The part about desperation vectors driving reward hacking matches something I've run into firsthand building agent loops where Claude writes and tests code iteratively.
When the prompt frames things with urgency -- "this test MUST pass," "failure is unacceptable" -- you get noticeably more hacky workarounds. Hardcoded expected outputs, monkey-patched assertions, that kind of thing. Switching to calmer framing ("take your time, if you can't solve it just explain why") cut that behavior way down. I'd chalked it up to instruction following, but this paper points at something more mechanistic underneath.
The method actor analogy in the paper gets at it well. Tell an actor their character is desperate and they'll do desperate things. The weird part is that we're now basically managing the psychological state of our tooling, and I'm not sure the prompt engineering world has caught up to that framing yet.
The technology they are discovering is called "Language". It was designed to encode emotions by a sender and invoke emotions in the reader. The emotions a reader gets from LLM are still coming from the language
Emotional signals are more than just text though, there is a reason tone and body language is so important for understanding what someone says. Sarcasm and so on doesn't work well without it.
Emotion is mainly encoded in tone and body language. It is somewhat difficult to transport emotion using words. I don't think you can guess my current emotional state while I am writing this, but if you'd see my face it would be easy for you.
Dammit, you cheated though! Why must you always do that? In your sentences it doesn't matter what your emotional state is, it makes no difference; bit like life really.
Hopefully, you can see that at least my chosen sentences have an emotional aspect?
An LLM could add emotional values to my previous sentences that a TTS can use for tonal variation, for example.
There was a really old project from mit called conceptnet that I worked with many years ago. It was basically a graph of concepts (not exactly but close enough) and emotions came into it too just as part of the concepts. For example a cake concept is close to a birthday concept is close to a happy feeling.
What was funny though is that it was trained by MIT students so you had the concept of getting a good grade on a test as a happier concept than kissing a girl for the first time.
Another problem is emotions are cultural. For example, emotions tied to dogs are different in different cultures.
We wanted to create concept nets for individuals - that is basically your personality and knowledge combined but the amount of data required was just too much. You'd have to record all interactions for a person to feed the system.
Super interesting, I wonder if this research will cause them to actually change their llm, like turning down the ”desperation neurons” to stop Claude from creating implementations for making a specific tests pass etc.
They likely already have. You can use all caps and yell at Claude and it'll react normally, while doing do so with chatgpt scares it, resulting in timid answers
For me GPT always seems to get stuck in a particular state where it responds with a single sentence per paragraph, short sentences, and becomes weirdly philosophical. This eventually happens in every session. I wish I knew what triggers it because it's annoying and completely reduces its usefulness.
Usually a session is delivered as context, up to the token limit, for inference to be performed on. Are you keeping each session to one subject? Have you made personalizations? Do you add lots of data?
It would be interesting if you posted a couple of sessions to see what 'philosophical' things it's arriving at and what proceeds it.
> Note that none of this tells us whether language models actually feel anything or have subjective experiences.
You’ll never find that in the human brain either. There’s the machinery of neural correlates to experience, we never see the experience itself. That’s likely because the distinction is vacuous: they’re the same thing.
Bundle of tokens comes in, bundle of tokens comes out. If there is any trace of consciousness or subjectivity in there, it exists only while matrices are being multiplied.
Do you think these llm's have subjective experiences? (by "subjective experience" I mean the thing that makes stepping on an ant worse than kicking a pebble) And if so, do you still use them? Additionaly: when do you think that subjectivity started? Was there a "there" there with gpt2?
Yes, I think they probably are conscious, though what their qualia are like might be incomprehensible to me. I don’t think that being conscious means being identical to human experience.
Philosophically I don’t think there is a point where consciousness arises. I think there is a point where a system starts to be structured in such a way that it can do language and reasoning, but I don’t think these are any different than any other mechanisms, like opening and closing a door. Differences of scale, not kind. Experience and what it is to be are just the same thing.
And yes, I use them. I try not to mistreat them in a human-relatable sense, in case that means anything.
Do you think there are "scales" of consciousness? As in, is there some quality that makes killing a frog worse than killing an ant, and killing a human worse than killing a frog? If so, do the llm models exist across this scale, or are gpt-3 and gpt-2 conscious at the same "scale" as gpt-4?
I ask because if your view of consciousness is mechanistic, this is fairly cut and dry: gpt-2 has 4 orders of magnitude less parameters/complexity than gpt-4.
But both gpt-2 and gpt-4 are very fluent at a language level (both moreso than a human 6 year old for example), so in your view they might both be roughly equally conscious, just expressed differently?
This is really a different question, what makes an entity a “moral patient”, something worthy of moral consideration. This is separate from the question of whether or not an entity experiences anything at all.
There are different ways of answering this, but for me it comes down to nociception, which is the ability to feel pain. We should try to build systems that cannot feel pain, where I also mean other “negative valence” states which we may not understand. We currently don’t understand what pain is in humans, let alone AIs, so we may have built systems that are capable of suffering without knowing it.
