I find it hard to imagine which areas in the economy and/or culture are most likely to experience the technology -- which I am defining as Artificial *General* Intelligence -- over the next few years. The problem of course is that the very concept of "General" makes it hard to focus on specific applications.
So I would be very interested in your speculations as to where AGI is likely to matter over the next few years.
If I were to assume that AGI means a literal artificial "general" intelligence, I'd say it would matter everywhere. A system that could adapt to any situation or circumstance and act optimally every time would revolutionize everything.
However, I don't think AGI means that for most people. AGI can also be understood as generally intelligent (in contrast to intelligent in every specific domain). One example of this definition of AGI could be C-3PO from Star Wars. You can consider it intelligent in many domains but it still lacks social skills. Still, a system like that could impact heavily many areas.
Adding to my original comment I can say that, regardless of your definition, technology usually goes first to the military and medicine. So those would probably be the first applications for AGI too.
Jul 28, 2022·edited Jul 28, 2022Liked by Alberto Romero
>> A system that could adapt to any situation or circumstance and act optimally every time
>> would revolutionize everything.
This is not as impossible as it might sound. Spread sheets were an early form of artificial general intelligence -- they might not have revolutionized *everything* but they cast quite a shadow.
Anyway, that is how I am defining AGI and that is the focus of my interests. A natural answer to my question might be Level 5 self-driving cars. Being able to do self-driving means having solved the problems of 1) extracting a complicated shape from a dynamic background, 2) empowering it to move through a complicated dynamic medium, and
3) inserting itself in a complex dynamic destination. Once you have that you have a program that can solve a very wide range of problems, way exceeding just self-driving cars. And maybe that is the right answer to my question. But it is reasonable to think that self-driving is a much harder problem than some other issue. But maybe not. Maybe all the examples out there will be solved at the same time by the same program. Imagine how hard it will be for the culture to come to terms with change on that scale.
"Imagine how hard it will be for the culture to come to terms with change on that scale." 100% agree. And I think we won't even need AGI to feel that. People isn't ready for the speed at which AI is advancing.
One factor will be the price of the technology. If AGI software costs a million $ then the social impact will probably be seen in class terms -- as the rich making life hard for the poor. But we are talking about a general technology, one that would solve problem throughout the society, and general technologies come
pretty cheap. And the ill effects of cheap technology tend to be blamed on their owners. I think.
Hi again Fred! I think we're assuming too much here. For instance, maybe AGI can't be just software. I don't think most people would consider something that can't act in the world AGI.
Still, you're right about AGI creating class problems. In the sense that it'll be those with more resources who will be able to access the benefits of these types of technologies -- even if they're generally applied to society. This is what usually happens, the already discriminated minorities, be it in terms of social class, race, etc.., end up suffering the most.
The AI application that would matter the most to me personally would be a significant upgrade in the NLR level achieved by voice and/or text assistants. That would make a deep and immediate change in how I spend my day. I have no idea when anything like this is likely to happen. One year? Ten?
Well, I'm not sure how to answer your question. I'd say the "why" is pretty straightforward: Because that will generate more money for the company, right? How they'll do it is another question for which I don't have an answer.
I'll keep digging into social media in future articles, though. I'd say these apps embody the most direct effect of AI on people and that's what I want this newsletter to be about.
Thanks the reason I ask is simply b/c I'm super curious about ByteDance's AI advantage. They are able to scale many different kinds of apps, get into social commerce, E-commerce and other things due to it. That Meta wants to pivot to their approach (too late) is also interesting.
From my point of view, the writings of AI could be on an ample set of perspectives such as: a) research and development trends, b) current state of the art innovation and technologies, c) speculative and futuristic, d) essays or deep writings on AI impact on philosophy, psychology, cognitive science, and society at all, and many more. I have been a fan reader of your writings, and I found you indeed have covered all these perspectives.
My personal interest is more oriented to (d), on how AI is already impacting our life and our society on a day-to-day basis, how it is affecting how we perceive, how we socialize and interact, how do we do our jobs and tasks, how do we benefit in our health and many more. (I occasionally write on the cyberpsychology subject)
I understand that you write for an ample audience with ample interests and preferences. But my feedback, and ideas on what I want you to write about, it would like you write on AI actual and foreseeable impact in society, personal life, and the body of knowledge on philosophy, psychology, cognitive science, digital health just to name a few. AI algorithms are really everywhere in our lives, without even us noticing that.
that matches with my interest. In addition, I would love to what happens in the old industry where predictive maintenance is THE buzzword.
