I'm not a fan of comparing modern AI to markets at all, as LLMs are nothing like a market built from real thinking human beings with actual broad scale intelligence. But to make the point that the...
I'm not a fan of comparing modern AI to markets at all, as LLMs are nothing like a market built from real thinking human beings with actual broad scale intelligence.
But to make the point that the AI control problem is one we are very familiar with, controlling non human intelligence so it meets our needs, is a good one.
Society is exactly such a monster, and humans are general intelligences that we have to fight to control every single day.
AI isn't new and we are hilariously well equipped to handle the control problem. We're only scared of it thanks to hype from people who think AI will explode into hyperinteligence with no proof or reason as to why.
And, in the long term, we've already lost the control problem. Assuming our society continues to progress give it a few thousand centuries and we will be more akin to cells than individuals, so specialized we will be.
What makes you say we are "hilariously well equipped to handle the control problem"? It's not obvious to me that there exists a known upper bound to AI capability (not only LLMs), and that this...
What makes you say we are "hilariously well equipped to handle the control problem"?
It's not obvious to me that there exists a known upper bound to AI capability (not only LLMs), and that this bound is low enough to have full confidence in our control of the AI.
Because it's a problem we handle every day with beings that are dozens to hundreds of times more capable than AI, and humans have grown more powerful as well, thanks to our access to tools and...
Because it's a problem we handle every day with beings that are dozens to hundreds of times more capable than AI, and humans have grown more powerful as well, thanks to our access to tools and such.
Basically all of our social structures exist to control general intelligence. We just have to tweak them a bit as the AI is a new class of being with different incentives.
It's not obvious to me that there exists a known upper bound to AI capability
There is a very harsh upper bound right now.
Electricity.
Data.
Ability to train on that data in an intelligent way.
Processors and their scale (the bigger the datacenter gets the further apart the processors are and the lower the AI runs).
The technological singularity is a theory, and one very much not backed by any practical observations. It's designed to be scary, like the Christian promise of hell, to get you to take it more seriously than you should.
And it ignores how hilariously power efficient humans are. In my opinion the future of intelligence (humanity in general) is going to be superhuman, enabled by genetic modification. Not chips doing matrix multiplication.
Those humans won't look like us or be... Human, but that's what evolution/life does. In the long term or humanity is doomed no matter what we do.
I don't know much about LLMs, but I think a common feature is that they are both complex systems nobody understands in such a way that they can be designed. A human can design a toaster. No single...
LLMs are nothing like a market built from real thinking human beings with actual broad scale intelligence.
I don't know much about LLMs, but I think a common feature is that they are both complex systems nobody understands in such a way that they can be designed. A human can design a toaster. No single human could design a space station, but a coordinated group of humans can. If you want to know why a certain ISS subsystem has a certain function and how it works exactly, it is possible to find someone who can explain it to you. But if you want to find out why an LLM identifies a STOP sign as an elephant if you change a pixel, you're screwed.
Similarly, global markets are so complex that nobody can explain exactly why they do what they do. This is the supply chain for a simple computer mouse. Maybe it is possible for a single human to have a complete understanding of every aspect of it, but that would be extremely hard and it would be pointless. The market tells the mouse manufacturer what to do, and if it doesn't, it is out of business and someone else takes its place.
To a large company with thousands of employees, your general intelligence is not why you got the job. You are a cogwheel in a machine. You have a very small set functions that are useful to the company, and you are rewarded if you function correctly and replaced if you don't. (That's the general idea at least.) If the company can replace you with a robot without general intelligence that can function identically but cheaper, it will.
I love Internet Archive as an organization, but their site is often slow as molasses. Archive.is is usually better for mirroring paywalled articles, IMO. https://archive.is/196L0
I love Internet Archive as an organization, but their site is often slow as molasses. Archive.is is usually better for mirroring paywalled articles, IMO. https://archive.is/196L0
Yeah, I used to use bypass-paywalls too... but the only reason I don't anymore is so I can know when something is paywalled, tag it accordingly, and then provide a mirror for people. ;)
Yeah, I used to use bypass-paywalls too... but the only reason I don't anymore is so I can know when something is paywalled, tag it accordingly, and then provide a mirror for people. ;)
Please don't copy-paste entire articles into a comment like this. That's the kind of thing that can get the site in trouble for copyright infringement.
