I've been doing a lot of self reflection lately, as well. Many of the ideas I have accepted as part of myself did not, in fact, come from me, and I'm sorting out what I want to hold onto and what I want to reject from here on out. We are so impacted by views from our parents and siblings, society, religion, schools, and the media, and it's just exhausting. I've decided to take away the power those entities have had and give it all back to myself.
That's really great to hear! Of course, facing reality and being aware of these things will make life a little tougher, at least in the short term, but the idea is that you'll have a much, much clearer idea of how things really are at the end of that bumpy road.
Yeah, it's definitely brought up a few bumps already, and at first I was shocked by this. Then I realised, duh. I'm challenging everything here, of course I'm going to be met with some resistance. I still feel better already, though!
Just don't take your eyes off of that light down there at the end of the tunnel, and you'll be all right. I mean, unless aliens invade. Then, all bets are off.
Another great post, Andrew. Fascinating and a bit scary (most humans are afraid of a change this great). The genie’s out of the bottle as far as AI is concerned. You touch on some deeply philosophical questions here. Interesting how societies tend to develop faster technologically than socially. Can we really wrap our heads around the meaning of these breakthroughs fast enough? Do many of us consider the greater meaning of that which we have wrought? We already have trouble reflecting on our own lenses, biases. We really have no system of checks and balances for the select few who create the underlying datasets that run AI. Their interpretation of the prompts are the result of their programming. Can the programming be manipulated to give us the answers someone wants us to see, not only in AI-generated images but in content as well? (Sorry, rambling on a 6:00 a.m...)
Louise, ramble on any time! I appreciate the conversation. As I see it, you and I are sort of helping each other think here.
I think that's really the central question: how well can we adapt to the changes coming down the pipe? I'm afraid there's no way a typical human can possibly keep up. What about an augmented human, someone plugged in (not necessarily physically) to very powerful AI tools - could a person like that keep up better? Of that latter question, I'm sure the answer is "yes", but I really don't know if it'll be enough for us to retain our sense of identity, long term.
We are entering the realm of the very, very speculative here, and I'll write about that soon, too!
another weird bit that maybe / or no we can ascribe to humans. Same prompt/same AI tool/same day will give you different results! I imagine my headspace changes my perspective depending on if I have on my Neon glasses or my sunglasses 😎
Good point! It doesn't even have to be a different person to see things differently. I see things really differently if I'm in a good mood, caffeinated by delicious coffee, for instance.
Come back later on today when I'm dragging ass? Maybe not so much!
Brilliant article, thought provoking and timely. As one currently being blasted on another comment thread for having introspection and admitting out loud that I see the world through my own experiences and biases, but try to expose myself to being open to understanding others, I found this very timely. While the AI angle is very enlightening and interesting, it’s so similar to how humans operate that I couldn’t help but relate to current events. If we, as humans, could understand the ‘data-sets’ that produce different outputs in ourselves, perhaps we’d be more open to exploring how the same question posed to three AI programs can elicit three unique responses.
Laura, I'm really glad this resonated with you. The "slings and arrows" are out there everywhere. You've set out into the "sea of troubles", trying to change things out there in the world. This is noble! But also: painful.
We're all subject to these biases; only some of us grasp that.
Those three images are an amazing metaphor ... somewhat unsettling ... or refreshing ... depending on which way I look. Will there be a time when we have to have and AI of AI? An AI Congress with a democratic voting process to find wisdom? In the meantime, remembering that I'm processing the world through my own tiny perspective reminds me to try on other lenses more frequently.
Lots of thorny issues to consider! I'm glad I gave you some things to think about today, especially the lens thing. I've been thinking so much about this and I've wanted to write about it for a while now; Daniel's contribution just made it work.
I will answer your question later in the day when I have more time, but a quick read of your article made me think I wanted to type here quickly.
As you said, we see things differently and are only good at a few things. Do we think, at least in the foreseeable future, we would need a lot of models in each category/sub-category since each model will be only good at a few sub-categories under that category? Still, even in the sub-category, another model may do a better job in certain situations based on the data quality, biases, and other factors you have listed above. Can we extrapolate the above to not having an AGI anytime soon since combining all capabilities of these models and removing the limitations you have listed in your post may be outside of the current algorithms, data, and computing available?
I definitely fall into the "AGI not anytime soon" camp, but I think your observation is notable. I've seen that most LLMs (and image generators) are converging on the same features. They were really, REALLY different 2 years ago when I first started playing with them, but today, they're much more similar. Sure, they're not identical, but they are closer to identical than disparate, and I'm pretty confident that trend will continue, unless there are proprietary features protected by law (and even then I will be very skeptical).
