Thanks for the essay. I’ll have to think a bit more whether ‘AI isn’t “worth much in areas of the arts and humanities where creativity and expression are paramount”’ is compatible that early claim in the essay that actors and musicians might soon be replaced.
Sora is being discontinued (partly because it was a commercial failure), so… it looks like even Disney characters might be safe from AI for now :)
My working hypothesis is that the bottleneck wasn’t that people lacked drawing or animation skills, the scarce attribute really is something like creativity…
That line about AI not being worth much in the arts and humanities -- in hindsight, that does seem like an overstatement. What we had in mind was some kind of difference between the arts and the sciences. In the sciences, largely, the results matter, primarily. In the arts, expression and authenticity matter more.
A valuable piece of clarity in a muddied conversation. One thread worth pulling further: throughout the debate you're analyzing — and occasionally within the piece itself — AI appears as the grammatical subject. AI creates slop. AI deskills students. AI pollutes the epistemic environment.
But AI doesn't do any of those things unprompted. A human asked for the slop and published it without judgment. A teacher assigned work that made thinking optional and a student made the rational calculation. A platform rewarded volume over quality and humans on both ends of that transaction obliged.
This isn't pedantry. The subject of the sentence determines the subject of the solution. If AI is doing these things, you regulate the tool. If humans are making these choices, you're looking at incentive structures, editorial judgment, institutional design — harder problems that implicate everyone in the room, including the critics.
The motivated reasoning you diagnose so well on both sides is partly sustained by this grammatical convenience. It lets tech CEOs call the tool neutral and critics call it a villain without either party having to look at the humans operating it. The passive construction is ideologically useful precisely because it's so easy to miss.
Your 'better arguments' section is where the piece earns its keep — but even there, framing these as things AI might do rather than choices humans are already making softens the diagnosis more than the evidence warrants.
Quite thoughtful, but there’s one or two things which are important to consider.
People have a hard time understanding probabilistic and deterministic responses, called in other writing “Thinking, fast and slow”. LLM’s are probabilistic, deterministic answers come from fact repositories.
People have been trained by google to construct terrible questions - “be more vague and you get more answers”. That technique gets the worst answers from LLM’s and people compare it to google mentally. Garbage query garbage response.
LLM’s work particularly well when you iterate to converge on answers. The public GUIs barely do that, while Claude Code seems to exist only to do that. It lies somewhere between thinking fast (one-shot) and slow (iterating over a plan constructed by executive function).
Most of the perception of LLM’s is an artifact of the GUI, they are “aligned” to recognize question form and aligned to what are acceptable response styles. It’s not in them (full model) but rather conditioned output.
LLM’s will create more work for certain people because it can mass customize text at a level that’s unprecedented. It creates any textual work 1000x the speed if humans; so humans dependent on exchanging text will find instantaneous responses increase workloads. A procurement portfolio manager may handle 50-75 major products. Suddenly it will be 1000? Swamped.
Iterative tools with self-testing will be then next plateau - ask for a list of cities and capitals, it plans a retrieval, retrieves from a fact base, and tests the result. But it will be much more interesting. I assume Claude code codes: it’s revolutionary, far beyond other innovations (OOP, libraries, compilers).
I asked Claude code to create a documentary film about a friends artwork (an artist friend who committed suicide a few years back). Certain of his subjects had antecedents which went back thousands of years; I had given a presentation on it a few years back at a special forum. Claude then found the presentation; his biography I had been writing: found my writing generation tools (books of any type; musicals, white papers, contacts, code…) and proceeded to setup a database (structured text) for a film - first outline; then narrative, then storyboard, then gabbed images I used for illustrations, identified their content, then made a Ken burns structure to storyboard and voiceover, found my elevenlabs key and proposed narrator voices to use, as we speak it is rendering chapter by chapter the complete documentary, it found my Suno key, and will ask about music soon (it made its own plan, approved before starting)…
It’s not really writing code - it wrote a lot of python helpers - it’s taking a goal and mimicking executive function at a massive level, very close to what I would have done and using the computer, code and external API’s to make the finished work, which isn’t computer code.
That cycle - break the problem down, iterate until you make a commit (write, call an API; test/reject a response, discover tools)… not general AI by far but a combination which is not LLM GUI- bound with replicating human conversation.
People don’t fully grasp what this means. We have not mimicry of speech but mimicry of higher cortical functions. We will start to see even more curious behaviors emerge that we never considered.
As I tell my clients, don’t have an AI replicate what a human does, have it drive what a human can’t do. Then you will enter a place of amazing things.
