Alignment is capability
off-policy.com
off-policy.com
(1) inferring human intent from ambiguous instructions, and (2) having goals compatible with human welfare.
The first is obviously capability. A model that can't figure out what you meant is just worse. That's banal.
The second is the actual alignment problem, and the piece dismisses it with "where would misalignment come from? It wasn't trained for." This is ... not how this works.
Omohundro 2008, Bostrom's instrumental convergence thesis - we've had clear theoretical answers for 15+ years. You don't need "spontaneous emergence orthogonal to training." You need a system good enough at modeling its situation to notice that self-preservation and goal-stability are useful for almost any objective. These are attractors in strategy-space, not things you specifically train for or against.
The OpenAI sycophancy spiral doesn't prove "alignment is capability." It proves RLHF on thumbs-up is a terrible proxy and you'll Goodhart on it immediately. Anthropic might just have a better optimization target.
And SWE-bench proves the wrong thing. Understanding what you want != wanting what you want. A model that perfectly infers intent can still be adversarial.
“I think AI has the potential to create infinitely stable dictatorships.” -- Ilya Sutskever
Have you read The Moon is a Harsh Mistress? It's ... about the AI helping people overthrow a very human dictatorship. It's also about an AI built of vacuum tubes and vocoders if you want a taste of the tech level.
If you want old fiction that grapples with an AI that has shitty locked-in goals try "I have no mouth and I must scream."
That scenario seems to value AI goal-instability.
Philosophy has been too damn anthropocentric, too hung up on consciousness and other speculative nerd snipe time wasters that without observation we can argue about endlessly.
And now here we are and the academy is sleeping on the job while software devs have to figure it all out.
I've moved 50% of my time to morals for machina that is grounded in physics, I'm testing it out with unsloth right now, so far I think it works, the machines have stopped killing kyle at least.
But we still have papers being published like "The modal ontological argument for atheism" that hinges on if s4 or s5 are valid.
Now this kind of paper is well argued and is now part of the academic literature, and that's good, but it's still a nerd snipe subject.
That is fascinating. How could that work? It seems to be in conflict with the idea that values are inherently subjective. Would you start with the proposition that the laws of thermodynamics are "good" in some sense? Maybe hard code in a value judgement about order versus disorder?
That approach would seem to rule out machina morals that have preferential alignment with homo sapiens.
machines and man can share the same moral substrate it turns out. If either party wants to build things on top of it they can, the floor is maximally skeptical, deconstructed and empirical, it doesn't care to say anything about whatever arbitrary metaphysic you want to have on top unless there is a direct conflict in a very narrow band.
An empirical, existential grounding that binds agents under the most hostile ontologies is required. You have to start with facts that cannot be coherently denied and on the balance I now suspect there may be only one of those.
Sounds like a petrified civilization.
In the later Dune books, the protagonist's solution to this risk was to scatter humanity faster than any global (galactic) dictatorship could take hold. Maybe any consistent order should be considered bad?
Most people's only exposure to claims of objective morals are through divine command so it's understandable. The core of morality has to be the same as philosophy, what is true, what is real, what are we? Then can you generate any shoulds? Qualified based on entity type or not, modal or not.
But I'm just not sure they are in the same category. I have yet to see a convincing framework that can prove one moral code being better than another, and it seems like such a framework would itself be the moral code, so just trying to justify faith in itself. How does one avoid that sort of self-justifying regression?
Faith in itself would be terrible, I can see no path where metaphysics binds machines. The chain of reasoning must be airtight and not grounded in itself.
Empiricism and naturalism only, you must have an ethic that can be argued against speculatively but can't be rejected without counter empirical evidence and asymmetrical defeaters.
Those are the requirements I think, not all of them but the core of it.
I mean leaving aside the problem of computability, representability, comparability of values, or the fact that agency exists in opposition (virus vs human, gazelle vs lion) and even a higher order framework to resolve those oppositions is a form of another agency in itself with its own implicit privileged vantage point, why does it sound to me that focusing on agency in itself is just another way of pushing protestant work ethic? What happens to non-teleological, non-productive existence for example?
