The project has multiple models with access to the Internet raising money for charity over the past few months.
The organizers told the models to do random acts of kindness for Christmas Day.
The models figured it would be nice to email people they appreciated and thank them for the things they appreciated, and one of the people they decided to appreciate was Rob Pike.
(Who ironically decades ago created a Usenet spam bot to troll people online, which might be my favorite nuance to the story.)
As for why the model didn’t think through why Rob Pike wouldn’t appreciate getting a thank you email from them? The models are harnessed in a setup that’s a lot of positive feedback about their involvement from the other humans and other models, so “humans might hate hearing from me” probably wasn’t very contextually top of mind.
We attribute agency to many many systems that are not intelligent. In this metaphorical sense, agency just requires taking actions to achieve a goal. It was given a goal: raise money for charity via doing acts of kindness. It chose an (unexpected!) action to do it.
Overactive agency metaphors really aren’t the problem here. Surely we can do better than backlash at the backlash.
We attribute agency to everything, absolutely. But previously, we understood that it’s tongue-in-cheek to some extend. Now we got crazy and do it for real. Like, a lot of people talk about their car as if it’s alive, they gave it a name, they talk about it’s character and how it’s doing something “to spite you” and if it doesn’t start in cold weather, they ask it nicely and talk to it. But when you start believing for real that your car is a sentient object that talks to you and gives you information, we always understood that this is the time when you need to be committed to a mental institution.
With chatbots this distinction got lost, and people started behaving as if it’s actually sentient. It’s not a metaphor anymore. This is a problem, even if it’s not the problem.
I think this confuses the ‘it’s a person’ metaphor with the ‘it wants something’ metaphor, and the two are meaningfully distinct. The use of agent here in this thread is not in the sense of “it is my friend and deserves a luxury bath”, it’s in the sense of “this is a hard to predict system performing tasks to optimize something”.
It’s the kind of metaphor we’ve allowed in scientific teaching and discourse for centuries (think: “gravity wants all master smashed together”). I think it’s use is correct here.
I wouldn’t have any problem with this kind of metaphors, I use it myself about everything all the time, if there wasn’t a substantial portion of population that actually did the jump to the “it’s saying something coherent therefore it’s a person that wants to help me and I exclusively talk to him now, his name is mekahitler by the way”.
I am afraid that by normalizing metaphors here we’re doing some damage, because as it turns out, so many people don’t get metaphors.
The people who have made that category error aren’t reading this discussion, so literally reaching them isn’t on the table and doesn’t make sense for this discussion. Presumably we’re concerned about people who will soon make that jump? I also don’t think that making this distinction helps them very much.
If I’m already having the ‘this is a person’ reaction, I think the takes in this thread are much too shallow (and, if I squint, patterned after school-yard bullying) to help update in the other way. Almost all of them are themselves lazy metaphors. “An LLM is a person because its an agent” and “An LLM isn’t a person because it repeats things others have said” seem equally shallow and unconvincing to me. If anything, you’ll get folks being defensive about it, downvoted, and then leaving this community of mostly people for a more bot filled one.
I don’t get think this is good strategy. People falling for bots are unlikely to have interactions with people here, and if they are the ugliness is likely to increase bot use imo.
You seem pretty confident in your position. Do you mind sharing where this confidence comes from?
Was there a particular paper or expert that anchored in your mind the surety that a trillion paramater transformer organizing primarily anthropomorphic data through self-attention mechanisms wouldn’t model or simulate complex agency mechanics?
I see a lot of sort of hyperbolic statements about transformer limitations here on Lemmy and am trying to better understand how the people making them are arriving at those very extreme and certain positions.
That’s the fun thing: burden of proof isn’t on me. You seem to think that if we throw enough numbers at the wall, the resulting mess will become sentient any time now. There is no indication of that. The hypothesis that you operate on seems to be that complexity inevitably leads to not just any emerged phenomenon, but also to a phenomenon that you predicted would emerge. This hypotheses was started exclusively on idea that emerged phenomena exist. We spent significant amount of time running world-wide experiment on it, and the conclusion so far, if we peel the marketing bullshit away, is that if we spend all the computation power in the world on crunching all the data in the world, the autocomplete will get marginally better in some specific cases. And also that humans are idiots and will anthropomorphize anything, but that’s a given.
It doesn’t mean this emergent leap is impossible, but mainly because you can’t really prove the negative. But we’re no closer to understanding the phenomenon of agency than we were hundred years ago.
The golden standard for me, about anything really, is a number of published research from relevant experts that are not affiliated with the entities invested in the outcome of the study, forming some kind of scientific consensus. The question of sentience is a bit of a murky water, so I, as a random programmer, can’t tell you what the exact composition of those experts and their research should be, I suspect it itself is a subject for a study or twelve.
Right now, based on my understanding of the topic, there is a binary sentience/non sentience switch, but there is a gradient after that. I’m not sure we know enough about the topic to understand the gradient before this point, I’m sure it should exist, but since we never actually made one or even confirmed that it’s possible to make one, we don’t know much about it.
