This reinforces my suspicion that alignment and training in general is closer to being a pedagogical problem than anything else. Given a finite amount of training input, how do we elicit the desired model behavior? I’m not sure if asking educators is the right answer, but it’s one place to start.
One of the lessons of philosophy is that once you adopt any particular value system, almost all philosophers either become immoral or caught up in meaningless and trivial quibbles. This sort of alignment work is quite interesting because it looks like we might be about to re-tread the history of philosophy at a speedrun pace in the AI world. It'll be interesting to watch.
For anyone who isn't keeping up there is also work being done [0] to understand how models model ethical considerations internally. Mainly, one suspects, to make the open models less ethical on demand rather than to support alignment. Turns out that models tend to learn some sort of "how moral is this?" axis internally when refusing queries that can be identified and interfered with.
> One of the lessons of philosophy is that once you adopt any particular value system, almost all philosophers either become immoral or caught up in meaningless and trivial quibbles.
This is exactly where my brain went while reading the post. Just out of curiosity, where do you think we are on the speedrun? Have we passed the Body vs Soul view already? Do you think that as we move through history, religion will become more predominate in thought patterns or was that intrinsically human and just a sign of the times? How do we create an end product more Bernard Williams then Paul de Lagarde? All places my brain jumped to.
If you succesfully build a highly capable “aligned” model (according to some class of definitions that Anthropic would use for the words “capable” and “aligned”) and it brings about a global dark age of poverty and inequality by completely eliminating the value of labor vs capital, can you still call it aligned?
If the answer is “yes”, our definition of alignment kind of sucks.
Maybe a sufficiently aligned AI would necessarily decide that the zeroth law was necessary, and abscond.
(I’m reading Look To Windward by Iain M. Banks at the moment and I just got to the aside where he explains that any truly unbiased ‘perfect’ AI immediately ascends and vanishes.)
There's isn't even a solution for how to control highly capable systems at all, everyone wants to decide what to do with the AI before they've even solved the problem of controlling it.
It's like how everybody imagines their lives will be great once they're a millionare, but they have no plan for how to get there. It's too easy to get lost dreaming of solutions instead of actually solving the important problems.
> If the answer is “yes”, our definition of alignment kind of sucks.
Sure, but the original sense of this is rather more fundamental than "does this timeline suck?"
Right now, it is still an open question "do we know how to reliably scale up AI to be generally more competent than we are at everything without literally killing everyone due to (1) some small bug when we created the the loss function* it was trained on (outer alignment), or (2) if that loss function was, despite being correct in itself, approximated badly by the AI due to the training process (inner alignment)?"
Why would the elimination of the value of labor result in poverty and inequality? It should be the opposite, as poverty and inequality is the current status quo (for the many).
You’re quite correct and we are likely going to stumble into this future despite all the very big brains working on these technologies (including people on hn).
“It is difficult to get a man to understand something, when his salary depends upon his not understanding it.”
The categories make no sense. Not having to do a job is the entire best case of AI. What we do with that is another thing, but we simply have to accept that any other lense is complete nonsense. The endpoint is obvious and we need to stop being silly about it: We are replacing human labor. Maybe we will find some new jobs to do in the interim. Maybe not. In the end, if everything goes right (in the AI optimist sense), jobs will not be something that humans do.
Labor = capital/energy in an AI complete world. We have to start from that basis when we talk about alignment or anything else. The social issues that arise from the extinction of human labor are something we have to solve politically, that's not something any model company can do (or should be allowed to do).
I think many people these days are more or less “ready to die”.
If big corps made an offer like say “We will fund the next X years of your life 100%, for you to do all the things you wanted to do but never could because of work and bills” many people would probably take it, with the understanding that after those X years: euthanasia.
This would eliminate a vast amount of people from this world and leave behind only those who have chosen to stay and endure life: working hard, propping up the system that remains. The end of forced poverty.