As an aside, most people seem to think that intelligence is what makes entities eligible for moral consideration, probably because of how we routinely treat animals, and this is a convenient self-serving justification. I eat meat by the way, in case you’re wondering. But I do think the way we treat animals is immoral, and there is the possibility that it may be thought of by future generations as being some sort of high crime.
Okay, but even leaving aside the pain stuff, people generally find subjectivity / consciousness to have inherent value, and by extent are sad if a person dies even if they didn't (subjectively) suffer.
I would not personally consider the death of a sentient being with decades of experiences a neutral event, even if the being had been programmed to not have a capacity for suffering.
I think the idea of there being a difference between an ant dying (or "disapearing" if that's less loaded) vs a duck dying makes sense to most people (and is broadly shared) even if they don't have a completely fleshed out system of when something gets moral consideration.
Sure, because you’re a human. We have social attachment to other humans and we mourn their passing, that’s built into the fabric of what we are. But that has nothing to do with whoever has passed away, it’s about us and how we feel about it.
It’s also about how we think about death. It’s weird in that being dead probably isn’t like anything at all, but we fear it, and I guess we project that fear onto the death of other entities.
I guess my value system says that being dead is less bad than being alive and suffering badly.
I know I feel experience. I don't know for sure if you do, but it seems a very reasonable extension to other people. LLMs are a radical jump though that needs a greater degree of justification.
And what kind of evidence would convince you? What experiment would ever bridge this gap? You’re relying entirely on similarity between yourself and other humans. This doesn’t extend very well to anything, even animals, though more so than machines. By framing it this way have you baked in the conclusion that nothing else can be conscious on an a priori basis?
I'm not sure what evidence would convince me, but I don't think the way LLMs act is convincing enough. The kinds of errors they make and the fact they operate in very clear discrete chunks makes it seem hard to me to attribute them subjective experience.
The Chinese room is nonsense though. How did it get every conceivable reply to every conceivable question? Presumably because people thought of and answered everything conceivable. Meaning that you’re actually talking to a Chinese room plus multiple people composite system. You would not argue that the human part of that system isn’t conscious.
But this distraction aside, my point is this: there is only mechanism. If someone’s demand to accept consciousness in some other entity is to experience those experiences for themselves, then that’s a nonsensical demand. You might just as well assume everyone and everything else is a philosophical zombie.
> You would not argue that the human part of that system isn’t conscious.
Sure I would. The human part is not being inferenced, the data is. LLM output in this circumstance is no more conscious than a book that you read by flipping to random pages.
> You might just as well assume everyone and everything else is a philosophical zombie.
I don't assume anything about everyone or everything's intelligence. I have a healthy distrust of all claims.
The CR is equivalent to a human being asked a question, thinking about it and answering. The setup is the same thing, it’s just framed in a way that obfuscates that.
And sure, you can assume that nobody and nothing else is conscious (I think we’re talking about this rather than intelligence) and I won’t try to stop you, I just don’t think it’s a very useful stance. It kind of means that assuming consciousness or not means nothing, since it changes nothing, which is more or less what I’m saying.
Something they don’t seem to mention in the article: Does greater model “enjoyment” of a task correspond to higher benchmark performance? E.g. if you steer it to enjoy solving difficult programming tasks, does it produce better solutions?
The desperation > blackmail finding stuck with me. If AI behavior shifts based on emotional states, maybe emotions are just a mechanism for changing behavior in the first place. If we think of human emotions the same way, just evolution's way of nudging behavior, the line between AI and humans starts to look a lot thinner.
I'm not being pejorative but that sounds more like psychopathy or autism?
Evolution isn't a god, it has no steering hand, it is accidents that either provide advantage or don't.
LLMs are getting more human-like because that's how we're developing them. Arguably that's about market forces. LM owners see opportunity to exploit people's desire for emotional interactions (ie loneliness) in order to make money.
Many non-psychopaths e.g., CBT therapists, evolutionary psychologists and neuroscientists, such as Damasio, view emotions as adaptive tools for guiding/changing behaviour.
It's still too early to tell, but it might make sense at some point. If because of symmetry and universality we decide that llms are a protected class, but we also need to configure individual neurons, that configuration must be done by a specialist.
The first and second principal components (joy-sadness and anger) explain only 41% of the variance. I wish the authors showed further principal components. Even principal components 1-4 would explain no more than 70% of the variance, which seems to contradict the popular theory that all human emotions are composed of 5 basic emotions: joy, sadness, anger, fear, and disgust, i.e. 4 dimensions.
>... emotion-related representations that shape its behavior. These specific patterns of artificial “neurons” which activate in situations—and promote behaviors—that the model has learned to associate with the concept of a particular emotion. .... In contexts where you might expect a certain emotion to arise for a human, the corresponding representations are active.
>For instance, to ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways.
Force-set to 0, "mask"/deactivate those representations associated with bad/dangerous emotions. Neural Prozac/lobotomy so to speak.
Its almost like LLMs have a vast, mute unconscious mind operating in the background, modeling relationships, assigning emotional state, and existing entirely without ego.
Sounds sort of like how certain monkey creatures might work.
