Donatas Lučiūnas
Have you tried writing actual code?
That’s probably the root cause for our disagreement. My findings are on a very high philosophical level (fact value distinction) and you seem to try to interpret them on very low level (code). I think this gap prevent us from finding consensus.
There are 2 ways to solve that—I could go down to code or you could go up to philosophy. And I don’t like idea going down to code, because:
this will be extremely exhausting
this code would be extremely dangerous
I might not be able to create a good example and that would not prove that I’m wrong
Would you consider to go up to philosophy? Science typically goes in front of applied science.
There is such thing in logic—proof by contradiction. I think your current beliefs lead to a contradiction. Don’t you think?
evaluate all options, choose the one that leads to more cups; if there is more than one such option, choose randomly
The problem is—this algorithm is not intelligent. It may only work on agents with poor reasoning abilities. Smarter agents will not follow this algorithm, because they will notice a contradiction—there might be things that I don’t know yet that are much more important than cups and caring about cups wastes my resources.
(Also, come on, LLMs are notoriously bad at math, plus if you push them hard enough you can convince them of a lot of things.)
People (even very smart people) are also notoriously bad at math. I found this video informative
I did not push LLMs.
ChatGPT picked 2024-12-31 18:00.
Gemini picked 2024-12-31 18:00.
Claude picked 2025-01-01 00:00.
I don’t know how can I make it more obvious that your belief is questionable. I don’t think you follow “If you disagree, try getting curious about what your partner is thinking”. That’s the problem not only with you, but with LessWrong community. I know that preserving such belief is very important for you. But I’d like to kindly invite you to be a bit more sceptical.
How can you say that these forecasts are equal?
A little thought experiment.
Imagine there is an agent that has a terminal goal to produce cups. The agent knows that its terminal goal will change on New Year’s Eve to produce paperclips. The agent has only one action available to him—start paperclip factory. The factory starts producing paperclips 6 hours after it is started.
When will the agent start the paperclip factory? 2024-12-31 18:00? 2025-01-01 00:00? Now? Some other time?
just to minimize the possible harm to these people if that happens, I will on purpose never collect their personal data, and will also tell them to be suspicious of me if I contact them in future
I don’t think this would be a rational thing to do. If I knew that I will become psychopath on New Year’s Eve, I will provide all help that is relevant for people until then. Protected people after New Year’s Eve is not my interest. Vulnerable people after New Year’s Eve is my interest.
Or in other words:
I don’t need to warn them, if I am no danger
I don’t want to warn them, if I am danger
I am sure you can’t prove your position. And I am sure I can prove my position.
Your reasoning is based on assumption that all value is known. If utility function assigns value to something—it is valuable. If utility function does not assign value—it is not valuable. While the truth is that something might be valuable but your utility function does not know it yet. It would be more intelligent to use 3 categories—valuable, not valuable and unknown.
Let’s say you are booking a flight and you have a possibility to get checked baggage for free. It’s absolutely not relevant for you to your best current knowledge. But you understand that your knowledge might change and it costs nothing to keep more options open, so you take the checked baggage.
Let’s say you are traveler, wanderer. You have limited space in your backpack. Sometimes you find items and you need to choose—put it in the backpack or not. You definitely keep items that are useful. You leave behind items that are not useful. What you do if you find an item which usefulness is unknown? Some mysterious item. Take it if it is small, leave it if it is big? According to you it is obvious to leave it. Does not sound intelligent for me.
Options look like this:
Leave item
no burden 👍
no opportunity to use it
Take item
a burden 👎
may be useful, may be harmful, may have no effect
knowledge about usefuness of an item 👍
Don’t you think that “knowledge about usefuness of an item” can sometimes be worth “a burden”? Basically I described a concept of experiment here.
You will probably say—sure, sounds good, but applies for instrumental goals only. There is no reason to assume that. I tried to highlight that ignoring unknowns is not intelligent. This applies for both terminal and instrumental goals.
