I apologize. Those links don’t answer the question at all. I dashed off an answer in the middle of a social occasion and completely missed the most relevant piece of your question—quite possibly motivated reasoning led me to assume it was on my favorite topic, whether language model agents will be our first AGIs and how to align them.
Anyway, here’s my answer to your actual question. This isn’t something I’ve thought about before, because it’s not clear it will play a central role in alignment. (I’m curious if you see a more direct link to alignment than I’m seeing)
In sum, they’ll probably have some MR and CB; they can use the same strategies as humans can to correct them, but more reliably if it can be included in scripted prompts as part of the scaffolding that makes the LLMs into cognitive architectures (or real AGI). That’s basically to notice when you’re in a situation that would cause MR or CB, and do some extra cognitive work to counteract it.
Language model AGI will probably have MR, but less than humans:
Language model agents won’t have as much motivated reasoning as humans do, because they’re not probably going to use the same very rough estimated-value-maximization decision-making algorithm. (this is probably good for alignment; they’re not maximizing anything, at least directly. They are almost oracle-based agents).
But they may have a substantial amount of MR, to the extent that language encodes linked beliefs that were created by humans with lots of MR. I wonder if anyone has run tests that could indicate how much MR or CB they have, if that can somehow be disentangled from sycophancy.
And they will probably have some amount of confirmation bias, because language probably encodes the same sort of associative links that make it easier to think of confirming evidence than disconfirming.
How LMAs could correct for MR and CB (imperfectly but better-than-human—at a cost):
Like humans, they could correct for MR and CB by doing some compensatory cognition. A human could correct for MR by just sort of weight their beliefs against what they want to believe. That’s probably a good start for humans, but not nearly enough; see below.
I was initially thinking this part wouldn’t be necessary for an LMA since it probably won’t directly use a reward-predictive decision-making algorithm. But to the extent it’s trained with RLHF/RLAIF, it’s using a policy shaped by reward prediction (as are humans when they don’t explicitly predict consequences). This is an interesting distinction; the model is biased not to believe what it “wants”, but what the process that trained it “wants” in response to that prompt. Gauging that bias in order to compensate for it would be tricky. But I think the right scripted prompting could approximate it—at a risk of overcompensating, since we don’t have a good way to model exactly how much bias you’d have in a given circumstance.
Second and probably more important is compensating the compounding effect of MR changing how much evidence you’ve considered for and against beliefs you “like”. How far off your beliefs are is only an indirect result of how much you want to believe them. The direct cause is (I think) mostly how much you’ve looked at evidence and logic supporting that belief vs. evidence and logic that would disconfirm it. That makes compensating for them a lot harder after-the-fact; you’ve got to go back and consider as much evidence against as you considered for the hypothesis.
That leads us to the same correction for MR you’d do for confirmation bias: forcing yourself to look at evidence and arguments against your favored hypothesis.
Once again, it wouldn’t be easy to figure out just how much you’d need to compensate, so it’s not going to be a bias-free belief system, just less-biased-on-average.
And it would take a bunch of extra computation to go looking for and weighing evidence against all of the beliefs the system is biased toward. So this process would probably only be deployed when an answer is particularly important.
Adjusting for biases as a third function of a scripted internal review:
The scripted process for judging this and then performing that extra cognition would look a lot like the “internal review” I described in that post. That has at least a small tax for just making another call (perhaps to a separate model or instance) to evaluate whether this decision (including adopting a new belief) is important enough to carry out another whole scripted set of calls to evaluate its costs (including ethical costs) and then in the case of compensating for bias, do a bunch more cognition looking at disconfirming evidence/arguments.
This all stacks up to being pretty costly. I’d expect LMAs to have a good bit less MR and CB; they would sort of only have the “echoes” of them captured by standard linguistic patterns. They won’t directly have the feelings (pride, competitiveness, shame) that result in strong MR and thereby CB.
