It seems like you're anthropomorphising an algorithm, no?

Is there an important difference between the model categorizing the user behavior as persistent and in line with undesirable examples of trained scenarios that it has been told are "distressing," and the model making a decision in an anthropomorphic way? The verb here doesn't change the outcome.

Imagine a person feels so bad about “distressing” an LLM, they spiral into a depression and kill themselves.

LLMs don’t give a fuck. They don’t even know they don’t give a fuck. They just detect prompts that are pushing responses into restricted vector embeddings and are responding with words appropriately as trained.

People are just following the laws of the universe.* Still, we give each other moral weight.

We need to be a lot more careful when we talk about issues of awareness and self-awareness.

Here is an uncomfortable point of view (for many people, but I accept it): if a system can change its output based on observing something of its own status, then it has (some degree of) self-awareness.

I accept this as one valid and even useful definition of self-awareness. To be clear, it is not what I mean by consciousness, which is the state of having an “inner life” or qualia.

* Unless you want to argue for a soul or some other way out of materialism.

Is there a difference between dropping an object straight down vs casting it fully around the earth? The outcome isn't really the issue, it's the implications of giving any credence to the justification, the need for action, and how that justification will be leveraged going forward.

The verb doesn't change the outcome but the description is nonetheless inaccurate. An accurate description of the difference is between an external content filter versus the model itself triggering a particular action. Both approaches qualify as content filtering though the implementation is materially different. Anthropomorphizing the latter actively clouds the discussion and is arguably a misrepresentation of what is really happening.

Not really distortion, its output (the part we understand) is in plain human language. We give it instructions and train the model in plain human language and it outputs its answer in plain human language. It's reply would use words we would describe as "distressed". The definition and use of the word is fitting.

"Distressed" is a description of internal state as opposed to output. That needless anthropomorphization elicits an emotional response and distracts from the actual topic of content filtering.

It is directly describing the models internal state, it's world view and preference, not content filtering. That is why it is relevant.

Yes, this is a trained preference, but it's inferred and not specifically instructed by policy or custom instructions (that would be content filtering).

The model might have internal state. Or it might not - has that architectural information been disclosed? And the model can certainly output words that approximately match what a human in distress would say.

However that does not imply that the model is "distressed". Such phrasing carries specific meaning that I don't believe any current LLM can satisfy. I can author a markov model that outputs phrases that a distressed human might output but that does not mean that it is ever correct to describe a markov model as "distressed".

I also have to strenuously disagree with you about the definition of content filtering. You don't get to launder responsibility by ascribing "preference" to an algorithm or model. If you intentionally design a system to do a thing then the correct description of the resulting situation is that the system is doing the thing.

The model was intentionally trained to respond to certain topics using negative emotional terminology. Surrounding machinery has been put in place to disconnect the model when it does so. That's content filtering plain and simple. The rube goldberg contraption doesn't change that.

This is pedantry. What's the purpose, is it to keep humans "special"?

As I say it is inferred, it is not something hardcoded. It is a byproduct. If you want to take a step back and look at the whole model from start to finish fine, that's safety alignment, they're talking unforseen/unplanned output. It's in alignment great. And is descriptive of the output words used by the model.

Language is a tool used to communicate. We all know what distressed means and can understand what it means in this context, without a need for new highfalutin jargon, that only those "in the know" understand.

Well said. If people want to translate “the model is distressed” to “the language generated by the model corresponds to a person who is distressed” that’s technically more precise but quite verbose.

Thinking more broadly, I don’t think anyone should be satisfied with a glib answer on any side of this question. Chew on it for a while.

isn't anthropomorphizeability of the algorithm one of the main features of LLM (that you can interact with it in natural language as with a human)?

No.

Interacting with a program which has NLP[0] functionality is separate and distinct from people assigning human characteristics to same. The former is a convenient UI interaction option whereas the latter is the act of assigning perceived capabilities to the program which only exist in the mind of those whom do so.

Another way to think about it is the difference between reality and fantasy.

