"“Large language models do not, cannot, and will not ‘understand’ anything at all,” argued Tyler Austin Harper,"
Maybe others have had better luck, but I've never gotten anyone who thinks that explain to me how they think the electro-chemical computer between their ears "knows" anything. It's not even that they are ghosts in the machine people - the thought seems to never have occurred to them.
It’s pretty simple, really. The electro-chemical computer between your ears is not a large language model (LLM), just as, a century or more ago, it is not the clockwork it might have been compared to.The large language model doesn’t have experiences with the world (it has only read rumors of experiences with the world), and does not appear to have a model of the world against which it can check its dice-roll output (figuring out whether this is true or not is a contentious area of research).
This does not preclude the LLM from being supplemented with models — this is how the coding assistants are making the progress they are making, and it is how mathematicians get results out of the LLMs (the statistical model produces input for proof-checking software, which gives feed back where the input is lacking). There is some work being done to supplement LLMs with models of the physical world, as well.
"The electro-chemical computer between your ears is not a large language model (LLM)"
From what I've seen of the way a lot of humans reason, I'm not entirely sure of that. Heck, I'm not entirely sure there's that much daylight between human mythmaking (whether it be classic religious myths or more modern urban legends) and AI hallucinations.
Can you explain your terminology when you say coding assistants are making progress by being "supplemented with models"? They're making progress using RLVR tuning; are you saying the reward verification system is a "model"; e.g. a model of how Python programs actually operate?
Basically, yes. Though I’ll admit that the term “model” is heavily overloaded. The verification system has a model (i.e., a mapping from representation to truth values) that it uses to verify that the steps taken achieve their goal. Though I don’t want to claim great expertise (RLVR was a new term for me, for example — I’m in somewhat former ground w.r.t. the work using LEAN to verify steps in mathematics).
And as to incorporating models of physical behavior, that’s happening both in universities and industry.
Functionally, "understanding" or "acquiring knowledge" means updating an internal representative map of truth values.
Von Neumann machines may struggle with anything that is vague and fuzzy instead of strictly binary, but otherwise the difference is not that large. The fundaments of the LLM's model of the world are mathematical axioms and scientific laws, with everything else just ranked highly likely or less likely.
Of course, the LLM instance a user actually interacts with doesn't learn, this happens in the training process and then gets frozen in time.
People don't have "world models". The concept of a "world model" is entirely made up and comes from GOFAI research where they'd just invent a term and then decide it must exist because they said so.
I knew Phil though I was late coming to read Dreyfus. You are correct that people don’t have a “model” instantiated in the technical sense I alluded to above (as a mapping between representations and truth values), but people most certainly do have models of the world that they gain through embodied experience: the behavior of a glass set on the edge of a table, for example, or the arc of a ball thrown through the air, or how a person will respond to one choice of phrase versus another.
So, we see these AI tools being applied most effectively in those domains where the path between symbols and validity is simple, clear, and verifiable: code, mathematics, and I think we see them growing less effective the further one moves from those domains into the world of (embodied) experience.
Sure, you may not compare your own brain to an LLM, but by "the problem of other minds", how confident are you that my brain actually has "a model of the world against which it can check its dice-roll output"? Heck, by how many times I have hit my toes on furnitures, I am not that confident in my brain checking its dice-roll output against its world model automatically.
Or going the other direction, I do not think lack of neuroplasticity ("the world that they gain through embodied experience") precludes the existence of a world model. Neuroplasticity is true and I agree that people with training can do amazing things such as juggling and gymnastics and super precise manufacturing. But it is also clear that a baby's brain contains a world model already at birth.
If we trace the evolution of symbols and languages, as I have happened upon Joseph Heath's recent article, according to Henrich, each language evolves as part of a cooperative culture system. That is, the domain of symbols is naturally outside of embodied experience, but rather have evolved as bridges among different embodied experiences. This situates LLM in a long evolutionary progress of human attempts of cooperating among minds blind to each others' inner workings but nevertheless inclined to imitate each other. According to Henrich, it is a feature that the language-driven cooperative structures evolve so much faster than biological evolution, that human brains in fact struggle to keep up. The world models contained in the cultures, expressed in their languages, are what the LLMs actually learn.
Whether LLMs have internal models or not is an active, and bit contentious, area of research. For example, there’s some work that claims to have located specific locations that represent the state of the game board in a net trained on sequences of legal game-moves (Othello). Changing the state of that area causes the LLM to start producing output consistent with the new game-state the model represents. In less constrained “worlds”, the evidence for the existence of a model is frequently harder to find (and there’s plenty of evidence that such models don’t exist (because they produce outputs that are not consistent with something we might call a model about the reality we want the LLM to “comment on”).