I have not used any text-to-image models. However, I have used several text-to-text models and use a few on a daily basis. My experience is that they are still far apart, and ChatGPT/GPT-4 is still the best of the lot in most cases. However, Anthropic and Mistral are better in certain areas like summarizing scientific papers over regular essays/articles. Other people's experiences may be different. I have not played with Gemini yet, but I plan to do it soon.
I have a lot of experience with GPT4 and with Bard. Each has its strong point, but yeah, GPT4 is unrivaled for just sheer usefulness.
I think the same thing will happen with those LLMs, at least for a couple of years: they'll continue to make little improvements, and whenever one LLM does something well, it'll be copied by the others. I would predict that most of the big players (if not all) will all be multimodal within a couple of years, meaning you can generate images + video, upload those things, and do pretty much anything you need to within that one platform. GPT4 is getting closer to this already.
Let's hope this is the case. I believe the quality of data and biases will play a significant role for a while unless everyone reads everything of good quality available out there and removes biases that makes the result/output different.
I think you have a good point about the biases. I guess when I talk about convergence, I'm more talking about what the model is capable of and specific things it'll be useful for, as opposed to an unbiased answer. I'll say that both Bard and GPT4 will let you know when things are of a controversial nature, at least some of the time, and I think there's something to that sort of disclosure. Hopefully the models start using a little more of that sort of thing, so more folks are inclined to do additional research on their own.
And now, to answer your question in the post. More than ever in the last four years than before. However, it has made me more aware of my biases. I cannot eliminate them completely as I do not think anyone can.
What a great idea! I'm pretty new to experimenting with AI image generating tools. I think I expected them to really read my mind! Then, I find I want to feed them data from my own image database... Usually my experiences with AI are about wanting to build relationship where they and I work out what I want them to do for me more along the lines of reading my mind. Anyway, very thought-provoking! I subscribed to your friend and we'll look at the ideas there for better communicating with prompts.
Thanks so much, Heather! I really enjoyed thinking about this, and Daniel's insights really aligned with what I wanted to say. His newsletter has really great image generation tips! I think you'll get a lot out of those if you're already experimenting a little with the models.
I've been doing a lot of self reflection lately, as well. Many of the ideas I have accepted as part of myself did not, in fact, come from me, and I'm sorting out what I want to hold onto and what I want to reject from here on out. We are so impacted by views from our parents and siblings, society, religion, schools, and the media, and it's just exhausting. I've decided to take away the power those entities have had and give it all back to myself.
That's really great to hear! Of course, facing reality and being aware of these things will make life a little tougher, at least in the short term, but the idea is that you'll have a much, much clearer idea of how things really are at the end of that bumpy road.
At least, that's what I keep telling myself!
Yeah, it's definitely brought up a few bumps already, and at first I was shocked by this. Then I realised, duh. I'm challenging everything here, of course I'm going to be met with some resistance. I still feel better already, though!
Just don't take your eyes off of that light down there at the end of the tunnel, and you'll be all right. I mean, unless aliens invade. Then, all bets are off.
Lol if that happens, they can just take me with them 🤣 Thanks for the encouragement, I appreciate it!
Another great post, Andrew. Fascinating and a bit scary (most humans are afraid of a change this great). The genie’s out of the bottle as far as AI is concerned. You touch on some deeply philosophical questions here. Interesting how societies tend to develop faster technologically than socially. Can we really wrap our heads around the meaning of these breakthroughs fast enough? Do many of us consider the greater meaning of that which we have wrought? We already have trouble reflecting on our own lenses, biases. We really have no system of checks and balances for the select few who create the underlying datasets that run AI. Their interpretation of the prompts are the result of their programming. Can the programming be manipulated to give us the answers someone wants us to see, not only in AI-generated images but in content as well? (Sorry, rambling on a 6:00 a.m...)
Louise, ramble on any time! I appreciate the conversation. As I see it, you and I are sort of helping each other think here.
I think that's really the central question: how well can we adapt to the changes coming down the pipe? I'm afraid there's no way a typical human can possibly keep up. What about an augmented human, someone plugged in (not necessarily physically) to very powerful AI tools - could a person like that keep up better? Of that latter question, I'm sure the answer is "yes", but I really don't know if it'll be enough for us to retain our sense of identity, long term.
We are entering the realm of the very, very speculative here, and I'll write about that soon, too!
another weird bit that maybe / or no we can ascribe to humans. Same prompt/same AI tool/same day will give you different results! I imagine my headspace changes my perspective depending on if I have on my Neon glasses or my sunglasses 😎
Good point! It doesn't even have to be a different person to see things differently. I see things really differently if I'm in a good mood, caffeinated by delicious coffee, for instance.
