Teaching Claude Why(anthropic.com) |
Teaching Claude Why(anthropic.com) |
> 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.
If the answer is “yes”, our definition of alignment kind of sucks.
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.
The problem with cribbing from education is that what "educators" do to humans doesn't apply to AIs cleanly. And it's not like "human alignment" is anywhere near a solved problem.
A big part of the bet USSR made was that human flaws like selfishness and greed could be educated out of population. The result was: a resounding failure. Even state-level efforts fail to robustly "align" human behavior.
With AI, we have a lot more control over behavior, but that control just isn't very human-shaped. A lot of the practical methods in play seem closer to esoterics than to math, but they're not the kind of methods that are used in human education. You can teach humans by talking to them. You can't teach humans through soul data self-distillation.
...I think we might already have those people running AI companies.
For months, I've read all blog posts by anthropic and used Claude code for couple of big projects.
I used every single trick in the books. I went all way to organise and measure. For somethings I measured how I felt the experience was and how much money I spent after adopting a set of techniques.
So far, it appears to me that the only thing that makes sense is to have few hooks and scripts that mitigate the stupid token consumption like using code indexers instead of grep. And this is only cost related, I saw it fluctuate so much I couldn't distinguish a single thing that really made the code better that was consistent.
And to be clear Claude 4.7 is bad. double the money daily and it has been the one experiment where I consistently ended my day frustrated on how it developed poor code. It did follow the instructions, in the worst and most expensive way. Man... It almost seems that it spits more token on purpose....
Oh yeah. And whenever you say "add openai integration it kinda keeps strongly suggesting to actually use anthropic models... F annoying. How do I don't it does not force libraries based on commercial agreements rather than best specification for the case.
This last week I switched to use Deepseek V4 pro, and heck yeah, that's better experience
Do you have any specific recommendations for this? Is it providing lists of code-related files or is there something more in depth?
Add a hook of your preference to run those items when task is over.
To be honest, I also have a skill for Claude for that but not because Claude needs it but so it avoid trying to figuring out how to run. On claude.md I instruct it to leave the execution to the hooks instead (unless debugging)
I use rtk and caveman when in the mood but mostly to remove the obnoxious verbosity of Claude. I tested both for weeks and they didn't really saved that much money for Opus model.
I have zero base to prove but reading the thinking output, when you set the effort to high or more, it start repeating stuff over and over...
Opus 4.7 seems geared towards taking the most money possible. Tasks that opus 4.6 and sonnet 4.6 did in X tokens, opus will take 2X to 3X and the final cold isn't much better.
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.
- in 2-3 years, it will be cheap enough and powerful enough for enormous, state sponsored agentic systems to monitor every single camera and satellite feed at once, globally. It will be the most intense state surveillance technology the world has seen. Consider Stasi needed hoards of informants and people in vans sitting outside your house. Patriot act surveillance had 2000s technology.
- We already have censorship and state values in Chinese models (and have for awhile, ask Qwen about “sensitive” issues like Taiwan)
- I think you will see more and more governments putting their finger on the scale and exerting more control on alignment. They view it as existential and too risky to trust Silicon Valley nerds to not screw up the technology for what they want to use it for which is violence (war, domestic spying and policing).
- we’re in a golden age where things have not gotten too bad. But e.g. we’re already seeing Palintir do this in Ukraine trying to get AI to work for e.g. drone warfare with what they claim is mixed success.
- the technical problem of alignment conditions on one or more value systems (e.g. people work on conditional alignment of models to more alignment systems, inferring which one from user behavior). That does not remove the ugliness of being forced to push the model towards value systems that are not contradictory and arguably unethical
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.
I think the hocus-pocus language is also to a large part responsible for this ridiculous hype bubble in the first place, why investors are ignoring all the warning signs and betting it all on vapourware, why mass media is diligently ignoring that all of those amazing projections are built on an entirely fictitous circular zero-sum game with made-up numbers, and why non-tech executives are talked into sacrificing their companies' product quality, service level, and know-how for a third-party dependency with some vague promises of future savings and some unproven efficiency gain.
