This is one of the most useful things I've read about AI. Working as a scientist and programmer I work with AI a lot, and one thing that I notice all the time is despite how incredibly powerful it is in some domains, it just doesn't have any "taste" at all. I've struggled to explain what I mean by this to people and now I can just send them your post which perfectly encapsulates the issue.
Yeah - the machine has no taste, so its output can't have flavour, or even mimic it really. It doesn't *enjoy* anything. Informed pleasure must be one of the subjective kinds of intelligence Mastroianni's talking about
The other thing that's important here is that LLMs make a _kind_ of mistake that humans don't. They are not merely quantitatively different, they are qualitatively different.
A human might mess up the recipe for pizza, but they are never going to tell you to put rocks on it*. A human driver is never going to come to a screeching halt on a freeway because it saw a Stop sign on an adjacent billboard [true story] because they have context about where Stop signs belong.**
Another recent, personal example. I was double-checking some facts about Britain's female prime ministers, and the unsolicited AI Summary told me that the three names I was looking for were "Margaret Thatcher, Theresa May, and current Prime Minister Keir Starmer". While many of us would like to forget Liz Truss, no human would mistake Keir Starmer's gender. An LLM on the other hand can confidently espouse this because it has no concept of "gender"; the only thing it knows, ultimately, is what words came close enough to each other in its training set.
Now, future AIs that do have "world models" might overcome this limitation; but I content that LLMs never will; it is inherent to their architecture.
----
*Because a human recognizes that there are things that are food, things that are not food, and things that might be food under certain circumstances, like insects or other people. An LLM does not have any categories like this - the term of art is a "world model - it just has words.
**Equally, if some prankster put a Stop sign on the shoulder of the interstate, any human driver would go "oh look, somebody lost a Stop sign" because they know it doesn't belong there. An LLM does not know anything.
I think what you're saying is that journalists mess up about 8% of the time (or more). That is the conventional wisdom, because mistakes can often be high-profile and produce tons of commentary, but it's not really the case for the industry as a whole, and there are other high-profile industries where I'm sure it's the same, I'm just not as familiar. We're not talking about bad opinions or things people disagree with or predictions that turn out to be wrong. A starting reporter on, say, a local daily newspaper, would be expected to get all facts right in their reports - dates, times, names - nearly all the time. More than one or two corrections a month would raise huge red flags, and this person is producing dozens of stories. Again, that's just for the most junior entry-level person.
I worked for a local daily newspaper for a while (before entering uni). We certainly didn't get all facts right nearly all the time. People tend to pay attention (look mum, I'm in the newspaper!) and complain when you get their names or other personal facts wrong. Such complaints were a constant background hum in the office. Now, this may just have been an exceptionally bad local newspaper - but systematic research seems to suggest otherwise
In the few cases a paper wrote something about me, or someone I know personally, misspelled names or other mistakes have been relatively common. My wife is a lawyer and was once interviewed by a local reporter who managed to butcher her legal assessment in a pretty impressive way. This wasn't a matter of opinion, either.
So I don't think your newspaper was particularly bad.
I think there’s a big difference between "not being right" and "getting the literal facts wrong“. The latter is indeed rare, albeit more frequent by orders of magnitude than a pilot messing up the landing.
You can state nothing but "facts" but still be completely wrong about everything. For example, you can omit important information, you can give one particular angle undue weight, you can take quotes out of context, you can state two things next to each other in a way that implies a logical connection where there is none, you can oversimplify and so on and so on.
In other words, it’s possible to never get corrected on the facts, while making your audience dumber with your content.
I don’t know what your political leanings are, and I don’t want to bring politics into this - but I assume you could name some outlets and individual journalists you categorize as "rarely say things that are completely untrue but are still wrong about a bunch of stuff“
That's a completely separate debate though. I'm talking about facts, the very basic stuff. I have read that ACX post, and I have also worked as a journalist for a long time, so I will spare my thoughts. I get it, nobody likes the media. It reports on things that people don't like and doesn't include what they do like, and on occasion more valid objections are raised. Many such cases. However, ChatGPT would get fired a week into the job, which is why any legit newsroom using AI today always has a human in the loop. I'm not saying the AI isn't right most of the time and I'm also not saying it might someday be better than a human reporter. But good right now? No.