It is going to have an impact on our life as well as AI is going to decide how a machine shall be maintenance (this is the vision) not to mentioned the abuse of AI in the war industry.
I'm involved in AI supported machine monitoring but public discussion are really rare compared to discussion around the AI of social media.
Thanks for your feedback Ricardo! I know you've been following me since the very beginning on Medium and I appreciate it a lot! I think we share a lot of interests. I'll do my best to cover all that I like as well as all that I think matters for the lives of people in general. :)
If AI is supposed to become more intelligent than humans, how are we supposed to start imagining what AI can do in future? Is it beyond comprehending based on current human Intelligence?
Let's say you're a violin virtuoso. I, being significantly less intelligent than you in that sense, can easily understand that you can play the violin very well. I can't do what you can, but I can imagine that you can.
The intelligence required to imagine that something is possible is generally lower than the intelligence required to do the thing (I say generally because instinctive actions don't require any imagination). For instance, I can imagine a future AI capable of finding a unifying theory of gravity and quantum mechanics but I certainly can't do it.
Still, if we extend this argument (i.e. we keep increasing AI's intelligence, assuming there's no upper limit), we'll get to a point where our limited intelligence wouldn't be enough to even start imagining what it can do. At that point, what you say becomes forever true.
You are right but partly as you are missing the bigger picture. My point is that a chimp might understand that a human can count more bananas. But can that chimp realise that human brain is so evolved that in can create something like Calculus? Your violin virtuoso allegory is based on our existing knowledge of a musical instrument. Likewise, the unifying theory is in the realm of a possibility as thought of by Einstein. But what about those questions and possibilities which we dont even know might exists?
Right. If we see this as a 'unknown unknown', then a very important 'known unknown' would be achieving the Principles of Determinism as explored in recent TV series Westworld and Devs.
Jul 28, 2022·edited Jul 28, 2022Liked by Alberto Romero
I noticed that you started generating your blogpost images using Midjourney. Do you see yourself permanently switching from stock photography to generative AI? Also, on a scale from 1 to 10, how happy are you with the controllability of current text-to-image models like DALLE-2 or Midjourney? How much time do you spend prompt engineering?
Great question and very relevant for my article of tomorrow. I don't want to reveal anything, so I'll leave it at that.
To answer your questions.
I'll probably never subscribe to another stock photo library ever again, as long as it's possible to create my images without incurring legal problems (I'll write an article on this soon, too, as I think many people will start doing the same thing and the legal repercussions are largely unknown).
I've tried a few different text-to-image models. From those in Google Colab notebooks for which you have to be comfortable coding to the easiest ones in which you only have to put a sentence and the model does everything else.
I don't yet have access to DALL·E 2, but about Midjourney I can say the control is pretty good. You can play with the prompts a lot to get very different results, styles, genres, media, etc. But there are two limitations.
First, each text-to-image model is different. You can't get DALL·E 2-like images in Midjourney and vice versa (veteran AI artists know this very well, as they're very capable to recognize the source of an AI image only by looking at it). This means that the algorithms somehow constrain the generations. That's good because the creations are more recognizable, but it's bad because it limits the possibilities.
Second, and answering also your last question, prompt engineering is a very real skill that needs a lot of practice. People are still discovering new concepts or ideas that give unique results even after many months (e.g. adding unreal engine or artstation to the prompt).
And also (and this differentiates AI art from traditional art) we don't have a direct feedback loop between our learnings and the results we get. If you paint oil on canvas, you have perfect information about what you're doing and how to improve. Prompt skills are more like a blind search. A discovery process combined with unwanted stochasticity.
For instance, for tomorrow's article, I've spent many hours defining the prompts and trying different ideas.
In case it helps, a friend and I trained a prompt generator based on scraped queries issued by Midjourney users. It helps by adding all those magic terms like "unreal engine" to your initial prompt. It's available on HuggingFace: https://huggingface.co/succinctly/text2image-prompt-generator
I've tried it and it's awesome, especially for people who are new to the AI art scene. I'll consider writing an article with your prompt generator. Great work!