I don't really see why it would be necessary for people inclined to agree with Mr. Harper to first explain how consciousness/subjectivity exists and can "know." Neuroscientists and philosophers of mind regard this as a hard enough problem that they refer to it as "the hard problem."
Having come to appreciate how interesting the hard problem is, I think it's a mistake to suggest that organic awareness can be analogized to a computer, electro-chemical or otherwise. When you are recovering from general anesthesia there is a point at which you are aware of (for example) the doctor speaking to you. The moment before you are "awake" and making a memory of the doc you are "awake" and not making a memory of the doc -- awake but not aware; conscious but not knowing. That is a state that we have no reason to believe a computer will ever have because it is "awake" only in that it is computing. Of course, electro-chemical processes continue even when we are not awake (or even in an unconscious dream state), but that is because the same processes keep us alive. Unplug a computer and it will be ready-to-go when you plug it back in. Unplug us and the only computation left to make is casket or jar.
That said, acknowledging that AI "cannot understand" says almost nothing about its technological or social importance, and is no argument for dismissing it.
Very true. The molecules and atoms in the brain are just like the ones found everywhere else. Even the functional 'intelligence' which cells or larger structures in the brain have is not what we're thinking or talking about when we refer to human intelligence.
And just like AI the human mind is comprised of a range of task-specific modules responding to particular inputs, yielding particular outputs. Some of the modules are encapsulated with one sensory route in but most have multiple feeds including feeds from downstream modules which have already undergone some processing. Sometimes these modules are localised in areas of the brain, sometimes they are distributed globally across the brain, though often with a single trigger source. These modules and the larger functions they contribute to may do their own thing under the radar of conscious experience or may come together in a concerted way to produce conscious experience.
As animals get more complex they are able to bring more of these modules together in different ways. But the killer app that sets humans apart from all other organisms and accounts for our phenomenal success, is language which enables us to connect domains of thought in ways that wouldn't otherwise be possible. When we talk to others or talk to ourselves (out loud or silently) we are opening up connections between modules that enable executive functioning to be extended into long-term planning and complex problem-solving. AI apps like ChatGPT and Claude appear to do something very similar to this reflexive, self-instruction process. When you set them a task they break it down into parts and create a plan for tackling it, check the fit of the plan to the problem, checks their own logic and working etc.
AI may end up not meeting the highest expectations that people have for it and whether it ever develops into Artificial Consciousness (AC ?) is another matter (and probably dependent on us wanting to develop AC by giving AI interests, urges to meet these interests and maybe even emotions related to these urges). But I don't think we can say at this point that there's something about AI that completely rules it out of being genuinely intelligent in a recognisably human way.
A Toto washlet in a bathroom at NRT basically saved my life after a rough flight over from the US back in 2016 and I’ve never looked back. Have them in upstairs and downstairs bathrooms of the house. I think each cost around $400. I’ve probably converted a dozen people over the years but have also felt perplexed by the slow adoption.
If you live with a partner or share a bathroom, the most overlooked benefit is the charcoal deodorizer. You just never smell each others 💩 ever again. It’s absolutely amazing. The first time sharing a hotel bathroom after having a Toto for a while will be a harsh reminder of the before times.
Biggest piece of advice if you’re looking into them is to figure out where the nearest outlet is in your bathroom because they require 24/7 electricity.
FWIW, I bought a high-voltage extension cord and a plastic junction box (for a waterproof connection with the Washlet), cemented it across my bathroom mirror's bottom trim. and plugged into a grounded outlet at the other end of the sink. Extra cost (beyond $340 for the S2 Washlet): ~$20, and total of an hour's time doing my own full installation.
It's kept my ass clean and happy ever since; even helps to ease constipation!
I’ve had multiple across multiple apartments and homes over the years and never had to hire someone for install. Did have to get creative with an extension cord at one place but figured it out.
Toto needs to put its products in mid-priced hotels. The slow adoption is because people don’t know they need it until they use it. A hotel is the place for people to use it without having to buy it first. Give those hotels a big volume discount.
There are issues with washlets in US commercial buildings because the water in toilets isn't clean enough for it. It's why Google took them out of their offices.
I installed a Toto seat washlet after visiting Japan and I don't think I'll ever be without one again. It makes everything so much better.
Worth noting that it's often fairly cheap and easy to add an outlet near a toilet in an existing house. 100% worth it to avoid an awkward extension cord.
