After using all these AI systems I still don't believe we will have AGI anytime soon. At least by the definition stated here in the article.
Uber put a huge part of it's future value into self-driving cars replacing those expensive drivers. Didn't work out for them.
Tesla's system won't pass Level 2 with vision only.
Google's setup works good enough until it doesn't.
I see ChatGPT and others as the next step up from factory robots. Those robots replaced certain workers and repetitive tasks. Now ChatGPT can do that for copy and certain things like that.
However there are still a lot of physical jobs that a robot can not do even though we have tried very hard even on a factory floor. Elon had to learn that lesson the hard way. The same will apply to these systems. It will program something for you to a certain point, then you need humans.
>It will program something for you to a certain point, then you need humans.
I'm more concerned about reaching near-AGI where the "you need humans" parts are more and more limited (while the number of humans is less and less limited) but our economies refuse to adapt.
This is an issue with the productivity gains from automation being siphoned off in general, and we've been boiling that frog for a century plus. We're gonna have to figure out a more sophisticated economic model before too long but it will probably have to be a painful lesson.
The number of humans is actually becoming more limited; global population growth is slowing and will cease not too long from now. So building productive systems that need fewer humans will be good for the economy in the long run.
Personally I hope we never make it to AGI as I think it would raise moral and social concerns about issues like compulsion and slavery. Best to stay under that line, where tools are extremely productive but not arguably conscious beings.
Their previous statements included "diversity of thought". Someone probably complained about that being a right wing dog whistle so they got rewritten.
It is hard to write down things nobody will complain about without them becoming meh.
I have been arguing for a decade: The only moral thing for any of us to be doing is that we mandate every single human available stops what they are doing and immediately puts 100% of their effort into helping to develop AGI.
The logic is that AGI will save countless lives and arguably make EVERY job easier and better. So not developing it is utterly immoral. Doing anything else you are literally (not literally) killing people.
Its like we are on a sinking life raft, 1 guy is bailing out water while the other 8 billion people are fishing, check their phones trying to call for help, making plans about who sleep where, whos going to be the captain...
I keep seeing people say “AGI” but no one will define it for me. It’s always “you know it when you see it” or “when software can do anything”.
I’m alright with “things will continue to advance” but people assume it’s a single event that happens, not a goal post that represents “the dream” that moves 1 step further away every 1 step we advance.
Just shouting random platitudes and religious grandiosity at people is just annoying.
Because we'll never have something indistinguishable from human. Once AI is as good as a human at everything, it'll also be far better than human at something. Possibly most things.
No, it’s proof that human intelligence is possible. Since we don’t know how that works either we can’t really draw any conclusions about the viability of replicating it artificially (if that’s the definitions of AGI).
This is pretty funny. Anytime you throw a theoretically infinite payoff into the calculation it blows up immediately. It's like a religious argument. If this god is real than everyone must worship them and achieve ultimate paradise.
Another comparison actually is the GameStop stocks phenomenon. At least some of them expected such high returns (in some cases nearly infinite), that it would be fiscally insane to invest in anything but GME.
> Anytime you throw a theoretically infinite payoff into the calculation it blows up immediately.
The payoff doesn't even need to be theoretically infinite. A sufficiently large finite payoff is enough. The real problem, as you say, is that these theoretically large payoffs are just pulled out of thin air (or someplace less savory...).
In the case of AGI, there is also the well known viewpoint that AGI will result in a huge negative payoff, as soon as the AGI realizes that it can use the atoms you are made of for something that, in its view, is more productive. AGI evangelists like the GP fail to factor that in as well. When such considerations are taken into account, the expected payoff from AGI is basically zero, even without factoring in the issue I described above.
I could not disagree with this position more. But ignoring that, I think there's no reason currently to think that AGI is something that's even possible[1], let alone something that will happen in our lifetimes.
[1] I'm not saying it isn't possible, just that I see no solid reason to think it is.
1) Humans are extremely poor predictors of the future and even worse when it comes to disruptive patterns.
2) They often try to drive scenarios that are merely an extension of existing systems.
3) Very few people are in a position to review AI progress accurately as most advanced systems aren’t public.
4) Those that are in that position speak of a great deal of ways that AI systems are failing, while simultaneously positioning their companies for the advent AGI.
Is it a land grab in preparation as the market would clearly demand? Yes. You can see this in the intense lobby work happening with government around the world.
But there is obviously a feeling of inevitability, or at least a desire to create the impression of the feeling of inevitability, for an AGI.
I would imagine this technology wave will be similar to previous waves, but faster, as they have been getting faster, making mature try for AI a 2040-2045 proposition. That leaves only another few years until we enter early mainstream adoption — at most. I assume there is a lot of discussion about Smihula / Kondratiev waves and turning points happening and really the turning point is what everyone is positioning for. Got to be in at that point and have the strategy set if you want to win this wave, and then get ready for the emergent second generation (salesforce, Google, etc.) competitors to come at you.
Anyone else see that as generally correct? What am I missing?
The most advanced system is the raw GPT-4 foundational model, which is not public.
What is public is the RLHF instruct/chat/safety fine tuned version of the foundational model.
As could be seen with DaVinci vs 3.5-chat, foundational models - while more difficult to work with - give much more variety and quality to responses than the chat tuned versions.
The average person with access to the public GPT-4 models actually has no idea of what the current SotA can actually do or can't do. They only know the capabilities of the version OpenAI deemed acceptable to give them access to.
