Hi Jason, thank you for giving a quick response. Both points are very reasonable.
The contest announcement post outlined "several ways an essay could substantively inform the thinking of a panelist", namely, changing the central estimate or shape of the probability distribution of AGI / AGI catastrophe, or clarifying a concept or identifying a crux.
It would be very interesting to hear if any of the submissions did change any of the panelists' (or other Open Phil employees') mind in these ways, and how so. If not, whether because you learned an unanticipated kind of a thing, or because the contest turned out to be less useful than you initially hoped, I think that might also be very valuable for the community to know.
Many people live in an area (or country) where there isn't even a single AI safety organization, and can't or don't want to move. In that sense - no they can't even join an existing organization (in any level).
(I think founding an organization has other advantages over joining an existing one, but this is my top disagreement.)
I am not a cryptographer (though I do have knowledge in cryptography) and did not try to run an explicit cost-benefit analysis. Nevertheless, I think that being a researcher doing similar things to the other researchers in this field is not likely to be impactful (for various reasons, e.g., quantum resistant ciphers exist, while quantum computers [at scale] do not; and this research field is quite active so I don't think there will be any shortage of solutions in the future). I think that it is possible (but not plausible) that one could come up with an alternative path within this field that would be impactful, but I don't have any such ideas. Moreover, that would have to be a case-by-case analysis rather than an a priori cost-benefit analysis, so I am not sure carrying such an analysis would be helpful, and I would rather try to think of such alternative paths. (One alternative would be working on cryptography in the industry to earn-to-give. I actually think this is not a bad idea if one is a good fit.)
Glad to hear this was helpful! In short, the answer is no, it was much less structured (though also thought of).
These decisions were made under very different circumstances, and I'm not sure there is much value in explaining how I arrived at them. Very quickly though - I chose math because I highly enjoyed it and had the time (I did consider alternatives, such as physics and CS); cyber-security was part of my military service, which did involve some choice, but it is a very different situation and would drag me into a long digression.
Apologies if this doesn't help much. Feel free to ask any other questions (publicly or privately).
Thanks Cameron. I think that I understand our differences in views. My understanding is that you argue that language agents might be a safe path (I am not sure I fully agree with this, but I am willing to be on board so far).
Our difference then is, as you say, in whether there are models which are not safe and whether this is relevant. In Section 5, on the probability of misalignment, and in your last comment, you suggest that it is highly likely that language agents are the path forward. I am not at all convinced that this is correct (e.g., I think that it is more likely that systems like I mentioned will be more useful/profitable or just work better somehow, even in the near future) - you would have to convince a lot of people to use language agents alone, and that wouldn't happen easily. Therefore, I think that it is relevant that there are other models which do not exhibit the sort of safety guarantees you think language agents have. Hope this clears our differences.
(I would like to mention again that I appreciate your thoughts on language agents, and your engagement with my criticism.)
I think the following might be a difference in our views: I expect that people will(/are) trying to train LLM variants that are RLHFed to express agentic behavior. There's no reason to have one model to rule them all - it only makes sense to have a distinct models for short conversations and for autonomous agents. Maybe the agentic version would get a modified prompt including some background. Maybe it will be given context from memory as you specified.
Do you disagree with this?
Given all of the above, I don't see a big difference between this and how other agents
(humans/RL systems/what have you) operate, aside maybe from the fact that the memory is more external.
In other words - I expect your point (i) to be in the prompt/LLM weights variant (via RLHF or some other modification, (ii) this is the standard convergent instrumental goals argument (which is relevant to these systems as much as to others, a priori), and (iii) again by this external memory (which could for example be a chain of thought or otherwise).
I, for one, think that it is good that climate change was not mentioned. Not necessarily because there are no analogies and lessons to be drawn, but rather because it can more easily be misinterpreted. I think that the kind of actions and risks are much more similar to bio and nuclear, in that there are way less actors and, at least for now, it is much less integrated to day-to-day life. Moreover, in many scenarios, the risk itself is of more abrupt and binary nature (though of course not completely so), rather than a very long and gradual process. I'd be worried that comparing AI safety to climate change would be easily misinterpreted or dismissed by irrelevant claims.
Thank you, Cameron and Simon, for writing this. It articulates some thoughts I've ben pondering. However, I would like to give one pushback, which I think is fairly significant. The relevant paragraph, which summarizes what I think is a wrong assumption, is the following:
We think this worry is less pressing than it might at first seem. The LLM in a language agent is integrated into the architecture of the agent as a whole in a way that would make it very difficult for it to secretly promote its own goals. The LLM is not prompted or otherwise informed that its outputs are driving the actions of an agent, and it does not have information about the functional architecture of the agent. This means that it has no incentive to answer prompts misleadingly and no understanding of what sorts of answers might steer the agent’s behavior in different ways. Moreover, since the model weights of the LLM are not updated in the process of operating a language agent, the only way for it to pursue a long-term plan by manipulating an agent would be to store information about that plan in the agent’s memory. But information stored in this way would not be secret.