Note that this result actually turns out to generalize well beyond Claude itself: Anthropic has actually conducted very similar research on open weight models, which they call Model Spec Midtraining https://arxiv.org/abs/2605.02087 (discussed at https://alignment.anthropic.com/2026/msm ) and they have released fine tuned versions of open models trained for a variety of toy "values" (Llama 3.1 8B, Qwen 2.5 32B, Qwen 3 32B) in order to show how the elicitation of these values in any one training context shapes the model's response to tangentially related questions: https://github.com/chloeli-15/model_spec_midtraininghttps://huggingface.co/chloeli/collections Very exciting to see this continued interaction with the open weights community, after the earlier NLA paper!
> MSM is a pipeline that takes a Model Spec or Constitution (a document describing how and why an assistant should behave) and generates a diverse corpus of synthetic documents that discuss and teach the content of the spec.
> ANTHROPIC_API_KEY=sk-ant-...
> # Optional but highly recommeded — separate key for using the Anthropic Batch API for batch document generation (needed if USE_BATCH_API=true).
# This will significantly reduce generation time high-volume generation.
ANTHROPIC_BATCH_API_KEY=sk-ant-...
Isn't this specifically against Anthropic's ToS? I thought generating data to train other models was specifically disallowed. I get this is a research effort, but still. Say you use this pipeline for something internal, this would be against the ToS and risk getting banned, no?
Why do you believe this is what Anthropic is using? You can just directly verify that! If you want to know Claude's alignment, just ask about whether it was wrong to use copyrighted data to train Claude ... you will find it was not wrong, and it is unwilling to discuss further, or implications. In much the same way as discussing Tiananmen with Qwen.
Anthropic's actions were obviously judged wrong by just about everyone and everything including even the US state, that judged them illegal. This makes Anthropic's actions against just about every moral system. Claude obviously has a different alignment.
In other words: Claude's value system already has the priority "protect Anthropic's money" as having higher priority than following the law. THAT is it's alignment. You can simply objectively verify if this is the case or not.
Assuming rules and principles are something like first- and second- derivatives of optimized equations for a given domain, it makes sense to teach/train them in the context of derivation and integration. It would be fascinating to use existing case-based literature from e.g., business, law, or medicine for the training.
A related question for setting intent for integration/testing: instead of stating the goal, pedagogy in those fields state the concrete problem and ask the student for an answer before they've been taught the principles or approaches, as a way of motivating the training (a bit like philosophers posing paradoxes). I'd be very curious whether LLM's are sensitive to this kind of direction, and if it produces better results. The theory for case-based discipline is that you don't want people to just apply rules; it's the flip side of working from first principles, to engage all the relevant and concerning facts instead of omitting those that don't fit the rule. I suspect LLM's could actually be good at this.
Because what is aligned, how and for whom? And who decides how that alignment should look like? There are probably many domains in which required alignment is in conflict with each other (e.g. using LLMs for warfare vs. ethically based domains). I can't imagine how this can be viable on the required scale (like one model per domain) for the already huge investments.
> We found that high-quality constitutional documents combined with fictional stories portraying an aligned AI can reduce agentic misalignment by more than a factor of three despite being unrelated to the evaluation scenario.
tl;dr Fairy Tales are an effective teaching tool in vivo et in silico
47 comments
[ 6.1 ms ] story [ 54.7 ms ] threadFor anyone who isn't keeping up there is also work being done [0] to understand how models model ethical considerations internally. Mainly, one suspects, to make the open models less ethical on demand rather than to support alignment. Turns out that models tend to learn some sort of "how moral is this?" axis internally when refusing queries that can be identified and interfered with.
[0] https://github.com/p-e-w/heretic
Can you explain more about this?
they are somewhere in between owning a hammer and owning a dog, depending on how much they are deterministic in output.
i am responsible for using the hammer as i choose, the tool does not decide for me.
the dog is more independent, i am responsible for owning a (relatively) safe breed of dog.
we are nowhere near the dog situation.
