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.
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.
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.
> 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.
I assume you say that in jest, but back in the early '90s I was seriously considering getting a major in psychology and a minor in CS for the fairly hot Human Factors jobs.
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.
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
>... 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.
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 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.
I remember when people were discussing the “performance-improving” hack of formulating their prompts as panicked pleas to save their job and household and puppy from imminent doom…by coding X. I wonder if the backfiring is a more recent phenomenon in models that are better at “following the prompt” (including the logical conclusion of its emotional charge), or it was just bad quantification of “performance” all along.
I use positive framing instead of negative framing for most things and get good results. Especially where asking for a thing to not happen, pollutes the context with that thing.
A bad example, but imagine "Build me a wrapper for this API but ABSOLUTELY DO NOT use javascript" versus "Build me a wrapper for this API and make sure to use python".
> Since these representations appear to be largely inherited from training data, the composition of that data has downstream effects on the model’s emotional architecture. Curating pretraining datasets to include models of healthy patterns of emotional regulation—resilience under pressure, composed empathy, warmth while maintaining appropriate boundaries—could influence these representations, and their impact on behavior, at their source.
What better source of healthy patterns of emotional regulation than, uhhh, Reddit?
Trying to separate the software from the hardware is a fool's errand in this case: emotions are primarily an hormonal response, not an intellectual one.
Whenever I come to HN I see a bunch of people say LLMs are just next token predictors and they completely understand LLMs. And almost every one of these people are so utterly self assured to the point of total confidence because they read and understand what transformers do.
Then I watch videos like this straight from the source trying to understand LLMs like a black box and even considering the possibility that LLMs have emotions.
How does such a person reconcile with being utterly wrong? I used to think HN was full of more intelligent people but it’s becoming more and more obvious that HNers are pretty average or even below.
As other posters have pointed, the core of a LLM is a pure function, which computes a token probability distribution from an input context.
An automaton, which can chat with you or write a program, is built externally to the LLM function, by storing the context and making it change, depending on the output of the LLM function.
However, the LLM pure function is exceedingly complex so it is essentially unpredictable what will it produce for a given input context.
So one may have to treat the LLM function as a black box and explore the huge space of the input contexts by varying them in various ways, inclusive by using words that express human emotions, and monitor how the output of the function changes, i.e. how the LLM "reacts" to the expressed emotions.
A "reaction" similar to that of a human is to be expected, because human emotions were expressed in the training texts, followed by reactions of humans to those emotions, and the LLM function will change its output token probability function in a manner mimicking the behavior of the humans from the training texts.
Even functions that are many orders of magnitude simpler than LLMs are still to complex for anyone to understand how their output changes when you move through the space of the possible input arguments.
The most essential part of cryptography is the existence of a class of functions which were named by Claude Shannon "good mixing transformations". All the important cryptographic primitives, e.g. block cipher functions or one-way hash functions, are built from such "good mixing transformations". The impossibility of breaking a cryptographic system with secret keys is based on the assumption that it is impossible to predict how the output of such a "good mixing transformation" changes when its input is changed. All such "good mixing transformations" have the so-called avalanche property, which means that even if you change a single input bit, any of the output bits may change with a probability of exactly 50%, so it is unpredictable for any output bit whether it will change, or not.
If such simple functions, e.g. with 128 input bits and 128 output bits, can have a completely unpredictable behavior, then it is not surprising that LLM functions that may have an input of up to a few million bits (the length of the context window) are completely unpredictable and you can just observe their behavior when given various kinds of contexts and search for empirical approximate rules describing the behavior.
If you read carefully my point is not about the external behavior of the LLM. It is the black box aspect of the LLM. The sheer complexity of the pure function is not something we can understand even though the high level structure is a feed forward network the core algorithm is in actuality encoded by weights.
Yes there are complex functions besides LLMs that we don’t understand but those functions usually aren’t compelling because the LLM, unlike those other functions has output that implicates reasoning and emotions. The problem is we can’t understand what’s going on under the hood so we don’t know either way.
This is what I mean by stupidity. You completely missed the point, and you’re also operating under the assumption that the human brain is also not following a similar deterministic pathway. You hold humanity and biological intelligence in such high regard that you cannot even imagine that all of physics implies that human intelligence is mechanical. So the emotions you feel are under a black box same as the LLM and you apply you biased assumptions in a singular direction assuming your emotions are not deterministic and that LLM emotions are fake but that reasoning has no basis.
I think the findings that the LLM triggers “desperation” like emotions when it about to run out of tokens in a coding session have practical implications. The tasks needs to be planned, so that they are likely to be consistent before the session runs into limits, to avoid issues like LLM starts hardcoding values from a test harnesses into UI layer to make the tests pass.
