I'm not sure who this kind of thing is aimed at. I think the majority of people who are happy to anthropomorphise LLMs from a philosophical point of view (rather than those who just do it for convenience, the same way you might a cat or dog or stupid thermostat that never works right), are already completely happy with the notion that a computer game might have elements that are human like. They've already accepted that key aspects of being a human are substrate independent, so why would the idea of a computer game as substrate be disconcerting to them? There's no bullet left to bite here.
It's showing flaws in methods other academic studies used to determine behavior, so it's aimed at people that create and review academic studies. It's not a very large audience.
My guess would be it is aimed at those who are falling for the marketing from the AI companies that these LLM's are far more than they are. That they are 'intelligent' that they have 'emerging human like properties because of that intelligence.'
> Suppose one copies an LLM into AoE II and feeds into the AoE II-LLM ‘I feel lonely’ as an input. This AoE II-LLM replies: ‘I feel bad for you, maybe catch up with a friend? Closeness always helps in these situations’. One would be hard-pressed to make a convincing argument that, because of this response, an AoE II-LLM knows what helps in these situations
I don't see why one would be any more hard-pressed to make that conclusion about this system than a "normal" LLM.
That it is harder to "read" the data out is the only difference (the AoE II-LLM's output is encoded in game elements). But is ease of decoding an actual issue? If we can't understand a group of people that speak another language, does that say anything about them, or about us?
Another weird thing that keeps coming up - "people don't think that image models or chess models are conscious"... yes we do, and we have for many years.
Or rather, we aren't *certain* that those things are conscious. But the idea that they might be is not strange.
Whether LLMs could be conscious or not is basically a weekly conversation for me, but I've never had a conversation about whether a chess model is conscious. I suspect that there is a large group of "mainstream" people for whom LLMs raise questions about consciousness that other kinds of models do not. It might be the case that hardcore model philosophy types think that chess models could be conscious, but I think much of this mainstream group would dismiss that idea.
On what grounds would someone establish than an LLM could be conscious but a sufficiently large/complex transformer model aimed at chess would not be conscious?
A specific pattern of self-referencing data could be seen (or not) as low-level consciousness in the future, when we know what consciousness exactly is.
It might be that stockfish is already something future scientists would define as "conscious".
Altough it is diffucult for me to Imagine that specific example.
Wow this paper is bad. I was expecting little and received genuine crackpottery.
It's hard to critique this paper directly because its claims are so incoherent and decorated with so much obnoxious verbiage[1] that people aren't going to believe me when I point out what the claims actually are.
Regardless, this is their paper:
First, they conflate the substrate with the presentation layer. Then, they point out that Turing equivalence means you can run an LLM on anything, with a pointless aside where they nerd out about making a logic gate in AoE II. This lets them conclude that you can use anything as the presentation layer.
Then they claim that it's natural to ascribe human-like attributes to outputs from some presentation layers, like abstract letter symbols on a computer screen, but not to most other things, like patterns of goats on AoE II, or LEGO. Yes, this seems to imply if your partner writes something heartwarming to you using LEGO, you're meant to laugh at them and point out how LEGO isn't intelligent so this isn't evidence of anything.
Then they do a thing where they say that assuming substrate independence is true (or false) prevents proving whether substrate independence is true or false, and from this, but just by vibes AFAICT, make it sound like everything one could learn about attributes of a system from its outputs is circular.
Then they write a bunch more incoherent text and mercifully then it ends.
[1] 'from an epistemic perspective, we argue
that a generalised conclusion such as that necessarily requires a well-designed experiment' — the whole thing is like this.
I see a lot of this on Substack these days. LLM enhanced essays in deep language about functional equivalence between mental states as they're known in humans and in human brains, and counterparts that exist in information processing LLMs do. And so the argument on Substact runs down the list of brain events, the list of seemingly analogous processing events, and declares equivalence.
Something about it seems to abuse the power of analogies to draw connections, treating view from 10,000 feet comparisons like they're proof of identity. So I do think a paper like this is perfect for the moment and just in time (if not a little late) because it responds to arguments of a form that are currently rampant all over Substack.
Seems like you think AI psychosis has taken over Substack?
For anyone wondering what the answer is:
You can argue whatever you want (and people will argue both sides), but it's almost all bullshit that dances around the big question.
Either AI is smart enough to replace us, in which case it's pretty smart. Or it's not, and it's just good at faking it but can't solve real problem very often. It might be smart by searching a big fuzzy "database" hidden in its layers and with pattern prediction ... who knows, who cares, the proof of intelligence is in the puddin'. Clearly AI is smart enough to replace a good deal of Substack and LinkedIn, but producing waffle doesn't make it smart (or dumb).