I believe it is ancient Samurai magic built into every Toto C2+ washlet. I don’t understand how it works and I’ve never changed any of the filters and don’t even know where they are, all I know is my bathroom never smells like poo
Seems, er, wasteful. I'll pass. At least until the electrical generation system is decarbonized (or my house is so oversubscribed on solar and storage that I'll never know the difference).
Oy vey, "The Pwannet!" That really triggers me; you sound like NPR...
Please! I might be a New York Jew (a gay one, at that), but I'm a Ruben Gallego Democrat:
"Every Latino man wants a big-ass truck, which, nothing wrong with that.... We use terms like ‘bring more economic stability.’ These guys don’t want that. They don’t want ‘economic stability.’ They want to really live the American dream.”
In other words, I take my cues from the KKK: Kerouac, Kesey, and Kafka. (So much for "behavioral health"!)
FWIW, my electric bills haven't noticeably increased since I installed the Washlet. (The electric draw is barely intermittent.)
As for the warm water? I'd tell you to kiss my ass, but the Washlet does that for me... and for a gay guy, in a pulsating stream? Sheer heaven!...
I’m skeptical that most of the members of the public using AI are doing that much useful with it! Part of the problem with the reflexive AI-skepticism is that seeing a lot of the common things people are doing with AI really does suggest it’s useless. It’s hard to realize you don’t have to use it to just generate slop when most of what you directly see people doing with it is generating slop. I gave a talk a few weeks ago to try to encourage the grad students in my department to figure out all the useful things AI can do for them, and gave five examples from my own research, teaching, and service to get them thinking. But if you just see slop doggerel and slop images, you might not realize that you now have the ability to write computer programs that do things you’ve always wanted to do.
Exactly. In Halt and Catch Fire, the character Joe McMillan - definitely more of a Steve Jobs than a Steve Wozniak - says:
“The computer isn’t The Thing; it’s the thing that gets you to The Thing.”
AI is the same way; I suspect that all the general models will become commodified - more specialized ones perhaps not so much - and the real money will be made by folks who figure out how to wrangle all that capability into, “killer apps.”
1. describe mathematical diagrams to get the LLM to write TikZ code to include diagrams in LaTeX (this has sometimes required a bit of editing by hand afterward)
2. generating interactive applets for in-class activities (I made a little app to do Claude Shannon's "guess the next letter" activity to have a bit of sense of what predict-the-next-token really is like)
3. talk with ChatGPT to better understand all the steps in the proof of the Caratheodory extension theorem and why each condition is there, as well as what goes wrong with certain couterexamples if you miss them
4. create slides for class (both a timeline of LLMs and doing a little original research on what fraction of movies in each year were about AI from 1972 to 2012 - the years I got a list of wide-release movies in the US for)
5. generate images to illustrate slides (both in academic talks and class lectures)
I also mentioned some things I definitely *don't* use AI to do, which includes writing text that will be published under my name and writing summaries of papers I should be reading. (Having it write summaries of a bunch of papers I *might* want to read seems like it could be reasonable, though I haven't done it much yet.)
I share the fear of a wet behind as noted in one of the early comments.
I’m floored by the AI comments. I’m 73 and almost universally among my circle of suburban friends (most of whom are to the left of me) the value of AI is a given. While there is concern re the impact on our kids and grandkids future, not developing and using it has not come up in our discussions.
I've had the same type of experience about AI discussions, Mr. Attermann (and I'm a bit older than you). The only exception is people who read about the Singularity and imagine AI is going to decide to destroy human civilization -- perhaps out of envy of the human pleasure of Toto washlets, which, I can assure you, have excellent aim once you choose the rear-view setting that fits.
It is interesting to realize, as I just have, that an AI can be but never have a coglione (an Italian word that I think many of us learned only today).
I have a Toto C100, one of the washlet seats that can be added to an existing toilet that Noah mentions. There's no wet behind issue, you just wipe after using it.
I think people who have never used one overestimate the amount of water it sprays; it's just a little bit of water.
ok... was I the only one who wondered.... so how is Noah going to connect Pizza Wheel, Japanese Toilets, and AI ....... until I saw this sentence and it all made sense....
"Unlike in the case of pizza cutters and washlets, Americans have correctly identified the most useful technology, and are adopting it."
For the record, lawyers are absolutely adopting AI. In biglaw, at least, AI use is increasingly dramatically, and clients are expecting if not mandating it for efficiency. The observation that it's about as good as a second year associate is high praise.
I suspect the lawyers will just expand the amount of work to meet the new capacity. Just like they have done with contracts over the last few decades. The old timers tell me about how 500 page credit agreements used to be 50 pages.