This is exactly what I am looking for and curious about. Julia all but predicts that DALLE-2-type programs will kill off stock photography companies. (I apologize if I read too much into her sentence, but if she doesn't want to go that far I am happy to.) If this happens it should happen soon (Shouldn't it?) Like in a year or two? Anyway, Alberto, if you were to look into how the stock photography people are dealing with the prospect of DALLE-2 stealing their market I would consider my subscriber money well spent!
Well, she just asked me what I think! And yes, I do think stock photography companies will have a hard time in the coming years.
I have a draft about that exact topic. A person that works at one of those companies contacted me with some interesting thoughts. Legal issues should be one of the main concerns for people using DALL·E 2 to sell art. Everyone should carefully read the terms of service and even then it's not entirely clear what's legal and what isn't, because this technology is ahead of the corresponding regulation.
Excellent point. I think it will be a long time -- by which I mean at least two decades -- before the copyright issues get straightened out. Look at the problems contemporary songwriters are having when they "quote" just a couple of notes from some other writer's composition. So perhaps that is one area that AGI will not change much.
The side of our culture where I would really like to see AGI make some serious headway is middle management.
I find it hard to imagine which areas in the economy and/or culture are most likely to experience the technology -- which I am defining as Artificial *General* Intelligence -- over the next few years. The problem of course is that the very concept of "General" makes it hard to focus on specific applications.
So I would be very interested in your speculations as to where AGI is likely to matter over the next few years.
I recommend reading my other comment on AGI (Here: https://thealgorithmicbridge.substack.com/p/gpt-3-wrote-a-paper-about-itself/comment/7982526#comment-8026311).
If I were to assume that AGI means a literal artificial "general" intelligence, I'd say it would matter everywhere. A system that could adapt to any situation or circumstance and act optimally every time would revolutionize everything.
However, I don't think AGI means that for most people. AGI can also be understood as generally intelligent (in contrast to intelligent in every specific domain). One example of this definition of AGI could be C-3PO from Star Wars. You can consider it intelligent in many domains but it still lacks social skills. Still, a system like that could impact heavily many areas.
Adding to my original comment I can say that, regardless of your definition, technology usually goes first to the military and medicine. So those would probably be the first applications for AGI too.
>> A system that could adapt to any situation or circumstance and act optimally every time
>> would revolutionize everything.
This is not as impossible as it might sound. Spread sheets were an early form of artificial general intelligence -- they might not have revolutionized *everything* but they cast quite a shadow.
Anyway, that is how I am defining AGI and that is the focus of my interests. A natural answer to my question might be Level 5 self-driving cars. Being able to do self-driving means having solved the problems of 1) extracting a complicated shape from a dynamic background, 2) empowering it to move through a complicated dynamic medium, and
3) inserting itself in a complex dynamic destination. Once you have that you have a program that can solve a very wide range of problems, way exceeding just self-driving cars. And maybe that is the right answer to my question. But it is reasonable to think that self-driving is a much harder problem than some other issue. But maybe not. Maybe all the examples out there will be solved at the same time by the same program. Imagine how hard it will be for the culture to come to terms with change on that scale.
"Imagine how hard it will be for the culture to come to terms with change on that scale." 100% agree. And I think we won't even need AGI to feel that. People isn't ready for the speed at which AI is advancing.
One factor will be the price of the technology. If AGI software costs a million $ then the social impact will probably be seen in class terms -- as the rich making life hard for the poor. But we are talking about a general technology, one that would solve problem throughout the society, and general technologies come
pretty cheap. And the ill effects of cheap technology tend to be blamed on their owners. I think.
Hi again Fred! I think we're assuming too much here. For instance, maybe AGI can't be just software. I don't think most people would consider something that can't act in the world AGI.
Still, you're right about AGI creating class problems. In the sense that it'll be those with more resources who will be able to access the benefits of these types of technologies -- even if they're generally applied to society. This is what usually happens, the already discriminated minorities, be it in terms of social class, race, etc.., end up suffering the most.
