For me personally, I like to hear more of the criticism of AI, simply because its so lacking out there. Everyone is talking about the greatness of AI😀。So for chatgpt, people are are again saying it will replace google search. Bing try to incorporate chatgpt into their search in perplexity.ai. I would like to hear what’s your view of this possibilities ?
I would never stop writing those articles! That's TAB's essence--not being what's already out there.
It was more a question like: do you want, from time to time, something exclusively useful (and therefore optimistic)?
About ChatGPT (or language models more generally) replacing Google search: I don't think the perspective is the right one. Here's an analogy: would you consider a super-fast train to replace planes because it seems to be even faster? No, because a train can't fly. This is the same thing. The only reason people are even considering this in the first place is, as I said because no one (companies, news outlets) is doing their job adequately to frame this technology for what it is (some people are, of course talking about this day and night, but most others don't listen).
ChatGPT shines the moment we learn to take it for what it is. That's the key lesson here: It's awesome, but it isn't just anything we'd like it to be.
I like your goal of "not being what's already out there". Intellectually, that sounds great. Business-wise, I'm not as sure. As example...
My personal interests are more big picture than the day to day details of particular emerging technologies. For example, my brain leaps from ChatGPT, to AI generally, to the knowledge explosion. So I'll slave over an article about the knowledge explosion, share it everywhere I go for years, and close to nobody is interested.
MAYBE a big picture focus on the knowledge explosion is intellectually useful, maybe, but as a business model it would seem to pretty much suck.
Point being, "not being what's already out there" comes with some risk. The cliche that seems to fit here might be... The more insightful the writing, the smaller the audience.
This is an interesting perspective on the business of writing in general. But I disagree. There are almost always more than enough people interested in any given thing. The main bottlenecks to growth are usually other variables.
Anyway, in my specific case, I try to combine being interesting with being useful. The first without the second can be too abstract. The second without the first is too undifferentiated.
I would like you to be right here. Should you come across writers focused on the knowledge explosion itself (and not just products of that process) I'd welcome a pointer. It could very well be true the problem is that I'm failing at finding those who are already interested.
I also like your point about "interesting without being useful" can be too abstract. I find myself guilty of that one more than a little.
Oh, I've been guilty of that a lot! I actually tend to think that some of my articles, even after editing to avoid that, still fall into that category. Not many writers are exploring such abstract, philosophical, and distant (either in time or space) topics. I still like to explore those questions and will continue to do so.
However, usefulness is key. This newsletter was born to bridge the gap between what I can see so clearly (even if not real yet) and what's concrete in the now. Only through that, I can succeed in conveying my ideas and making them useful to you--and fulfill the promise of TAB.
The secret sauce lies somewhere in between those two things: Too centered on usefulness and it lacks depth. Too centered in the grandiose questions and you lose contact with the here and now.
I enjoy your reflections on writing Alberto. Your secret sauce concept seems quite wise. It's a wisdom that I too often ignore myself. I'll be hopefully learning from your example.
Does this help? Maybe you could tell us about versions of AI that we're already using, perhaps without realizing it. As example, if Google Search is powered by AI, and I use it 39 times a day, then I guess I don't hate AI entirely. :-). What AI tools have we already embraced without controversy?
"What AI tools have we already embraced" A lot. "Without controversy?" Not many.
AI is a very big scientific and technological discipline that involves so many types of systems. When we talk about ChatGPT, we're zooming from AI to ML, then DL, then transformers, then LMs, then large LMs, then GPT-like large LMs...
There are sooo many things that we're leaving out when we go from AI (using the traditional usage of the term) to DL. Also, when you look closely at the boundaries that try to define what's AI and what isn't, they dissolve before your eyes.
Maybe we're getting too accustomed to talking about transformer-based systems all day...
If we look at AI since the beginning of the so-called DL revolution starting 2012, the other two large groups of AI systems that we see are classification and recommender systems. Those are all around us at all times (typical examples: the system that recognizes your face to unlock an iPhone belongs to the former category. The system that decides what pages/products/videos Google/Amazon/YouTube show you belongs to the latter).
And I'm not even considering there non-DL machine learning (e.g. AlphaZero and CICERO fall in between DL and ML) and non-DL AI (besides ML/DL, there are rule-based systems, symbolic AI--anything non-statistical/non-connectionist. These aren't dominant today, but there are systems implemented following those principles).