I think we don’t disagree that much, actually. I don’t hate/dislike "the media", and I have friends at prestigious outlets in the US and Germany who adhere to very high standards.
But at the same time, there’s a huge market for journalism that isn’t that accurate at all, while the market for pilots who mess up their landings doesn’t exist.
Like it or not, OAN and the National Enquirer are journalistic outlets. (I guess my plan to not mention specific ones didn’t hold up)
Eek. Yes. For 15 years, I taught engineers how to find the last 3% of patentable ideas they [& others in their industry] had not already found. Obviously it's difficult or everyone would do it. Still, it's all good when it's just about ideas. But when you're a pilot? Hmmm.
Perhaps I missed it but my concern is that society will discover that the majority of activities (jobs, research, what-have-you) actually don't require much subjective intelligence. Wholly agree that faster problem solving just hastens the next roadblock.
We'll see! I am more optimistic about the amount of subjective intelligence required for most activities, mainly because that part is much harder to perceive. Anyone can tell whether the words on the slides are spelled correctly, but it is much harder to tell whether they inspire confidence.
When I read AI-generated writing, I usually don't get past the first sentence unless I'm reading something meant to be an objective report (such as meeting minutes). It's the verbal equivalent of the uncanny valley. It "sounds" good but there is always something subtly "off."
AI is very, very helpful for a lot of tasks, but I wish people would continue to take responsibility to write their own stuff. It doesn't save you time or money if no one is engaged enough to read what you write and take action. Unless maybe you're phishing...
I'm wondering if it's better or worse when (presumably) human-generated writing makes the writer sound like an idiot. Because, dear reader, that is undeniably true of modern publishing, and it *has* gotten worse since my adolescence in the first Bush administration.
"A Marriage at Sea," a bestseller recommended by such publications as the New York Times, reads like the product of an earnest eighth-grader with mild aphasia, and it's by no means unique. I think it's human...it's of a badness that a machine can't match yet. Better or worse? More importantly: can humans tell, or care?
Your comment reminded me of my favorite King horror movie, the Langoliers, where the travelers land in an abandoned airport where everything is subtly "off".
I love that story and was just thinking about it recently. Yes the airport looks normal but the sound is off, food tastes wrong, drinks have no flavor etc. It's a very cool scifi idea.
I just read (reread) Chandler's six Marlowe novels. It's very straight-forward writing. Simple, clear, and yet the skill and creativity he provided was amazing. I wait to see if AI can ever do this.
Sallis writes very well, but, IMHO, there are no notches above Chandler. The prose is almost as spare as Hemmingway's and yet the pictures he paints can be as baroque and lurid as Nabokov's. One of the great stylists in any genre.
You might be interested in paying attention to the Un-slop Fiction Prize https://www.hyperstitionai.com/unslop $10k for best purely AI-generated short story (500-10k words), >$100 in token usage strongly recommended ("It's up to you how you do so; hundreds of generations, elaborate multi-pass pipelines, whatever"), judged by gwern, roon, Alexander Wales, and Jamie Wahls. The application deadline just passed today, so the winners should be out soon.
I'm interested to see how it turns out! I'll be impressed if the winner meets four criteria: the people who prompted the winning entries are incapable of writing a good short story themselves, if whatever is compelling about the winner is not at all present in the prompt, and if the winner isn't a cheap but difficult-to-foresee trick like a la ELIZA, and if I personally like it.
I find it highly unlikely they'll meet most of these criteria! I already suspect the winners will come from prompters with taste and a ton of context-dumping.
