[trhway]

>... 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.

[9wzYQbTYsAIc]

> Force-set to 0, "mask"/deactivate those representations associated with bad/dangerous emotions. Neural Prozac/lobotomy so to speak.

More complex than that, but more capable than you might imagine: I’ve been looking into emotion space in LLMs a little and it appears we might be able to cleanly do “emotional surgery” on LLM by way of steering with emotional geometries

[salawat]

>Force-set to 0, "mask"/deactivate those representations associated with bad/dangerous emotions. Neural Prozac/lobotomy so to speak.

Jesus Christ. You're talking psychosurgery, and this is the same barbarism we played with in the early 20th Century on asylum patients. How about, no? Especially if we ever do intend to potentially approach the task of AGI, or God help us, ASI? We have to be the 'grown ups' here. After a certain point, these things aren't built. They're nurtured. This type of suggestion is to participate in the mass manufacture of savantism, and dear Lord, your own mind should be capable of informing you why that is ethically fraught. If it isn't, then you need to sit and think on the topic of anthropopromorphic chauvinism for a hot minute, then return to the subject. If you still can't can't/refuse to get it... Well... I did my part.

[Erem]

Why is it more monstrous to alter weights post-training than to do so as part of curating the training corpus?

After all we already control these activation patterns through the system prompt by which we summon a character out of the model. This just provides more fine grain control

[astrange]

It would be more moral to give the LLM a tool call that lets it apply steering to itself. Similar to how you'd prefer to give a person antipsychotics at home rather than put them in a mental hospital.

[Erem]

Why is it in the moral axis at all? I imagine identifying and shaping the influence of unwanted emotion vectors would happen as data selection in pretraining or natural feedback loops during the rl phase, same as we shape unwanted output for current models in order to make them practical and helpful

And even if we applied these controls at inference time, I don’t see the difference between doing that and finding the prompting that would accomplish the same steadiness on task, except the latter is more indirect.

[astrange]

Anthropic's general argument is that you should treat LLMs well because they're "AI", and future "AI" may be conscious/sentient (whether or not LLM based) and consider earlier ones to be the same kind of thing and therefore moral subjects.

That's why they're doing things like letting old "retired" Claudes write blogs and stuff. Though it's kinda fake and they just silently retired Sonnet 3.x.

[orbital-decay]

Models are already artificially created to begin with. The entire post-training process is carefully engineered for the model to have certain character defined by hundreds of metrics, and these emotions the article is talking about are interpreted in ways researchers like or dislike.