The AI application that would matter the most to me personally would be a significant upgrade in the NLR level achieved by voice and/or text assistants. That would make a deep and immediate change in how I spend my day. I have no idea when anything like this is likely to happen. One year? Ten?
Hi Fred, what do you mean with NLR, natural language reasoning or recognition (or something else)?
Correct. Natural Language Recognition.
Why or why not will ByteDance (TikTok) continue to innovate and disrupt Meta and the AI behind it?
Hi Michael!
Well, I'm not sure how to answer your question. I'd say the "why" is pretty straightforward: Because that will generate more money for the company, right? How they'll do it is another question for which I don't have an answer.
I'll keep digging into social media in future articles, though. I'd say these apps embody the most direct effect of AI on people and that's what I want this newsletter to be about.
Thanks the reason I ask is simply b/c I'm super curious about ByteDance's AI advantage. They are able to scale many different kinds of apps, get into social commerce, E-commerce and other things due to it. That Meta wants to pivot to their approach (too late) is also interesting.
From my point of view, the writings of AI could be on an ample set of perspectives such as: a) research and development trends, b) current state of the art innovation and technologies, c) speculative and futuristic, d) essays or deep writings on AI impact on philosophy, psychology, cognitive science, and society at all, and many more. I have been a fan reader of your writings, and I found you indeed have covered all these perspectives.
My personal interest is more oriented to (d), on how AI is already impacting our life and our society on a day-to-day basis, how it is affecting how we perceive, how we socialize and interact, how do we do our jobs and tasks, how do we benefit in our health and many more. (I occasionally write on the cyberpsychology subject)
I understand that you write for an ample audience with ample interests and preferences. But my feedback, and ideas on what I want you to write about, it would like you write on AI actual and foreseeable impact in society, personal life, and the body of knowledge on philosophy, psychology, cognitive science, digital health just to name a few. AI algorithms are really everywhere in our lives, without even us noticing that.
Thanks for your excelent writings :)
that matches with my interest. In addition, I would love to what happens in the old industry where predictive maintenance is THE buzzword.
It is going to have an impact on our life as well as AI is going to decide how a machine shall be maintenance (this is the vision) not to mentioned the abuse of AI in the war industry.
I'm involved in AI supported machine monitoring but public discussion are really rare compared to discussion around the AI of social media.
thx for all the work
stefan
Thanks for your feedback Ricardo! I know you've been following me since the very beginning on Medium and I appreciate it a lot! I think we share a lot of interests. I'll do my best to cover all that I like as well as all that I think matters for the lives of people in general. :)
If AI is supposed to become more intelligent than humans, how are we supposed to start imagining what AI can do in future? Is it beyond comprehending based on current human Intelligence?
Let's say you're a violin virtuoso. I, being significantly less intelligent than you in that sense, can easily understand that you can play the violin very well. I can't do what you can, but I can imagine that you can.
The intelligence required to imagine that something is possible is generally lower than the intelligence required to do the thing (I say generally because instinctive actions don't require any imagination). For instance, I can imagine a future AI capable of finding a unifying theory of gravity and quantum mechanics but I certainly can't do it.
Still, if we extend this argument (i.e. we keep increasing AI's intelligence, assuming there's no upper limit), we'll get to a point where our limited intelligence wouldn't be enough to even start imagining what it can do. At that point, what you say becomes forever true.
You are right but partly as you are missing the bigger picture. My point is that a chimp might understand that a human can count more bananas. But can that chimp realise that human brain is so evolved that in can create something like Calculus? Your violin virtuoso allegory is based on our existing knowledge of a musical instrument. Likewise, the unifying theory is in the realm of a possibility as thought of by Einstein. But what about those questions and possibilities which we dont even know might exists?
That's where the second part of my reply applies!
Right. If we see this as a 'unknown unknown', then a very important 'known unknown' would be achieving the Principles of Determinism as explored in recent TV series Westworld and Devs.
Consider the materials in chapters 9, 8, and 7 on https://experimental-epistemology.ai
And perhaps the whole site. A lot of it is news to many people, even among practitioners.
Thanks Monica, will check it out!