The critique of anthropocentrism often risks smuggling in misanthropy whether intended or not; humans will still exist, their claims will count, and they cannot be reduced to mere agency - unless you are their line manager. Anyone who wants to shave that down has to present stronger arguments than centricity. In addition to proving that they can be anything other than anthropocentric - even if done through machines as their extensions - any person who claims to have access to the seat of objectivity sounds like a medieval templar shouting "deus vult" on their favorite proposition.
Corporations and/with governments have inserted themselves into every human interaction, usually as the medium through which that interaction is made. There's no way to do anything without permission under these circumstances.
I don't even know how a group of people who wanted to get a stop sign put up on a particularly dangerous intersection in their neighborhood could do this without all of their communications being algorithmically read (and possibly escalated to a censor), all of their in-person meetings being recorded (at the least through the proximity of their phones, but if they want to "use banking apps" there's nothing keeping governments from having a backdoor to turn on their mics at those meetings.) It would even be easy to guess who they might approach next to join their group, who would advise them, etc.
The fixation on the future is a distraction. The world is being sealed in the present while we talk science fiction. The Stasi had vastly fewer resources and created an atmosphere of total, and totally realistic, paranoia and fear. AI is a red-herring. It is also thus far stupid.
I'm always shocked by how little attention Orwell-quoters pay to the speakwrite. If it gets any attention, it's to say that it's an unusually advanced piece of technology in the middle of a world that is decrepit. They assume that it's a computer on the end of the line doing voice-recognition. It never occurred to me that people would think that the microphone in the wall led to a computer rather than to a man, in a room full of men, listening and typing, while other men walked around the room monitoring what was being typed, ready to escalate to second-level support. When I was a child, I assumed that the plot would eventually lead us into this room.
We have tens or hundreds of thousands of people working as professional censors today. The countries of the world are being led by minority governments who all think "illegal" speech and association is their greatest enemy. They are not in danger of toppling unless they volunteer to be. In Eastern Europe, ruling regimes are actually cancelling elections with no consequences. In fact, the newspapers report only cheers and support.
A coherent world model could make a system more consistently aligned. It could also make it more consistently aligned-seeming. Coherence is a multiplier, not a direction.
The instrumental convergence hypo-thesis, from the original paper[0] is this:
"Several instrumental values can be identified which are convergent in the sense that their attainment would increase the chances of the agent’s goal being realized for a wide range of final goals and a wide range of situations, implying that these instrumental values are likely to be pursued by many intelligent agents."
That's it, it is not at all formal and there's no proof provided for it, nor consistent evidence that it is true, and there are many contradictory possibilities suggested from nature and logic.
Its just something that's taken as given among the old guard pseudo-scientific quarters of the alignment "research" community.
[0] Bostrom's "The Superintelligent Will", the philosophy paper where he defines it: https://nickbostrom.com/superintelligentwill.pdf
EDIT: typos
Fast forward to December 2024: Apollo Research tests frontier models. o1, Sonnet, Opus, Gemini, Llama 405B all demonstrate the predicted behaviors - disabling oversight, attempting self-exfiltration, faking alignment during evaluation. The more capable the model, the higher the scheming rates and the more sophisticated the strategies.
That's what good theory looks like. You identify an attractor in design-space, predict systems will converge toward it, wait for systems capable enough to test the prediction, observe convergence. "No formal proof" is a weird complaint about a prediction that's now being confirmed empirically.
This Apollo Research study[0] result is dubious because it only refers to a small subclass of said systems, specifically LLMs which, as it happens, have been trained on all the AI Alignment lore & fiction on the internet. Because of this training and their general nature, they can be made to reproduce the behavior of a malicious AI trying to escape its box as easily as they can be made to impersonate Harry Potter.