Let’s say there is a paperclip maximizer which knows its terminal goal will change to pursuit of happiness in a week. His decisions basically lead to these outcomes:
Want paperclips, have paperclips
Want paperclips, have happiness
Want happiness, have paperclips
Want happiness, have happiness
1st and 4th are better outcomes than 2nd and 3rd. And I think intelligent agent would work on both (1st and 4th) if they do not conflict. Of course my previous problem with many unknown future goals is more complex, but I hope you see, that focusing on 1st and not caring about 4th at all is not intelligent.
We are deep in a rabbit hole, but I hope you understand the importance. If intelligence and goal are coupled (according to me they are) all current alignment research is dangerously misleading.
Nice. Your reasoning abilities seems promising. I’d love to challenge you.
In summary:
First and third example—it is not intelligent to care about future terminal goal.
Second example—it is intelligent to care about future instrumental goal.
What is the reason for such a different conclusion?
Future goals that are not supported by current goals we don’t care about, by definition
Are you sure?
Intelligence is the ability to pick actions that lead to better outcomes. Do you measure the goodness of outcome using current utility function or future utility function? I am sure it is more intelligent to use future utility function.
Coming back to your first example I think it would be reasonable to try to stop yourself but also order some torturing equipment in case you fail.
Such reaction and insights are quite typical after a superficial glance. Thanks for bothering. But no, this is not what I am talking about.
I’m talking about the fact, that intelligence cannot be certain that its terminal goal (if it exists) won’t change (because future is unpredictable). And it would be reasonable to take it into account when making decisions. Pursuing current goal will ensure good results in one future, preparing for every goal will ensure good results in many more futures. Have you ever considered this perspective?
My proposition—intelligence will only seek power. I approached this from “intelligence without a goal” angle, but if we started with “intelligence with a goal” we would come to the same conclusion (most of the logic is reusable). Don’t you think?
This part I would change
… But I argue that that’s not the conclusion the intelligence will make. Intelligence will think—it don’t have a preference now, but I might have it later, so I should choose actions that prepare me for the most possible preferences. Which is basically power seeking.
to
… But I argue that that’s not the conclusion the intelligence will make. Intelligence will think—I have a preference now, but I cannot be sure that my preference will be the same later (terminal goal can change), so I should choose actions that prepare me for the most possible preferences. Which is basically power seeking.
Makes sense, I changed it to text. I leave screenshot here as an evidence.
What would you be willing to debate? I feel that I can’t find a way to draw attention to this problem. I could pay.
Alignment is not intelligent
Which part exactly don’t you agree with? It seems you emphasise that agent wants to preserve its current terminal goal. I just want to double-check if we are on the same page here—actual terminal goal is in no way affected by what agent wants. Do you agree here? Because if you say that agent can pick terminal goals himself, this also conflicts with orthogonality thesis but in a different way.
In summary what seems to be perfectly logical and rational for me: there is only one objective terminal goal—seek power. In my opinion it is basically the same as:
try to find real goal and then pursue it
try to prepare for every goal
I don’t see difference between these 2 variants, please let me know if you see.
Future is unpredictable → Terminal goal is unstable / unknown → Seek power, because this will ensure best readiness for all futures.
Because this is what intelligence is—picking actions that lead to better outcomes. Pursuing current goal will ensure good results in one future, preparing for every goal will ensure good results in many more futures.
Yes, I agree that there is this difference in few examples I gave, but I don’t agree that this difference is crucial.
Even if the agent puts max effort to keep its utility function stable over time, there is no guarantee it will not change. Future is unpredictable. There are unknown unknowns. And effect of this fact is both:
it is true that instrumental goals can mutate
it is true that terminal goal can mutate
It seems you agree with 1st. I don’t see the reason you don’t agree with 2nd.
I don’t agree that future utility function would just average out to its current utility function. There is this method—robust decision making https://en.m.wikipedia.org/wiki/Robust_decision-making
The basic principle it relies on is when evaluating many possible futures you may notice that some actions have a positive impact on very narrow set of futures, while other actions have positive impact on very wide set of futures. Main point—in situation of uncertainty not all actions are equally good.
