Re:LLM Powered Autonomous Agents
by ilaksh on Tuesday June 27, 2023 @06:54AM
(#36491238)
Good point. Part of the problem is the gap between GPT-4 and what people imagine to be dangerous is so large (on the surface).
The term "AGI" for many has morphed into meaning something that's like simulation of a person, but with godlike infinite IQ and powers. So it's a special type of ASI now. Not my definition, just the sloppy way most people seem to think about it these days.
So for those people, to get something dangerous, we have to create a full emulation of a person , give it a 1000 IQ, maybe make it operate at 1 million times human speed, etc.
But the reality is that for a group of AI agents to be dangerous they just need to be very effective problem solvers connected to the world.
I will say that the idea that many AI agents will not be connected to the internet or embodied is unrealistic. But I do think it's critical as you point out that we have awareness of the dangerous of full autonomy and effectors.
The trick is that right now we have a very strong economic motive to increase the performance and efficiency of LLMs. And I believe that we should because there is huge upside and the state of the world is horrendous.
But within only a few years (certainly less than ten) of optimizing this, we will be walking a fine line between tools that are extraordinarily empowering on the one side and completely out of our control on the other.
All we need for GPT to be dangerous is for it to get maybe 25% smarter and much faster. We should anticipate 50-100 faster than humans output within a few years. Within 5-10 years those systems will be extremely affordable and widely deployed.
The danger starts to emerge because the the extreme performance difference between human and AI. You can't tell the AI to wait for you to make decisions, because while it is waiting a few hours for you to figure out what it's talking about, the equivalent of days worth of activity has been accomplished by your competitor's AI agent.
For people to see the danger it could help to explain the exponential track record of computing efficiency performance and also create demonstrations of what it's like to compete with something that thinks 50 times faster than you. One would think that people could extrapolate from video games or something. But I guess the problem is the open world video game agents are still largely unintelligent so it's hard to imagine.
The term "AGI" for many has morphed into meaning something that's like simulation of a person, but with godlike infinite IQ and powers. So it's a special type of ASI now. Not my definition, just the sloppy way most people seem to think about it these days.
So for those people, to get something dangerous, we have to create a full emulation of a person , give it a 1000 IQ, maybe make it operate at 1 million times human speed, etc.
But the reality is that for a group of AI agents to be dangerous they just need to be very effective problem solvers connected to the world.
I will say that the idea that many AI agents will not be connected to the internet or embodied is unrealistic. But I do think it's critical as you point out that we have awareness of the dangerous of full autonomy and effectors.
The trick is that right now we have a very strong economic motive to increase the performance and efficiency of LLMs. And I believe that we should because there is huge upside and the state of the world is horrendous.
But within only a few years (certainly less than ten) of optimizing this, we will be walking a fine line between tools that are extraordinarily empowering on the one side and completely out of our control on the other.
All we need for GPT to be dangerous is for it to get maybe 25% smarter and much faster. We should anticipate 50-100 faster than humans output within a few years. Within 5-10 years those systems will be extremely affordable and widely deployed.
The danger starts to emerge because the the extreme performance difference between human and AI. You can't tell the AI to wait for you to make decisions, because while it is waiting a few hours for you to figure out what it's talking about, the equivalent of days worth of activity has been accomplished by your competitor's AI agent.
For people to see the danger it could help to explain the exponential track record of computing efficiency performance and also create demonstrations of what it's like to compete with something that thinks 50 times faster than you. One would think that people could extrapolate from video games or something. But I guess the problem is the open world video game agents are still largely unintelligent so it's hard to imagine.