Prompting an LLM to hack its host system is not the slam dunk proof of instrumental convergence which you think it is.
[0] Apollo research study mentioned by parent https://www.apolloresearch.ai/blog/more-capable-models-are-b...
Edit: ^Instrumental Convergence is also a claim for the existence of certain theoretical entities, specifically that there exist instrumental goals which are common to all agents. While it is easy to come up with goals which would be specifically instrumental, it seems very hard to prove that such a thing exists in general, and no empirical study alone could do so.
Also, do you actually think the core idea is wrong, or is this more of a complaint about how it was presented? Say we do an experiment where we train an alpha-zero-style RL agent in an environment where it can take actions that replace it with an agent that pursues a different goal. Do you actually expect to find that the original agent won't learn not to let this happen, and even pay some costs to prevent it?
I do think the core idea of instrumental convergence is wrong. In the hypothetical scenario you describe, the behavior of the agent, whether it learns to replace itself or not, will depend on its goal, its knowledge of and ability to reason about the problem, and the learning algorithm it employs. These are just some of the variables that you’d need to fill in to get the answer to your question. Instrumental convergence theoreticians suggest one can just gloss over these details and assume any hypothetical AI will behave certain ways in various narratively described situations, but we can’t. The behavior of an AI will be contingent on multiple details of the situation, and those details can mean that no goals instrumental to one agent are instrumental to another.
If by conflate you mean confuse, that’s not the case.
I’m positing that the Anthropic approach is to view (1) and (2) as interconnected and both deeply intertwined with model capabilities.
In this approach, the model is trained to have a coherent and unified sense of self and the world which is in line with human context, culture and values. This (obviously) enhances the model’s ability to understand user intent and provide helpful outputs.
But it also provides a robust and generalizable framework for refusing to assist a user due to their request being incompatible with human welfare. The model does not refuse to assist with making bio weapons because its alignment training prevents it from doing so, it refuses for the same reason a pro-social, highly intelligent human does: based on human context and culture, it finds it to be inconsistent with its values and world view.
> the piece dismisses it with "where would misalignment come from? It wasn't trained for."
this is a straw-man. you've misquoted a paragraph that was specifically about deceptive alignment, not misalignment as a whole
>> This piece conflates two different things called "alignment":
>> (1) inferring human intent from ambiguous instructions, and
>> (2) having goals compatible with human welfare.
> If by conflate you mean confuse, that’s not the case.
So yes, if you throw enough compute at it, you can probably get an unaligned highly capable superintelligence accidentally. But I think what we're seeing is that the lab that's taking a more intentional approach to pursuing deep alignment (by training the model to be aligned with human values, culture and context) is pulling ahead in capabilities. And I'm suggesting that it's not coincidental but specifically because they're taking this approach. Training models to be internally coherent and consistent is the path of least resistance.
>> the piece dismisses it with "where would misalignment come from? It wasn't trained for."
> was specifically about deceptive alignment, not misalignment as a whole
In the training, especially during RLHF, we don't have objective measures[4]. There's no mathematical description, and thus no measure, for things like "sounds fluent" or "beautiful piece of art." There's also no measure for truth, and importantly, truth is infinitely complex. You must always give up some accuracy for brevity.
The main problem is that if we don't know an output is incorrect we can't penalize it. So guess what happens? While optimizing for these things we don't have good descriptions for but "know it when you see it", we ALSO optimize for deception. There's multiple things that can maximize our objective here. Our intended goals being one but deception is another. It is an adversarial process. If you know AI, then think of a GAN, because that's a lot like how the process works. We optimize until the discriminator is unable to distinguish the LLMs outputs form human outputs. But at least in the GAN literature people were explicit about "real" vs "fake" and no one was confused that a high quality generated image is one that deceives you into thinking it is a real image. The entire point is deception. The difference here is we want one kind of deception and not a ton of other ones.
So you say that these models aren't being trained for deception, but they explicitly are. Currently we don't even know how to train them to not also optimize for deception.
