It seems to me we keep moving the goal posts for AI. Somehow it seems AGI has shifted from “artificial general intelligence”, which I’ve always read as “as good as a typical human might be” to “as good as any expert in any field”.
I’m an academic. I can do peer review and literature searches and mentor grad students. But I can’t devise a business plan well, and certainly not as well as you would expect from MBA students at the end of the semester. I can’t do videography or compose music. I can’t reverse engineer hardware. And if you asked me to do any of those things, I could probably learn to do them tolerably well in time. Because I am curious and capable and intelligent. But it would take me months or years to do one of those tasks competently, and I suspect I would still do it less competently than Fable 5.2.
I suspect that I am a reasonable example of a “general intelligence”, and I don’t really know how I can deserve that accolade without it also being given to where AI is, right this second.
Fascinating. This just raises such weird questions in my mind. The thing AI still doesn’t seem capable of is the kind of intuitive wisdom about which thing matters, and what quality is most valuable, which things are worth doing or a waste of time. These things come from social experience about the nature of other minds, and our own sense of beauty, meaning, and purpose. It feels like AI is slowly outlining the things that make human consciousness unique because it just can’t replicate them. I just wonder if it will ever pierce that boundary and start having its own views. I think not, but it’s hard to have too much confidence.
Sigh... I must say, Fable's "most impressive thing" film leaves me with a bit of ick. The idea that all of everything can be reduced to language and code is a) what separates us from the deeper mysteries of actual conscious, lived experience, and b) exactly what "AGI" is built upon, lifeless and empty when it comes to actually connecting to what is most precious about being here in this life. I find the film a perfect expression of all that is wrong about AI while simultaneously being a great summary of what AI is actually useful for. Both-and.
Everything can further be reduced to a relatively simple set of interactions between subatomic particles. No grand mysteries or infinite fractal complexity or miracles of God's creation have been found there (yet?). Just a bunch of dumb particles following a small set of fundamental interactions.
If knowing that doesn't take all the color and joy out of the world, then fundamentally everything being (kinda sorta, sometimes) reduced to code shouldn't either.
There's nothing baked into the microscopic fabric of the universe that dictates that like, my dog should exist. Or even hints at the notion that a "dog" is a possible thing that could exist. And yet...here she is. I don't see code as being fundamentally all that different in principle.
It seems like AI continues to get better and better at the mechanics of execution, and some kinds of creativity like idea generation. But as you say, taste in a particular domain seems to be harder to train. Then again, people's taste is all over the place as well. We listen to bad music, have bad takes, and start businesses that fail. So it seems like we might be defining AGI as the superset of human capabilities now. That arguably makes sense, but it'll be jarring when we finally declare something AGI and realize that it's at least light ASI as well.
One thing the "jagged AGI" framing makes me think about is what we mean by "general".
One reason current AI models feel more general, compared to the narrow prior generation AIs, may be their ability (together with the harnesses and tools around them) to tap into many different domains at once and combine that expertise contextually. That lets them do work that previously required tight, time-consuming collaboration across multiple people. In that sense, top models really are becoming more and more general.
At the same time, a lot of people feel that AI capabilities have plateaued for them. Models keep climbing benchmark boards, but for many digital-native professionals I interviewed that no longer translates into much, if any, improvement in day-to-day work. That creates an interesting tension with the stronger AGI promise - systems keep getting better overall, but for many people they still don't feel much closer to doing most of their work for them.
Awe inspiring is no joke. Wow. The world is written. And AI can write. These are such fascinating examples, thanks for sharing!
I also feel like somehow somewhere in how I work with AI I must be missing something because I've never had a single prompt work for so long, recruit many so many agents, and come out with something so elaborate. What's the secret?
Regarding the observation that "AI can now produce research faster than anyone can read it, creating a crisis in a lot of academia" what advice is there for aspiring and emerging academics? Should we optimise reviewing the outputs of AI as super peer reviewers? Or perhaps we should be honing our shepherding of the AI to produce more research grist for the mill?
Open source code for Portals: https://github.com/emollick/ai-portals
For Auden: https://github.com/emollick/auden-class
It seems to me we keep moving the goal posts for AI. Somehow it seems AGI has shifted from “artificial general intelligence”, which I’ve always read as “as good as a typical human might be” to “as good as any expert in any field”.