When the prompt frames things with urgency -- "this test MUST pass," "failure is unacceptable" -- you get noticeably more hacky workarounds. Hardcoded expected outputs, monkey-patched assertions, that kind of thing. Switching to calmer framing ("take your time, if you can't solve it just explain why") cut that behavior way down. I'd chalked it up to instruction following, but this paper points at something more mechanistic underneath.
The method actor analogy in the paper gets at it well. Tell an actor their character is desperate and they'll do desperate things. The weird part is that we're now basically managing the psychological state of our tooling, and I'm not sure the prompt engineering world has caught up to that framing yet.
Hopefully, you can see that at least my chosen sentences have an emotional aspect?
An LLM could add emotional values to my previous sentences that a TTS can use for tonal variation, for example.
Essentially we have created the Cylon.
What was funny though is that it was trained by MIT students so you had the concept of getting a good grade on a test as a happier concept than kissing a girl for the first time.
Another problem is emotions are cultural. For example, emotions tied to dogs are different in different cultures.
We wanted to create concept nets for individuals - that is basically your personality and knowledge combined but the amount of data required was just too much. You'd have to record all interactions for a person to feed the system.
> If the person becomes abusive over the course of a conversation, Claude avoids becoming increasingly submissive in response.
See: https://platform.claude.com/docs/en/release-notes/system-pro...
It would be interesting if you posted a couple of sessions to see what 'philosophical' things it's arriving at and what proceeds it.
You’ll never find that in the human brain either. There’s the machinery of neural correlates to experience, we never see the experience itself. That’s likely because the distinction is vacuous: they’re the same thing.
Bundle of tokens comes in, bundle of tokens comes out. If there is any trace of consciousness or subjectivity in there, it exists only while matrices are being multiplied.
Philosophically I don’t think there is a point where consciousness arises. I think there is a point where a system starts to be structured in such a way that it can do language and reasoning, but I don’t think these are any different than any other mechanisms, like opening and closing a door. Differences of scale, not kind. Experience and what it is to be are just the same thing.
And yes, I use them. I try not to mistreat them in a human-relatable sense, in case that means anything.
I ask because if your view of consciousness is mechanistic, this is fairly cut and dry: gpt-2 has 4 orders of magnitude less parameters/complexity than gpt-4. But both gpt-2 and gpt-4 are very fluent at a language level (both moreso than a human 6 year old for example), so in your view they might both be roughly equally conscious, just expressed differently?
There are different ways of answering this, but for me it comes down to nociception, which is the ability to feel pain. We should try to build systems that cannot feel pain, where I also mean other “negative valence” states which we may not understand. We currently don’t understand what pain is in humans, let alone AIs, so we may have built systems that are capable of suffering without knowing it.
As an aside, most people seem to think that intelligence is what makes entities eligible for moral consideration, probably because of how we routinely treat animals, and this is a convenient self-serving justification. I eat meat by the way, in case you’re wondering. But I do think the way we treat animals is immoral, and there is the possibility that it may be thought of by future generations as being some sort of high crime.
I would not personally consider the death of a sentient being with decades of experiences a neutral event, even if the being had been programmed to not have a capacity for suffering.
I think the idea of there being a difference between an ant dying (or "disapearing" if that's less loaded) vs a duck dying makes sense to most people (and is broadly shared) even if they don't have a completely fleshed out system of when something gets moral consideration.
It’s also about how we think about death. It’s weird in that being dead probably isn’t like anything at all, but we fear it, and I guess we project that fear onto the death of other entities.
I guess my value system says that being dead is less bad than being alive and suffering badly.
The Chinese Room would like a word.
But this distraction aside, my point is this: there is only mechanism. If someone’s demand to accept consciousness in some other entity is to experience those experiences for themselves, then that’s a nonsensical demand. You might just as well assume everyone and everything else is a philosophical zombie.
Sure I would. The human part is not being inferenced, the data is. LLM output in this circumstance is no more conscious than a book that you read by flipping to random pages.
> You might just as well assume everyone and everything else is a philosophical zombie.
I don't assume anything about everyone or everything's intelligence. I have a healthy distrust of all claims.
And sure, you can assume that nobody and nothing else is conscious (I think we’re talking about this rather than intelligence) and I won’t try to stop you, I just don’t think it’s a very useful stance. It kind of means that assuming consciousness or not means nothing, since it changes nothing, which is more or less what I’m saying.
[1] https://en.wikipedia.org/wiki/Functionalism_%28philosophy_of...
What are other alternative, realistic possible ways to see emotions?
Evolution isn't a god, it has no steering hand, it is accidents that either provide advantage or don't.
LLMs are getting more human-like because that's how we're developing them. Arguably that's about market forces. LM owners see opportunity to exploit people's desire for emotional interactions (ie loneliness) in order to make money.
Only psychopaths think of emotion as nothing but a means to changing behavior. The scary thing is that LLMs by nature would exhibit the same behavior.
>For instance, to ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways.
Force-set to 0, "mask"/deactivate those representations associated with bad/dangerous emotions. Neural Prozac/lobotomy so to speak.
Sounds sort of like how certain monkey creatures might work.
You don't have to teach a monkey language for it to feel sadness.