I think that the highlighted part is wrong already today in an implicit way, and might be more explicitly broken in the (near) future. Processes like RLHF (used for GPT-4) or RLAIF (used for Claude) change the LLM's weights (admittedly during training) by evaluating its behavior on relatively long tasks (in comparison to next token prediction or similar tasks). Loosely speaking, this essentially informs the LLM that it is being used as a foundation for an agent. This, at least in principle, reintroduces the pressure to steer the agent, and raises again the problem of goal misgeneralization, as policies might be learned into the LLM's weights during this process which generalize poorly to other contexts.
(Of lesser importance, I think that the last two sentences in the quoted paragraph are also assuming that not only would the agent's memory be non-secret, but also interpretable [e.g., written in English], and I don't see why this has to be the case.)
I might have missed a pointed in your argument where you address this points. In any case, I would appreciate hearing your thoughts on this.
Not answering the question, but I would like to quickly mention a few of the benefits of having confidence/credible intervals or otherwise quantifying uncertainty. All of these comments are fairly general, and are not specific criticisms of GiveWell's work.
Decision making under risk aversion - Donors (large or small) may have different levels of risk aversion. In particular, some donors might prefer having higher certainty of actually making an impact at the cost of having a lower expected value. Moreover, (mostly large) donors could build a portfolio of different donations in order to achieve a better risk profile. To that end, one needs to know more about the distribution rather than a point-estimate.
Point-estimates are many times done badly - It is fairly easy to make many kinds of mistakes when doing point-estimates, some of which are more noticeable when quantifying uncertainties. To name one example, point-estimates of cost-effectiveness typically try to estimate the expected value, and is many times calculated as a product of different factors. While it is true that expected value is multiplicative (assuming that the factors are uncorrelated or, more generally, independent, which is also sometimes not the case but that's another problem), this is not true for other statistics, such as the median. I think it is a common mistake to use an estimate of the median for the mean, or something in between, which in many cases are wildly different.
Sensitivity analysis - Quantifying uncertainty allows for sensitivity analysis, which serves many purposes, one of which is to get more accurate (point-)estimate and reduce uncertainty. One example is by understanding which parameters are the most uncertain, and focus further (internal and external) research on improving their certainty.
In direct response to Hazelfire's comment, I think that even if the uncertainty spans only one order of magnitude (he mentioned 2-3, which seems reasonable to me), this could have a really larger effect on resource allocation. The bar for funding is currently 8x relative to GiveDirectly IIRC, which is one order of magnitude, so gaining a better understanding of the certainty could be really important. For instance, we could learn that some interventions which are currently above the bar, are not very clearly so, whereas other interventions which seem to be under the bar but very close to it, could turn out to be fairly certain and thus perhaps a very safe bet.
I think that all of these effects could have a large influence on GiveWell's recommendations and donors choices, future research, and directly on getting more accurate point-estimates (which could potentially be fairly big).
Yeah, that makes sense, and is fairly clear selection bias. Since here in Israel we have a very strong tech hub and many people finishing their military service in elite tech units, I see the opposite selection bias, of people not finding too many EA (or even EA-inspired) opportunities that are of interest to them.
I failed to mention that I think your post was great, and I would also love to see (most of) these critiques flashed out.
Hi Jason, thank you for giving a quick response. Both points are very reasonable.
The contest announcement post outlined "several ways an essay could substantively inform the thinking of a panelist", namely, changing the central estimate or shape of the probability distribution of AGI / AGI catastrophe, or clarifying a concept or identifying a crux.
It would be very interesting to hear if any of the submissions did change any of the panelists' (or other Open Phil employees') mind in these ways, and how so. If not, whether because you learned an unanticipated kind of a thing, or because the contest turned out to be less useful than you initially hoped, I think that might also be very valuable for the community to know.
Thanks!
Many people live in an area (or country) where there isn't even a single AI safety organization, and can't or don't want to move. In that sense - no they can't even join an existing organization (in any level).
(I think founding an organization has other advantages over joining an existing one, but this is my top disagreement.)
I am not a cryptographer (though I do have knowledge in cryptography) and did not try to run an explicit cost-benefit analysis. Nevertheless, I think that being a researcher doing similar things to the other researchers in this field is not likely to be impactful (for various reasons, e.g., quantum resistant ciphers exist, while quantum computers [at scale] do not; and this research field is quite active so I don't think there will be any shortage of solutions in the future). I think that it is possible (but not plausible) that one could come up with an alternative path within this field that would be impactful, but I don't have any such ideas. Moreover, that would have to be a case-by-case analysis rather than an a priori cost-benefit analysis, so I am not sure carrying such an analysis would be helpful, and I would rather try to think of such alternative paths. (One alternative would be working on cryptography in the industry to earn-to-give. I actually think this is not a bad idea if one is a good fit.)