This was a fun read! I think ur characterization of intelligence is too narrow, but then again, to give it its full treatment would have this essay longer than the Bible. I think that kind of hints at the distinction you are trying to illustrate with your objective/subjective intelligence. Accuracy is not the most important part of rhetoric - what matters is whether the reader, when they follow the recipe you give them, cooks up with more or less the same dish.
What also hints at that distinction is your implicit conflating of "intelligence" with "success." The idea that someone with alot of objective intelligence is successful at solving alot of kinds of problems seems totally reasonable on its face (and for the purposes of the point you are making, it is sufficient) but the true nature of even "objective" intelligence is far more nebulous. If a person is only successfully able to solve math problems because they have a Wolfram Alpha subscription, that's very different than someone who can calculate square roots of large numbers with a pencil and paper. At the end of the day, even these "objective" domains are far more subjective than they initially seem, and that is exactly what you are getting at with your points about being bored by the right things.
Ultimately, I decided against going into graduate school because I didn't want to spend my time locked in a room with knowledge that I wouldn't be able to talk to anyone else about. I wanted to be able to share my knowledge with the world and ordinary people. Communicating ideas to people, ie actual education, is a far different skill than "objective intelligence." Your point about data accumulation (that it should be "no contest" between you and the machines) is also fascinating, because again, it shows that objective intelligence isn't even necessarily correlated with lots of data points. In fact, I often find it is subjective knowledge that requires the accumulation of data. But you can't just look at "every single thing" and magic a thru line. You have to gain experience, build a mental model at each stage, and slowly tweak it over time. In this sense, I think people forget that the scientific method is a form of humanistic inquiry. So is mathematics! It's just that mathematicians tend to agree on things.
Anyway, I am just rambling at this point. Thank u for the read!
Hereafter when asked about “my views on LLMs and AI as a writer” I will just forward people this article instead of filling my pockets with rocks and walking into the sea as they seem to be suggesting I should consider when they ask.
I had some notes on a piece with the current zeitgeist around the failure of LLMs to improve their ability to write (and the small resurgence of “I-told-you-I-still-matter” talk from writers like me.) Thankfully you already wrote the piece much better than I could and now I don’t have to try. <insert applause sounds here>
This is a lovely post as always. There is one problem, I feel. Yes, I agree with you: AI writing tends to be boring and lack something, which you call 'subjective intelligence' (clearly just because you fail to push your D&D creds of calling it WIS). But how many people agree? Look at social media feeds being spammed with long-form essays these days, all designed to trigger people's engagement. Read the comments below, discussing the matter with the AI OP as if they actually have any clue. Clearly the AI achieved what it was set out to do, and somehow it 'lit the spark', with no real wisdom, but sufficient amounts of simulated subjectivity to fool most people out there (I don't have the numbers to be fair, but it's not just the elderly anymore!). Sooner or later, I fear, our writings on general interest topics may well find as many readers as your average Q1 social cognition paper.
I think this was true before AI––95% of everything was always crap. I open up Netflix and almost everything looks like slop, even though I assume it's mainly still made by humans. Automating the production of slop just means "look at all options from the top of the list to the bottom" is not a viable search strategy, but then it never really was.
“Writing is a task that takes both objective and subjective intelligence… good writing requires an additional bit of juju that makes the prose live and breathe..” Yes! The mojo is in the juju!
Adam, your writing is always so informative and entertaining. It's fantastic! I love how you take a subject and break it down into sections with easy-to-comprehend analogies and, of course, just the right amount of humor. If LLM's ever displace this type of writing, the future is bleak.
Another take on this distinction you're drawing between objective and subjective intelligence. What if we frame it like this: text versus testimony, or maybe rationality versus phenomenology? I think some surprising things fall out of this. Truths about conversation and shared context with minds running on blood and bone.
I'd be hard pressed to draw the distinction between the conversations I have with the LLMs in rationality space and those of many humans. The boundary seems to be precisely where experience picks up. I know the discernment of writing flavor you're talking about, but I also know that I go off into general uselessness when I don't touch grass every day.
