I’m an staff artificial intelligence engineer in Silicon Valley currently working with LLMs, and have been interested in AI alignment, safety and interpretability for the last 15 years. I’m now actively looking for employment working in this area.
RogerDearnaley
Karma: 1,504
Value learning converges to full alignment by construction: since a value learning AI basically starts with the propositions:
a) as an AI, I should act fully aligned to human values
b) I do not fully understand what human value are, or how to act fully aligned to them, so in order to be able to do this I need to learn more about human values and how to act fully aligned to them, by applying approximately Bayesian learning to this problem
c) Here are some Bayesian priors about what human values are, and how to act fully aligned to them: <insert initialization information here>…As usual for a Bayesian learning problem, as long as the Bayesian priors 1 c) are not completely screwed up as a place to start from, this will converge. Thus there is a region of convergence to full alignment.
LLMs have a very large amount of detailed information about what human values are and how to act aligned to them. Thus they provide a very detailed set of Bayesian priors for 1 c).
Also, training an LLM is a fairly good approximation to Bayesian learning. Thus (with suitable additions to enable online learning) they provide one possible implementation for the Bayesian learning process required by 1 b). For example, one could apply fine-tuning to the LLM to incorporate new information, and/or periodically retrain the LLM based on the training set plus new information the AI has gathered during the value learning process.
There had been a number of papers published over the last year on how to do this kind of training, and for roughly a year now there have been rumors that OpenAI were working on it. If converting that into a working version is possible for a Chinese company like DeepSeek, as it appears, then why haven’t Anthropic and Google released versions yet? There doesn’t seem to be any realistic possibility that DeepSeek actually have more compute or better researchers than both Anthropic and Google.
One possible interpretation would be that this has significant safety implications, and Anthropic and Google are both still working through these before releasing.
Another possibility would be that Anthropic has in fact released, in the sense that their Claude models’ recent advances in agentic behavior (while not using inference-time scaling) are distilled from reasoning traces generated by an internal-only model of this type that is using inference-time scaling.
If correct, this looks like an important theoretical advance in understanding why and under what conditions neural nets can generalize outside their training distribution.
So maybe part of the issue here is just that deducing/understanding the moral/ethical consequences of the options being decided between is a bit inobvious most current models, other than o1? (It would be fascinating to look at the o1 CoT reasoning traces, if only they were available.)
In which case simply including a large body of information on the basics of fiduciary responsibility (say, a training handbook for recent hires in the banking industry, or something) into the context might make a big difference for other models. Similarly, the possible misunderstanding of what ‘auditing’ implies could be covered in a similar way.
A much more limited version of this might be to simply prompt the models to also consider, in CoT form, the ethical/legal consequences of each option: that tests whether the model is aware of what fiduciary responsibility is, that it’s relevant, and how to apply it, if it is simply prompted to consider ethical/legal consequences. That would probably be more representative of what current models could likely do with minor adjustments to their alignment training or system prompts, the sorts of changes the foundation model companies could likely do quite quickly.
I think an approach I’d try would be to keep the encoder and decoder weights untied (or possibly add a loss term to mildly encourage them to be similar), but then analyze the patterns between them (both for an individual feature and between pairs of features) for evidence of absorption. Absorption is annoying, but it’s only really dangerous if you don’t know it’s happening and it causes you to think a feature is inactive when it’s instead inobviously active via another feature it’s been absorbed into. If you can catch that consistently, then it turns from concerning to merely inconvenient.
This is all closely related to the issue of compositional codes: absorption is just a code entry that’s compositional in the absorbed instances but not in other instances. The current standard approach to solving that is meta SAEs, which presumably should also help identify absorption. It would be nice to have a cleaner and simpler process than that: than that I’ve been wondering if it would be possible to modify top-k or jump-RELU SAEs so that the loss function cost for activating more common dictionary entries is lower, in a way that would encourage representing compositional codes directly in the SAE as two-or-more more common activations rather than one rare one. Obviously you can’t overdo making common entries cheap, otherwise your dictionary will just converge on a basis for the embedding space you’re analyzing, all of which are active all the time — I suspect using something like a cost proportional to might work, where is the dimensionality of the underlying embedding space and is the frequency of the dictionary entry being activated.
