So on lesswrong, the esteemed Eric Drexler came up with a proposal for controlling AI systems. In short, rather than taking big monolithic AI systems who get internal memory and continuity of existence, you chop up the tasks you assign to them into the smallest, shortest duration subtasks you can, and you erase the systems memory every time it finishes a subtask or the time limit expires.
This fixes most AI alignment issues and makes almost all of them less probable.
Eliezers post in response was heavily downvoted and he essentially blew the idea off.
I have noticed that this proposal is equivalent to stateless microservices where each service communicates using some schema like Json/protobuf. This is a well known software technique and something that is convergent—all the tech giants end up with some flavor of it. This is how Netflix and Google both work in their backend.
This changes the estimates of long term AI risk by a lot, because if a plausible technical method to keep even superintelligent systems doing their jobs as part of a larger system created by humans, we should build ASI as fast as possible.
Technically if you think the benefits are positive—since you can use ASI to control robots to trap CO2, or mass produce space suits and bunkers to deal with bioterrorism attacks using AI designed weapons, or build missile defense in the quantities needed to stop an all out nuclear attack from Russia, or methodically construct larger and larger assemblies of human cells, manipulating them using the same molecular signals from embryology, and eventually discovering a way to make adult human organs in vitro …
Ok in short obviously asi systems you can prompt and assign any task to solve or mitigate any potential problem humans have in the future. Same way gpt-4 is a general solution to any homework problem below a certain level of difficulty.
So it’s a bipolar scenario.
If you think long term benefits of AGI/ASI are net + and near future, the EA answer is 100 percent investment into acceleration. This is what the stock market currently believes.
If you think they are net -, you should be 100 percent invested into slowing down AI by any means you have available.
No EA cause has any worth besides AI except ongoing efforts (like reducing lead exposure) that are cheap and proven and already happening.
So AI is all that matters, one way or another. (Unless you believe it is many years away). Just everyone on lesswrong posts this total confidence view that the outcome will be negative, which isn’t even evidence based—there are not many human technologies we have ever found that are net negative. I mean not everyone, there are a handful of credible dissidents, but usually I just see it as implied assumption. “Build ASI, it does whatever it wants, if it’s not aligned we die”.
The open agency model doesn’t give the ASI any agency to do anything but what we want. It’s just another form of computer system. (Obviously there will be bugs and sometimes it will cause outcomes that we didn’t want but this is also why you limit the scope of subtasks)
This is how current deployed models all work. Gpt-4 retains no state between API calls. Nobody else’s model does either. What Eliezer talks about as a threat does not yet exist and he openly admits that. He just believes it is coming within his remaining lifetime.
As geohot pointed out in the recent debate, timing matters. Details matter.
So on lesswrong, the esteemed Eric Drexler came up with a proposal for controlling AI systems. In short, rather than taking big monolithic AI systems who get internal memory and continuity of existence, you chop up the tasks you assign to them into the smallest, shortest duration subtasks you can, and you erase the systems memory every time it finishes a subtask or the time limit expires.
This fixes most AI alignment issues and makes almost all of them less probable.
Eliezers post in response was heavily downvoted and he essentially blew the idea off.
I have noticed that this proposal is equivalent to stateless microservices where each service communicates using some schema like Json/protobuf. This is a well known software technique and something that is convergent—all the tech giants end up with some flavor of it. This is how Netflix and Google both work in their backend.
This changes the estimates of long term AI risk by a lot, because if a plausible technical method to keep even superintelligent systems doing their jobs as part of a larger system created by humans, we should build ASI as fast as possible.
Technically if you think the benefits are positive—since you can use ASI to control robots to trap CO2, or mass produce space suits and bunkers to deal with bioterrorism attacks using AI designed weapons, or build missile defense in the quantities needed to stop an all out nuclear attack from Russia, or methodically construct larger and larger assemblies of human cells, manipulating them using the same molecular signals from embryology, and eventually discovering a way to make adult human organs in vitro …
Ok in short obviously asi systems you can prompt and assign any task to solve or mitigate any potential problem humans have in the future. Same way gpt-4 is a general solution to any homework problem below a certain level of difficulty.
So it’s a bipolar scenario.
If you think long term benefits of AGI/ASI are net + and near future, the EA answer is 100 percent investment into acceleration. This is what the stock market currently believes.
If you think they are net -, you should be 100 percent invested into slowing down AI by any means you have available.
No EA cause has any worth besides AI except ongoing efforts (like reducing lead exposure) that are cheap and proven and already happening.
So AI is all that matters, one way or another. (Unless you believe it is many years away). Just everyone on lesswrong posts this total confidence view that the outcome will be negative, which isn’t even evidence based—there are not many human technologies we have ever found that are net negative. I mean not everyone, there are a handful of credible dissidents, but usually I just see it as implied assumption. “Build ASI, it does whatever it wants, if it’s not aligned we die”.
The open agency model doesn’t give the ASI any agency to do anything but what we want. It’s just another form of computer system. (Obviously there will be bugs and sometimes it will cause outcomes that we didn’t want but this is also why you limit the scope of subtasks)
Thanks a lot for writing this thoughtful comment. (I hope you won’t be unreasonably downvoted here)
I wonder how much do top AI labs use these techniques? Do they know about them? What if they knew, but they didn’t use them for some reason?
So far this sounds good!
This is how current deployed models all work. Gpt-4 retains no state between API calls. Nobody else’s model does either. What Eliezer talks about as a threat does not yet exist and he openly admits that. He just believes it is coming within his remaining lifetime.
As geohot pointed out in the recent debate, timing matters. Details matter.