Glad to hear this was helpful! In short, the answer is no, it was much less structured (though also thought of).
These decisions were made under very different circumstances, and I'm not sure there is much value in explaining how I arrived at them. Very quickly though - I chose math because I highly enjoyed it and had the time (I did consider alternatives, such as physics and CS); cyber-security was part of my military service, which did involve some choice, but it is a very different situation and would drag me into a long digression.
Apologies if this doesn't help much. Feel free to ask any other questions (publicly or privately).
Thanks Cameron. I think that I understand our differences in views. My understanding is that you argue that language agents might be a safe path (I am not sure I fully agree with this, but I am willing to be on board so far).
Our difference then is, as you say, in whether there are models which are not safe and whether this is relevant. In Section 5, on the probability of misalignment, and in your last comment, you suggest that it is highly likely that language agents are the path forward. I am not at all convinced that this is correct (e.g., I think that it is more likely that systems like I mentioned will be more useful/profitable or just work better somehow, even in the near future) - you would have to convince a lot of people to use language agents alone, and that wouldn't happen easily. Therefore, I think that it is relevant that there are other models which do not exhibit the sort of safety guarantees you think language agents have. Hope this clears our differences.
(I would like to mention again that I appreciate your thoughts on language agents, and your engagement with my criticism.)
Thanks for responding so quickly.
I think the following might be a difference in our views: I expect that people will(/are) trying to train LLM variants that are RLHFed to express agentic behavior. There's no reason to have one model to rule them all - it only makes sense to have a distinct models for short conversations and for autonomous agents. Maybe the agentic version would get a modified prompt including some background. Maybe it will be given context from memory as you specified. Do you disagree with this?
Given all of the above, I don't see a big difference between this and how other agents (humans/RL systems/what have you) operate, aside maybe from the fact that the memory is more external.
In other words - I expect your point (i) to be in the prompt/LLM weights variant (via RLHF or some other modification, (ii) this is the standard convergent instrumental goals argument (which is relevant to these systems as much as to others, a priori), and (iii) again by this external memory (which could for example be a chain of thought or otherwise).
I, for one, think that it is good that climate change was not mentioned. Not necessarily because there are no analogies and lessons to be drawn, but rather because it can more easily be misinterpreted. I think that the kind of actions and risks are much more similar to bio and nuclear, in that there are way less actors and, at least for now, it is much less integrated to day-to-day life. Moreover, in many scenarios, the risk itself is of more abrupt and binary nature (though of course not completely so), rather than a very long and gradual process. I'd be worried that comparing AI safety to climate change would be easily misinterpreted or dismissed by irrelevant claims.
Thank you, Cameron and Simon, for writing this. It articulates some thoughts I've ben pondering. However, I would like to give one pushback, which I think is fairly significant. The relevant paragraph, which summarizes what I think is a wrong assumption, is the following:
I think that the highlighted part is wrong already today in an implicit way, and might be more explicitly broken in the (near) future. Processes like RLHF (used for GPT-4) or RLAIF (used for Claude) change the LLM's weights (admittedly during training) by evaluating its behavior on relatively long tasks (in comparison to next token prediction or similar tasks). Loosely speaking, this essentially informs the LLM that it is being used as a foundation for an agent. This, at least in principle, reintroduces the pressure to steer the agent, and raises again the problem of goal misgeneralization, as policies might be learned into the LLM's weights during this process which generalize poorly to other contexts.
(Of lesser importance, I think that the last two sentences in the quoted paragraph are also assuming that not only would the agent's memory be non-secret, but also interpretable [e.g., written in English], and I don't see why this has to be the case.)
I might have missed a pointed in your argument where you address this points. In any case, I would appreciate hearing your thoughts on this.
Not answering the question, but I would like to quickly mention a few of the benefits of having confidence/credible intervals or otherwise quantifying uncertainty. All of these comments are fairly general, and are not specific criticisms of GiveWell's work.
In direct response to Hazelfire's comment, I think that even if the uncertainty spans only one order of magnitude (he mentioned 2-3, which seems reasonable to me), this could have a really larger effect on resource allocation. The bar for funding is currently 8x relative to GiveDirectly IIRC, which is one order of magnitude, so gaining a better understanding of the certainty could be really important. For instance, we could learn that some interventions which are currently above the bar, are not very clearly so, whereas other interventions which seem to be under the bar but very close to it, could turn out to be fairly certain and thus perhaps a very safe bet.
I think that all of these effects could have a large influence on GiveWell's recommendations and donors choices, future research, and directly on getting more accurate point-estimates (which could potentially be fairly big).
Yeah, that makes sense, and is fairly clear selection bias. Since here in Israel we have a very strong tech hub and many people finishing their military service in elite tech units, I see the opposite selection bias, of people not finding too many EA (or even EA-inspired) opportunities that are of interest to them.
I failed to mention that I think your post was great, and I would also love to see (most of) these critiques flashed out.