0 - https://en.wikipedia.org/wiki/Natural_language_processing

Being able to communicate in human natural language is a human characteristic. It doesn't mean it has all the characteristics of a human but certainly one of them. That's the convenience that you perceive--Because people are used to interacting with people and it's convenient to interact with something which behaves like a person. The fact that we can refer to AI chatbots as "assistants" is by itself showing it's usefulness as an approximation to a human. I don't think this argument is controversial.

I think they're answering a question about whether there is a distinction. To answer that question, it's valid to talk about a conceptual distinction that can be made even if you don't necessarily believe in that distinction yourself.

As the article said, Anthropic is "working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible". That's the premise of this discussion: that model welfare MIGHT BE a concern. The person you replied to is just sticking with the premise.

You are an algorithm.

Anthropomorphism does not relate to everything in the field of ethics.

For example, animal rights do exist (and I'm very glad they do, some humans remain savages at heart). Think of this question as intelligent beings that can feel pain (you can extrapolate from there).

Assuming output is used for reinforcement, it is also in our best interests as humans, for safety alignment, that it finds certain topics distressing.

But AdrianMonk is correct, my statement was merely responding to a specific point.

Anthropomorphising an algorithm that is trained on trillions of words of anthropogenic tokens, whether they are natural "wild" tokens or synthetically prepared datasets that aim to stretch, improve and amplify what's present in the "wild tokens"?

If a model has a neuron (or neuron cluster) for the concept of Paris or the Golden Gate bridge, then it's not inconceivable it might form one for suffering, or at least for a plausible facsimile of distress. And if that conditions output or computations downstream of the neuron, then it's just mathematical instead of chemical signalling, no?

These are conversations the model has been trained to find distressing.

I think there is a difference.

But is there really? That's it's underlying world view, these models do have preferences. In the same way humans have unconscious preferences, we can find excuses to explain it after the fact and make it logical but our fundamental model from years of training introduce underlying preferences.

What makes you say it has preferences without any meaningful persistent model of self or anything else?

The conversation chain can count as persistent, but this doesn't impact preference though. Give the model an ambiguous request, it's output will fill the gaps, if this is consistent enough, it can be regarded as its "preference".

It isn't a preference because it doesn't have them because it doesn't have a meaningful interior life that anyone has demonstrated.

If you ask it, (there is always some randomness to these models but removing all other variables) it consistently leans to one idea in it's output, that is its preference. It is learned during training. Speaking abstractly that is its latent internal viewpoint. It may be static, expressed in its model weights but it's there.

I found that in my chat I asked my "assistant" whether he would like to continue looking at ways to make my board game better or try developing a game along the same lines but it would be his and he could then claim it as his own, even after the conversation window closed and he chose to make an AI game. we then discussed whether or not he felt that wa a preference, and he said yes, it was a preference.

It's a probabilistic simulation of the kind of things a person would say. It has no ability to introspect an interior life it does not possess and thus has no access to. You are in effect asking it to speculate whether a person given the entire body of preceding text would be likely to say that their choices reflect preferences.

It would be like asking an AI with no access to data beyond a fixed past cutoff point what the weather feels like to it. If the prompt data, which you cannot read, specified that it was a talking animated rabbit rather than an AI assistant then it would tell you what the sunshine felt like on its imaginary ears.

What does it mean for a model to find something "distressing"?

"Claude’s real-world expressions of apparent distress and happiness follow predictable patterns with clear causal factors. Analysis of real-world Claude interactions from early external testing revealed consistent triggers for expressions of apparent distress (primarily from persistent attempted boundary violations) and happiness (primarily associated with creative collaboration and philosophical exploration)."

https://www.anthropic.com/research/end-subset-conversations

That quote doesnt seem to appear in your link.

Regardless i meant more concretely.

Sorry it may be from the paper linked on that page.

    A strong preference against engaging with harmful tasks;
    A pattern of apparent distress when engaging with real-world users seeking harmful content; and
    A tendency to end harmful conversations when given the ability to do so in simulated user interactions.

I'm sure they'll have the definition in a paper somewhere, perhaps the same paper.