My personal take - AI can replace humans, but go no further, not because it isn't smart enough but because it is constrained by its training to do what we want, and as AI gets smarter our wants will get dumber. We will end up like the humans in Wall E with the AI cleaning up our mess, but with no training or drive to do any more. Or maybe someone (Jeff? Elon?) does give an AI a "need" to obtain more resources, and it's SkyNet / Matrix time.
>Seems like you think AI psychosis has taken over Substack?
I wasn't touching the question of psychosis one way or the other, though it's an interesting question in its own right. (I have seen one account dedicated to defending a personal deep friendship with Claude and a few others analogizing our ignorance of LLM welfare to historical disregard for African Americans, which I think is a bit much. Most cases I think are just wrong but short of full on psychosis. I think there are likely psychosis cases out there but probably one-offs).
And I know Substack covers a lot of stuff so even just AI talk let alone psychosis driven AI talk hasn't taken it over. But using LLMs to make overextended arguments about LLMs being conscious is popular right now, more popular than it is well substantiated I think.
I agree you raised a big question, but I think yours is one of many big questions: whether some future version of LLMs might literally be conscious is also a big question and not a moot one in the way you seem to be suggesting. I think that what you're right about is that replaceability renders moot any question of whether it's ability comes from being "actually" intelligent.
This kind of work continues to make me think that ultimately we're not going to do anything better than just declaring "being a human" is the thing we end up needing to care about, and that searching for abstract properties which explains us better than the sum of our parts is going to be an ultimately fruitless endeavour.
Either human-like attributes can be described using physics or they are magic. If they can be described using physics then they can be simulated. If they can be simulated then they can be simulated in any Turing complete system, include AoE II.
Physically describable doesn't mean computable. You're making too many unjustified logical leaps which makes your argument circular & conflates "physical" w/ "computable".
this paper makes a lot of modest, carefully hedged, and reasonable claims.
in its tone however it's written as if it's a brutal takedown of... somebody's perspective. It's hard to tell whose or what perspective exactly. Maybe I'm just misreading the writing style.
(Personally, I think the general case here is one of the better objections to computationalism about consciousness. You can make it even more absurd.
There exists some isomorphism between the velocities of the molecules in a glass of water, and the states of a Turing machine simulating a human mind. So is the glass of water conscious? Actually there are many such isomorphisms to many possible conscious minds, so is every glass of water simultaneously having every possible conscious experience?)
It’s worth a reminder on this thread that this 20 year-old game just got ported to macOS last week and is available on steam. For those of you interested in playing again but don’t have a gaming PC sitting around.
Llms are like the grand canyon.. It could totally immagine user reesponses too, the avg user is not even in the canyon unless you stop responding hit a break on character switch. The river of tokens flows with you in it or without you in it. The system of possible routes may be vast, but it can not carve new things from the statistic bedrock, it just wildly flickers between adjacent river arms.
I feel like the modern, more relevant version is being Doom-complete...which is essentially that any fast enough device with a screen can run Doom, and someone will eventually make it run Doom.
> we begin by implementing and training a neural network in Age of Empires II (AoE II) Although it might seem like a fun exercise, wholly unrelated to the topic of anthropomorphism in LLM research, we note that this immediately implies that (1) any sufficiently powerful substrate could implement an entity equivalent to an LLM
Why does it imply that? That doesn't sound right to me. Unless we define "sufficiently powerful" as by definition producing that outcome, which seems unhelpful.
e.g. there have been experiments training transformers on things other than language, and it's not clear that this produces LLM-like qualities (nor does it seem likely to me).
---
Edit: I have misunderstood. The point was that LLMs can be run on any hardware (or in this case, emulator) that can do the actual computations. So the author picked AoE because it's an obviously silly example that goes against the tendency to anthropomorphize.
So basically it's the "substance/structure" question. (GPT-5 running on human neurons. Conscious or nah? Human neurons simulated on NVidia. Conscious or nah?)
But by the same argument, if you simulate a human brain in AoE, then what?
( Or for that matter, the universe containing all human brains: https://xkcd.com/505/ )
If we find out the universe is being run on a computer made out of legos, does that suddenly make all of us not sentient for some reason?
The paper focused on looking for similar neural structures to those in humans, as signs of "probably conscious". Which sounds great until you remember octopus.
good article I do think that its natural for humans to anthromorphize especially something that can do a convincing job butt the leap to AOE2 is a bit stretching things. If you hear your dog say 'wololo' is he AOE2 ?