Please don't copy-paste entire articles into a comment like this. That's the kind of thing that can get the site in trouble for copyright infringement.
I'm not a fan of comparing modern AI to markets at all, as LLMs are nothing like a market built from real thinking human beings with actual broad scale intelligence.
But to make the point that the AI control problem is one we are very familiar with, controlling non human intelligence so it meets our needs, is a good one.
Society is exactly such a monster, and humans are general intelligences that we have to fight to control every single day.
AI isn't new and we are hilariously well equipped to handle the control problem. We're only scared of it thanks to hype from people who think AI will explode into hyperinteligence with no proof or reason as to why.
And, in the long term, we've already lost the control problem. Assuming our society continues to progress give it a few thousand centuries and we will be more akin to cells than individuals, so specialized we will be.
Doomed either way.
What makes you say we are "hilariously well equipped to handle the control problem"?
It's not obvious to me that there exists a known upper bound to AI capability (not only LLMs), and that this bound is low enough to have full confidence in our control of the AI.
Because it's a problem we handle every day with beings that are dozens to hundreds of times more capable than AI, and humans have grown more powerful as well, thanks to our access to tools and such.
Basically all of our social structures exist to control general intelligence. We just have to tweak them a bit as the AI is a new class of being with different incentives.
There is a very harsh upper bound right now.
Electricity.
Data.
Ability to train on that data in an intelligent way.
Processors and their scale (the bigger the datacenter gets the further apart the processors are and the lower the AI runs).
The technological singularity is a theory, and one very much not backed by any practical observations. It's designed to be scary, like the Christian promise of hell, to get you to take it more seriously than you should.
And it ignores how hilariously power efficient humans are. In my opinion the future of intelligence (humanity in general) is going to be superhuman, enabled by genetic modification. Not chips doing matrix multiplication.
Those humans won't look like us or be... Human, but that's what evolution/life does. In the long term or humanity is doomed no matter what we do.
I don't know much about LLMs, but I think a common feature is that they are both complex systems nobody understands in such a way that they can be designed. A human can design a toaster. No single human could design a space station, but a coordinated group of humans can. If you want to know why a certain ISS subsystem has a certain function and how it works exactly, it is possible to find someone who can explain it to you. But if you want to find out why an LLM identifies a STOP sign as an elephant if you change a pixel, you're screwed.
Similarly, global markets are so complex that nobody can explain exactly why they do what they do. This is the supply chain for a simple computer mouse. Maybe it is possible for a single human to have a complete understanding of every aspect of it, but that would be extremely hard and it would be pointless. The market tells the mouse manufacturer what to do, and if it doesn't, it is out of business and someone else takes its place.
To a large company with thousands of employees, your general intelligence is not why you got the job. You are a cogwheel in a machine. You have a very small set functions that are useful to the company, and you are rewarded if you function correctly and replaced if you don't. (That's the general idea at least.) If the company can replace you with a robot without general intelligence that can function identically but cheaper, it will.
Be good to get the full article, as that one is hidden behind The Economists paywall.
Archive.org depaywaller:
https://web.archive.org/web/20230621175505/https://www.economist.com/by-invitation/2023/06/21/artificial-intelligence-is-a-familiar-looking-monster-say-henry-farrell-and-cosma-shalizi
I love Internet Archive as an organization, but their site is often slow as molasses. Archive.is is usually better for mirroring paywalled articles, IMO. https://archive.is/196L0
Personally I use a browser extension that bypasses it automatically.
For Chromium based browsers, and for Firefox based browsers.
Yeah, I used to use bypass-paywalls too... but the only reason I don't anymore is so I can know when something is paywalled, tag it accordingly, and then provide a mirror for people. ;)
I like that line of thinking!
Please don't copy-paste entire articles into a comment like this. That's the kind of thing that can get the site in trouble for copyright infringement.
Makes sense, sorry about that.