The study cited for the 10-25ml water usage figure has a very severe flaw: they used statistics for GPT-3. That’s an ancient model which predates ChatGPT. This ignores all the optimizations to the models and improvements in GPUs. Official statistics say ChatGPT uses only 0.32ml of water per prompt, and if you don’t trust the official statistics you can’t use that study because it also used official OpenAI statistics just from back when GPT-3 was their main model
LLMs are the final resting place of the cybernetic functionalist glass ceiling. There was never a function to anything evolution handed us, everything offered is a temporary fix best as possible fit. And arbitrariness was that ceiling limit. Language was the outcome of that reduction and it worked as well as it could, and then, instead of studying its limits, we optimized it. First in alphabets, then in binary and now we automate that optimization. But don't fool yourselves. LLMs aren't viable tools, they're guessing games of arbitrariness that never solves the initial lossy condition. They unravel from the illusion anything arbitrary can reach a specific reality.
Thank you for proving one of the points of the article.
The point is that critics grab whatever intellectual weapons are at hand regardless of quality — and this is a rhetorical weapon dressed as philosophy. Sounds resistant to AI hype, signals the right tribal affiliation, can't actually be falsified because it never quite commits to a specific claim.
The end of arbitrary signals? Sorry, we've been veering at this since Aristotle. Re articulated variously by Kant, Bergson, Cassirer and plenty more.
The quality issue is inherent in the symbol — It never fully functions. This isn't a "weapon on hand" it's the falsity of linguistics and Turing. Turing's fatal error was language's: theres complete separation between language/code and reality. In technical terms linguistics and code are really a UFOlogy, not a signaling system. LLMs are simply the final outcome of that failure to solve perception.
You're obviously not even aware enough to recognize this approach's resistance to linguistic hype, and that LLMs are simply collapse prone from that inherent lossiness.
And oddly, I developed ELIZAs for lots of money, I know the intrinsic weights, parameters, and especially, the illusions of its learning.
Clearly philosophers like you writing this AI hand wave have no idea what this technology is. How it functions.
“We refute (based on empirical evidence) claims that humans use linguistic representations to think.” Ev Fedorenko
I'd carefully read this before your IQs drop so low, you'll be mistaken for machines.
Thanks for the essay. I’ll have to think a bit more whether ‘AI isn’t “worth much in areas of the arts and humanities where creativity and expression are paramount”’ is compatible that early claim in the essay that actors and musicians might soon be replaced.
Sora is being discontinued (partly because it was a commercial failure), so… it looks like even Disney characters might be safe from AI for now :)
My working hypothesis is that the bottleneck wasn’t that people lacked drawing or animation skills, the scarce attribute really is something like creativity…
That line about AI not being worth much in the arts and humanities -- in hindsight, that does seem like an overstatement. What we had in mind was some kind of difference between the arts and the sciences. In the sciences, largely, the results matter, primarily. In the arts, expression and authenticity matter more.
A valuable piece of clarity in a muddied conversation. One thread worth pulling further: throughout the debate you're analyzing — and occasionally within the piece itself — AI appears as the grammatical subject. AI creates slop. AI deskills students. AI pollutes the epistemic environment.
But AI doesn't do any of those things unprompted. A human asked for the slop and published it without judgment. A teacher assigned work that made thinking optional and a student made the rational calculation. A platform rewarded volume over quality and humans on both ends of that transaction obliged.
This isn't pedantry. The subject of the sentence determines the subject of the solution. If AI is doing these things, you regulate the tool. If humans are making these choices, you're looking at incentive structures, editorial judgment, institutional design — harder problems that implicate everyone in the room, including the critics.
The motivated reasoning you diagnose so well on both sides is partly sustained by this grammatical convenience. It lets tech CEOs call the tool neutral and critics call it a villain without either party having to look at the humans operating it. The passive construction is ideologically useful precisely because it's so easy to miss.
Your 'better arguments' section is where the piece earns its keep — but even there, framing these as things AI might do rather than choices humans are already making softens the diagnosis more than the evidence warrants.
Quite thoughtful, but there’s one or two things which are important to consider.
People have a hard time understanding probabilistic and deterministic responses, called in other writing “Thinking, fast and slow”. LLM’s are probabilistic, deterministic answers come from fact repositories.
People have been trained by google to construct terrible questions - “be more vague and you get more answers”. That technique gets the worst answers from LLM’s and people compare it to google mentally. Garbage query garbage response.
LLM’s work particularly well when you iterate to converge on answers. The public GUIs barely do that, while Claude Code seems to exist only to do that. It lies somewhere between thinking fast (one-shot) and slow (iterating over a plan constructed by executive function).