But I’m not sure. Again, I’m curious if anyone sees direct links to alignment. I’m currently worried about correct-but-unexpected changes in an agent’s beliefs and how that changes its functional alignment. Biases might make that worse, but I don’t see it opening up totally new dangers.
Question: Wouldn’t these imperfect bias-corrections for LMA’s also work similarly well for humans? E.g. humans could have a ‘prompt’ written on their desk that says “Now, make sure you spend 10min thinking about evidence against as well...” There are reasons why this doesn’t work so well in practice for humans (though it does help); might similar reasons apply to LMAs? What’s your argument that the situation will be substantially better for LMAs?
I’m particularly interested in elaboration on this bit:
Language model agents won’t have as much motivated reasoning as humans do, because they’re not probably going to use the same very rough estimated-value-maximization decision-making algorithm. (this is probably good for alignment; they’re not maximizing anything, at least directly. They are almost oracle-based agents).
I think there is an important reason things are different for LMAs than humans: you can program in a check for whether it’s worth correcting for motivated reasoning. Humans have to care enough to develop a habit (including creating reminders they’ll actually mind).
Whether a real AGI LMA would want to remove that scripted anti-bias part of their “artificial conscience” is a fascinating question; I think it could go either way, with them identifying it as a valued part of themselves, or an external check limiting their freedom of thought (same logic applies to internal alignment checks).
This also would substitute for a motivation that humans mostly don’t have. People, particularly non-rationalists, just aren’t trying very hard to arrive at the truth—because taking the effort to do that doesn’t serve their perceived interests.
Most often, humans don’t even want to correct for motivated reasoning. Firmly believing the same things as their friends and family serves them.
In important life decisions, they can benefit by countering MR and CB. I just added Staring into the abyss as a core life skill to the footnote, since it seems to be about exactly that.
Spending an extra ten minutes thinking about the counterevidence is usually a huge waste of time—unless you hugely value reaching correct conclusions on abstract matters that are likely irrelevant to your life success (I expect you do, and I do too—but it’s not hard to see why that’s a minority position).
Finally, there is no common knowledge of how big a problem MR/CB is, or how one might correct them.
I couldn’t find any study where they told people “try to compensate for this bias”, at least as of ~8 years ago when I was actively researching this.
Oracle-based agent is a term I’m toying with to intuitively capture how a language model agent still isn’t directly motivated by RL based on a goal. They are trained to have an accurate world model, and largely to answer questions as they were intended (although not necessarily accurately—sycophancy effects are large). They (in current form) are made agentic by having someone effectively ask “what would an agent do to accomplish this goal, given these tools?” and getting a correct-enough answer (which is then converted to actions by tools).
Sure, there are ways that the goals implicit in RLHF could deeply influence the LMA, giving them alien shoggoth-goals. That could happen if we optimize a lot more—including having a real AGI LMA reflect on its goals and beliefs for a long time.
But currently we’re actually training mostly for instruction-following. If we use that moderately wisely, it seems like we could head off the disasters of strongly optimizing for goals. That’s a brief diversion into the alignment implications of oracle-based (language model) agents; I’m not sure if that’s part of what you’re asking about, but there you go anyway.
So LMAs are currently selecting actions by trying to answer questions as their training encouraged. It seems LMAs are pretty strongly influenced by motivated reasoning based on their RL-based policy—but it isn’t their interests/desires that motivate their reasoning, but that of the RLHF respondents (or the constitution for RLAIF). They are sycophantic instead of motivated by their own predicted rewards as humans are.
That will cause them to be inaccurate but not misaligned, which seems more important.
Did that get at your interest in that passage, or am I once again misinterpreting your question?
This is helping, thanks. I do buy that something like this would help reduce the biases to some significant extent probably.
Will the overall system be trained? Presumably it will be. So, won’t that create a tension/pressure, whereby the explicit structure prompting it to avoid cognitive biases will be hurting performance according to the training signal? (If instead it helped performance, then shouldn’t a version of it evolve naturally in the weights?)