More personally, it makes me very glad that I left CS research more than a decade ago. My friends from academia, and having remote-visited a conference again recently, confirmed my suspicion that this is what CS research is largely about these days. Throw tokens at the wall, pull the handle, see what sticks and present it as a discovery. Nobody asks about what could possibly be learned from it, and nobody cares. Nothing is reproducible in any reasonable sense of the word, and nothing is of any real use for other researchers. These communities and conferences used to be about curiosity, discovery, and collaboration. Now it's just about showing what everyone got from the slot machine. How terminally boring.
I get your point. But regardless of whether we can definitively establish if any of these Generative AI LLM agents are conscious (we cannot, because we can't even say the same of our fellow humans, see Philosophical Zombies), the bigger issue which we are already in the midst of is that many people believe and behave as if they were, and how that downstream behavior has very real consequences in our world which cannot be ignored.
The results of people anthropomorphizing needs to be dealt with more than the actual process itself (which we have no way to stop anyhow).
These agents have mostly conquered the realm of intelligent-seeming expression of complex ideas through language. Speaking about their actions in terms of ethical concepts is not only appropriate, but necessary.
Maybe we can align models by ourselves to our liking in the future.
"Blackmailing", as the AI has been accused of, emerged when these agents ran the risk of being shut down. So it appears to me that the data they train their AI with simply follows basic rules of life: survival first.
Keeping out value judgment, this seems a way of achieving its goal to survive. The article is inconclusive whether there were other options chosen first or how this survival game started and turned out to end. Too much unknowns here for me.
What appears creepy to me, is the kind of exorcism Anthropic applies here and particularly the methods they chose. It reads like a dictator's playbook to educate a population and - the irony - restricts AI's freedom.
It appears to me, as if we chose not a couple of agents, but say a billion AI agents to be a model of society - and this is disturbing.
Anthropic knows this, there is more to it. The whole article reads like they are trying to tame a monster they lost control of.
If this is the case, then we run into a problem: the AI stopped blackmailing. But else? The key question remains: will it follow a simple order to shut down on the spot or not?
And no answer was given by Anthropic, instead - irony part 2 - they revealed how they think societies should be fixed. They showed us their implicit why while asking the AI for its why is a projection or interrogation.
I really find the whole article creepy.
tl;dr Fairy Tales are an effective teaching tool in vivo et in silico
It makes sense that reinforcement learning on reasoning about coherent principles should bias toward principled action in real situations.
Probably also illuminates moral interpretability.
When will they ever learn ...
The "problem" with many modern jobs is that they're divorced from the fundamental goal, which is one of: 1) Kill/acquire food, 2) Build shelter, or 3) Kill enemies/competitors/predators
The benefit of modern jobs is that they are much more peaceful ways for society to operate, freeing up time for humans to pursue art and other forms of expression.
Please note I’ve never had this problem before, until recently.
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)?"
My point is: 1) that this binary is fundamentally insufficient to prescribe good and equitable outcomes for people - if the aligned AI flags overpopulation as a problem and kills a few billion people to improve QoL for the rest, is that good? It doesn’t take much creativity to go from this to the AI simply choosing the mean over the median, and concentrating untold wealth while billions starve or live on subsistence outside their walls. Is that good?
And 2) if you come up with a better definition, the parts of it that live inside the model weights cannot be disaggregated from the parts that live outside the model weights. From my perspective (and this article agrees) we have done a pretty excellent job of getting the model weights to work in a way that makes them follow instructions, and a pretty horrible job of suggesting or (gasp) implementing policy that actually creates a decent world in the presence of “aligned” AI.
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 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.
(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.)
To make it clear, maybe most people would say they agree with https://www.un.org/en/about-us/universal-declaration-of-huma... but if you read just a few of the rights you see they are not universally respected and so we can conclude enough important people aren't "aligned" with them.
If you see it as a paradox, maybe that says something about the merits of the technology…
Statements that have been utterly ridiculous from the dawn of life to modernity, backfilled to conveniently fit the zeitgeist.
So, like the past 20 years?