I don't think learning is either sufficient or necessary to consider a system an AI. ML/DL obviously focus on learning and the reason is simple: it works. Even if learning can't solve all problems and even if AI's learning processes don't resemble ours, it has helped us to solve problems that seemed impossible just a couple of decades ago.
About recsys: I think so, yes, recsys power every platform that has to transform a gigantic database of content into a digestible feed customized for every person. All social media, all streaming platforms...
And yes, about the boundaries of things...it gets too philosophical quickly.
Great stuff. Through a summation prompt, I compared this article to how I use ChatGPT to enhance business processes and this is what ChatGPT generated:
The main difference between using ChatGPT as a creative tool, as explained by The Algorithmic Bridge, and using it for enhancing businesses, as suggested by XYZ company is the intended outcome. The Algorithmic Bridge suggests using ChatGPT as a tool to enhance creativity by exploring possibilities and generating ideas that are not bound to the constraints of reality. On the other hand, XYZ company suggests using ChatGPT to improve businesses by providing solutions to specific problems and generating content that is aligned with the goals and objectives of the business. In other words, the use of ChatGPT for creativity is focused on exploration and experimentation, whereas its use for enhancing businesses is focused on achieving specific, practical results.
I pretty much agree. Two caveats: first, pure experimentation can be very practical. Second, using chatGPT to "provid[e] solutions to specific problems and generat[e] content that is aligned with the goals and objectives of the business" is a problematic application of the tech (unless you carefully check afterward). It was not built to do that or anything like it.
Again, it may feel appealing to use it to save time, but please, do it carefully. One of the points of the first section is that current language models aren't built for that (and don't do it well or reliably).
I see. Yes, using ChatGPT (or any other language model) for generating solutions to specific problems or creating content aligned with the goals and objectives of a business can be tempting, but it is important to use it carefully and understand its limitations. I agree Language models, like ChatGPT, are not designed for that purpose and may not provide accurate or reliable results. Instead, it is better to use them as a tool for exploration and experimentation, to generate ideas and inspire creativity. It is also important to carefully review any output generated by a language model before using it in a business context, to ensure that it aligns with the goals and objectives of the company. So glad to have discovered Alberto and your blog. I tweet about GPT and ChatGPT @scaffeoa :)
Quick note: the person who tried to summarize their book (and make additions to it) using ChatGPT in the linked tweet was using ChatGPT incorrectly. The tweeter said they pasted their entire book into the bot, and went off of ChatGPT’s response to that input. Problem being, ChatGPT’s max input is 4096 tokens which is probably ~4000 wordsish. The policy in cases like this is to just take the first 4096 tokens and ignore the rest. So, the bot probably had access to something like the table of contents and a few pages of the first chapter (depending on intro/preface/etc.). There are better ways to do book summaries with ChatGPT.
I’ve been really enjoying your essays, keep up the great work! I especially liked your piece on the four filters of ai knowledge... I use that framework daily and have shared that essay with friends.
Briefly, as a general principle, we're now in an era when a healthy objective balanced review weighing the benefits and costs of emerging technologies is less useful, due to the scale of the emerging tools. You know, it doesn't matter if a technology has many amazing incredible benefits if the price tag is the collapse of civilization.
More on point, your post has me wondering if technologies like ChatGPT will serve to promote writers to editors. We'll spend less time typing and more time reviewing, considering, evaluating, and correcting? As example, factory workers now spend less time doing mindless grunt work, and more time managing the overall operations, presumably a higher order of thinking.
On the positive side, you've got me wondering the degree to which technologies like ChatGPT might serve as companions to lonely old people.
I'm particularly interested in how interactive text generators might be married to animated faces to deepen the illusion of real world human conversation. Face photos can already be animated with text, but doing so in a real time interactive manner would seem to require a lot of processing power.
"we're now in an era when a healthy objective balanced review weighing the benefits and costs of emerging technologies is less useful." I disagree. I think it's now more useful and necessary than ever. The more attention AI gets the more incentives exist to not portray advancements strictly as they are. When no one is looking (and there's nothing else to gain) science works better.
Well, ChatGPT-like AIs don't write well. They write pretty bad prose, as I've said a few times. For now, I don't see good writers relying too much on these products. Writing is super useful beyond typing words on the page (e.g. to clarify the mind). It's only in the cases when conveying something is less important than causing a reaction that I see these technologies having a large impact--when content and not intent is what matters.