I’m not at all sure the overall system will be trained. Interesting that you seem to expect that with some confidence.
I’d expect the checks for cognitive biases to only call for extra cognition when a correct answer is particularly important to completing the task at hand. As such, it shouldn’t decrease performance much.
But I’m really not sure that training the overall system end-to-end is going to play a role. The success and relatively faithful CoT from r1 and QwQ give me hope that end-to-end training won’t be very useful.
Certainly people will try end-to-end training, but given the high compute cost for long-horizon tasks, I don’t think that’s going to play as large a role as piecewise and therefore fairly goal-agnostic training.
I think humans’ long-horizon performance isn’t mostly based on RL training, but our ability to reason and to learn important principles (some from direct success/failure at LTH tasks, some from vicarious experience or advice). So I expect the type of CoT RL training used in o1 to be used, as well as extensions to general reasoning where there’s not a perfectly check-able correct answer. That allows good System 2 reasoning performance, which I think is the biggerst basis of humans’ ability to perform useful LTH tasks.
Combining that with some form of continuous learning (either better episodic memory than vector databases and/or fine-tuning for facts/skills judged as useful) seems like all we need to get to human level.
Probably there will be some end-to-end performance RL, but that will still be mixed with strong contributions from reasoning about how to achieve a user-defined goal.
Gauging how much goal-directed RL is too much isn’t an ideal situation to be in, but it seems like if there’s not too much, instruction-following alignment will work.
WRT to cognitive biases, end-to-end training would increase some desired biases while decreasing some that are hurting performance (sometimes correct answers are very useful).
MR as humans experience it is only optimial within our very sharp cognitive limitations, and the types of tasks we tend to take on. So optimal MR for agents will be fairly different.
I’m curious about your curiousity; is it just that, or are you seeing a strong connection between biases in LMAs and their alignment?
But I’m really not sure that training the overall system end-to-end is going to play a role. The success and relatively faithful CoT from r1 and QwQ give me hope that end-to-end training won’t be very useful.
Huh, isn’t this exactly backwards? Presumably r1 and QwQ got that way due to lots of end-to-end training. They aren’t LMPs/bureaucracies.
...reading onward I don’t think we disagree much about what the architecture will look like though. It sounds like you agree that probably there’ll be some amount of end-to-end training and the question is how much?
My curiosity stems from: 1. Generic curiosity about how minds work. It’s an important and interesting topic and MR is a bias that we’ve observed empirically but don’t have a mechanistic story for why the structure of the mind causes that bias—at least, I don’t have such a story but it seems like you do! 2. Hope that we could build significantly more rational AI agents in the near future, prior to the singularity, which could then e.g. participate in massive liquid virtual prediction markets and improve human collective epistemics greatly.
Ah- now now I see how we’re using end-to-end training a little differently.
I’m thinking of end-to-end training as training a CoT, as in o1 etc, AND training an outer-loop control scheme for organizing complex plans and tasks. That outer-loop training is what I’m not sure will be helpful. You can script roughly how humans solve complex tasks, I think in a way that will help and not get in the way of the considerable power of the trained CoT. So I still expect this to play a role; I just don’t know how big a role it will play.
I agree that training CoT will be done, and has important implications for alignment, particularly in making CoTs opaque or unfaithful.
Reason 1: makes sense. One reason I never got around to publishing my work in comprehensible form is that it’s just adding mechanistic detail to an earlier essentially-correct theory that it’s just the result of applying self-interest (or equivalently, RL) to decisions about beliefs, in the context of our cognitive limitations. I’ll try to dig up the reference when I get a little time.
Reason 2: Interesting. I do hope we can align early agents well enough to get nontrivial help from them in solving the remaining alignment problems, both for AGI agents and for the power structures that are deploying and controlling them. Biases in their thinking have been less of a concern for me than their capabilities, but they will play a role. Sycophancy in particular seems like the language model equivalent of human motivated reasoning, and it could really wreck agents’ usefulness in making their user’s/controllers actions more sane.