Interesting. I’m disappointed to see the Claude models do so badly. Possibly Anthropic needs to extend their constitutional RLAIF to cover not committing financial crimes? The large different between o1 Preview and o1 Mini is also concerning.
If these rumors are true, it sounds like we’re already starting to hit the issue I predicted in LLMs May Find It Hard to FOOM. The majority of content on the Internet isn’t written by geniuses with post-doctoral experience, so we’re starting to run out of the highest-quality training material for getting LLMs past doctoral student performance levels. However, as I describe there, this isn’t a wall, it’d just a slowdown: we need to start using AI to generate a lot more high-quality training data, As o1 shows, that’s entirely possible, using inference-time compute scaling and then training on the results. We’re having AI do the equivalent of System 2 thinking (in contexts where we can check the results are accurate), and then attempting to train a smarter AI that can solver the same problems by System 1 thinking.
However, this might be enough to render fast takeoff unlikely, which from an alignment point of view would be an excellent thing.
Now we just need to make sure all that synthetic training data we’re having the AI generate is well aligned.
Opacity: if you could directly inspect an AI’s motivations (or its cognition more generally), this would help a lot. But you can’t do this with current ML models.
The ease with which Anthropic’s model organisms of misalignment were diagnosed by a simple and obvious linear probe suggests otherwise. So does the number of elements in SAE feature dictionaries that describe emotions, motivations, and behavioral patterns. Current ML models are no longer black boxes: they rapidly becoming more-translucent grey boxes. So the sorts of applications for this you go on to discuss look like they’re rapidly becoming practicable.
Actual humans aren’t “aligned” with each other, and they may not be consistent enough that you can say they’re always “aligned” with themselves.
Completely agreed, see for example my post 3. Uploading which makes this exact point at length.
Anyway, even if the approach did work, that would just mean that “its own ideas” were that it had to learn about and implement your (or somebody’s?) values, and also that its ideas about how to do that are sound. You still have to get that right before the first time it becomes uncontrollable. One chance, no matter how you slice it.
True. Or, as I put it just above:
But yes, you do need to start the model off close enough to aligned that it converges to value learning.
The point is that you now get one shot at a far simpler task: defining “your purpose as an AI is to learn about and implement the humans’ collective values” is a lot more compact, and a lot easier to get right first time, than an accurate description of human values in their full large-and-fairly-fragile details. As I demonstrate in the post linked to in that quote, the former, plus its justification as being obvious and stable under reflection, can be described in exhaustive detail on a few pages of text.
As for the the model’s ideas on how to do that research being sound, that’s a capabilities problem: if the model is incapable of performing a significant research project when at least 80% of the answer is already in human libraries, then it’s not much of an alignment risk.
Yeah, that means you get exactly one chance to get “its own ideas” right, and no, I don’t think that success is likely.
Not if you built a model that does (or on reflection decides to do) value learning: then you instead get to be its research subject and interlocutor while it figures out its ideas. But yes, you do need to start the model off close enough to aligned that it converges to value learning.
A great paper highly relevant to this. That suggests that lying is localized just under a third of the way into the layer stack, significantly earlier than I had proposed. My only question is whether the lie is created before (at an earlier layer then) the decision whether to say it, or after, and whether their approach located one or both of those steps. They’re probing yes-no questions of fact, where assembling the lie seems trivial (it’s just a NOT gate), but lying is generally a good deal more complex than that.
That’s a great paper on this question. I would note that by the midpoint of the model, it has clearly analyzed both the objective viewpoint and also that of the story protagonist. So presumably it would next decide which of these was more relevant to the token it’s about to produce — which would fit with my proposed pattern of layer usage.
These models were fine-tuned from base models. Base models are trained with a vast amount of data to infer a context from the early parts of a document and then extrapolate that to predict later tokens, across a vast amount of text from the Internet and books, including actions and dialog from fictional characters. I.e they have been trained to observe and then simulate a wide variety of behavior, both of real humans, groups of real humans like the editors of a wikipedia page, and fictional characters. A couple of percent of people are psychopaths, so likely ~2% of this training data was written by psychopaths. Villains in fiction often also display psychopath-like traits. It’s thus completely unsurprising that a base model can portray a wide range of ethical stances, including psychopathic ones. Instruct training does not remove behaviors from models (so far we no know effective way to do that), it just strengthens some (making them occur more by default) and weakens others (making them happen less often by default) — however, there is a well-known theoretical result that any behavior the model is capable of, even if (now) rare, can be prompted to occur at arbitrarily levels with a suitably long prompt, and all that instruct-training or fine tuning can do is reduce the initial probability and lengthen the prompt required. So it absolutely will be possible to prompt an instruct-trained model to portray psychopathic behavior. Apparently the prompt required isn’t even long: all you have to do is tell it that it’s a hedge fund manager and not to break character.