Yet, if you ask me if future AI systems could write much better, I couldn't deny it. It may happen.
"Are you aware of any developments in this arena?" Replika (https://replika.com/) is probably the best-known example. There are a few others, but all rely on GPT-3-like tech, which isn't built to develop "personality" over time. It's a first approach but lacks too many things--the facade is fragile.
I was trying to say, it doesn't matter if a technology has many benefits if the price tag is completely unacceptable. When any technology reaches that scale, a process of comparing pros to cons tends to become meaningless.
Thanks for the reminder of Replica, I will return to investigate again. Maybe I didn't give it a fair inspection the first time?
I use old software called CrazyTalk which allows you to animate a face photo using text or audio inputs. The output isn't perfect, but it's still pretty convincing. And it's more engaging than a text only chatbot. But CrazyTalk is not real time interactive, a major limitation.
If the output from GPT could be delivered in the form of an animated face, and if that could be done at a speed approaching real world dialog, wouldn't that create a whole new industry? Best I can tell, the technology needed already exists in both text and face image generation, but the horsepower needed to make it work in real time interactive may be unavailable?
Point being, if old people in nursing homes could talk to a face on the screen and enjoy real-like conversations, that would seem a major benefit, an optimistic story.
Agree with your conclusions. It'd be great for a lot of people who, sadly, don't have fulfilled their needs for human connection. And also, the most probable reason why there's nothing like that is that processing it in real time is unfeasible for now (as you say, the technologies work well separately: text gen, text to speech, speech to face movements, etc.)
For me personally, I like to hear more of the criticism of AI, simply because its so lacking out there. Everyone is talking about the greatness of AI😀。So for chatgpt, people are are again saying it will replace google search. Bing try to incorporate chatgpt into their search in perplexity.ai. I would like to hear what’s your view of this possibilities ?
I would never stop writing those articles! That's TAB's essence--not being what's already out there.
It was more a question like: do you want, from time to time, something exclusively useful (and therefore optimistic)?
About ChatGPT (or language models more generally) replacing Google search: I don't think the perspective is the right one. Here's an analogy: would you consider a super-fast train to replace planes because it seems to be even faster? No, because a train can't fly. This is the same thing. The only reason people are even considering this in the first place is, as I said because no one (companies, news outlets) is doing their job adequately to frame this technology for what it is (some people are, of course talking about this day and night, but most others don't listen).
ChatGPT shines the moment we learn to take it for what it is. That's the key lesson here: It's awesome, but it isn't just anything we'd like it to be.
I like your goal of "not being what's already out there". Intellectually, that sounds great. Business-wise, I'm not as sure. As example...
My personal interests are more big picture than the day to day details of particular emerging technologies. For example, my brain leaps from ChatGPT, to AI generally, to the knowledge explosion. So I'll slave over an article about the knowledge explosion, share it everywhere I go for years, and close to nobody is interested.
MAYBE a big picture focus on the knowledge explosion is intellectually useful, maybe, but as a business model it would seem to pretty much suck.
Point being, "not being what's already out there" comes with some risk. The cliche that seems to fit here might be... The more insightful the writing, the smaller the audience.
This is an interesting perspective on the business of writing in general. But I disagree. There are almost always more than enough people interested in any given thing. The main bottlenecks to growth are usually other variables.
Anyway, in my specific case, I try to combine being interesting with being useful. The first without the second can be too abstract. The second without the first is too undifferentiated.
I would like you to be right here. Should you come across writers focused on the knowledge explosion itself (and not just products of that process) I'd welcome a pointer. It could very well be true the problem is that I'm failing at finding those who are already interested.
I also like your point about "interesting without being useful" can be too abstract. I find myself guilty of that one more than a little.
Oh, I've been guilty of that a lot! I actually tend to think that some of my articles, even after editing to avoid that, still fall into that category. Not many writers are exploring such abstract, philosophical, and distant (either in time or space) topics. I still like to explore those questions and will continue to do so.
However, usefulness is key. This newsletter was born to bridge the gap between what I can see so clearly (even if not real yet) and what's concrete in the now. Only through that, I can succeed in conveying my ideas and making them useful to you--and fulfill the promise of TAB.