I apologize. Those links don’t answer the question at all. I dashed off an answer in the middle of a social occasion and completely missed the most relevant piece of your question—quite possibly motivated reasoning led me to assume it was on my favorite topic, whether language model agents will be our first AGIs and how to align them.
Anyway, here’s my answer to your actual question. This isn’t something I’ve thought about before, because it’s not clear it will play a central role in alignment. (I’m curious if you see a more direct link to alignment than I’m seeing)
In sum, they’ll probably have some MR and CB; they can use the same strategies as humans can to correct them, but more reliably if it can be included in scripted prompts as part of the scaffolding that makes the LLMs into cognitive architectures (or real AGI). That’s basically to notice when you’re in a situation that would cause MR or CB, and do some extra cognitive work to counteract it.
Language model AGI will probably have MR, but less than humans:
Language model agents won’t have as much motivated reasoning as humans do, because they’re not probably going to use the same very rough estimated-value-maximization decision-making algorithm. (this is probably good for alignment; they’re not maximizing anything, at least directly. They are almost oracle-based agents).
But they may have a substantial amount of MR, to the extent that language encodes linked beliefs that were created by humans with lots of MR. I wonder if anyone has run tests that could indicate how much MR or CB they have, if that can somehow be disentangled from sycophancy.
And they will probably have some amount of confirmation bias, because language probably encodes the same sort of associative links that make it easier to think of confirming evidence than disconfirming.
How LMAs could correct for MR and CB (imperfectly but better-than-human—at a cost):
Like humans, they could correct for MR and CB by doing some compensatory cognition. A human could correct for MR by just sort of weight their beliefs against what they want to believe. That’s probably a good start for humans, but not nearly enough; see below.
I was initially thinking this part wouldn’t be necessary for an LMA since it probably won’t directly use a reward-predictive decision-making algorithm. But to the extent it’s trained with RLHF/RLAIF, it’s using a policy shaped by reward prediction (as are humans when they don’t explicitly predict consequences). This is an interesting distinction; the model is biased not to believe what it “wants”, but what the process that trained it “wants” in response to that prompt. Gauging that bias in order to compensate for it would be tricky. But I think the right scripted prompting could approximate it—at a risk of overcompensating, since we don’t have a good way to model exactly how much bias you’d have in a given circumstance.
Second and probably more important is compensating the compounding effect of MR changing how much evidence you’ve considered for and against beliefs you “like”. How far off your beliefs are is only an indirect result of how much you want to believe them. The direct cause is (I think) mostly how much you’ve looked at evidence and logic supporting that belief vs. evidence and logic that would disconfirm it. That makes compensating for them a lot harder after-the-fact; you’ve got to go back and consider as much evidence against as you considered for the hypothesis.
That leads us to the same correction for MR you’d do for confirmation bias: forcing yourself to look at evidence and arguments against your favored hypothesis.
Once again, it wouldn’t be easy to figure out just how much you’d need to compensate, so it’s not going to be a bias-free belief system, just less-biased-on-average.
And it would take a bunch of extra computation to go looking for and weighing evidence against all of the beliefs the system is biased toward. So this process would probably only be deployed when an answer is particularly important.
Adjusting for biases as a third function of a scripted internal review:
The scripted process for judging this and then performing that extra cognition would look a lot like the “internal review” I described in that post. That has at least a small tax for just making another call (perhaps to a separate model or instance) to evaluate whether this decision (including adopting a new belief) is important enough to carry out another whole scripted set of calls to evaluate its costs (including ethical costs) and then in the case of compensating for bias, do a bunch more cognition looking at disconfirming evidence/arguments.