I disagree just a tiny bit that seems significant. I don’t really know much though.
I certainly agree that subjective intelligence and objective intelligence are different things, but I think there is just a tiny sliver of subjective intelligence that helps you decide whether to become an artist or go on holiday to Tuscany or have bananas and custard for dessert — but after that, it’s objective intelligence all the way down. There’s not much call for subjective intelligence after that initial choice.
With writing, we need that tiny bit of subjective intelligence to write the very best books and poems; but only a tiny proportion of readers have enough subjective intelligence to tell the difference. AI poems will be good enough for most people, and those great writers will have an ever-shrinking audience. And perhaps the computer can’t taste the meal, but just waiting to hear back for the meal-tasting robots to come along though.
A question for someone who knows a lot more than I do: I’m thinking of Kahneman’s Thinking Fast and Slow. I wonder if there is any correlation between objective/subjective thinking and System 1/System 2 thinking. Perhaps AI is stuck on System 1 thinking?
This is one of the most useful things I've read about AI. Working as a scientist and programmer I work with AI a lot, and one thing that I notice all the time is despite how incredibly powerful it is in some domains, it just doesn't have any "taste" at all. I've struggled to explain what I mean by this to people and now I can just send them your post which perfectly encapsulates the issue.
Yeah - the machine has no taste, so its output can't have flavour, or even mimic it really. It doesn't *enjoy* anything. Informed pleasure must be one of the subjective kinds of intelligence Mastroianni's talking about
Imagine a plane pilot who sticks 92% of landings.
Or let's lower the stakes. A journalist who is accurate 92% of the time.
A baseball pitcher who throws the ball towards home plate 92% of the time.
Near-perfection is the goal in a lot of industries. Even creative ones. That last 8% might be really, really difficult.
I see your point, but I will say it doesn’t require any imagination at all to picture a journalist who is right 92% of the time.
The other thing that's important here is that LLMs make a _kind_ of mistake that humans don't. They are not merely quantitatively different, they are qualitatively different.
A human might mess up the recipe for pizza, but they are never going to tell you to put rocks on it*. A human driver is never going to come to a screeching halt on a freeway because it saw a Stop sign on an adjacent billboard [true story] because they have context about where Stop signs belong.**
Another recent, personal example. I was double-checking some facts about Britain's female prime ministers, and the unsolicited AI Summary told me that the three names I was looking for were "Margaret Thatcher, Theresa May, and current Prime Minister Keir Starmer". While many of us would like to forget Liz Truss, no human would mistake Keir Starmer's gender. An LLM on the other hand can confidently espouse this because it has no concept of "gender"; the only thing it knows, ultimately, is what words came close enough to each other in its training set.
Now, future AIs that do have "world models" might overcome this limitation; but I content that LLMs never will; it is inherent to their architecture.
----
*Because a human recognizes that there are things that are food, things that are not food, and things that might be food under certain circumstances, like insects or other people. An LLM does not have any categories like this - the term of art is a "world model - it just has words.
**Equally, if some prankster put a Stop sign on the shoulder of the interstate, any human driver would go "oh look, somebody lost a Stop sign" because they know it doesn't belong there. An LLM does not know anything.
I think what you're saying is that journalists mess up about 8% of the time (or more). That is the conventional wisdom, because mistakes can often be high-profile and produce tons of commentary, but it's not really the case for the industry as a whole, and there are other high-profile industries where I'm sure it's the same, I'm just not as familiar. We're not talking about bad opinions or things people disagree with or predictions that turn out to be wrong. A starting reporter on, say, a local daily newspaper, would be expected to get all facts right in their reports - dates, times, names - nearly all the time. More than one or two corrections a month would raise huge red flags, and this person is producing dozens of stories. Again, that's just for the most junior entry-level person.