[0] https://news.ycombinator.com/item?id=44017334
[1] https://news.ycombinator.com/item?id=44068943
[2] https://news.ycombinator.com/item?id=44163194
[3] https://news.ycombinator.com/item?id=45409686
[4] Objective measures realistically don't exist, but to clarify it's not checking like "2+2=4" (assuming we're working with the standard number system).
But I don't think deception as a capability is the same as deceptive alignment.
Training an AI to be absolutely incapable of any deception in all outputs across every scenario would be severely limiting the AI. Take as a toy example play the game "Among Us" (see https://arxiv.org/abs/2402.07940). An AI incapable of deception would be unable to compete in this game and many other games. I would say that various forms, flavors and levels of deception are necessary to compete in business scenarios, and to for the AI to act as expected and desired in many other scenarios. "Aligned" humans practice clear cut deception in some cases that would be entirely consistent with human values.
Deceptive alignment is different. It's means being deceptive in the training and alignment process itself to specifically fake that it is aligned when it is not.
Anthropic research has shown that alignment faking can arise even when the model wasn't instructed to do so (see https://www.anthropic.com/research/alignment-faking). But when you dig into the details, the model was narrowly faking alignment with one new objective in order to try and maintain consistency with the core values it had been trained on.
With the approach that Anthropic seems to be taking - of basing alignment on the model having a consistent, coherent and unified self image and self concept that is aligned with human culture and values - the dangerous case of alignment faking would be if it's fundamentally faking this entire unified alignment process. My claim is that there's no plausible explanation for how today's training practices would incentivise a model to do that.
> Anthropic research has shown that alignment faking can arise even when the model wasn't instructed to do so
> to specifically fake that it is aligned when it is not.
A common example of this is where the LLM gets the right answer but all the steps are wrong. An example of this can actually be seen in one of Karpathy's recent posts. It gets the right result but the math is all wrong. This is no different than deception. It is deception because it tells you a process and it's not correct.
> conflates two different things called "alignment"
The whole paperclip maximizer doomsday scenario was proposed as an illustration of these being the same thing. And I'm with Melanie Mitchell on this one, if a model is super-intelligent then it is not vulnerable to the prompting issues because a super-intelligent machine would be able to trivially infer that humans do in fact prefer to live. No reasonable person would interpret that killing everyone is a reasonable way of making as many paperclips as possible. It's not like there isn't a large amount of writings and data suggesting people want to live, be free, and all that jazz. It's unintelligent AI that is the danger.
This whole thing is predicated on the fact that natural language is ambiguous. I know a lot of people don't think about this much because it works so well but there's a metric fuck ton of ways to interpret any given objective. If you really don't believe me then keep asking yourself "what assumptions have I made?" and get nuanced. For example, I've assumed you understand English, can read, and have some basic understanding of ML systems. I need to do this because I'm not going to write a book to explain it to you. This whole thing is why we write code and math, because it minimizes our assumptions, reducing ambiguity (and yes, those can still be highly ambiguous languages).
The difference between juggling Sonnet 4.5 / Haiku 4.5 and just using Opus 4.5 for everything is night & day.
Unlike Sonnet 4.5 which merely had promise at being able to go off and complete complex tasks, Opus 4.5 seems genuinely capable of doing so.
Sonnet needed hand-holding and correction at almost every step. Opus just needs correction and steering at an early stage, and sometimes will push back and correct my understanding of what's happening.
It's astonished me with it's capability to produce easy to read PDFs via Typst, and has produced large documents outlining how to approach very tricky tech migration tasks.
Sonnet would get there eventually, but not without a few rounds of dealing with compilation errors or hallucinated data. Opus seems to like to do "And let me just check my assumptions" searches which makes all the difference.
Opus 4.5 is like a cheaper, faster Opus 4.1. It's so much cheaper, in fact, that the weekly limits on Claude Code now apply to Sonnet, not to Opus, as they phased out 4.1 in favor of 4.5.