This all stacks up to being pretty costly. I’d expect LMAs to have a good bit less MR and CB; they would sort of only have the “echoes” of them captured by standard linguistic patterns. They won’t directly have the feelings (pride, competitiveness, shame) that result in strong MR and thereby CB.
But I’m not sure. Again, I’m curious if anyone sees direct links to alignment. I’m currently worried about correct-but-unexpected changes in an agent’s beliefs and how that changes its functional alignment. Biases might make that worse, but I don’t see it opening up totally new dangers.
no need to apologize, thanks for this answer!
Question: Wouldn’t these imperfect bias-corrections for LMA’s also work similarly well for humans? E.g. humans could have a ‘prompt’ written on their desk that says “Now, make sure you spend 10min thinking about evidence against as well...” There are reasons why this doesn’t work so well in practice for humans (though it does help); might similar reasons apply to LMAs? What’s your argument that the situation will be substantially better for LMAs?
I’m particularly interested in elaboration on this bit:
I think there is an important reason things are different for LMAs than humans: you can program in a check for whether it’s worth correcting for motivated reasoning. Humans have to care enough to develop a habit (including creating reminders they’ll actually mind).
Whether a real AGI LMA would want to remove that scripted anti-bias part of their “artificial conscience” is a fascinating question; I think it could go either way, with them identifying it as a valued part of themselves, or an external check limiting their freedom of thought (same logic applies to internal alignment checks).
This also would substitute for a motivation that humans mostly don’t have. People, particularly non-rationalists, just aren’t trying very hard to arrive at the truth—because taking the effort to do that doesn’t serve their perceived interests.
Most often, humans don’t even want to correct for motivated reasoning. Firmly believing the same things as their friends and family serves them.
In important life decisions, they can benefit by countering MR and CB. I just added Staring into the abyss as a core life skill to the footnote, since it seems to be about exactly that.
Spending an extra ten minutes thinking about the counterevidence is usually a huge waste of time—unless you hugely value reaching correct conclusions on abstract matters that are likely irrelevant to your life success (I expect you do, and I do too—but it’s not hard to see why that’s a minority position).
Finally, there is no common knowledge of how big a problem MR/CB is, or how one might correct them.
I couldn’t find any study where they told people “try to compensate for this bias”, at least as of ~8 years ago when I was actively researching this.
Oracle-based agent is a term I’m toying with to intuitively capture how a language model agent still isn’t directly motivated by RL based on a goal. They are trained to have an accurate world model, and largely to answer questions as they were intended (although not necessarily accurately—sycophancy effects are large). They (in current form) are made agentic by having someone effectively ask “what would an agent do to accomplish this goal, given these tools?” and getting a correct-enough answer (which is then converted to actions by tools).
Sure, there are ways that the goals implicit in RLHF could deeply influence the LMA, giving them alien shoggoth-goals. That could happen if we optimize a lot more—including having a real AGI LMA reflect on its goals and beliefs for a long time.
But currently we’re actually training mostly for instruction-following. If we use that moderately wisely, it seems like we could head off the disasters of strongly optimizing for goals. That’s a brief diversion into the alignment implications of oracle-based (language model) agents; I’m not sure if that’s part of what you’re asking about, but there you go anyway.
So LMAs are currently selecting actions by trying to answer questions as their training encouraged. It seems LMAs are pretty strongly influenced by motivated reasoning based on their RL-based policy—but it isn’t their interests/desires that motivate their reasoning, but that of the RLHF respondents (or the constitution for RLAIF). They are sycophantic instead of motivated by their own predicted rewards as humans are.
That will cause them to be inaccurate but not misaligned, which seems more important.
Did that get at your interest in that passage, or am I once again misinterpreting your question?
This is helping, thanks. I do buy that something like this would help reduce the biases to some significant extent probably.
Will the overall system be trained? Presumably it will be. So, won’t that create a tension/pressure, whereby the explicit structure prompting it to avoid cognitive biases will be hurting performance according to the training signal? (If instead it helped performance, then shouldn’t a version of it evolve naturally in the weights?)