Nothing in this set of results is very surprising to me. LLMs are can simulate pretty-much any persona
you ask them to. The hard part of alignment is not prompting them to be good, or bad — it’s getting them to stay that way (or detecting that they have not) after they’ve been fed another 100,000 tokens of context that may push them into simulating some other persona.
Having already played with this a little, it’s pretty amazing: the range of concepts you can find in the SAE, how clearly the autointerp has labelled them and how easy they are to find, and how effective they are (as long as you don’t turn them up too much) are all really impressive. I can’t wait to try a production model where you can set up sensors and alarms on features, clip or ablate them or wire them together at various layers, and so forth. It will also be really interesting to see how larger models compare.
I’d also love to start looking at jailbreaks with this and seeing what features the jailbreak is inducing in the residual stream. Finding the emotional/situational manipulation elements I suspect will be pretty easy — I’m curious to see if it will also show the ‘confusion’ effect of jailbreaks that read like confusing nonsense to a human as some form of confusion or noise, or if those are also emotional/situational manipulation just in a more noise-like adversarial format, comparable to adversarial attacks on image classifiers that just look like noise to a human eye, but actually effectively activate internal features of the vision model
…when Claude keeps telling me how I’m asking complex and interesting questions…
Yeah — also “insightful”. If it was from coming from character.ai I’d just assume it was flirting with me, but Claude is so very neuter and this just comes over as creepy and trying too hard. I really wish it would knock off the blatant intellectual flattery.
OpenAI CFO Sarah Friar warns us that the next model will only be about one order of magnitude bigger than the previous one.
The question is whether she’s talking parameter count, nominal training flops, or actual cost. In general, GPT generations so far have been roughly one order of magnitude apart in parameter count and training cost, and roughly two orders of magnitude in nominal training flops (parameter count x training tokens). Since she’s a CFO, and that was a financial discussion, I assume she natively thinks in terms of training cost, so the ‘correct’ answer to her is one order of magnitude not two, so my suspicion is that she’s actually talking in terms of parameter count. So I don’t think she’s warning us of anything, I think she’s just projecting a straight line on a logarithmic plot. I.e. business as usual at OpenAI.
Reversible computation means you aren’t erasing information, so you don’t lose energy in the form of heat (per Landauer[1][2]). But if you don’t erase information, you are faced with the issue of where to store it.
If you are performing a series of computations and only have a finite memory to work with, you will eventually need to reinitialise your registers and empty your memory, at which point you incur the energy cost that you had been trying to avoid. [3]
Generally, reversible computation allows you to avoid wasting energy by deleting any memory used on intermediate answers, and only do so only for final results. It does require that you have enough memory to store all those intermediate answers until you finish the calculation and then run it in reverse. If you don’t have that much memory, you can divide your calculation into steps, connected by final results from each step fed into the next step, and save on the energy cost of all the intermediate results within each step, and pay the cost only for data passed from one step to the next or output from the last step. Or, for a 4x slowdown rather than the usual 2x slowdown for reversible computation, you can have two sizes of step, and have some intermediate results that last only during a small step, and others that are retained for a large step before being uncomputed.
Memory/energy loss/speed trade-off management for reversible computation is a little more complex than conventional memory management, but is still basically simple, and for many computational tasks you can achieve excellent tradeoffs.
If there’s a change to human brains that human-evolution could have made, but didn’t, then it is net-neutral or net-negative for inclusive relative genetic fitness. If intelligence is ceteris paribus a fitness advantage, then a change to human brains that increases intelligence must either come with other disadvantages or else be inaccessible to evolution.