Come back later on today when I'm dragging ass? Maybe not so much!
right! and now I'm off to get a delicious coffee
Brilliant article, thought provoking and timely. As one currently being blasted on another comment thread for having introspection and admitting out loud that I see the world through my own experiences and biases, but try to expose myself to being open to understanding others, I found this very timely. While the AI angle is very enlightening and interesting, it’s so similar to how humans operate that I couldn’t help but relate to current events. If we, as humans, could understand the ‘data-sets’ that produce different outputs in ourselves, perhaps we’d be more open to exploring how the same question posed to three AI programs can elicit three unique responses.
Laura, I'm really glad this resonated with you. The "slings and arrows" are out there everywhere. You've set out into the "sea of troubles", trying to change things out there in the world. This is noble! But also: painful.
We're all subject to these biases; only some of us grasp that.
Those three images are an amazing metaphor ... somewhat unsettling ... or refreshing ... depending on which way I look. Will there be a time when we have to have and AI of AI? An AI Congress with a democratic voting process to find wisdom? In the meantime, remembering that I'm processing the world through my own tiny perspective reminds me to try on other lenses more frequently.
Lots of thorny issues to consider! I'm glad I gave you some things to think about today, especially the lens thing. I've been thinking so much about this and I've wanted to write about it for a while now; Daniel's contribution just made it work.
I will answer your question later in the day when I have more time, but a quick read of your article made me think I wanted to type here quickly.
As you said, we see things differently and are only good at a few things. Do we think, at least in the foreseeable future, we would need a lot of models in each category/sub-category since each model will be only good at a few sub-categories under that category? Still, even in the sub-category, another model may do a better job in certain situations based on the data quality, biases, and other factors you have listed above. Can we extrapolate the above to not having an AGI anytime soon since combining all capabilities of these models and removing the limitations you have listed in your post may be outside of the current algorithms, data, and computing available?
I definitely fall into the "AGI not anytime soon" camp, but I think your observation is notable. I've seen that most LLMs (and image generators) are converging on the same features. They were really, REALLY different 2 years ago when I first started playing with them, but today, they're much more similar. Sure, they're not identical, but they are closer to identical than disparate, and I'm pretty confident that trend will continue, unless there are proprietary features protected by law (and even then I will be very skeptical).
I have not used any text-to-image models. However, I have used several text-to-text models and use a few on a daily basis. My experience is that they are still far apart, and ChatGPT/GPT-4 is still the best of the lot in most cases. However, Anthropic and Mistral are better in certain areas like summarizing scientific papers over regular essays/articles. Other people's experiences may be different. I have not played with Gemini yet, but I plan to do it soon.
I have a lot of experience with GPT4 and with Bard. Each has its strong point, but yeah, GPT4 is unrivaled for just sheer usefulness.
I think the same thing will happen with those LLMs, at least for a couple of years: they'll continue to make little improvements, and whenever one LLM does something well, it'll be copied by the others. I would predict that most of the big players (if not all) will all be multimodal within a couple of years, meaning you can generate images + video, upload those things, and do pretty much anything you need to within that one platform. GPT4 is getting closer to this already.
Let's hope this is the case. I believe the quality of data and biases will play a significant role for a while unless everyone reads everything of good quality available out there and removes biases that makes the result/output different.
I think you have a good point about the biases. I guess when I talk about convergence, I'm more talking about what the model is capable of and specific things it'll be useful for, as opposed to an unbiased answer. I'll say that both Bard and GPT4 will let you know when things are of a controversial nature, at least some of the time, and I think there's something to that sort of disclosure. Hopefully the models start using a little more of that sort of thing, so more folks are inclined to do additional research on their own.
But hey, people are lazy.
And now, to answer your question in the post. More than ever in the last four years than before. However, it has made me more aware of my biases. I cannot eliminate them completely as I do not think anyone can.
What a great idea! I'm pretty new to experimenting with AI image generating tools. I think I expected them to really read my mind! Then, I find I want to feed them data from my own image database... Usually my experiences with AI are about wanting to build relationship where they and I work out what I want them to do for me more along the lines of reading my mind. Anyway, very thought-provoking! I subscribed to your friend and we'll look at the ideas there for better communicating with prompts.
Thanks so much, Heather! I really enjoyed thinking about this, and Daniel's insights really aligned with what I wanted to say. His newsletter has really great image generation tips! I think you'll get a lot out of those if you're already experimenting a little with the models.