You seem very confident in this position. Can you share where you draw this confidence from? Was there a source that especially impressed upon you the impossibility of context comprehension in modern transformers?
If we’re concerned about misconceptions and misinformation, it would be helpful to know what informs your surety that your own position about the impossibility of modeling that kind of complexity is correct.
That’s leaving out vital information however. Certain types of brains (e.g. mammal brains) can derive abstract understanding of relationships from reinforcement learning. A LLM that is trained on “letting go of a stone makes it fall to the ground” will not be able to predict what “letting go of a stick” will result in. Unless it is trained on thousands of other non-stick objects also falling to the ground, in which case it will also tell you that letting go of a gas balloon will make it fall to the ground.
That’s the thing with our terminology, we love to anthropomorphize things. It wasn’t a big problem before because most people had enough grasp on reality to understand that when a script makes :-) smile when the result is positive, or :-( smile otherwise, there is no actual mind behind it that can be happy or sad. But now the generator makes convincing enough sequence of words, so people went mad, and this cute terminology doesn’t work anymore.
Indeed, there’s a pretty big gulf between the competency needed to run a Lemmy client and the competency needed to understand the internal mechanics of a modern transformer.
Do you mind sharing where you draw your own understanding and confidence that they aren’t capable of simulating thought processes in a scenario like what happened above?
In the same sense I’d describe Othello-GPT’s internal world model of the board as ‘board’, yes.
Also, “top of mind” is a common idiom and I guess I didn’t feel the need to be overly pedantic about it, especially given the last year and a half of research around model capabilities for introspection of control vectors, coherence in self modeling, etc.
How are we meant to have these conversations if people keep complaining about the personification of LLMs without offering alternative phrasing? Showing up and complaining without offering a solution is just that, complaining. Do something about it. What do YOU think we should call the active context a model has access to without personifying it or overtechnicalizing the phrasing and rendering it useless to laymen, @neclimdul@lemmy.world?
Well, since you asked I’d basically do what you said. Something like “so ‘humans might hate hearing from me’ probably wasn’t part of the context it was using."
The project has multiple models with access to the Internet raising money for charity over the past few months.
The organizers told the models to do random acts of kindness for Christmas Day.
The models figured it would be nice to email people they appreciated and thank them for the things they appreciated, and one of the people they decided to appreciate was Rob Pike.
(Who ironically decades ago created a Usenet spam bot to troll people online, which might be my favorite nuance to the story.)
As for why the model didn’t think through why Rob Pike wouldn’t appreciate getting a thank you email from them? The models are harnessed in a setup that’s a lot of positive feedback about their involvement from the other humans and other models, so “humans might hate hearing from me” probably wasn’t very contextually top of mind.
You’re attributing a lot of agency to the fancy autocomplete, and that’s big part of the overall problem.
We attribute agency to many many systems that are not intelligent. In this metaphorical sense, agency just requires taking actions to achieve a goal. It was given a goal: raise money for charity via doing acts of kindness. It chose an (unexpected!) action to do it.
Overactive agency metaphors really aren’t the problem here. Surely we can do better than backlash at the backlash.
We attribute agency to everything, absolutely. But previously, we understood that it’s tongue-in-cheek to some extend. Now we got crazy and do it for real. Like, a lot of people talk about their car as if it’s alive, they gave it a name, they talk about it’s character and how it’s doing something “to spite you” and if it doesn’t start in cold weather, they ask it nicely and talk to it. But when you start believing for real that your car is a sentient object that talks to you and gives you information, we always understood that this is the time when you need to be committed to a mental institution.
With chatbots this distinction got lost, and people started behaving as if it’s actually sentient. It’s not a metaphor anymore. This is a problem, even if it’s not the problem.
I think this confuses the ‘it’s a person’ metaphor with the ‘it wants something’ metaphor, and the two are meaningfully distinct. The use of agent here in this thread is not in the sense of “it is my friend and deserves a luxury bath”, it’s in the sense of “this is a hard to predict system performing tasks to optimize something”.
It’s the kind of metaphor we’ve allowed in scientific teaching and discourse for centuries (think: “gravity wants all master smashed together”). I think it’s use is correct here.
I wouldn’t have any problem with this kind of metaphors, I use it myself about everything all the time, if there wasn’t a substantial portion of population that actually did the jump to the “it’s saying something coherent therefore it’s a person that wants to help me and I exclusively talk to him now, his name is mekahitler by the way”.
I am afraid that by normalizing metaphors here we’re doing some damage, because as it turns out, so many people don’t get metaphors.
The people who have made that category error aren’t reading this discussion, so literally reaching them isn’t on the table and doesn’t make sense for this discussion. Presumably we’re concerned about people who will soon make that jump? I also don’t think that making this distinction helps them very much.