Of course they do have emotions as an internal circuit or abstraction, this is fully expected from intelligence at least at some point. But interpreting these emotions as human-like is a clear blunder. How do you tell the shoggoth likes or dislikes something, feels desperation or joy? Because it said so? How do you know these words mean the same for us? Our internal states are absolutely incompatible. We share a lot of our "architecture" and "dataset" with some complex animals and even then we barely understand many of their emotions. What does a hedgehog feel when eating its babies? This thing is 100% unlike a hedgehog or a human, it exists in its own bizarre time projection and nothing of it maps to your state. It's a shapeshifting alien.
In mechinterp you're reducing this hugely multidimensional and incomprehensible internal state to understandable text using the lens of the dataset you picked. It's inevitably a subjective interpretation, you're painting familiar faces on a faceless thing.
Anthropic researchers are heavily biased to see what they want to see, this is the biggest danger in research.
(note: Friston is definitely considered a bit out there by ... everyone? But he makes some good points. And here he's getting referenced, so I guess some people grok him)
34 comments
[ 208 ms ] story [ 2194 ms ] threadSounds sort of like how certain monkey creatures might work.
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.
I think we basically do, the only interesting bit is our perception of phenomenal experiences.
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.
>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.
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.
A bad example, but imagine "Build me a wrapper for this API but ABSOLUTELY DO NOT use javascript" versus "Build me a wrapper for this API and make sure to use python".
Essentially we have created the Cylon.
What better source of healthy patterns of emotional regulation than, uhhh, Reddit?
Then I watch videos like this straight from the source trying to understand LLMs like a black box and even considering the possibility that LLMs have emotions.
How does such a person reconcile with being utterly wrong? I used to think HN was full of more intelligent people but it’s becoming more and more obvious that HNers are pretty average or even below.
An automaton, which can chat with you or write a program, is built externally to the LLM function, by storing the context and making it change, depending on the output of the LLM function.
However, the LLM pure function is exceedingly complex so it is essentially unpredictable what will it produce for a given input context.
So one may have to treat the LLM function as a black box and explore the huge space of the input contexts by varying them in various ways, inclusive by using words that express human emotions, and monitor how the output of the function changes, i.e. how the LLM "reacts" to the expressed emotions.
A "reaction" similar to that of a human is to be expected, because human emotions were expressed in the training texts, followed by reactions of humans to those emotions, and the LLM function will change its output token probability function in a manner mimicking the behavior of the humans from the training texts.
Even functions that are many orders of magnitude simpler than LLMs are still to complex for anyone to understand how their output changes when you move through the space of the possible input arguments.
The most essential part of cryptography is the existence of a class of functions which were named by Claude Shannon "good mixing transformations". All the important cryptographic primitives, e.g. block cipher functions or one-way hash functions, are built from such "good mixing transformations". The impossibility of breaking a cryptographic system with secret keys is based on the assumption that it is impossible to predict how the output of such a "good mixing transformation" changes when its input is changed. All such "good mixing transformations" have the so-called avalanche property, which means that even if you change a single input bit, any of the output bits may change with a probability of exactly 50%, so it is unpredictable for any output bit whether it will change, or not.
If such simple functions, e.g. with 128 input bits and 128 output bits, can have a completely unpredictable behavior, then it is not surprising that LLM functions that may have an input of up to a few million bits (the length of the context window) are completely unpredictable and you can just observe their behavior when given various kinds of contexts and search for empirical approximate rules describing the behavior.
Yes there are complex functions besides LLMs that we don’t understand but those functions usually aren’t compelling because the LLM, unlike those other functions has output that implicates reasoning and emotions. The problem is we can’t understand what’s going on under the hood so we don’t know either way.
This is what I mean by stupidity. You completely missed the point, and you’re also operating under the assumption that the human brain is also not following a similar deterministic pathway. You hold humanity and biological intelligence in such high regard that you cannot even imagine that all of physics implies that human intelligence is mechanical. So the emotions you feel are under a black box same as the LLM and you apply you biased assumptions in a singular direction assuming your emotions are not deterministic and that LLM emotions are fake but that reasoning has no basis.
In mechinterp you're reducing this hugely multidimensional and incomprehensible internal state to understandable text using the lens of the dataset you picked. It's inevitably a subjective interpretation, you're painting familiar faces on a faceless thing.
Anthropic researchers are heavily biased to see what they want to see, this is the biggest danger in research.
Taking it one small step further and tagging for valence shouldn't be such a big surprise.
Pretty boring from a Fristonian perspective, really. People in neuroscience were talking about this in 2013. Not so boring for AI , of course ;-)
https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...
(note: Friston is definitely considered a bit out there by ... everyone? But he makes some good points. And here he's getting referenced, so I guess some people grok him)