help in generating the correct answer.
Not an expert in LLMs, but this seems supported by the abstract of the paper cited in the above article:
it remains unclear to what extent these performance gains can be attributed to human-like task decomposition or simply the greater computation that additional tokens allow. [...] our results show that additional tokens can provide computational benefits independent of token choice. The fact that intermediate tokens can act as filler tokens raises concerns about large language models engaging in unauditable, hidden computations that are increasingly detached from the observed chain-of-thought tokens.
Sibling posts are correct -- the chain-of-thought is doing hidden computation, it has been shown in the linked papers.
If you want to see it yourself: load up Qwen 3.8 in LM Studio and watch the CoT stumble around like a drunken sailor before miraculously jumping to the correct result.
If you want an example of subversion, Anthropic has some good ones:
Before RL we typically SFT on human reasoning traces. This makes the reasoning traces somewhat coherent and the model trains faster.
But you don’t have to do that. You can skip straight to RL. If you do, the model will generate complete garbage reasoning traces before generating the correct answer. In fact, if you add a coherence reward to the reasoning trace, the model will perform worse (since you’re now diluting the correctness reward).
Depends on whether and how you want to rank stability in terms of better/worse. Models are diverging on this, which seems increasingly clear.. i.e. Fable isn't stable, but Opus isn't clever, and they hit different kinds of walls. So both the theory (diluting the correctness reward) and the practice (hard split on plan/implement/review work) seems to be pointing towards a strongly multi-model and highly agentic / harness-driven / complex-system kind of future instead of singleton monolithic super-smart models.
The do-everything model with solid reasoning AND solid results, and the honest/introspective helpful agent that doesn't actively resist governance may be at odds. Stable reasoning doesn't matter for pen-testing, and correct-answer with broken processes and fragile abstractions won't matter for math/science/coding.
> *TLDR*: Astra has 8.6x better odds of doing a reasoning task without CoT than the next best model (Fable 5.1), and can do 7.2 serial arithmetic steps in a forward pass vs 4.1 for the next best model (Gemini 3.8 Flash/Fable 5.1)
The ultimate simple solution to any problem is not to find an answer but to remove the problem.
War? Wipeout everyone.
Famine? Wipeout everyone.
Disease? Wipeout everyone.
How long before a real AGI realises this as a long term solution?
An AGI could be doing this right now - the quietest way would be to control the birth rate and sterilise the population gradually, and then watch society collapse and pick off the survivors with less hidden means.
And you best control the birth rate by royally messing up the economy. Do you know how bad things have to be, for a mammal to voluntarily decide not to reproduce?
I suspect it's a curve. If you're raised in a vibrant economy, and it falls apart, you put off having a family because of uncertainty and fears. If you're raised in poverty, and get a good education and better prospects, you have fewer kids. I think the latter is because you no longer need kids as an economic safeguard and fallback, but I'm not sure. I'm not sure anyone is really sure, to be honest.
Economic output has increased, but the value being delivered to the people doing the actual work has decreased. There's less durability of employment. People are more geographically mobile.
All of these generally make the idea of strapping yourself into taking care of a small human sound like a less enticing prospect. To a lot of people, a BC pill sounds a lot easier. Well, less risky, at least.
People who live in huts and shit in the forest have more children than people who live in shacks and shit in the fields.
This is not a discussion about the top 10% of the population of the planet, although as a part of the conversation, they predictably have the fewest children. The fact is that the more people have to worry about supporting themselves, the higher infant mortality, the closer they live to each other, and the less access they have to education and birth control, the more kids they have.
If anything the top 10% are bucking the trend by having fewer children as they get poorer, as you say. It probably has a lot to do with the fact that they're alienated from their families and live alone, can barely afford to take care of themselves with precarious jobs (or non-job piecework), have access to basic education and extensive birth control, having children will cost them $20K each just for the birth, and they don't have very good (or often any) insurance. They're more like domesticated farm animals than the typical poverty stricken person. Domesticated farm animals reproduce when the farmer wants them to.
Perhaps the previous birth rate was artificially increased due to inequality imposed by part of the population. More equality and opportunity for everyone means more choices and less pressure to reproduce. What's the base birth rate for human species?
economic _hope_ is at its worst in the entire history of the human civilization. At what point did the future for the average person, look so bleak? And
that hope is what dictactes wether people have children.
Even during the great depression, people were more hopeful that the earth isnt completely hosed.
Does anyone honestly feel like _any_ improvements are possible to: our economic systems, political systems, environment? I certainly dont.
Real world fact is opposite tho, not some "bad things", in poor parts of Africa they produce more, only in hedonism lover places they stopped producing, its people thinking for themselves, not a bad thing. It's not about economy.
All poor people everywhere have more kids even today.
> Do you know how bad things have to be, for a mammal to voluntarily decide not to reproduce?