I’m not at all sure the overall system will be trained. Interesting that you seem to expect that with some confidence.
I’d expect the checks for cognitive biases to only call for extra cognition when a correct answer is particularly important to completing the task at hand. As such, it shouldn’t decrease performance much.
But I’m really not sure that training the overall system end-to-end is going to play a role. The success and relatively faithful CoT from r1 and QwQ give me hope that end-to-end training won’t be very useful.
Certainly people will try end-to-end training, but given the high compute cost for long-horizon tasks, I don’t think that’s going to play as large a role as piecewise and therefore fairly goal-agnostic training.
I think humans’ long-horizon performance isn’t mostly based on RL training, but our ability to reason and to learn important principles (some from direct success/failure at LTH tasks, some from vicarious experience or advice). So I expect the type of CoT RL training used in o1 to be used, as well as extensions to general reasoning where there’s not a perfectly check-able correct answer. That allows good System 2 reasoning performance, which I think is the biggerst basis of humans’ ability to perform useful LTH tasks.
Combining that with some form of continuous learning (either better episodic memory than vector databases and/or fine-tuning for facts/skills judged as useful) seems like all we need to get to human level.
Probably there will be some end-to-end performance RL, but that will still be mixed with strong contributions from reasoning about how to achieve a user-defined goal.
Gauging how much goal-directed RL is too much isn’t an ideal situation to be in, but it seems like if there’s not too much, instruction-following alignment will work.
WRT to cognitive biases, end-to-end training would increase some desired biases while decreasing some that are hurting performance (sometimes correct answers are very useful).
MR as humans experience it is only optimial within our very sharp cognitive limitations, and the types of tasks we tend to take on. So optimal MR for agents will be fairly different.
I’m curious about your curiousity; is it just that, or are you seeing a strong connection between biases in LMAs and their alignment?
Huh, isn’t this exactly backwards? Presumably r1 and QwQ got that way due to lots of end-to-end training. They aren’t LMPs/bureaucracies.
...reading onward I don’t think we disagree much about what the architecture will look like though. It sounds like you agree that probably there’ll be some amount of end-to-end training and the question is how much?
My curiosity stems from:
1. Generic curiosity about how minds work. It’s an important and interesting topic and MR is a bias that we’ve observed empirically but don’t have a mechanistic story for why the structure of the mind causes that bias—at least, I don’t have such a story but it seems like you do!
2. Hope that we could build significantly more rational AI agents in the near future, prior to the singularity, which could then e.g. participate in massive liquid virtual prediction markets and improve human collective epistemics greatly.
Ah- now now I see how we’re using end-to-end training a little differently.
I’m thinking of end-to-end training as training a CoT, as in o1 etc, AND training an outer-loop control scheme for organizing complex plans and tasks. That outer-loop training is what I’m not sure will be helpful. You can script roughly how humans solve complex tasks, I think in a way that will help and not get in the way of the considerable power of the trained CoT. So I still expect this to play a role; I just don’t know how big a role it will play.
I agree that training CoT will be done, and has important implications for alignment, particularly in making CoTs opaque or unfaithful.
Reason 1: makes sense. One reason I never got around to publishing my work in comprehensible form is that it’s just adding mechanistic detail to an earlier essentially-correct theory that it’s just the result of applying self-interest (or equivalently, RL) to decisions about beliefs, in the context of our cognitive limitations. I’ll try to dig up the reference when I get a little time.
Reason 2: Interesting. I do hope we can align early agents well enough to get nontrivial help from them in solving the remaining alignment problems, both for AGI agents and for the power structures that are deploying and controlling them. Biases in their thinking have been less of a concern for me than their capabilities, but they will play a role. Sycophancy in particular seems like the language model equivalent of human motivated reasoning, and it could really wreck agents’ usefulness in making their user’s/controllers actions more sane.