'A survey of 4,800 news sources cited in fourteen newspapers [...] found errors in 61% of local news and feature stories'
https://www.researchgate.net/publication/241655705_Accuracy_Matters_A_Cross-Market_Assessment_of_Newspaper_Error_and_Credibility
I worked for a local daily newspaper for a while (before entering uni). We certainly didn't get all facts right nearly all the time. People tend to pay attention (look mum, I'm in the newspaper!) and complain when you get their names or other personal facts wrong. Such complaints were a constant background hum in the office. Now, this may just have been an exceptionally bad local newspaper - but systematic research seems to suggest otherwise
In the few cases a paper wrote something about me, or someone I know personally, misspelled names or other mistakes have been relatively common. My wife is a lawyer and was once interviewed by a local reporter who managed to butcher her legal assessment in a pretty impressive way. This wasn't a matter of opinion, either.
So I don't think your newspaper was particularly bad.
I think there’s a big difference between "not being right" and "getting the literal facts wrong“. The latter is indeed rare, albeit more frequent by orders of magnitude than a pilot messing up the landing.
You can state nothing but "facts" but still be completely wrong about everything. For example, you can omit important information, you can give one particular angle undue weight, you can take quotes out of context, you can state two things next to each other in a way that implies a logical connection where there is none, you can oversimplify and so on and so on.
In other words, it’s possible to never get corrected on the facts, while making your audience dumber with your content.
I don’t know what your political leanings are, and I don’t want to bring politics into this - but I assume you could name some outlets and individual journalists you categorize as "rarely say things that are completely untrue but are still wrong about a bunch of stuff“
I liked this piece about the topic:
https://www.astralcodexten.com/p/the-media-very-rarely-lies
That's a completely separate debate though. I'm talking about facts, the very basic stuff. I have read that ACX post, and I have also worked as a journalist for a long time, so I will spare my thoughts. I get it, nobody likes the media. It reports on things that people don't like and doesn't include what they do like, and on occasion more valid objections are raised. Many such cases. However, ChatGPT would get fired a week into the job, which is why any legit newsroom using AI today always has a human in the loop. I'm not saying the AI isn't right most of the time and I'm also not saying it might someday be better than a human reporter. But good right now? No.
I think we don’t disagree that much, actually. I don’t hate/dislike "the media", and I have friends at prestigious outlets in the US and Germany who adhere to very high standards.
But at the same time, there’s a huge market for journalism that isn’t that accurate at all, while the market for pilots who mess up their landings doesn’t exist.
Like it or not, OAN and the National Enquirer are journalistic outlets. (I guess my plan to not mention specific ones didn’t hold up)
Eek. Yes. For 15 years, I taught engineers how to find the last 3% of patentable ideas they [& others in their industry] had not already found. Obviously it's difficult or everyone would do it. Still, it's all good when it's just about ideas. But when you're a pilot? Hmmm.
Perhaps I missed it but my concern is that society will discover that the majority of activities (jobs, research, what-have-you) actually don't require much subjective intelligence. Wholly agree that faster problem solving just hastens the next roadblock.
We'll see! I am more optimistic about the amount of subjective intelligence required for most activities, mainly because that part is much harder to perceive. Anyone can tell whether the words on the slides are spelled correctly, but it is much harder to tell whether they inspire confidence.
AI has never made me burst out in laughter (I think).
I love that you are a human that wrote stuff before the AI age, best sanity check ever.
I have a family and a full time job. I barely find time for anything. But I find time for your texts because they are great. Thank you, fellow human.
„ “Well, let’s just get some data!” and then we waste a few months being like “hmm what does this data mean, so many numbers, so mysterious”“
This sounds like much of biology these days. And then they hope that AI will help make sense of all the data…
It's also not a good way to do science. p-hacking is one negative manifestation
When I read AI-generated writing, I usually don't get past the first sentence unless I'm reading something meant to be an objective report (such as meeting minutes). It's the verbal equivalent of the uncanny valley. It "sounds" good but there is always something subtly "off."