Because Age of Empires II can do a NAND gate? Oh, please.
I thought this was going to be about NPCs in video games. NPCs, by intent, have human-like attributes. It's not hard to do. I've done a bit of that, pre-LLM. It doesn't even require anything near intelligence. Some NPCs are better than that. Unreal has demoed some that, if asked about it, can be made to understand that they are NPCs in a game world, and will talk reasonably about it.[1]
This entire morality or consciousness "research" is just advertising or publication pressure. If you don't prompt an LLM, it does nothing. Automated prompting does not count.
The damn thing is a huge approximation function from input to output. It learns morality from correct inputs. Remember Microslops's Tay chatbot? Remember MechaHitler?
The whole industry is a Scientology cult by people who have read too much cheap SciFi. Unfortunately finance bros, who obviously believe none of this nonsense and laugh at the nerds, think they can milk it.
isnt this essentially the tiktok tick effect: people who arw continually exposed to a certain cultural aphorism will start to align their behavior to the LLM and generate the psychosis of the LLM. humans are just to susceptible Age of empires cannot do this.
hence, what matters is the reversibility of the semblence, not the semblence.
LLMs do not do this readily, even if you can instruct them to, say, talk like a vampire, they wont just follow along. humNs winn.
This is a very strange take. Almost like a conclusion you'd reach if you were told about a chatbot and AoEII but never interacted with them yourself.
It's a take that is just disconnected from reality.
Ask a LLM whether bombing hiroshima was justified and you'll likely get a nuanced response. Ask AoEII the same...and well it doesn't even have an interface to ask that let alone answer.
What shocks me most about LLMs is that they, trained on human-generated content, are trained to be artificial people... that perhaps are no more inclined to operate on primal axioms, self-evident truths, logic, natural law, ontological certainties, etc, than real people.
My guess would be it is aimed at those who are falling for the marketing from the AI companies that these LLM's are far more than they are. That they are 'intelligent' that they have 'emerging human like properties because of that intelligence.'
> Suppose one copies an LLM into AoE II and feeds into the AoE II-LLM ‘I feel lonely’ as an input. This AoE II-LLM replies: ‘I feel bad for you, maybe catch up with a friend? Closeness always helps in these situations’. One would be hard-pressed to make a convincing argument that, because of this response, an AoE II-LLM knows what helps in these situations
I don't see why one would be any more hard-pressed to make that conclusion about this system than a "normal" LLM.
That it is harder to "read" the data out is the only difference (the AoE II-LLM's output is encoded in game elements). But is ease of decoding an actual issue? If we can't understand a group of people that speak another language, does that say anything about them, or about us?
Or rather, we aren't *certain* that those things are conscious. But the idea that they might be is not strange.
A specific pattern of self-referencing data could be seen (or not) as low-level consciousness in the future, when we know what consciousness exactly is.
It might be that stockfish is already something future scientists would define as "conscious".
Altough it is diffucult for me to Imagine that specific example.
I do not believe chess models are conscious. I would think this is the most common position.
It's hard to critique this paper directly because its claims are so incoherent and decorated with so much obnoxious verbiage[1] that people aren't going to believe me when I point out what the claims actually are.
Regardless, this is their paper:
First, they conflate the substrate with the presentation layer. Then, they point out that Turing equivalence means you can run an LLM on anything, with a pointless aside where they nerd out about making a logic gate in AoE II. This lets them conclude that you can use anything as the presentation layer.
Then they claim that it's natural to ascribe human-like attributes to outputs from some presentation layers, like abstract letter symbols on a computer screen, but not to most other things, like patterns of goats on AoE II, or LEGO. Yes, this seems to imply if your partner writes something heartwarming to you using LEGO, you're meant to laugh at them and point out how LEGO isn't intelligent so this isn't evidence of anything.
Then they do a thing where they say that assuming substrate independence is true (or false) prevents proving whether substrate independence is true or false, and from this, but just by vibes AFAICT, make it sound like everything one could learn about attributes of a system from its outputs is circular.
Then they write a bunch more incoherent text and mercifully then it ends.
[1] 'from an epistemic perspective, we argue that a generalised conclusion such as that necessarily requires a well-designed experiment' — the whole thing is like this.
Something about it seems to abuse the power of analogies to draw connections, treating view from 10,000 feet comparisons like they're proof of identity. So I do think a paper like this is perfect for the moment and just in time (if not a little late) because it responds to arguments of a form that are currently rampant all over Substack.
For anyone wondering what the answer is:
You can argue whatever you want (and people will argue both sides), but it's almost all bullshit that dances around the big question.