I've watched core values get rewritten at multiple companies. Usually happens when new leadership comes in, there's some kind of culture clash and failed change management, managers can't get people to behave how they want so they go deep into a rewrite of the cultural tenets to shake out longtime employees and develop new mantras to enforce. New core values set almost always ends up being a word salad that's less precise and meaningful than the original list.
If you look at OpenAI's original list, they were all adjectives to describe how the team operates. Nice and clean. Consistent.
The new list is a vague, unfocused mish-mash of concepts ("Team spirit"), adjective pairs ("Intense AND scrappy"), and phrases ("Make something people love"). Some of these things should have been initiatives or objectives for the quarter instead of core values for the whole org.
31 comments
[ 3.0 ms ] story [ 33.2 ms ] threadUber put a huge part of it's future value into self-driving cars replacing those expensive drivers. Didn't work out for them.
Tesla's system won't pass Level 2 with vision only.
Google's setup works good enough until it doesn't.
I see ChatGPT and others as the next step up from factory robots. Those robots replaced certain workers and repetitive tasks. Now ChatGPT can do that for copy and certain things like that.
However there are still a lot of physical jobs that a robot can not do even though we have tried very hard even on a factory floor. Elon had to learn that lesson the hard way. The same will apply to these systems. It will program something for you to a certain point, then you need humans.
I'm more concerned about reaching near-AGI where the "you need humans" parts are more and more limited (while the number of humans is less and less limited) but our economies refuse to adapt.
Personally I hope we never make it to AGI as I think it would raise moral and social concerns about issues like compulsion and slavery. Best to stay under that line, where tools are extremely productive but not arguably conscious beings.
It is hard to write down things nobody will complain about without them becoming meh.
I don't hate that as a way to describe a similar sentiment without invoking comparisons to DEI, which is a separate thing entirely.
The logic is that AGI will save countless lives and arguably make EVERY job easier and better. So not developing it is utterly immoral. Doing anything else you are literally (not literally) killing people.
Its like we are on a sinking life raft, 1 guy is bailing out water while the other 8 billion people are fishing, check their phones trying to call for help, making plans about who sleep where, whos going to be the captain...
Not to mention that it seems immoral to mandate what every single human being does with their lives.
I keep seeing people say “AGI” but no one will define it for me. It’s always “you know it when you see it” or “when software can do anything”.
I’m alright with “things will continue to advance” but people assume it’s a single event that happens, not a goal post that represents “the dream” that moves 1 step further away every 1 step we advance.
Just shouting random platitudes and religious grandiosity at people is just annoying.
https://towardsdatascience.com/beyond-the-turing-test-two-sc...
Another comparison actually is the GameStop stocks phenomenon. At least some of them expected such high returns (in some cases nearly infinite), that it would be fiscally insane to invest in anything but GME.
The payoff doesn't even need to be theoretically infinite. A sufficiently large finite payoff is enough. The real problem, as you say, is that these theoretically large payoffs are just pulled out of thin air (or someplace less savory...).
In the case of AGI, there is also the well known viewpoint that AGI will result in a huge negative payoff, as soon as the AGI realizes that it can use the atoms you are made of for something that, in its view, is more productive. AGI evangelists like the GP fail to factor that in as well. When such considerations are taken into account, the expected payoff from AGI is basically zero, even without factoring in the issue I described above.
[1] I'm not saying it isn't possible, just that I see no solid reason to think it is.
2) They often try to drive scenarios that are merely an extension of existing systems.
3) Very few people are in a position to review AI progress accurately as most advanced systems aren’t public.
4) Those that are in that position speak of a great deal of ways that AI systems are failing, while simultaneously positioning their companies for the advent AGI.
Is it a land grab in preparation as the market would clearly demand? Yes. You can see this in the intense lobby work happening with government around the world.
But there is obviously a feeling of inevitability, or at least a desire to create the impression of the feeling of inevitability, for an AGI.
I would imagine this technology wave will be similar to previous waves, but faster, as they have been getting faster, making mature try for AI a 2040-2045 proposition. That leaves only another few years until we enter early mainstream adoption — at most. I assume there is a lot of discussion about Smihula / Kondratiev waves and turning points happening and really the turning point is what everyone is positioning for. Got to be in at that point and have the strategy set if you want to win this wave, and then get ready for the emergent second generation (salesforce, Google, etc.) competitors to come at you.
Anyone else see that as generally correct? What am I missing?
The most advanced system (GPT4) is public. Same goes for image generation and most others. I certainly can’t think of a counter example.
The most advanced system is the raw GPT-4 foundational model, which is not public.
What is public is the RLHF instruct/chat/safety fine tuned version of the foundational model.
As could be seen with DaVinci vs 3.5-chat, foundational models - while more difficult to work with - give much more variety and quality to responses than the chat tuned versions.
The average person with access to the public GPT-4 models actually has no idea of what the current SotA can actually do or can't do. They only know the capabilities of the version OpenAI deemed acceptable to give them access to.
If you look at OpenAI's original list, they were all adjectives to describe how the team operates. Nice and clean. Consistent.
The new list is a vague, unfocused mish-mash of concepts ("Team spirit"), adjective pairs ("Intense AND scrappy"), and phrases ("Make something people love"). Some of these things should have been initiatives or objectives for the quarter instead of core values for the whole org.
Feels AGI it’s getting closer!
They might think of lots of stupid things with nothing to back it up other than their feelings.