I’m an academic. I can do peer review and literature searches and mentor grad students. But I can’t devise a business plan well, and certainly not as well as you would expect from MBA students at the end of the semester. I can’t do videography or compose music. I can’t reverse engineer hardware. And if you asked me to do any of those things, I could probably learn to do them tolerably well in time. Because I am curious and capable and intelligent. But it would take me months or years to do one of those tasks competently, and I suspect I would still do it less competently than Fable 5.2.
I suspect that I am a reasonable example of a “general intelligence”, and I don’t really know how I can deserve that accolade without it also being given to where AI is, right this second.
“I have no technical ability. And I know nothing about music.”
When asked what he gets paid for, he replied:
“The confidence that I have in my taste. And my ability to express what I feel has proven helpful for artists.”
-Rick Rubin
Fascinating. This just raises such weird questions in my mind. The thing AI still doesn’t seem capable of is the kind of intuitive wisdom about which thing matters, and what quality is most valuable, which things are worth doing or a waste of time. These things come from social experience about the nature of other minds, and our own sense of beauty, meaning, and purpose. It feels like AI is slowly outlining the things that make human consciousness unique because it just can’t replicate them. I just wonder if it will ever pierce that boundary and start having its own views. I think not, but it’s hard to have too much confidence.
Sigh... I must say, Fable's "most impressive thing" film leaves me with a bit of ick. The idea that all of everything can be reduced to language and code is a) what separates us from the deeper mysteries of actual conscious, lived experience, and b) exactly what "AGI" is built upon, lifeless and empty when it comes to actually connecting to what is most precious about being here in this life. I find the film a perfect expression of all that is wrong about AI while simultaneously being a great summary of what AI is actually useful for. Both-and.
Everything can further be reduced to a relatively simple set of interactions between subatomic particles. No grand mysteries or infinite fractal complexity or miracles of God's creation have been found there (yet?). Just a bunch of dumb particles following a small set of fundamental interactions.
If knowing that doesn't take all the color and joy out of the world, then fundamentally everything being (kinda sorta, sometimes) reduced to code shouldn't either.
There's nothing baked into the microscopic fabric of the universe that dictates that like, my dog should exist. Or even hints at the notion that a "dog" is a possible thing that could exist. And yet...here she is. I don't see code as being fundamentally all that different in principle.
It seems like AI continues to get better and better at the mechanics of execution, and some kinds of creativity like idea generation. But as you say, taste in a particular domain seems to be harder to train. Then again, people's taste is all over the place as well. We listen to bad music, have bad takes, and start businesses that fail. So it seems like we might be defining AGI as the superset of human capabilities now. That arguably makes sense, but it'll be jarring when we finally declare something AGI and realize that it's at least light ASI as well.
One thing the "jagged AGI" framing makes me think about is what we mean by "general".
One reason current AI models feel more general, compared to the narrow prior generation AIs, may be their ability (together with the harnesses and tools around them) to tap into many different domains at once and combine that expertise contextually. That lets them do work that previously required tight, time-consuming collaboration across multiple people. In that sense, top models really are becoming more and more general.
At the same time, a lot of people feel that AI capabilities have plateaued for them. Models keep climbing benchmark boards, but for many digital-native professionals I interviewed that no longer translates into much, if any, improvement in day-to-day work. That creates an interesting tension with the stronger AGI promise - systems keep getting better overall, but for many people they still don't feel much closer to doing most of their work for them.
Awe inspiring is no joke. Wow. The world is written. And AI can write. These are such fascinating examples, thanks for sharing!
I also feel like somehow somewhere in how I work with AI I must be missing something because I've never had a single prompt work for so long, recruit many so many agents, and come out with something so elaborate. What's the secret?
Regarding the observation that "AI can now produce research faster than anyone can read it, creating a crisis in a lot of academia" what advice is there for aspiring and emerging academics? Should we optimise reviewing the outputs of AI as super peer reviewers? Or perhaps we should be honing our shepherding of the AI to produce more research grist for the mill?