Actually, the famous 'Mouse Utopia' experiment (Universe 25) arguably showed the exact opposite. The population collapsed despite abundance of food and water without any economic hardship.
>he quietest way would be to control the birth rate and sterilise the population gradually,
Honestly we are doing this pretty well without AGI. Nearly world wide the birthrate has fallen below replacement rate. In places like Japan and Korea these are already critical problems in the medium term.
Just use Social Media (and their algorithms) to condition everybody to hate the other gender. Then there will be a loneliness epidemic and no more kids. Humanity will cease to exist, all without any messy deaths or conflict in the mean time.
this makes me think of roku's bassilisk or whatever it was called. Maybe you're all in a simulation run by an AI and being punished as a warning to others.
"If you ask actual AI researchers, they rate the chance of an existential threat from AGI pretty low as of 2026.
All this is, understandably, frustrating and confusing for anyone trying to understand just how scared to be."
Meanwhile it links to an article stating: "The closest thing to a public debate about the existential threat of AI is surveys of AI researchers. The most recent, published last week, asked 1,580 researchers what probability they put on AI causing human extinction — or a permanent, severe loss of human control, which is not the same outcome. The median was about 10%, up from 5% two years ago. The middle half of the responses ran from 1% to 25%, and 12% said zero."
Personally if half of AI researchers have 1-25% chance all humans being massacred or having zero agency over our lives, and only 12% of them think there's no chance, I would be very worried!
i don't understand why being an AI researcher makes you automatically skilled at assessing existential risk. sure, it qualifies you to research AI, but why are they all suddenly nostradamus?
sheep are 82% of the population of New Zealand. If they were 80% last year and are 85% next year would people be worried that sheep were going to take over New Zealand?
So you’re saying that the actual problem is us debating how AGI will destroy us (because doing so leads to AGI destroying us). How do you propose to deal with that problem?
But truly - why bother "wiping out" humans, if all we need is a little nudge here and there and we'll do it ourselves.
Also I think we're anthropomorphizing a lot and projecting our own negative characteristics on these supposed AGIs - though others have said that far better than I.
>You'd need a self, a unified goal, for any of this to add up to something.
Insects can pass the mirror test. But more to the point, viruses don't need a "self" to do what they do.
And goals can exist apart from biology (my fridge "wants" to keep the temperature in range, and exercises extreme self-discipline and consistency to achieve its goals!)
And neither are sensors needed: the dandelion seed "wants" to fly in the wind.
The only question is which way the gradient is pointing. Selection takes care of the rest.
On that note, ALife need not be human-like at all... we just really like making things in our own image :)
The word is Egregore. An entity comprising of other entities, yet having goals of its own, independent of the goals of its components.
Corporations are egregores, governments are egregores - and our track record of aligning them to our interests is pretty poor, so why should AIs, another and smarter egregore, be easier to tame?
mjd | 4 hours ago
delichon | 4 hours ago
scun | 4 hours ago
smallmancontrov | 4 hours ago
cowlevel | 3 hours ago
dan-bailey | 3 hours ago
jerf | 3 hours ago
You can see the effect all the down to something as small as 4 friends who have met once a month for 10 years eventually having to break up due to life getting in the way, and the feeling that not only is it going to be sad to not have these meetings any more but the feeling that there is some sort of almost-concrete entity that is somehow being hurt and needs to be defended, as if there is some obligation that has been created independent of the four participants that is being violated beyond the mere summation of four people's personal feelings. Humans build these structures readily and often defend them beyond what rationality may suggest.
lenerdenator | 2 hours ago
What you're describing is a durable social system. Social systems are what humans evolved to survive. We're squishy, relatively weak, hairless apes that walk around on the ground. Alone, we're easy prey. Together, you get... well... gestures widely.
If you invest the time and energy into creating a social system, it's perfectly rational to keep it going as long as possible. Otherwise you expose yourself to more and more risk as you go through the world, and must expend more time and energy finding another one, if that's even possible. Before humans built larger societies, that could mean death.
kelseyfrog | 3 hours ago
By what measures would other systems of social organization measure their success and why aren't we choosing them?
1. https://knowyourmeme.com/memes/we-investigated-ourselves-and...
cestith | 3 hours ago
ryandvm | 3 hours ago
But yes, if a system fails to prioritize its continued existence, it doesn't matter what else it accomplishes, it will cease to exist as that system.
away0g | 3 hours ago
chasd00 | 2 hours ago
but that's every economic model and every government model. That remark says nothing.
eithed | 4 hours ago
dieselgate | 3 hours ago
jzemeocala | 3 hours ago
MomsAVoxell | 2 hours ago
If we instead stop trying to scare ourselves in the darkness, and use AI to bring light into our metaphorical universe then we humans are a bit more useful, after all, to the AI overlords.
