What are the flaws in this argument about p(Doom)?
Technical alignment is hard
Technical alignment will take 5+ years
AI capabilities are currently subhuman in some areas (driving cars), about human in some areas (Bar exam), and superhuman in some areas (playing chess)
Capabilities scale with compute
The doubling time for AI compute is ~6 months
In 5 years compute will scale 2^(5÷0.5)=1024 times
In 5 years, with ~1024 times the compute, AI will be superhuman at most tasks including designing AI
Designing a better version of itself will increase an AI’s reward function
An AI will design a better version of itself and recursively loop this process until it reaches some limit
Such any AI will be superhuman at almost all tasks, including computer security, R&D, planning, and persuasion
The AI will deploy these skills to increase its reward function
Human survival is not in the AIs reward function
The AI will kill of most or all humans to prevent the humans from possibly decreasing its reward function
Therefore: p(Doom) is high within 5 years
Despite what the title says this is not a perfect argument tree. Which part do you think is the most flawed?
Edit: As per request the title has been changed from the humourous “An utterly perfect argument about p(Doom)” to “What are the flaws in this argument about p(Doom)?”
Edit2: yah Frontpage! Totally for the wrong reasons though
Edit3: added ”, with ~1024 times the compute,” to “In 5 years AI will be superhuman at most tasks including designing AI”
I kinda wanna downvote for clickbaity title.
Personally, I found it obvious that the title was being playful and don’t mind that sort of tongue-in-cheek thing. I mean “utterly perfect” is kind of a give away that they’re not being serious.
You are correct, I was not being serious. I was a little worried someone might think I was, but considered it a low probably.
Edit: this little stunt has cost me a 1 hour time limit on replies. I will reply to the other replies soon
Yes, I wanted to downvote too. But this is actually a good little argument to analyze. @William the Kiwi, please change the title to something like “What are the weaknesses in this argument for doom?”
As requested I have updated the title. How does the new one look?
Edit: this is a reply to the reply below, as I am commenting restricted but still want to engage with the other commenters: deleted
Edit2: reply moved to actual reply post
It’s fine. I have no authority here, that was really meant as a suggestion… Maybe the downvoters thought it was too basic a post, but I like the simplicity and informality of it. The argument is clear and easy to analyze, and on a topic as uncertain and contested as this one, it’s good to return to basics sometimes.
I think it was a helpful suggestion. I am happy that you liked the simplicity of the argument. The idea was it was meant to be as concise as possible to make the flaws seem more easy to spot. The argument relies on a range of assumptions but I deliberately left out the more confident assumptions. I find the topic of predicting AI development challenging, and was hoping this argument tree would be an efficient way of recognizing the more challenging parts.
Disagree vote this post if you disagree that the topic of predicting AI development is challenging.
Disagree vote this post if you disagree with liking the simplicity of the original post.
The above reply has two disagreement votes. I am trying to discern which reasons they are for. Disagree vote this post if you disagree that Mitchell_Porters suggestion was helpful.
Oops I realized I have used “flaws” rather than “weaknesses”. Do you consider these to be appropriate synonyms? I can update if not.
An AI doesn’t have to have a reward function, or one that implies self improvement. RFs often only apply at the training stage.
How would an AI be directed without using a reward function? Are there some examples I can read?
Current AIs are mostly not explicit expected-utility-maximizers. I think this is illustrated by RLHF (https://huggingface.co/blog/rlhf).
But isn’t that also using a reward function? The AI is trying to maximise the reward it receives from the Reward Model. The Reward Model that was trained using Human Feedback.
This does not follow, because subhuman AI can still accelerate R&D.
The OP’s argument can be modified to be immune to your objection:
This too seems like an improvement. However I would leave out the “kills us all” bit as this is meant to be the last line of the argument.
A fair comment. Would the following be an improvement? “Some technical alignment engineers predict with current tool and resources technical alignment will take 5+ years?”
For those who are downvoting this post: A short one sentence comment will help the original poster make better articles in the future.
Source?
This is a nitpick, but I think you meant 2^(5*2)=1024
This kind of clashes with the idea that AI capabilities gains are driven mostly by compute. If “moar layers!” is the only way forward, then someone might say this is unlikely. I don’t think this is a hard problem, but I thing its a bit of a snag in the argument.
I think you’ll lose some people on this one. The missing step here is something like “the AI will be able to recognize and take actions that increase its reward function”. There is enough of a disconnect between current systems and systems that would actually take coherent, goal-oriented actions that the point kind of needs to be justified. Otherwise, it leaves room for something like a GPT-X to just kind of say good AI designs when asked, but which doesn’t really know how to actively maximize its reward function beyond just doing the normal sorts of things it was trained to do.
I think this is a stronger claim than you need to make and might not actually be that well-justified. It might be worse than humans at loading the dishwasher bc that’s not important to it, but if it was important, then it could do a brief R&D program in which it quickly becomes superhuman at dish-washer-loading. Idk, maybe the distinction I’m making is pointless, but I guess I’m also saying that there’s a lot of tasks it might not need to be good at if its good at things like engineering and strategy.
Overall, I tend to agree with you. Most of my hope for a good outcome lies in something like the “bots get stuck in a local maximum and produce useful superhuman alignment work before one of them bootstraps itself and starts ‘disempowering’ humanity”. I guess that relates to the thing I said a couple paragraphs ago about coherent, goal-oriented actions potentially not arising even as other capabilities improve.
I am less and less optimistic about this as research specifically designed to make bots more “agentic” continues. In my eyes, this is among some of the worst research there is.
Thank you Jacob for taking the time for a detailed reply. I will do my best to respond to your comments.
Source: https://www.lesswrong.com/posts/sDiGGhpw7Evw7zdR4/compute-trends-comparison-to-openai-s-ai-and-compute. They conclude 5.7 months from the years 2012 to 2022. This was rounded to 6 months to make calculations more clear. They also note that “OpenAI’s analysis shows a 3.4 month doubling from 2012 to 2018”
I actually wrote it the (5*2) way in my first draft of this post, then edited it to (5÷0.5) as this is [time frame in years]÷[length of cycle in years], which is technically less wrong.
I think this is one of the weakest parts of my argument, so I agree it is definitely a snag. The move from “superhuman at some tasks” to “superhuman at most tasks” is a bit of a leap. I also don’t think I clarified what I meant very well. I will update to add ”, with ~1024 times the compute,”.
Would adding that suggested text to the previous argue step help? Perhaps “The AI will be able to recognize and take actions that increase its reward function. Designing a better version of itself will increase that reward function” But yea I tend to agree that there needs to be some sort of agentic clause in this argument somewhere.
Would this be an improvement? “Such any AI will be superhuman, or able to become superhuman, at almost all tasks, including computer security, R&D, planning, and persuasion”
I would speculate that most of our implemented alignment strategies would be meta-stable, they only stay aligned for a random amount of time. This would mean we mostly rely on strategies that hope to get x before we get y. Obviously this is a gamble.
I speculate that a lot of the x-risk probability comes from agentic models. I am particularly concerned with better versions of models like AutoGPT that don’t have to be very intelligent (so long as they are able to continuously ask GPT5+ how to act intelligent) to pose a serious risk.
Meta question: how do I dig my way out of a karma grave when I can only comment once per hour and post once per 5 days?
Meta comment: I will reply to the other comments when the karma system allows me to.
Edit: formatting