Most of the perception of LLM’s is an artifact of the GUI, they are “aligned” to recognize question form and aligned to what are acceptable response styles. It’s not in them (full model) but rather conditioned output.
LLM’s will create more work for certain people because it can mass customize text at a level that’s unprecedented. It creates any textual work 1000x the speed if humans; so humans dependent on exchanging text will find instantaneous responses increase workloads. A procurement portfolio manager may handle 50-75 major products. Suddenly it will be 1000? Swamped.
Iterative tools with self-testing will be then next plateau - ask for a list of cities and capitals, it plans a retrieval, retrieves from a fact base, and tests the result. But it will be much more interesting. I assume Claude code codes: it’s revolutionary, far beyond other innovations (OOP, libraries, compilers).
I asked Claude code to create a documentary film about a friends artwork (an artist friend who committed suicide a few years back). Certain of his subjects had antecedents which went back thousands of years; I had given a presentation on it a few years back at a special forum. Claude then found the presentation; his biography I had been writing: found my writing generation tools (books of any type; musicals, white papers, contacts, code…) and proceeded to setup a database (structured text) for a film - first outline; then narrative, then storyboard, then gabbed images I used for illustrations, identified their content, then made a Ken burns structure to storyboard and voiceover, found my elevenlabs key and proposed narrator voices to use, as we speak it is rendering chapter by chapter the complete documentary, it found my Suno key, and will ask about music soon (it made its own plan, approved before starting)…
It’s not really writing code - it wrote a lot of python helpers - it’s taking a goal and mimicking executive function at a massive level, very close to what I would have done and using the computer, code and external API’s to make the finished work, which isn’t computer code.
That cycle - break the problem down, iterate until you make a commit (write, call an API; test/reject a response, discover tools)… not general AI by far but a combination which is not LLM GUI- bound with replicating human conversation.
People don’t fully grasp what this means. We have not mimicry of speech but mimicry of higher cortical functions. We will start to see even more curious behaviors emerge that we never considered.
As I tell my clients, don’t have an AI replicate what a human does, have it drive what a human can’t do. Then you will enter a place of amazing things.
Terran incognita.
The study cited for the 10-25ml water usage figure has a very severe flaw: they used statistics for GPT-3. That’s an ancient model which predates ChatGPT. This ignores all the optimizations to the models and improvements in GPUs. Official statistics say ChatGPT uses only 0.32ml of water per prompt, and if you don’t trust the official statistics you can’t use that study because it also used official OpenAI statistics just from back when GPT-3 was their main model
Thanks for featuring work by Andy Masley and Hannah Ritchie!
LLMs are the final resting place of the cybernetic functionalist glass ceiling. There was never a function to anything evolution handed us, everything offered is a temporary fix best as possible fit. And arbitrariness was that ceiling limit. Language was the outcome of that reduction and it worked as well as it could, and then, instead of studying its limits, we optimized it. First in alphabets, then in binary and now we automate that optimization. But don't fool yourselves. LLMs aren't viable tools, they're guessing games of arbitrariness that never solves the initial lossy condition. They unravel from the illusion anything arbitrary can reach a specific reality.
Thank you for proving one of the points of the article.
The point is that critics grab whatever intellectual weapons are at hand regardless of quality — and this is a rhetorical weapon dressed as philosophy. Sounds resistant to AI hype, signals the right tribal affiliation, can't actually be falsified because it never quite commits to a specific claim.
The end of arbitrary signals? Sorry, we've been veering at this since Aristotle. Re articulated variously by Kant, Bergson, Cassirer and plenty more.
The quality issue is inherent in the symbol — It never fully functions. This isn't a "weapon on hand" it's the falsity of linguistics and Turing. Turing's fatal error was language's: theres complete separation between language/code and reality. In technical terms linguistics and code are really a UFOlogy, not a signaling system. LLMs are simply the final outcome of that failure to solve perception.
You're obviously not even aware enough to recognize this approach's resistance to linguistic hype, and that LLMs are simply collapse prone from that inherent lossiness.
And oddly, I developed ELIZAs for lots of money, I know the intrinsic weights, parameters, and especially, the illusions of its learning.
Clearly philosophers like you writing this AI hand wave have no idea what this technology is. How it functions.
“We refute (based on empirical evidence) claims that humans use linguistic representations to think.” Ev Fedorenko
I'd carefully read this before your IQs drop so low, you'll be mistaken for machines.
https://substack.com/@eventperception/p-182707220