AI is very, very helpful for a lot of tasks, but I wish people would continue to take responsibility to write their own stuff. It doesn't save you time or money if no one is engaged enough to read what you write and take action. Unless maybe you're phishing...
I'm wondering if it's better or worse when (presumably) human-generated writing makes the writer sound like an idiot. Because, dear reader, that is undeniably true of modern publishing, and it *has* gotten worse since my adolescence in the first Bush administration.
"A Marriage at Sea," a bestseller recommended by such publications as the New York Times, reads like the product of an earnest eighth-grader with mild aphasia, and it's by no means unique. I think it's human...it's of a badness that a machine can't match yet. Better or worse? More importantly: can humans tell, or care?
Indeed, and what about film and "man-made" techno music?
Your comment reminded me of my favorite King horror movie, the Langoliers, where the travelers land in an abandoned airport where everything is subtly "off".
I love that story and was just thinking about it recently. Yes the airport looks normal but the sound is off, food tastes wrong, drinks have no flavor etc. It's a very cool scifi idea.
I landed in Omaha once. Tornadoes were all around but the pilot saw a window. The airport was empty. No people. Felt very weird. Apocalyptic.
I just read (reread) Chandler's six Marlowe novels. It's very straight-forward writing. Simple, clear, and yet the skill and creativity he provided was amazing. I wait to see if AI can ever do this.
Kick it up a notch with James Sallis. Check out *The Long-Legged Fly* to get underway.
Sallis writes very well, but, IMHO, there are no notches above Chandler. The prose is almost as spare as Hemmingway's and yet the pictures he paints can be as baroque and lurid as Nabokov's. One of the great stylists in any genre.
Thanks. I just ordered 3 of his.
You might be interested in paying attention to the Un-slop Fiction Prize https://www.hyperstitionai.com/unslop $10k for best purely AI-generated short story (500-10k words), >$100 in token usage strongly recommended ("It's up to you how you do so; hundreds of generations, elaborate multi-pass pipelines, whatever"), judged by gwern, roon, Alexander Wales, and Jamie Wahls. The application deadline just passed today, so the winners should be out soon.
I'm interested to see how it turns out! I'll be impressed if the winner meets four criteria: the people who prompted the winning entries are incapable of writing a good short story themselves, if whatever is compelling about the winner is not at all present in the prompt, and if the winner isn't a cheap but difficult-to-foresee trick like a la ELIZA, and if I personally like it.
I find it highly unlikely they'll meet most of these criteria! I already suspect the winners will come from prompters with taste and a ton of context-dumping.
This was a fun read! I think ur characterization of intelligence is too narrow, but then again, to give it its full treatment would have this essay longer than the Bible. I think that kind of hints at the distinction you are trying to illustrate with your objective/subjective intelligence. Accuracy is not the most important part of rhetoric - what matters is whether the reader, when they follow the recipe you give them, cooks up with more or less the same dish.
What also hints at that distinction is your implicit conflating of "intelligence" with "success." The idea that someone with alot of objective intelligence is successful at solving alot of kinds of problems seems totally reasonable on its face (and for the purposes of the point you are making, it is sufficient) but the true nature of even "objective" intelligence is far more nebulous. If a person is only successfully able to solve math problems because they have a Wolfram Alpha subscription, that's very different than someone who can calculate square roots of large numbers with a pencil and paper. At the end of the day, even these "objective" domains are far more subjective than they initially seem, and that is exactly what you are getting at with your points about being bored by the right things.
Ultimately, I decided against going into graduate school because I didn't want to spend my time locked in a room with knowledge that I wouldn't be able to talk to anyone else about. I wanted to be able to share my knowledge with the world and ordinary people. Communicating ideas to people, ie actual education, is a far different skill than "objective intelligence." Your point about data accumulation (that it should be "no contest" between you and the machines) is also fascinating, because again, it shows that objective intelligence isn't even necessarily correlated with lots of data points. In fact, I often find it is subjective knowledge that requires the accumulation of data. But you can't just look at "every single thing" and magic a thru line. You have to gain experience, build a mental model at each stage, and slowly tweak it over time. In this sense, I think people forget that the scientific method is a form of humanistic inquiry. So is mathematics! It's just that mathematicians tend to agree on things.