I'm not in the business but I've read that clients are increasingly demanding fixed prices for chunks of work, not open-ended billable hours, and it's been squeezing the incomes of biglaw firms. If the clients keep doing that, and the actual cost of grinding out 500-page agreements goes down by a large factor, the clients will harvest much of the savings.
I have read stats elsewhere that number of lawsuits have gone up dramatically since the advent of LLm. I suspect previous ppl might just shrug something off, now they check with their LLM and say Hey, I have a case!!!
This is absolutely happening, although I think most of the effect is helping people without lawyers access the legal system.
Funny story: My friend is a city attorney, a job that requires him to go to court occasionally when people try to fight citations for things like off-leash dogs and littering. He recently had an incident in which someone fighting a citation tried to use a smartphone running an AI app to listen to what was going on the in the courtroom and tell them how to respond to the city/judge.
To some extent, it's good that AI is helping more people navigate complex and confusing legal systems. I do worry that a lot of our legal system is basically built on the assumption that things are rate-limited by the need to hire an expert of some sort, and that is going to quickly break. For example, I've heard from people in schools that parents are using AI to write letters demanding disability accommodations like extra test time for their children.
I haven’t read the AI part yet, but let me congratulate you on your inspired selection of pizza cutters and buttwash toilets for this post.
Regarding pizza cutters: the long curved cutter is 100% the way to go.
Regarding buttwash toilets: once you’ve used one, anything else seems positively barbarian. I REALLY don’t like to use the bathroom anywhere except home.
As an animator and game designer I have noticed this dismissal of AI a bit in my work with a little twist. Recently, artists and animators have been highly encouraged to start using AI tools to create useful scripts and fix workflow problems. I love using AI for these purposes. As someone with basically no coding skills at all but with lots of knowledge about animation pipelines these tools were a godsend. Now I didn’t have to put in a new tech debt request to fix something. I can just do it. The moment the tools were available I jumped in and started making things. All my managers and producers love it. However, most of the animators didn’t care. They seemed to just ignore the tools entirely, didn’t engage with it at all, and kinda just kept their heads down and kept doing their usual work. Whenever I explained how to use the tools they didn’t care. That is, until they actually made something. When one of my co workers had a problem that required a script fix or plugin tool and I could just show them exactly how it works how you could make this new thing…their eyes lit up. It was like magic to them. It’s amazing to see someone who doesn’t have the coding skills create a tool they’ve always wanted to make. Then, after that they fall in love with this technology and it dawns on them…wow this really really matters.
Now, I wonder sometimes if being part of the elite educated progressive crowd is that you acquire a smug confidence that really you know everything already and that magic doesn’t exist. For regular artists, animators, and producers…it does.
i think its insane that most google engineers are only able to use google AI coding tools. it probably really skews their perceptions of what's possible and the current state of the industry. I wonder if the productivity difference is measurable.
Going into the COVID-19 lockdown, when people were hoarding toilet paper, some genius at Costco put out a whole bunch of washlets for sale, and we picked one up, I installed it, and my wife was very pleased with it. She eventually ordered a higher-grade model and I moved the first one to the guest bathroom.
I am a software engineer at Google and have been trying out the elephant-goldfish model. https://research.google/pubs/elephants-goldfish-and-the-new-golden-age-of-software-engineering/ My impression is that LLMs are able to do things that have been done dozens of times before, but need careful supervision by a senior engineer to make sure that the output is readable enough for human comprehension, and it tempts a lot of engineers to laziness that they pay for later when it comes time to fix the bugs. LLMs are bad at doing anything where they don't have plenty of examples to work from, and make foolish mistakes because they don't actually understand anything; if I get sloppy and don't check the references when I ask Duckie (Google's internally trained version of Gemini) questions, it will frequently waste my time suggesting things that aren't possible. LLMs that provide references most effective when treated as search engines that handle natural language queries instead of keywords. Trusting the summary without following the links is an easy way to get into trouble.
I use Copilot. What you're describing sounds like my experience with GPT-4-era models; the results I would get would often feel like they were roughly adapted from other peoples' code.
Working with a GPT-5 era model is not like that *at all* - I'm guessing it's the RLVR that makes the difference. It shows (or emulates) clear understanding of the causal structure of the code is writing or analyzing, even when I'm asking for something extremely specific.
Where it still struggles is high-level planning; I have to give it clear instructions about the architecture of what it's creating or it will make decisions that are locally correct but globally messy.
I suspect software architecture will remain very “human-hands-on,” even as AI does more and more of the actual coding. The architecture design is arguably the most important part of it all.
Apparently Mythos is a step change up in architecture specifically (haven't used it; just heard). There's no objective function to train towards, but I don't think it's an inherently unsolvable problem.