I noticed that you started generating your blogpost images using Midjourney. Do you see yourself permanently switching from stock photography to generative AI? Also, on a scale from 1 to 10, how happy are you with the controllability of current text-to-image models like DALLE-2 or Midjourney? How much time do you spend prompt engineering?
Great question and very relevant for my article of tomorrow. I don't want to reveal anything, so I'll leave it at that.
To answer your questions.
I'll probably never subscribe to another stock photo library ever again, as long as it's possible to create my images without incurring legal problems (I'll write an article on this soon, too, as I think many people will start doing the same thing and the legal repercussions are largely unknown).
I've tried a few different text-to-image models. From those in Google Colab notebooks for which you have to be comfortable coding to the easiest ones in which you only have to put a sentence and the model does everything else.
I don't yet have access to DALL·E 2, but about Midjourney I can say the control is pretty good. You can play with the prompts a lot to get very different results, styles, genres, media, etc. But there are two limitations.
First, each text-to-image model is different. You can't get DALL·E 2-like images in Midjourney and vice versa (veteran AI artists know this very well, as they're very capable to recognize the source of an AI image only by looking at it). This means that the algorithms somehow constrain the generations. That's good because the creations are more recognizable, but it's bad because it limits the possibilities.
Second, and answering also your last question, prompt engineering is a very real skill that needs a lot of practice. People are still discovering new concepts or ideas that give unique results even after many months (e.g. adding unreal engine or artstation to the prompt).
And also (and this differentiates AI art from traditional art) we don't have a direct feedback loop between our learnings and the results we get. If you paint oil on canvas, you have perfect information about what you're doing and how to improve. Prompt skills are more like a blind search. A discovery process combined with unwanted stochasticity.
For instance, for tomorrow's article, I've spent many hours defining the prompts and trying different ideas.
Hope that answers everything!
Thanks Alberto for your very nuanced take!
In case it helps, a friend and I trained a prompt generator based on scraped queries issued by Midjourney users. It helps by adding all those magic terms like "unreal engine" to your initial prompt. It's available on HuggingFace: https://huggingface.co/succinctly/text2image-prompt-generator
Looking forward to your article tomorrow!
I've tried it and it's awesome, especially for people who are new to the AI art scene. I'll consider writing an article with your prompt generator. Great work!
I'm glad you found it helpful! Let me know if you have any questions about it.
Thanks Iulia for commenting and reading! I'll check it out now.
Got your name wrong there, sorry about that! Not familiar with it. Edited now :)
This is exactly what I am looking for and curious about. Julia all but predicts that DALLE-2-type programs will kill off stock photography companies. (I apologize if I read too much into her sentence, but if she doesn't want to go that far I am happy to.) If this happens it should happen soon (Shouldn't it?) Like in a year or two? Anyway, Alberto, if you were to look into how the stock photography people are dealing with the prospect of DALLE-2 stealing their market I would consider my subscriber money well spent!
Fred
Well, she just asked me what I think! And yes, I do think stock photography companies will have a hard time in the coming years.
I have a draft about that exact topic. A person that works at one of those companies contacted me with some interesting thoughts. Legal issues should be one of the main concerns for people using DALL·E 2 to sell art. Everyone should carefully read the terms of service and even then it's not entirely clear what's legal and what isn't, because this technology is ahead of the corresponding regulation.
We'll have to wait for it. Until then, in the case of a lawsuit, each situation will probably be judged individually. This Wired article is a nice read on the topic: https://www.wired.com/story/openai-dalle-copyright-intellectual-property-art/
Excellent point. I think it will be a long time -- by which I mean at least two decades -- before the copyright issues get straightened out. Look at the problems contemporary songwriters are having when they "quote" just a couple of notes from some other writer's composition. So perhaps that is one area that AGI will not change much.
The side of our culture where I would really like to see AGI make some serious headway is middle management.
Jobs That Will Be Fully Automated In The Next Few Years.
https://www.forbes.com/sites/forbestechcouncil/2019/03/01/tech-experts-predict-13-jobs-that-will-be-automated-by-2030/?sh=4c32c09f22bf
https://medium.com/@redmond.jenneke/jobs-that-will-be-fully-automated-in-the-next-10-years-745627145de0
How did Huggingface code help address NLP tasks?
Hi Andrea, can you be more specific?