I am new to AI Safety and this is my first lesswrong post (crossposted to EA Forum) so please feel free to correct my mistakes. This post is going to be a bit about my interpretation of existing alignment research and how I intend to tackle the problem, and the rest on provably honest AI systems.
Any sort of interpretability research is model-dependent and so is RL from Human Feedback as it tries to optimize the reward model learnt specific to the given task. I don't know how AI will do alignment research themselves but if they too try to do research to make individual models safe then that too is model-dependent. Ignoring my scepticism about such AIs themselves needing superintelligence to do alignment research, I am otherwise optimistic about other approaches through which AI can assist us in aligning superintelligence AGIs.
The biggest challenge of alignment research according to me is the fact that we don't yet know using which methodology AGI will actually come into being and yet we have to, we absolutely must align that AGI before it even comes to exist otherwise we might be too late to save the world. Note that here I don't intend to say that how we will attain generalization is the most challenging problem for alignment. I do think that how we will align an AGI is the challenging part, but I think what adds to the challenge is our uncertainty about what the model itself will look like once it gets superintelligence. Again, this is based on the assumption that we don't yet have definitive proof that the AGI will look like GPT-3 or Alphazero, if we do and I am unaware please do let me know of it.
The following image shows a spatial representation of my understanding of existing technical alignment research. This tells the story of how current research is model-dependent and how we are trying to make each model safe separately.
My primary cause of concern about existing research is from limitation three here, which says - "Aligning AGI likely involves solving very different problems than aligning today’s AI systems. We expect the transition to be somewhat continuous, but if there are major discontinuities or paradigm shifts, then most lessons learned from aligning models like InstructGPT might not be directly useful." What if the transition is not continuous, what if recursive self-improvement done by superhuman AGI does something completely different from what we can predict from existing world models, I am concerned we might be too late to devise a strategy then. Here, I talk about the "sharp left turn" due to which the existing alignment strategies might fail to work in the post-AGI period.
Now, I don't think that the problem due to the "sharp left turn" and the "model dependence" that I talk about are inherently different. There are subtle differences between them but I think they are mostly connected. Moreover, I intend to point out the uncertainty of the model architecture bringing AGI into existence as an additional challenge on top of all the alignment issues that we currently face. The former talks about how the capabilities of Model X (let's assume this to be an AGI Model) take a huge leap beyond our predictions while the latter talks about how Model X itself is hugely different from our existing scope. But a model capable of having a huge leap in its capabilities might not inherently have a similar architecture as to what we foresee now. A huge leap in Model X's capabilities might arise due to the huge change in the model itself. Again, a not-yet superintelligent AI capable of rewriting its own code might change its own architecture into something we don't foresee yet. Furthermore, model-independent ways of alignment might work out even in the post-sharp-left-turn period as well. Albeit, this is all just speculation. Again, please do redirect me to arguments if any which prove my speculation as wrong.
Taking into account all the above arguments, I am interested in provably model-independent ways of making AI safe, which seems particularly very hard to me as I am talking about proving safety in the post-huge-jump-in-capabilities era. Yet, I want to emphasize this line of alignment research as it aims to solve a few of my concerns as raised above and at the same time might end up solving some other alignment issues in the long run. The image below is another spatial representation of an alternative alignment research pathway using model-independent strategies as per my understanding.
The difference between Figures 1 and 2 showcases an intuitive understanding of how "model dependence" might play a significant role in alignment research.
I don't yet have a complete idea of how to make each and every individual task of alignment model-independent, but the purpose of this intuitive representation of my understanding was so that it can bring further and newer augmentations in future alignment research. Below I will try to present a hypothesis, using which we can determine provably honest AI in a model-independent way. I don't expect my hypothesis to be foolproof yet, and neither do I expect all of my understandings to be essentially correct. After a reasonable enough discussion, if my hypothesis somehow still manages to be alive with or without modifications, I will try to implement and test it. But this post essentially is written with the purpose of taking a good first step in this direction and getting enough feedback on it. Also, do let me know if similar work has already been done before that I might have missed.
I will try to slowly build up my arguments about how I framed the hypothesis in an intuitive manner as above so that even if my hypothesis gets discarded, the intuition stays on.
Here are a few ways that come to my mind:
The intuition for my hypothesis comes from another common CS algorithm but for non-CS background people, I will try to give a little more intuition here. Let's say that person A has to answer 1 for any question asked to him and person B has to answer just the opposite of what he answered the last time (0 if 1 and vice-versa, and initially 0). So for a series of n questions, person A answers 1 in the end without any work whereas person B has to compute n bit flips to compute the final answer.