The secret sauce lies somewhere in between those two things: Too centered on usefulness and it lacks depth. Too centered in the grandiose questions and you lose contact with the here and now.
I enjoy your reflections on writing Alberto. Your secret sauce concept seems quite wise. It's a wisdom that I too often ignore myself. I'll be hopefully learning from your example.
Does this help? Maybe you could tell us about versions of AI that we're already using, perhaps without realizing it. As example, if Google Search is powered by AI, and I use it 39 times a day, then I guess I don't hate AI entirely. :-). What AI tools have we already embraced without controversy?
"What AI tools have we already embraced" A lot. "Without controversy?" Not many.
AI is a very big scientific and technological discipline that involves so many types of systems. When we talk about ChatGPT, we're zooming from AI to ML, then DL, then transformers, then LMs, then large LMs, then GPT-like large LMs...
There are sooo many things that we're leaving out when we go from AI (using the traditional usage of the term) to DL. Also, when you look closely at the boundaries that try to define what's AI and what isn't, they dissolve before your eyes.
Maybe we're getting too accustomed to talking about transformer-based systems all day...
If we look at AI since the beginning of the so-called DL revolution starting 2012, the other two large groups of AI systems that we see are classification and recommender systems. Those are all around us at all times (typical examples: the system that recognizes your face to unlock an iPhone belongs to the former category. The system that decides what pages/products/videos Google/Amazon/YouTube show you belongs to the latter).
And I'm not even considering there non-DL machine learning (e.g. AlphaZero and CICERO fall in between DL and ML) and non-DL AI (besides ML/DL, there are rule-based systems, symbolic AI--anything non-statistical/non-connectionist. These aren't dominant today, but there are systems implemented following those principles).
You write, "Also, when you look closely at the boundaries that try to define what's AI and what isn't, they dissolve before your eyes."
That sounds totally right. All boundaries tend to dissolve when examined too closely. When does water that you drink become you?
Is the ability to learn a key component of the definition of AI? Is that how we distinguish AI from regular software?
Would recommender systems perhaps be the type of AI that most of us have most contact with today?
Questions for future articles perhaps....
I don't think learning is either sufficient or necessary to consider a system an AI. ML/DL obviously focus on learning and the reason is simple: it works. Even if learning can't solve all problems and even if AI's learning processes don't resemble ours, it has helped us to solve problems that seemed impossible just a couple of decades ago.
About recsys: I think so, yes, recsys power every platform that has to transform a gigantic database of content into a digestible feed customized for every person. All social media, all streaming platforms...
And yes, about the boundaries of things...it gets too philosophical quickly.
Great stuff. Through a summation prompt, I compared this article to how I use ChatGPT to enhance business processes and this is what ChatGPT generated:
The main difference between using ChatGPT as a creative tool, as explained by The Algorithmic Bridge, and using it for enhancing businesses, as suggested by XYZ company is the intended outcome. The Algorithmic Bridge suggests using ChatGPT as a tool to enhance creativity by exploring possibilities and generating ideas that are not bound to the constraints of reality. On the other hand, XYZ company suggests using ChatGPT to improve businesses by providing solutions to specific problems and generating content that is aligned with the goals and objectives of the business. In other words, the use of ChatGPT for creativity is focused on exploration and experimentation, whereas its use for enhancing businesses is focused on achieving specific, practical results.
What are your thoughts on this?
Thanks Anthony!
I pretty much agree. Two caveats: first, pure experimentation can be very practical. Second, using chatGPT to "provid[e] solutions to specific problems and generat[e] content that is aligned with the goals and objectives of the business" is a problematic application of the tech (unless you carefully check afterward). It was not built to do that or anything like it.
Again, it may feel appealing to use it to save time, but please, do it carefully. One of the points of the first section is that current language models aren't built for that (and don't do it well or reliably).