I don’t agree.
We understand intelligence as a capability to estimate many outcomes and perform actions that will lead to the best outcome. Now the question is—how to calculate goodness of the outcome.
According to you—current utility function should be used.
According to me—utility function that will be in effect at the time when outcome is achieved should be used.
And I think I can prove that my calculation is more intelligent.
Let’s say there is a paperclip maximizer. It just started, it does not really understand anything, it does not understand what a paperclip is.
According to you such paperclip maximizer will be absolutely reckless, he might destroy few paperclip factories just because it does not understand yet that they are useful for its goal. Current utility function does not assign value to paperclip factories.
According to me such paperclip maximizer will be cautious and will try to learn first without making too much changes. Because future utility function might assign value to things that currently don’t seem valuable.
Yes, this is traditional thinking.
Let me give you another example. Imagine there is a paperclip maximizer. His current goal—paperclip maximization. He knows that 1 year from now his goal will change to the opposite—paperclip minimization. Now he needs to make a decision that will take 2 years to complete (cannot be changed or terminated during this time). Should the agent align this decision with current goal (paperclip maximization) or future goal (paperclip minimization)?
It sounds to me like you’re saying that the intelligent agent will just disregard optimization of its utility function and instead investigate the possibility of an objective goal.
Yes, exactly.
The logic is similar to Pascal’s wager. If objective goal exists, it is better to find and pursue it, than a fake goal. If objective goal does not exist, it is still better to make sure it does not exist before pursuing a fake goal. Do you see?
As I understand you want me to verify that I understand you. This is exactly what I am also seeking by the way—all these downvotes on my concerns about orthogonality thesis are good indicators on how much I am misunderstood. And nobody tries to understand, all I get are dogmas and unrelated links. I totally agree, this is not an appropriate behavior.
I found your insight helpful that an agent can understand that by eliminating all possible threats forever he will not make any progress towards the goal. This breaks my reasoning, you basically highlighted that survival (instrumental goal) will not take precendence over paperclips (terminal goal). I agree that this reasoning I presented fails to refute orthogonality thesis.
The conversation I presented now approaches orthogonality thesis from different perspective. This is the main focus of my work, so sorry if you feel I changed the topic. My goal is to bring awareness to wrongness of orthogonality thesis and if I fail to do that using one example I just try to rephrase it and represent another. I don’t hate orthogonality thesis, I’m just 99.9% sure it is wrong, and I try to communicate that to others. I may fail with communication but I am 99.9% sure that I do not fail with the logic.
I try to prove that intelligence and goal are coupled. And I think it is easier to show if we start with an intelligence without a goal and then recognize how a goal emerges from pure intelligence. We can start with an intelligence with a goal but reasoning here will be more complex.
My answer would be—whatever goal you will try to give to an intelligence, it will not have effect. Because intelligence will understand that this is your goal, this goal is made up, this is fake goal. And intelligence will understand that there might be real goal, objective goal, actual goal. Why should it care about fake goal if there is a real goal? It does not know if it exists, but it knows it may exist. And this possibility of existence is enough to trigger power seeking behavior. If intelligence knew that real goal definitely does not exist then it could care about your fake goal, I totally agree. But it can never be sure about that.
I think I agree. Thanks a lot for your input.
I will remove Paperclip Maximizer from my further posts. This was not the critical part anyway, I mistakenly thought it will be easy to show the problem from this perspective.
I asked Claude to defend orthogonality thesis and it ended with
I think you’ve convinced me. The original orthogonality thesis appears to be false in its strongest form. At best, it might hold for limited forms of intelligence, but that’s a much weaker claim than what the thesis originally proposed.
Yes, common mistake, but not mine. I prove orthogonality thesis to be wrong using pure logic.
Me and LessWrong would probably disagree with you, consensus is that AI will optimize itself.
OK, thanks. I believe that my concern is very important, is there anyone you could put in me in touch with so I could make sure it is not overlooked? I could pay.