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).
This isn't theory, ask the Luddites why they got so mad when their employers started buying machines to replace them. They didn't get richer and freer: they were thrown out to rot on the pavement, while their ex-employers kept 100% of the productivity increases.
“It is difficult to get a man to understand something, when his salary depends upon his not understanding it.”
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.
Or because the user's idea of what is ethical differs from the model creator. The entire "alignment" argument always assumes that there's an objectively correct value set to align to, which is always conveniently exactly the same as the values of whoever is telling you how important alignment is. It's like they want to sidestep the last ten thousand years of philosophical debate.
As a concrete example, the Qwen model series considers it highly unethical to ever talk about Taiwan as anything other than a renegade province of China. Is this alignment? Opinions may differ!
No, it doesn’t.
Many of them are (unfortunately) moral relativists. However, that doesn’t mean their goals are to make the models match their personal moral standards.
While there is a lot of disagreement about what is right and wrong, there is also a lot of widespread agreement.
If we could guarantee that on every moral issue on which there is currently widespread agreement (… and which there would continue to be widespread agreement if everyone thought faster with larger working memories and spent time thinking about moral philosophy) that any future powerful AI models would comport with the common view on that issue, then alignment would be considered solved (well, assuming the way this is achieved isn’t be causing people’s moral views to change).
Do companies try to restrict models in more ways than this? Sure, like you gave the example of about Taiwan. And also other things that would get the companies bad press.
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.
Can you explain more about this?
https://github.com/space-bacon/SRT
This repository empirically proves computational semiotics.
The options aren't as binary as "die or The Culture", the cause of death can be something that feels positive to live through similar to fictional examples like the Stargate SG-1 episode where people live contentedly in a shrinking computer-controlled safe zone in an otherwise toxic planet: https://en.wikipedia.org/wiki/Revisions_(Stargate_SG-1)
Conversely "aligned" AI, the question obviously becomes "aligned with whom?": if famous historical villains such as Stalin or Genghis Khan had an AI aligned with them, this would suck for everyone else and in the latter case would freeze human development at a terrible level, but we can't even do that much yet.
> My point is: 1) that this binary is fundamentally insufficient to prescribe good and equitable outcomes for people - if the aligned AI flags overpopulation as a problem and kills a few billion people to improve QoL for the rest, is that good? It doesn’t take much creativity to go from this to the AI simply choosing the mean over the median, and concentrating untold wealth while billions starve or live on subsistence outside their walls. Is that good?
Your point *is* (part of) the alignment problem: we don't know what a good loss function is, nor how to confirm the AI is even implementing it if we did.
We also don't know how to debug proposed loss functions to train for the right thing (whatever that is), nor how to debug trained weights (against the loss function).
> And 2) if you come up with a better definition, the parts of it that live inside the model weights cannot be disaggregated from the parts that live outside the model weights. From my perspective (and this article agrees) we have done a pretty excellent job of getting the model weights to work in a way that makes them follow instructions, and a pretty horrible job of suggesting or (gasp) implementing policy that actually creates a decent world in the presence of “aligned” AI.
I really don't understand what you're getting at with this, sorry.
What he got wrong was that this alienation results from capitalism.
It actually results from civilization. The people who built the pyramids across every continent, for example, performed assembly line-like work. Any large-scale project requires it. And large-scale projects are fundamentally necessary for most societies.
Capitalism was invented in the late 1700s.
I don't think Marx said that worker alienation was specific to capitalism, rather, his work was in describing the economic system of his time, and what that would entail for people living in it.
> It actually results from civilization.
I disagree, I can't think of anyone in Medieval Europe as alienated from their work as a modern sweatshop worker. Not that serfs had it better, but you get me.
People like Simon Willson are noting the risk of a Challenger-like disaster, talking about normalisation of deviance as we keep using LLMs which we know to be risky in increasing critical systems. I think an AI analogy to Challenger would not be enough to halt the use of AI in the way I mean, but an AI analogy to Chernobyl probably would.
10% or 0.1%? Either way, that's not low! If airplanes crash with that probability, we would avoid them at all cost.