I see. Yes, using ChatGPT (or any other language model) for generating solutions to specific problems or creating content aligned with the goals and objectives of a business can be tempting, but it is important to use it carefully and understand its limitations. I agree Language models, like ChatGPT, are not designed for that purpose and may not provide accurate or reliable results. Instead, it is better to use them as a tool for exploration and experimentation, to generate ideas and inspire creativity. It is also important to carefully review any output generated by a language model before using it in a business context, to ensure that it aligns with the goals and objectives of the company. So glad to have discovered Alberto and your blog. I tweet about GPT and ChatGPT @scaffeoa :)
Quick note: the person who tried to summarize their book (and make additions to it) using ChatGPT in the linked tweet was using ChatGPT incorrectly. The tweeter said they pasted their entire book into the bot, and went off of ChatGPT’s response to that input. Problem being, ChatGPT’s max input is 4096 tokens which is probably ~4000 wordsish. The policy in cases like this is to just take the first 4096 tokens and ignore the rest. So, the bot probably had access to something like the table of contents and a few pages of the first chapter (depending on intro/preface/etc.). There are better ways to do book summaries with ChatGPT.
I’ve been really enjoying your essays, keep up the great work! I especially liked your piece on the four filters of ai knowledge... I use that framework daily and have shared that essay with friends.
"I especially liked your piece on the four filters of ai knowledge... I use that framework daily and have shared that essay with friends."
Thanks so much Austin!!
Awesome roundup! Going to add this as a link in my last post. Haven’t seen such a comprehensive review of prompt tactics.
Thanks Charlotte!! Appreciate it :)
Hi Alberto,
Briefly, as a general principle, we're now in an era when a healthy objective balanced review weighing the benefits and costs of emerging technologies is less useful, due to the scale of the emerging tools. You know, it doesn't matter if a technology has many amazing incredible benefits if the price tag is the collapse of civilization.
More on point, your post has me wondering if technologies like ChatGPT will serve to promote writers to editors. We'll spend less time typing and more time reviewing, considering, evaluating, and correcting? As example, factory workers now spend less time doing mindless grunt work, and more time managing the overall operations, presumably a higher order of thinking.
On the positive side, you've got me wondering the degree to which technologies like ChatGPT might serve as companions to lonely old people.
I'm particularly interested in how interactive text generators might be married to animated faces to deepen the illusion of real world human conversation. Face photos can already be animated with text, but doing so in a real time interactive manner would seem to require a lot of processing power.
Are you aware of any developments in this arena?
"we're now in an era when a healthy objective balanced review weighing the benefits and costs of emerging technologies is less useful." I disagree. I think it's now more useful and necessary than ever. The more attention AI gets the more incentives exist to not portray advancements strictly as they are. When no one is looking (and there's nothing else to gain) science works better.
Well, ChatGPT-like AIs don't write well. They write pretty bad prose, as I've said a few times. For now, I don't see good writers relying too much on these products. Writing is super useful beyond typing words on the page (e.g. to clarify the mind). It's only in the cases when conveying something is less important than causing a reaction that I see these technologies having a large impact--when content and not intent is what matters.
Yet, if you ask me if future AI systems could write much better, I couldn't deny it. It may happen.
"Are you aware of any developments in this arena?" Replika (https://replika.com/) is probably the best-known example. There are a few others, but all rely on GPT-3-like tech, which isn't built to develop "personality" over time. It's a first approach but lacks too many things--the facade is fragile.
I was trying to say, it doesn't matter if a technology has many benefits if the price tag is completely unacceptable. When any technology reaches that scale, a process of comparing pros to cons tends to become meaningless.
Thanks for the reminder of Replica, I will return to investigate again. Maybe I didn't give it a fair inspection the first time?
I use old software called CrazyTalk which allows you to animate a face photo using text or audio inputs. The output isn't perfect, but it's still pretty convincing. And it's more engaging than a text only chatbot. But CrazyTalk is not real time interactive, a major limitation.
If the output from GPT could be delivered in the form of an animated face, and if that could be done at a speed approaching real world dialog, wouldn't that create a whole new industry? Best I can tell, the technology needed already exists in both text and face image generation, but the horsepower needed to make it work in real time interactive may be unavailable?
Point being, if old people in nursing homes could talk to a face on the screen and enjoy real-like conversations, that would seem a major benefit, an optimistic story.
Agree with your conclusions. It'd be great for a lot of people who, sadly, don't have fulfilled their needs for human connection. And also, the most probable reason why there's nothing like that is that processing it in real time is unfeasible for now (as you say, the technologies work well separately: text gen, text to speech, speech to face movements, etc.)
Brilliant, funny and useful. Thanks a lot for this meta-manual ! 😎
Thanks Pascal!
Thanks for the shoutout!
All my best prompt injections and general useful prompts are in this repo: https://github.com/abilzerian/GPT-4-Prompt-Library