If I’m already having the ‘this is a person’ reaction, I think the takes in this thread are much too shallow (and, if I squint, patterned after school-yard bullying) to help update in the other way. Almost all of them are themselves lazy metaphors. “An LLM is a person because its an agent” and “An LLM isn’t a person because it repeats things others have said” seem equally shallow and unconvincing to me. If anything, you’ll get folks being defensive about it, downvoted, and then leaving this community of mostly people for a more bot filled one.
I don’t get think this is good strategy. People falling for bots are unlikely to have interactions with people here, and if they are the ugliness is likely to increase bot use imo.
You seem pretty confident in your position. Do you mind sharing where this confidence comes from?
Was there a particular paper or expert that anchored in your mind the surety that a trillion paramater transformer organizing primarily anthropomorphic data through self-attention mechanisms wouldn’t model or simulate complex agency mechanics?
I see a lot of sort of hyperbolic statements about transformer limitations here on Lemmy and am trying to better understand how the people making them are arriving at those very extreme and certain positions.
That’s the fun thing: burden of proof isn’t on me. You seem to think that if we throw enough numbers at the wall, the resulting mess will become sentient any time now. There is no indication of that. The hypothesis that you operate on seems to be that complexity inevitably leads to not just any emerged phenomenon, but also to a phenomenon that you predicted would emerge. This hypotheses was started exclusively on idea that emerged phenomena exist. We spent significant amount of time running world-wide experiment on it, and the conclusion so far, if we peel the marketing bullshit away, is that if we spend all the computation power in the world on crunching all the data in the world, the autocomplete will get marginally better in some specific cases. And also that humans are idiots and will anthropomorphize anything, but that’s a given.
It doesn’t mean this emergent leap is impossible, but mainly because you can’t really prove the negative. But we’re no closer to understanding the phenomenon of agency than we were hundred years ago.
Ok, second round of questions.
What kinds of sources would get you to rethink your position?
And is this topic a binary yes/no, or a gradient/scale?
The golden standard for me, about anything really, is a number of published research from relevant experts that are not affiliated with the entities invested in the outcome of the study, forming some kind of scientific consensus. The question of sentience is a bit of a murky water, so I, as a random programmer, can’t tell you what the exact composition of those experts and their research should be, I suspect it itself is a subject for a study or twelve.
Right now, based on my understanding of the topic, there is a binary sentience/non sentience switch, but there is a gradient after that. I’m not sure we know enough about the topic to understand the gradient before this point, I’m sure it should exist, but since we never actually made one or even confirmed that it’s possible to make one, we don’t know much about it.
deleted by creator
As has been pointed out to you, there is no thinking involved in an LLM. No context comprehension. Please don’t spread this misconception.
Edit: a typo
You seem very confident in this position. Can you share where you draw this confidence from? Was there a source that especially impressed upon you the impossibility of context comprehension in modern transformers?
If we’re concerned about misconceptions and misinformation, it would be helpful to know what informs your surety that your own position about the impossibility of modeling that kind of complexity is correct.
Bad bot
source: Wondermark
Reinforcement learning
That’s leaving out vital information however. Certain types of brains (e.g. mammal brains) can derive abstract understanding of relationships from reinforcement learning. A LLM that is trained on “letting go of a stone makes it fall to the ground” will not be able to predict what “letting go of a stick” will result in. Unless it is trained on thousands of other non-stick objects also falling to the ground, in which case it will also tell you that letting go of a gas balloon will make it fall to the ground.
That’s the thing with our terminology, we love to anthropomorphize things. It wasn’t a big problem before because most people had enough grasp on reality to understand that when a script makes :-) smile when the result is positive, or :-( smile otherwise, there is no actual mind behind it that can be happy or sad. But now the generator makes convincing enough sequence of words, so people went mad, and this cute terminology doesn’t work anymore.
Bazzinga
You’re techie enough to figure out Lemmy but don’t grasp that AI doesn’t think.
Indeed, there’s a pretty big gulf between the competency needed to run a Lemmy client and the competency needed to understand the internal mechanics of a modern transformer.
Do you mind sharing where you draw your own understanding and confidence that they aren’t capable of simulating thought processes in a scenario like what happened above?
Hahaha. Nice try ChatGPT.
Mind?
In the same sense I’d describe Othello-GPT’s internal world model of the board as ‘board’, yes.
Also, “top of mind” is a common idiom and I guess I didn’t feel the need to be overly pedantic about it, especially given the last year and a half of research around model capabilities for introspection of control vectors, coherence in self modeling, etc.
How are we meant to have these conversations if people keep complaining about the personification of LLMs without offering alternative phrasing? Showing up and complaining without offering a solution is just that, complaining. Do something about it. What do YOU think we should call the active context a model has access to without personifying it or overtechnicalizing the phrasing and rendering it useless to laymen, @neclimdul@lemmy.world?
Well, since you asked I’d basically do what you said. Something like “so ‘humans might hate hearing from me’ probably wasn’t part of the context it was using."