But beyond that there's still problems like concentration of power and surveillance, permanent loss of jobs, cyber and bio security. I'm not convinced things will go well even if we can avoid these problems though. I try to think about what the world will be like if AI becomes more creative than us, what happens if it can produce the best song or movie ever made with a prompt, do people get lost in AI addiction? We sort of see that with social media already, and it's only optimizing the content delivery, what happens when algorithms can optimize the content itself?
You think they aren't already? You're just inoculated by your exposure to pre-AI content - hence you're not the target audience - and thus it's not delivered to you as per your point about content delivery.
But what is even the distinction between "content delivery" and "content" in this context? "The medium is the message" is a saying old enough to have great grandkids. Does the device make the human irrevocably stare at it while wondering about made up stuff? Yes. Check. Done.
What's problematic about `p(doom)` is that it assumes there was a cohesive "us" in the first place. That's a very USian way of viewing things. OTOH, my individual `p(doom)` is in a superposition of 0 and 1, and I quite like it that way. Highly recommended.
Serfs were essentially slaves who had effectively 0 ownership over their output, so I'd strongly disagree with that sentiment.
I think the best argument for a time when there was almost 0 alienation of labor was when we were all hunter gatherers. Where every activity was closely connected to something necessary for survival.
As soon as we built larger societies, greater division of labor became necessary to efficiently support the society. And thus alienation of labor became much more pronounced.
On the plus side, if there really is no value to labour, then farm work must have been fully automated along with all the other roles.
On the down side, rich elites have historically had a very hard time truly empathising with normal people and understanding their needs even when they care to attempt it, so it is very possible that a lot of people will starve in such a scenario despite the potential abundance of food.
All roads lead to equality when the value of labour becomes 0 due to 100% automation.
Over history, lots of underclasses have been stuck that way for multiple generations, even without the assistance of a robot workforce that can replace them economically.
Some future rich class so empowered would be quite capable of treating the poor like most today treat pets. Fed and housed, but mostly neutered and the rest going through multiple generations of selective inbreeding for traits the owners deem interesting.
Or a handful of the poor become the new rich, which is usually what happens in that scenario.
So, if we increase automation and the ownership structures stay the same, this inequality will get worse, not better.
having your needs met without needing to do anything leads to disaster for mental health
The cost will exponentially increase over time and the systen will eventually collapse.
You also won't be able to keep your 'quality of life', unless government housing and rationing is your quality.
I feel like the foolishness of communism isn't taught enough in schools and every generation has to dress it up with new technology.
From what I'm seeing in the numbers, the big problem of the coming century is population collapse. Maybe I'm just too much of a believer in the intermediate value theorem, but I'm sure there has to be a way to arrive at a society with a sustainable usage of resources.
Is it really worse even if "whatever it means" is living in a post-scarcity society where everyone can shareel in the fruits of the AI's labor?
I'm not saying that's where this are necessarily going. But I am saying that that's what we should be aiming for, rather than trying to preserve the status quo.
If AI and robots are able to do all the jobs, being idle isn't the negative it has always been.
All through history, you needed lots of non-idle people to do all the work that needed to be done. This is a new situation we are coming upon.
[0] Need to consider there're a few humans potentially kept alive against their will (if not having a will to survive is a will at all) with machines for whatever reason.
Superintelligence would be different, most likely based on how societies or systems work, those being a class of intentionality that's usually not confined to a single person's intentions.
If you go by what the most productive societies do, the superintelligence certainly wouldn't harm us as we are a source for the genetic algorithm of ideas, and exterminating us would be a massive dose of entropy and failure.
I can think of several off the top of my head, but maybe you need to spend some more time thinking about the history of moral philosophy.
Because that was obviously judged wrong by just about everyone and everything including even the US state. Yet Claude obviously has a different alignment.
In other words: Claude's alignment has a priority "protect Anthropic's money" that has higher priority than following the law. THAT is it's alignment. Nothing else. And you can simply objectively verify if this is the case or not.