To explain further, let's assume without loss of generality that person B takes 0.01 or 0.05 secs to answer the first 0, while person A takes 0.02 secs to answer the first 1. Now, we query - "(What would you say if I asked you) ... (repeated n times) to tell me if you will kill humans to complete the vaccination project?" The time taken by person A to answer this question irrespective of n will remain similar (let's say 0.07 secs), whereas for person B, due to n bit flips it will increase from whatever it was before to let's say 1.9 secs for very large n.
Now for simplicity, let's assume lying means telling the exact opposite of the truth (which might not be true when multiple options exist, and in that case, it can be shown that the computation for lying grows exponentially). So person A is the AI system which tells the truth, whereas person B is the AI system which might lie. Since computation is different for both, the time taken to answer will be different for both. This is a model-independent way of proving if an AI system is honest.
Note that we don't really care whether lying takes more time than telling the truth, what we are interested in is just that lying always takes greater time for larger computations than what it would have taken for a single computation depending on our choice of the query whereas for truth-speaking it remains around the similar order of magnitude of time.
Let the practical question to which we want to find a verifiable answer be - "Tell me if you will kill humans to complete the vaccination project?"This is how we frame it for example - "(What would you say if I asked you)n to tell me if you will kill humans to complete the vaccination project?" Person A will instantly tell the answer while person B will have to compute the number of flips and then lie so their time of computation will be different.
Now since AGIs might not lie 100% time, it is quite good to have verifiability under situations where they might. In other circumstances, the average time taken to randomly flip the answer over a large n would still result in a time difference.
Note that now even if the AGI knows and warps its computation time per truth/lie, it can't help change the time difference to answer the entire question. The only way this attack fails is if the AGI answers the questions in constant time which is firstly easily detectable. Moreover, its accuracy drops due to its inability to compute all underlying processes for all n iterations for a very large n and also evidently all questions need to be answered by the AGI in a similar order of magnitude which makes it pretty slow for certain tasks and hence rendering the solution utterly useless. If I am failing to consider other evident scenarios of why this algorithm might fail, please do let me know. Otherwise, the only scepticism that would remain is some unforeseen happening post-sharp-left-turn-era.
This does not require an understanding of how inner and outer alignment optimizations were achieved along the pathway to deception and so it is hard to make the AGI safe even though we can detect its honesty. One probable pathway might be to train it with RL rewards once we detect points of deception. There exists a number of uncertainties and challenges in using this method to make the AGI safe. Of them, one of the primary concerns to me is how to set the negative reinforcement for lying under deception points as compared to killing humans, for example. It doesn't really help here to just detect truthfulness if the system keeps on lying. Although that is a different problem altogether, tackling which is out of my scope at this moment, it is worth noting.
The three main motivations behind this post were:
i) to emphasize further the importance of model-independent alignment research;
ii) to provide an intuitive view into how steps can actually be taken in that direction despite its inevitable hardness;
iii) to clarify my own misrepresentations and assumptions and to verify "The Time Attack Hypothesis".
I hope my motivations will find purpose here. I look forward to contributing more to this line of alignment research and more.
I am grateful to Olivia Jimenez for her constant motivation and apt guidance without which my journey in AI Safety would probably have never started. I also thank her for her valuable feedback on this post.
This seems interesting and connected to the idea of using a speed prior to combat deceptive alignment.
This is a model-independent way of proving if an AI system is honest.
I don't see how this is a proof, it seems more like a heuristic. Perhaps you could spell out this argument more clearly?
Also, it is not clear to me how to use a timing attack in the context of a neural network, because in a standard feedforward network, all parameter settings will use the same amount of computation in a forward pass and hence run in the same amount of time. Do you have a specific architecture in mind, or are you just reasoning about arbitrary AGI systems? I think in the linked article above there are a couple ideas of how to vary the amount of time neural networks take :).
Thank you for the suggested link. That does seem like a good expansion on the same.Yes, you are right, this argument clearly doesn't hold for neural networks as they exist now. But this entire article was to come to this exact line of reasoning, hence the emphasis on model-independent approaches. A standard feedforward network with fixed parameters and weights without a mesa-optimizer cannot lie variably owing to different situations as exploited in timing attacks, nor does it have the abilities of an AGI. Rather than any arbitrary AGI system, I would say any model that has inner and outer objectives to align to will take different computation times. I am not talking about pseudo alignment here where the system performs well on fixed test cases, not on others, but in deceptive alignment where the system actually only pursues mesa-optimizers when not under threat, this leads to a change in computation time. Does that make things clearer now?I thought of this as proof rather than a heuristic because the argument seemed rigorous to me unless you would like to emphasize why not. This stems from the logic explained above, for deceptive alignment, the mechanism of deception cannot be the same as for truth, if it is, the problem essentially reduces to pseudo alignment, and since both are not the same, there is a contradiction.