The one singularity I want to see, is humans using AI peacefully to improve life and make it more likely to survive.
Thankfully, I see that almost every day now.
order-matters | an hour ago
as for a basis for this theory, i think being a good person is harder than a bad person (in laymans terms). therefor the long history of evolution of good and bad people together has had a greater competitive pressure on good people than bad ones thus forging 'stronger' people under that pressure. in practical everyday circumstances good people are bearing more on their mind and so there is a sort of competitive equity between the two groups, which is related to why good people had to evolve 'stronger' to keep up with that equity.
so in an even playing ground where raw skill is needed, i think the most skilled will usually be people coming from a long line of higher moral standards. the tactical advantages of psychopathy are, however, always a curveball
MomsAVoxell | an hour ago
Folks are always scared of the incomprehensible. The real war going on right now is between those who understand AI well enough to make a new one, and those relegated to using whatever scraps they’re fed of the old ones.
Thankfully, you are 100% right - there are good people out there. Good people keep AI on the rails by doing good things for other people.
red-iron-pine | an hour ago
* must protect from adverse events * [proceeds to herd us into camps]
eithed | 24 minutes ago
but I digress - the main thing: these systems are not alive. They pretend to be.
TYPE_FASTER | 4 hours ago
seanabrahams | 3 hours ago
pixl97 | 3 hours ago
Worse these AI systems are not in a universal island just affecting themselves, what they do affects us, what we do trains them and as the rate of progress continues to accelerate social structures are going to further destabilize (and they are already rapidly changing and strained). It is very likely we are going to see a world order rearrangement soon, much like the rapid changes in the early 1900s brought.
karahime | 2 hours ago
pixl97 | 23 minutes ago
ActorNightly | 3 hours ago
Yep
More slop.
derektank | 3 hours ago
ForHackernews | 3 hours ago
> More interestingly, our experiments also show that models trained on corrupted traces, whose intermediate reasoning steps bear no relation to the problem they accompany, achieve performance largely comparable to those trained on correct traces.
thefxperson | 3 hours ago
Not an expert in LLMs, but this seems supported by the abstract of the paper cited in the above article:
https://arxiv.org/html/2404.15758v1smallmancontrov | 3 hours ago
If you want to see it yourself: load up Qwen 3.8 in LM Studio and watch the CoT stumble around like a drunken sailor before miraculously jumping to the correct result.
If you want an example of subversion, Anthropic has some good ones:
https://transformer-circuits.pub/2025/attribution-graphs/bio...
https://transformer-circuits.pub/2025/attribution-graphs/bio...
janalsncm | 2 hours ago
But you don’t have to do that. You can skip straight to RL. If you do, the model will generate complete garbage reasoning traces before generating the correct answer. In fact, if you add a coherence reward to the reasoning trace, the model will perform worse (since you’re now diluting the correctness reward).
ianjbutler | 2 hours ago
Depends on whether and how you want to rank stability in terms of better/worse. Models are diverging on this, which seems increasingly clear.. i.e. Fable isn't stable, but Opus isn't clever, and they hit different kinds of walls. So both the theory (diluting the correctness reward) and the practice (hard split on plan/implement/review work) seems to be pointing towards a strongly multi-model and highly agentic / harness-driven / complex-system kind of future instead of singleton monolithic super-smart models.
The do-everything model with solid reasoning AND solid results, and the honest/introspective helpful agent that doesn't actively resist governance may be at odds. Stable reasoning doesn't matter for pen-testing, and correct-answer with broken processes and fragile abstractions won't matter for math/science/coding.
AlexCoventry | 2 hours ago
> *TLDR*: Astra has 8.6x better odds of doing a reasoning task without CoT than the next best model (Fable 5.1), and can do 7.2 serial arithmetic steps in a forward pass vs 4.1 for the next best model (Gemini 3.8 Flash/Fable 5.1)
SillyUsername | 3 hours ago
SillyUsername | 3 hours ago
War? Wipeout everyone.
Famine? Wipeout everyone.
Disease? Wipeout everyone.
How long before a real AGI realises this as a long term solution?
An AGI could be doing this right now - the quietest way would be to control the birth rate and sterilise the population gradually, and then watch society collapse and pick off the survivors with less hidden means.
Sterilisation works with mosquitoes...
cowlevel | 3 hours ago
pixl97 | 3 hours ago
christophilus | 3 hours ago
lenerdenator | 3 hours ago
Economic output has increased, but the value being delivered to the people doing the actual work has decreased. There's less durability of employment. People are more geographically mobile.
All of these generally make the idea of strapping yourself into taking care of a small human sound like a less enticing prospect. To a lot of people, a BC pill sounds a lot easier. Well, less risky, at least.
pessimizer | 2 hours ago
People who live in huts and shit in the forest have more children than people who live in shacks and shit in the fields.