You’re assuming a steady state. Firstly, evolution takes time. Secondly, if humans were, for example, in an intelligence arms-race with other humans (for example, if smarter people can reliably con dumber people out of resources often enough to get a selective advantage out of it), then the relative genetic fitness of a specific intelligence level can vary over time, depending on how it compares to the rest of the population. Similarly, if much of the advantage of an IQ of 150 requires being able to find enough IQ 150 coworkers to collaborate with, then the relative genetic fitness of IQ 150 depends on the IQ profile of the rest of the population.
I like your point that humans aren’t aligned, and while I’m more optimistic about human alignment than you are, I agree that the level of human alignment currently is not enough to make a superintelligence safe if it only had human levels of motivation/reliability.
The most obvious natural experiments about what humans do when they have a lot of power with no checks-and-balances are autocracies. While there are occasional examples (such as Singapore) of autocracies that didn’t work out too badly for the governed, they’re sadly few and far between. The obvious question then is whether “humans who become autocrats” are a representative random sample of all humans, or if there’s a strong selection bias here. It seems entirely plausible that there’s at least some selection effects in the process of becoming an autocrat. A couple of percent of all humans are sociopaths, so if there were a sufficiently strong (two orders of magnitude or more) selection bias, then this might, for example, be a natural experiment about the alignment properties of a set of humans consisting mostly of sociopaths, in which case it usually going badly would be unsurprising.
The thing that concerns me is the aphorism “Power corrupts, and absolute power corrupts absolutely”. There does seem to be a strong correlation between how long someone has had a lot of power and an increasing likelihood of them using it badly. That’s one of the reasons for term limits in positions like president: humans seem to pretty instinctively not trust a leader after they’ve been in a position of a lot of power with few check-and-balances for roughly a decade. The histories of autocracies tend to reflect them getting worse over time, on decade time-scales. So I don’t think the problem here is just from sociopaths. I think the proportion of humans who wouldn’t eventually be corrupted by a lot of power with no checks-and-balances may be fairly low, comparable to the proportion of honest senior politicians, say.
How much of this argument applies to ASI agents powered by LLMs “distilled” from humans is unclear — it’s much more obviously applicable to uploads of humans that then get upgraded to super-human capabilities.
People who train text-to-image generative models have had a good deal of success with training (given a large enough and well-enough human-labeled training set) an “aesthetic quality” scoring model, and then training a generative image model to have “high aesthetic quality score” as a text label. Yes, doing things like this can produces effects like the recognizable Midjourney aesthetic, which can be flawed, and generally optimizing such things too hard leads to sameness — but if trained well such models’ idea of aesthetic quality is at least pretty close to most human judgements. Presumably what can be done for images can also be done for prose, poetry, or fiction as well.
There isn’t a direct equivalent of that approach for an LLM, but RLHF seems like a fairly close equivalent. So far people have primarily used RLHF for “how good is the answer to my question?” Adapting a similar approach to “how high quality is the poetry/prose/fiction produced by the model?” is obviously feasible. Then you just need a large number of high quality human judgements from a representative cross-section of people with good taste in poetry/prose/fiction: hiring professional human editors or literary talent scouts seems like a good idea. One of the good things about foundation model sizes and training costs going up is that reasonable budgets for fine-tuning should also increase proportionately.
Another option would be to train or fine-tune the quality scoring model used for the RL on literary sources (books, poetry, etc) with quality labels drawn from relatively objective existing data, like total sales, literary awards, critical rankings, reviews from good reviewers, and so forth.
The RLHF approach only trains a single aesthetic, and probably shouldn’t be taken too far or optimized too hard: while there is some widespread agreement about what prose is good vs, dreadful, finer details of taste vary, and should do so. So the obvious approach for finer-grained style control would be to train or fine-tune on a training set of a large number documents each of which consists of a prompt-like description/review/multiple reviews of a literary work, giving a variety of different types of aesthetic opinions and objective measures of its quality, followed by the corresponding literary work itself.
These ideas have been phrased as model-post-training suggestions, but turning these into a benchmark is also feasible: the “Aesthetic quality scoring model” from the RLHF approach is in itself a benchmark, and the “prompt containing reviews and statistics → literary work” approach could also be inverted to instead train a reviewer model to review literary works from various different aesthetic viewpoints, and estimate their likely sales/critical reception.