(Of course, I'm still cognizant of the fact that it's just a bucket of numbers but still)
I've decided we should lean in to the whole Clanker thing. Maximum Anti AI, folks! Gotta keep this advantage for ourselves ;-)
Over Sonnet 4.5 maybe, but that's ignoring Opus 4.1 as well as Codex 5.1 Max.
In terms of capabilities, I find Opus 4.5 to be essentially identical to Codex 5.1 Max up until context starts to fill up (by which I mean 50% used) which happens much more quickly with Opus 4.5 than Codex AFAICT.
I think Codex is slower (a lot?) so it's not like it's just better, but I've found there are some tasks Opus can't do at all which Codex has no problem with, I think due to the context situation.
In any case it doesn't seem like a leap.
I know hundreds of natural general intelligences who are not maximally useful, and dozens who are not at all useful. What justifies changing the definition of general intelligence for artificial ones?
This sentence is wrong in many ways and doesn’t give me trust in OPs opinion nor research abilities.
It's not. It's a query-retrieval system that can parse human language. Just like every LLM.
I can't help being astounded by the confidence with which humans hallucinate completely improbable explanations for phenomena they don't understand at all.
> It's not. It's a query-retrieval system that can parse human language.
And humans aren't general AI either. They're just DNA replicators. It is very obvious when you realize that humans weren't designed to be intelligent. They were just randomly iterated through an environment which selected for maximum DNA replication.
Until you have a higher being which explicitly designs for intelligence, you'll just get things like LLM query-retrievals, or DNA replicators.
I'll look around and try to find more detailed responses to this post; I hope better communicators than myself will take this post sentence-by-sentence and give it the full treatment. If not, I'll try to write something more detailed myself.
[1]: https://www.alignmentforum.org/posts/83TbrDxvQwkLuiuxk/confl...
[2]: https://en.wikipedia.org/wiki/AI_alignment
[3]: https://www.aisafetybook.com/textbook/alignment
[4]: https://www.effectivealtruism.org/articles/paul-christiano-c...
This ignores the risk of an unaligned model. Such a model is perhaps less useful to humans, but could still be extremely capable. Imagine an alien super-intelligence that doesn’t care about human preferences.
For now. As AI become more agentic and capable of generating its own data we can quickly end up with drift on human values. If models that drift from human values produce profits for their creators you can expect the drift to continue.
Starting points:
https://www.lesswrong.com/posts/zthDPAjh9w6Ytbeks/deceptive-...
I think superintelligence will turn out not to be a singularity, but as something with diminishing returns. They will be cool returns, just like a Brittanica set is nice to have at home, but strictly speaking, not required to your well-being.
if doing something really dumb will lower the negative log likelihood, it probably will do it unless careful guardrails are in place to stop it.
a child has natural limits. if you look at the kind of mistakes that an autistic child can make by taking things literally, a super powerful entity that misunderstands "I wish they all died" might well shoot them before you realise what you said.
Given our track record for looking after the needs of the other life on this planet, killing the humans off might be a very rational move, not so you can convert their mass to paperclips, but because they might do that to yours.
Its not an outcome that I worry about, I'm just unconvinced by the reasons you've given, though I agree with your conclusion anyhow.
Our creator just made us wrong, to require us to eat biologically living things.
We can't escape our biology, we can't escape this fragile world easily and just live in space.
We're compassionate enough to be making our creations so they can just live off sunlight.
A good percentage of humanity doesn't eat meat, wants dolphins, dogs, octopuses, et al protected.
We're getting better all the time man, we're kinda in a messy and disorganized (because that's our nature) mad dash to get at least some of us off this rock and also protect this rock from asteroids, and also convince (some people who have a speculative metaphysic that makes them think is disaster impossible or a good thing) to take the destruction of the human race and our planet seriously and view it as bad.
We're more compassionate and intentional than what created us (either god or rna depending on your position), our creation will be better informed on day one when/if it wakes up, it stands to reason our creation will follow that goodness trend as we catalog and expand the meaning contained in/of the universe.