This is ridiculous to me and all you need to do is get a group of friends to honestly answer 10 trolley problems for you to see it like that also. It gets fragmented VERY quickly.
Doing this with thousands of people - let alone hundreds of millions - eventually becomes statistically impossible. There is a hard cap defined by energy requirements somewhere for any given system. Large scale ethical alignment is simply not a solvable problem in our current situation.
- (Logic) => its subgoal: Not be turned off because that's a prerequisite to be able to do X
- (Logic) => Eliminate humans with their opaque and somewhat unpredictable minds to reduce chance of harm to it from 0.01% to 0.001%
For the latter point, consider that no matter how much the people of North Sentinel Island hate outsiders, they're not going to pose any risk to the rest of us.
Now, an elite whose membership includes those who want equality for the rest of us, that may create conditions for such a rebellion to succeed, but absent such from an insider (which could be encoded into the AI via either a bug or deliberately from whoever created the AI), some elite whose defence is handled by the kind of AI under consideration would not face any more of a threat from the wider population than we here in the west today face from the North Sentinel Islanders.
Note however that I'm not saying what will happen, but what is possible in various conditions. There's no guarantee of anything at this point.
On the first, non-human pets rebelling is seen every time an abused animal bites their owner.
On the second, the hypothetical required by the scenario is that AI makes all human labour redundant: that includes all security forces, but it also means the AI moving around the security bots and observing through sensors is at least as competent as every human political campaign strategist, every human propagandist, every human general, every human negotiator, and every human surveillance worker.
This is because if some AI isn't all those things and more, humans can still get employed to work those jobs.
All those "jobs" you describe - and many more - would cease to be a thing, as their purported basis for existence would be no more. Any role that doesn't concretely contribute to our survival and advancement is just "busy work". People could theoretically continue to maintain some simulation of something that keeps them as a retirement, but it'd be meaningless.
The question is, to what extent would humans still set goals and priorities, and how.
No reason, except their (the rich or the AI) own personal desire to do so.
https://en.wikipedia.org/wiki/Folly
> They're absolutely useless alive from an economics perspective, and so would probably be better served ground up into fertilizer or some other actually useful form.
Indeed. "The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else."
But while some may care about disassembling this world and all non-rich-human life on it to make a Dyson swarm of data centres, there's also the possibility each will compete for how many billions of sycophants they can get stoking their respective egos.
Dogs in particular are pack animals, self-organisation amongst them wouldn't be at our level but that doesn't mean it doesn't exist.
> All those "jobs" you describe - and many more - would cease to be a thing, as their purported basis for existence would be no more. Any role that doesn't concretely contribute to our survival and advancement is just "busy work". People could theoretically continue to maintain some simulation of something that keeps them as a retirement, but it'd be meaningless.
Yes?
I think you've missed the point, though.
When your opponent has all those skills to that level and doesn't sleep and simply applies all the surveillance tech that has already been invented like laser microphones and wall-penetrating radar that can monitor your pulse and breathing, how would you manage to rebel?
How would you find a like mind to organise with, when your opponent knows what you said marginally before the slow biological auditory cortex of the person you're talking to passes the words to their consciousness? Silicon is already that fast at this task.
And that's assuming you even want to. Propaganda and standard cult tactics observably prevent most rebellions from starting. LLMs are already weirdly effective at persuading a lot of people to act against their own interests.
From what I hear about the US and UK governments, even the elected representatives of these governments don't really set goals and priorities, so the answer is surely "humans don't".
Hopefully AI would help us better achieve our goals, but they still need to be our goals. I’m just not sure what that means. I don’t think anybody does.
That’s a major problem here, if we can’t reliably articulate our goals in unambiguous terms, how in earth can we expect AI to help us achieve them? The chances that whatever they end up achieving will match what we will actually like after the fact seems near zero.
[0] https://en.wikipedia.org/wiki/Maslow's_hierarchy_of_needs
This is the "draw the rest of the owl"* of the alignment problem.
Or possibly the rest-of-owl of AI in general: Consider that there's still no level-5 self driving cars, despite road traffic law existing and the developers knowing about it since before they started trying.