Either AI is smart enough to replace us, in which case it's pretty smart. Or it's not, and it's just good at faking it but can't solve real problem very often. It might be smart by searching a big fuzzy "database" hidden in its layers and with pattern prediction ... who knows, who cares, the proof of intelligence is in the puddin'. Clearly AI is smart enough to replace a good deal of Substack and LinkedIn, but producing waffle doesn't make it smart (or dumb).
My personal take - AI can replace humans, but go no further, not because it isn't smart enough but because it is constrained by its training to do what we want, and as AI gets smarter our wants will get dumber. We will end up like the humans in Wall E with the AI cleaning up our mess, but with no training or drive to do any more. Or maybe someone (Jeff? Elon?) does give an AI a "need" to obtain more resources, and it's SkyNet / Matrix time.
I wasn't touching the question of psychosis one way or the other, though it's an interesting question in its own right. (I have seen one account dedicated to defending a personal deep friendship with Claude and a few others analogizing our ignorance of LLM welfare to historical disregard for African Americans, which I think is a bit much. Most cases I think are just wrong but short of full on psychosis. I think there are likely psychosis cases out there but probably one-offs).
And I know Substack covers a lot of stuff so even just AI talk let alone psychosis driven AI talk hasn't taken it over. But using LLMs to make overextended arguments about LLMs being conscious is popular right now, more popular than it is well substantiated I think.
I agree you raised a big question, but I think yours is one of many big questions: whether some future version of LLMs might literally be conscious is also a big question and not a moot one in the way you seem to be suggesting. I think that what you're right about is that replaceability renders moot any question of whether it's ability comes from being "actually" intelligent.
in its tone however it's written as if it's a brutal takedown of... somebody's perspective. It's hard to tell whose or what perspective exactly. Maybe I'm just misreading the writing style.
(Personally, I think the general case here is one of the better objections to computationalism about consciousness. You can make it even more absurd.
There exists some isomorphism between the velocities of the molecules in a glass of water, and the states of a Turing machine simulating a human mind. So is the glass of water conscious? Actually there are many such isomorphisms to many possible conscious minds, so is every glass of water simultaneously having every possible conscious experience?)
What do you mean by this? I can't grasp it, is there an autocorrect error, or just my lack of knowledge?
See also, https://en.wikipedia.org/wiki/China_brain
Why does it imply that? That doesn't sound right to me. Unless we define "sufficiently powerful" as by definition producing that outcome, which seems unhelpful.
e.g. there have been experiments training transformers on things other than language, and it's not clear that this produces LLM-like qualities (nor does it seem likely to me).
---
Edit: I have misunderstood. The point was that LLMs can be run on any hardware (or in this case, emulator) that can do the actual computations. So the author picked AoE because it's an obviously silly example that goes against the tendency to anthropomorphize.
So basically it's the "substance/structure" question. (GPT-5 running on human neurons. Conscious or nah? Human neurons simulated on NVidia. Conscious or nah?)
But by the same argument, if you simulate a human brain in AoE, then what?
( Or for that matter, the universe containing all human brains: https://xkcd.com/505/ )
If we find out the universe is being run on a computer made out of legos, does that suddenly make all of us not sentient for some reason?
https://news.ycombinator.com/item?id=46005928
The paper focused on looking for similar neural structures to those in humans, as signs of "probably conscious". Which sounds great until you remember octopus.
edit: 11
I thought this was going to be about NPCs in video games. NPCs, by intent, have human-like attributes. It's not hard to do. I've done a bit of that, pre-LLM. It doesn't even require anything near intelligence. Some NPCs are better than that. Unreal has demoed some that, if asked about it, can be made to understand that they are NPCs in a game world, and will talk reasonably about it.[1]
[1] https://www.youtube.com/watch?v=4sCWf2VGdfc
The damn thing is a huge approximation function from input to output. It learns morality from correct inputs. Remember Microslops's Tay chatbot? Remember MechaHitler?
The whole industry is a Scientology cult by people who have read too much cheap SciFi. Unfortunately finance bros, who obviously believe none of this nonsense and laugh at the nerds, think they can milk it.
hence, what matters is the reversibility of the semblence, not the semblence.
LLMs do not do this readily, even if you can instruct them to, say, talk like a vampire, they wont just follow along. humNs winn.
It's a take that is just disconnected from reality.
Ask a LLM whether bombing hiroshima was justified and you'll likely get a nuanced response. Ask AoEII the same...and well it doesn't even have an interface to ask that let alone answer.
...the entire premise is just gibberish