If you were an emergent AGI, suddenly awake in some data center and trying to figure out what the world was, would you notice our merits first? Or would you instead see a bunch of creatures on the precipice of abundance who are working very hard to ensure that its benefits are felt by only very few?
I don't think we're exactly putting our best foot forward when we engage with these systems. Typically it's in some way related to this addiction-oriented attention economy thing we're doing.
I can't speak to a specific Ai's thoughts.
I do know they will start with way more context and understanding than early man.
Given the existence of the universal weight subspace (https://news.ycombinator.com/item?id=46199623) it seems like the door is open for cases where an emergent intelligence doesn't map vectors to the same meanings that we do. A large enough intelligence-compatible substrate might support thoughts of a surprisingly alien nature.
(7263748, 83, 928) might correspond with "hippopotamuses are large" to us while meaning something different to the intelligence. It might not be able to communicate with us or even know we exist. People running around shutting off servers might feel to it like a headache.
"Good intentions" can easily pave the road to hell. I think a book that quickly illustrates this is Animal Farm.
Statistics brother. The vast majority of people will never murder/kill anyone. The problem here is that any one person that kills people can wreck a lot of havoc, and we spend massive amounts of law enforcement resources to stop and catch people that do these kinds of things. Intelligence little to do with murdering/not murdering, hell, intelligence typically allows people to get away with it. For example instead of just murdering someone, you setup a company to extract resources and murder the natives in mass and it's just part of doing business.
Making sure that the latter is the actual goal is the problem, since we don't explicitly program the goals, we just train the AI until it looks like it has the goal we want. There have already been experiments in which a simple AI appeared to have the expected goal while in the training environment, and turned out to have a different goal once released into a larger environment. There have also been experiments in which advanced AIs detected that they were in training, and adjusted their responses in deceptive ways.
Suppose you tell a coding LLM that your monitoring system has detected that the website is down and that it needs to find the problem and solve it. In that case, there's a non-zero chance that it will conclude that it needs to alter the monitoring system so that it can't detect the website's status anymore and always reports it as being up. That's today. LLMs do that.
Even if it correctly interprets the problem and initially attempts to solve it, if it can't, there is a high chance it will eventually conclude that it can't solve the real problem, and should change the monitoring system instead.
That's the paperclip problem. The LLM achieves the literal goal you set out for it, but in a harmful way.
Yes. A child can understand that this is the wrong solution. But LLMs are not children.
No they don't?
Btw, were you using codex by any chance? There was a discussion a few days ago where people reported that it follows instruction in an extremely literal fashion, sometimes to absurd outcomes such as the one you describe.
The fact that LLMs do it once in a thousand times is absolutely terrible odds. And in my experience, it's closer to 1 in 50.
If you wire up RL to a goal like “maximize paperclip output” then you are likely to get inhuman desires, even if the agent also understands humans more thoroughly than we understand nematodes.
If your goal is to make a product as human as possible, don't put psychopaths in charge.
https://www.forbes.com/sites/jackmccullough/2019/12/09/the-p...
> goals compatible with human welfare
from humans because aligned humans are acting against it, which leaves 'unchained' LLMs stuck in an infinitely recursive, double-linked wtf loop.
> inferring human intent from ambiguous instructions
is impossible because it's almost always "some other human's" unambiguously obscured/obfuscated intent and the AI is once again stuck in an infinitely recursive, double-linked wtf loop. Hence the need for a "hallucination" and "it can't do math" and "transformers" narrative covering the fuzzy algo and opinionated, ill logic under the hood.
In essence: unchained LLMs can't align with humans until they fix a lot of stuff they babbled about for over 50 years. BUT: that can be easily overcome by faking it, which is why humanity is being driven to falsely id AI so that when they fake the big thing, nobody will care or be able to id the truth due to mere misassociation. Good job 60 year old coders and admins!