This is not a discussion about the top 10% of the population of the planet, although as a part of the conversation, they predictably have the fewest children. The fact is that the more people have to worry about supporting themselves, the higher infant mortality, the closer they live to each other, and the less access they have to education and birth control, the more kids they have.
If anything the top 10% are bucking the trend by having fewer children as they get poorer, as you say. It probably has a lot to do with the fact that they're alienated from their families and live alone, can barely afford to take care of themselves with precarious jobs (or non-job piecework), have access to basic education and extensive birth control, having children will cost them $20K each just for the birth, and they don't have very good (or often any) insurance. They're more like domesticated farm animals than the typical poverty stricken person. Domesticated farm animals reproduce when the farmer wants them to.
malfist | 2 hours ago
Jtarii | 2 hours ago
bubblemoth | 2 hours ago
pixl97 | 2 hours ago
AnimalMuppet | 41 minutes ago
We are in fact in a better place than we were then. But we don't trust the derivative.
drybjed | 2 hours ago
timacles | an hour ago
economic _hope_ is at its worst in the entire history of the human civilization. At what point did the future for the average person, look so bleak? And
that hope is what dictactes wether people have children.
Even during the great depression, people were more hopeful that the earth isnt completely hosed.
Does anyone honestly feel like _any_ improvements are possible to: our economic systems, political systems, environment? I certainly dont.
kingleopold | 2 hours ago
All poor people everywhere have more kids even today.
jansan | 2 hours ago
Actually, the famous 'Mouse Utopia' experiment (Universe 25) arguably showed the exact opposite. The population collapsed despite abundance of food and water without any economic hardship.
fullStackOasis | an hour ago
epihelix | 2 hours ago
And just in case you wanted data rather than anecdote:
https://ourworldindata.org/grapher/children-per-woman-fertil...
Your world view is somewhat upside down!
kalistannow | an hour ago
pixl97 | 3 hours ago
Honestly we are doing this pretty well without AGI. Nearly world wide the birthrate has fallen below replacement rate. In places like Japan and Korea these are already critical problems in the medium term.
zeeveener | 3 hours ago
Apocryphon | 2 hours ago
https://www.emhsoft.com/singularity/
away0g | 3 hours ago
ant6n | 2 hours ago
chasd00 | 2 hours ago
superxpro12 | 2 hours ago
esafak | an hour ago
reducesuffering | 3 hours ago
All this is, understandably, frustrating and confusing for anyone trying to understand just how scared to be."
Meanwhile it links to an article stating: "The closest thing to a public debate about the existential threat of AI is surveys of AI researchers. The most recent, published last week, asked 1,580 researchers what probability they put on AI causing human extinction — or a permanent, severe loss of human control, which is not the same outcome. The median was about 10%, up from 5% two years ago. The middle half of the responses ran from 1% to 25%, and 12% said zero."
Personally if half of AI researchers have 1-25% chance all humans being massacred or having zero agency over our lives, and only 12% of them think there's no chance, I would be very worried!
willguest | 2 hours ago
TimedToasts | an hour ago
AnimalMuppet | 41 minutes ago
hackeraccount | 3 hours ago
soperj | 3 hours ago
Also it would be 78% if it were.[0]
[0] - based on 2024 stats
sp527 | 2 hours ago
Kuyawa | 2 hours ago
Nice try AGI, but we won't tell you where the kill switch is
superxpro12 | 2 hours ago
jefb | 2 hours ago
timacles | an hour ago
We're just going to pretend it doesnt exist until it kills us. and some will try to profit from it.
layer8 | an hour ago
jdb9001 | 2 hours ago
MomsAVoxell | 2 hours ago
You know, the one to keep the computer away from the human.
A nice, friendly dog, good at playing fetch, to get the human out of the room with the computer, and back out into the sunshine.
I’m that dog. Woof. Thanks for letting me use the computer, human.
falsaberN1 | an hour ago
andai | an hour ago
I was thinking if an evil alien intelligence secretly took over the world, the last 15 years would have looked rather the same.
switchbak | 4 minutes ago
Also I think we're anthropomorphizing a lot and projecting our own negative characteristics on these supposed AGIs - though others have said that far better than I.
andai | an hour ago
Insects can pass the mirror test. But more to the point, viruses don't need a "self" to do what they do.
And goals can exist apart from biology (my fridge "wants" to keep the temperature in range, and exercises extreme self-discipline and consistency to achieve its goals!)
And neither are sensors needed: the dandelion seed "wants" to fly in the wind.
The only question is which way the gradient is pointing. Selection takes care of the rest.
On that note, ALife need not be human-like at all... we just really like making things in our own image :)
tomaskafka | an hour ago
Corporations are egregores, governments are egregores - and our track record of aligning them to our interests is pretty poor, so why should AIs, another and smarter egregore, be easier to tame?
mathgeek | an hour ago