Situational awareness requires extrapolation to the alarming side of plausible in order to prepare for it.
Perhaps the core prediction here is that AI will reach the capability of an AI engineer in just three years. If true, we might consider that to be the functional singularity, in the sense that extrapolations past that point have little relevant basis.
AI engineer/AI researcher, e.g. in three years we will see AI do what maybe Andrej Karpathy does today? A complete 100% replacement that we can clone an infinite number of times? That seems hard to believe. But yes, if true then extrapolating what happens after that seems like science fiction. We have no tools to even imagine what that looks like.
EDIT: The graph also shows GPT-4 as "smart high schooler". I don't think that's really true. GPT-4 can perform some tasks at the level of a high schooler but is definitely not a smart high schooler in the more general sense. A smart high schooler can get a driver's license and drive a car around better than any AI in existence. And that's just one thing they can do.
I had a trouble ticket on the job today that I couldn't fix with a bunch of regular searches on the error. So I gave it to GPT-4o and it came up with the right answer on the first try. I just needed to ask it a few more questions to understand the answer. If a smart high schooler could have answered it as well and as quickly, that would be extraordinary. It saved me real time on real work on a day when I didn't want to be working.
What you're describing sounds more like a knowledge base and less like a general intelligence. The high schooler might fare worst on something that requires memorizing an encyclopedia but might fare better on an original problem that nobody has seen before. Was GPT-4o just synthesizing something from its trained data or did it have to do real problem solving?
The prompt was the error: "Mysql2::Error: Incorrect string value: '\\xF0\\x9D\\x9C\\x86 i...' for column 'variables' at row 1". It quickly led me to understand that this character is from 4-byte UTF8 but my db is set to 3-byte UTF8. Whether this is "real" problem solving or not, it helped me solve a real problem.
By the way, looks like this topic/paper had a few other competing posts and one of them got more discussion a few days ago. One thing is sure, these are interesting times in AI.
I agree it's hard to say. It seems more likely we'll plateau, along the lines of Pareto Principle. At the very least we still need a few more data points on this curve to give a prediction. But it's worth thinking about.
8 comments
[ 137 ms ] story [ 127 ms ] threadBut I've got no strong evidence one way or another: it's an extrapolation.
Perhaps the core prediction here is that AI will reach the capability of an AI engineer in just three years. If true, we might consider that to be the functional singularity, in the sense that extrapolations past that point have little relevant basis.
EDIT: The graph also shows GPT-4 as "smart high schooler". I don't think that's really true. GPT-4 can perform some tasks at the level of a high schooler but is definitely not a smart high schooler in the more general sense. A smart high schooler can get a driver's license and drive a car around better than any AI in existence. And that's just one thing they can do.
What you're describing sounds more like a knowledge base and less like a general intelligence. The high schooler might fare worst on something that requires memorizing an encyclopedia but might fare better on an original problem that nobody has seen before. Was GPT-4o just synthesizing something from its trained data or did it have to do real problem solving?
By the way, looks like this topic/paper had a few other competing posts and one of them got more discussion a few days ago. One thing is sure, these are interesting times in AI.