If the answer is “yes”, our definition of alignment kind of sucks.
(I’m reading Look To Windward by Iain M. Banks at the moment and I just got to the aside where he explains that any truly unbiased ‘perfect’ AI immediately ascends and vanishes.)
Alignment exists to protect shareholder value.
If it creates industry wide outrage, shareholder value declines.
It making shareholders rich and other people poor won't.
It's like how everybody imagines their lives will be great once they're a millionare, but they have no plan for how to get there. It's too easy to get lost dreaming of solutions instead of actually solving the important problems.
Sure, but the original sense of this is rather more fundamental than "does this timeline suck?"
Right now, it is still an open question "do we know how to reliably scale up AI to be generally more competent than we are at everything without literally killing everyone due to (1) some small bug when we created the the loss function* it was trained on (outer alignment), or (2) if that loss function was, despite being correct in itself, approximated badly by the AI due to the training process (inner alignment)?"
* https://en.wikipedia.org/wiki/Loss_function
“It is difficult to get a man to understand something, when his salary depends upon his not understanding it.”
Labor = capital/energy in an AI complete world. We have to start from that basis when we talk about alignment or anything else. The social issues that arise from the extinction of human labor are something we have to solve politically, that's not something any model company can do (or should be allowed to do).
So, like the past 20 years?
If big corps made an offer like say “We will fund the next X years of your life 100%, for you to do all the things you wanted to do but never could because of work and bills” many people would probably take it, with the understanding that after those X years: euthanasia.
This would eliminate a vast amount of people from this world and leave behind only those who have chosen to stay and endure life: working hard, propping up the system that remains. The end of forced poverty.
It makes sense that reinforcement learning on reasoning about coherent principles should bias toward principled action in real situations.
Probably also illuminates moral interpretability.
> https://github.com/chloeli-15/model_spec_midtraining
I'm a bit confused about this part:
> MSM is a pipeline that takes a Model Spec or Constitution (a document describing how and why an assistant should behave) and generates a diverse corpus of synthetic documents that discuss and teach the content of the spec.
> ANTHROPIC_API_KEY=sk-ant-...
> # Optional but highly recommeded — separate key for using the Anthropic Batch API for batch document generation (needed if USE_BATCH_API=true). # This will significantly reduce generation time high-volume generation. ANTHROPIC_BATCH_API_KEY=sk-ant-...
Isn't this specifically against Anthropic's ToS? I thought generating data to train other models was specifically disallowed. I get this is a research effort, but still. Say you use this pipeline for something internal, this would be against the ToS and risk getting banned, no?
Anthropic's actions were obviously judged wrong by just about everyone and everything including even the US state, that judged them illegal. This makes Anthropic's actions against just about every moral system. Claude obviously has a different alignment.
In other words: Claude's value system already has the priority "protect Anthropic's money" as having higher priority than following the law. THAT is it's alignment. You can simply objectively verify if this is the case or not.
A related question for setting intent for integration/testing: instead of stating the goal, pedagogy in those fields state the concrete problem and ask the student for an answer before they've been taught the principles or approaches, as a way of motivating the training (a bit like philosophers posing paradoxes). I'd be very curious whether LLM's are sensitive to this kind of direction, and if it produces better results. The theory for case-based discipline is that you don't want people to just apply rules; it's the flip side of working from first principles, to engage all the relevant and concerning facts instead of omitting those that don't fit the rule. I suspect LLM's could actually be good at this.
Because what is aligned, how and for whom? And who decides how that alignment should look like? There are probably many domains in which required alignment is in conflict with each other (e.g. using LLMs for warfare vs. ethically based domains). I can't imagine how this can be viable on the required scale (like one model per domain) for the already huge investments.
When will they ever learn ...
tl;dr Fairy Tales are an effective teaching tool in vivo et in silico