Anyway, I am just rambling at this point. Thank u for the read!
And even subjective domains are far more objective than they might seem.
Yes, absolutely!
Hereafter when asked about “my views on LLMs and AI as a writer” I will just forward people this article instead of filling my pockets with rocks and walking into the sea as they seem to be suggesting I should consider when they ask.
I had some notes on a piece with the current zeitgeist around the failure of LLMs to improve their ability to write (and the small resurgence of “I-told-you-I-still-matter” talk from writers like me.) Thankfully you already wrote the piece much better than I could and now I don’t have to try. <insert applause sounds here>
Excellent thank you - I don't care if its right or wrong, that was a good read
This is a lovely post as always. There is one problem, I feel. Yes, I agree with you: AI writing tends to be boring and lack something, which you call 'subjective intelligence' (clearly just because you fail to push your D&D creds of calling it WIS). But how many people agree? Look at social media feeds being spammed with long-form essays these days, all designed to trigger people's engagement. Read the comments below, discussing the matter with the AI OP as if they actually have any clue. Clearly the AI achieved what it was set out to do, and somehow it 'lit the spark', with no real wisdom, but sufficient amounts of simulated subjectivity to fool most people out there (I don't have the numbers to be fair, but it's not just the elderly anymore!). Sooner or later, I fear, our writings on general interest topics may well find as many readers as your average Q1 social cognition paper.
I think this was true before AI––95% of everything was always crap. I open up Netflix and almost everything looks like slop, even though I assume it's mainly still made by humans. Automating the production of slop just means "look at all options from the top of the list to the bottom" is not a viable search strategy, but then it never really was.
“Writing is a task that takes both objective and subjective intelligence… good writing requires an additional bit of juju that makes the prose live and breathe..” Yes! The mojo is in the juju!
Adam, your writing is always so informative and entertaining. It's fantastic! I love how you take a subject and break it down into sections with easy-to-comprehend analogies and, of course, just the right amount of humor. If LLM's ever displace this type of writing, the future is bleak.
There may not be a light behind our eyes. But there is a light inside our hearts and yours is shining well.
Thanks Adam.
Another take on this distinction you're drawing between objective and subjective intelligence. What if we frame it like this: text versus testimony, or maybe rationality versus phenomenology? I think some surprising things fall out of this. Truths about conversation and shared context with minds running on blood and bone.
I'd be hard pressed to draw the distinction between the conversations I have with the LLMs in rationality space and those of many humans. The boundary seems to be precisely where experience picks up. I know the discernment of writing flavor you're talking about, but I also know that I go off into general uselessness when I don't touch grass every day.
I disagree just a tiny bit that seems significant. I don’t really know much though.
I certainly agree that subjective intelligence and objective intelligence are different things, but I think there is just a tiny sliver of subjective intelligence that helps you decide whether to become an artist or go on holiday to Tuscany or have bananas and custard for dessert — but after that, it’s objective intelligence all the way down. There’s not much call for subjective intelligence after that initial choice.
With writing, we need that tiny bit of subjective intelligence to write the very best books and poems; but only a tiny proportion of readers have enough subjective intelligence to tell the difference. AI poems will be good enough for most people, and those great writers will have an ever-shrinking audience. And perhaps the computer can’t taste the meal, but just waiting to hear back for the meal-tasting robots to come along though.
A question for someone who knows a lot more than I do: I’m thinking of Kahneman’s Thinking Fast and Slow. I wonder if there is any correlation between objective/subjective thinking and System 1/System 2 thinking. Perhaps AI is stuck on System 1 thinking?