I am a software engineer at not Google. I think you're way behind the curve on this. Your description sounds like models from almost a year ago that were good for unit tests but not much else. The current versions have their weaknesses and still need adult supervision, but the code they write is "readable enough for human comprehension" 100% of the time. I am now able to make changes to services all over the code base that I have never touched before. The productivity increase is big. Once the organization adapts to the new reality and I am not held back by obsolete process and procedure the difference will be massive.
Huh, I didn't realize elephant-goldfish was publicized externally (albeit, on Medium).
Your points are valid but the models have gotten better quite recently, enough that I would caution against treating them as "not understanding anything".
Duckie, since you mention it, is much better if you get access to bigger models (or use the newest default) and use "agentic mode". I also noticed a step function improvement in recent weeks in success with longer horizon coding tasks with the latest Gemini 3.5 Flash and the latest Jetski (Antigravity) harness. 3.5 seems to be better at instruction following and has better design sense.
Some of the biggest improvements to my own practice comes from asking the models to help me reason through an analysis of legacy code. It's better at being systematic and creating artifacts to trace system behavior than I am.
("Before we make any changes, tell me what happens if you change X to happen before Y? Track down all the references to X and provide a case by case analysis. Output a CSV with file names, line numbers, and conditions under which that line is executed. I will review it.")
In the case of Duckie, I might ask for an analysis of how the change came to be (it can integrate across code changes, historical docs, the org chart/team graph, etc).
Also, the Flash model is often a better choice if you're not doing frontier work or trying to one-shot a giant app, because it's much faster and can iterate quickly instead of spending a lot of time ratiocinating. And yet it doesn't lose the plot while it's churning (which was a failure mode of the older models).
Generally, I find that once I know what I want (the hard part, which can involve a sustained conversation that results in artifacts like those I mention above) I give a moderately long prompt and walk away and get a reasonable first cut at a submittable change. But, you're right that in that it's a first cut -- you have to review it yourself, first!
I’ve had better luck with Claude, but I agree that one must keep it on a relatively short leash — when it encounters problems, it can come up with some hair-brained approaches, when what is really needed is rethinking the approach. It has allowed me to do things in a day that would have taken weeks before, however.
Ok, there’s a thing that bothers me about this “the left is missing out on AI” discourse. The reasons for skepticism and dismissal that are rebutted here are certainly representative of commonly held beliefs, I’ll grant that. But I don’t think they do justice to the *best* arguments progressives have for being wary of the technology. If the reasons progressives commonly cite for not using AI are *wrong*, but, at the same time there exist much stronger arguments against using it than those typically proffered, then obviously they’d just need to adopt those better arguments, they wouldn’t need to change their actions.
And what bugs me is that it doesn’t seem that hard, with a minute of thought, to come up with far better justifications for progressives to be wary. There is no danger of becoming dependent on a pizza shear that then jacks up its subscription price. There is no danger of a company that produces washlets becoming so powerful that its unelected leadership is seated beside heads of state at the G7. Using these products does not contribute to the further precaritization of artists, or the delegitimization of other professions. While the “data centers are using all the water” stuff is massively overblown, and data centers don’t look so bad when you compare them to the impact of other commodities and conveniences we happily use without a second thought, there are still horror stories like the xAI’s Colossus data center in Memphis with its screaming smog-producing gas turbines—which is now being used by Anthropic!
At minimum, progressives just experienced a decade in which many major tech billionaires went from being purported progressive allies to enthusiastic Trump collaborators—about the worst thing someone can do from a progressive POV, and something that surely would set many folks against any activity likely to make that cohort richer and more powerful.
So I feel like, until the better arguments are really addressed, this discourse is going to remain an exercise in talking past one another, and is unlikely to produce real changes in the behavior of people on different sides of this growing divide.
Agree. It is not 1998. People have experienced over a decade of enshittification. They have seen enough. The new technology at best makes things we already had enough of (often too much of) more plentiful. Roping a bunch of chips together and filling them with stolen text turned out to be very useful in a brute force sort of way. But so was laying down a bunch of highways. The same amount of brute force effort could have improved the world in more valuable ways, and a looming sense that people have little control over how these efforts are directed is being communicated in all kinds of ways - some of them counterproductive!
Scolding AI skeptics while pointing out that smart phones are toxic suggests an opportunity for a synthesis around “what is going on”.
As to art: another example of LLM facilitating the production of something we already have plenty of. Does anyone think there are too few novels and paintings and little cutesy videos in the world? Did anyone think we should build a server farm the size of Phoenix to improve our supply of “art?”
And in practice at real offices: this stuff can be super annoying. It produces more emails when we want fewer, better emails. It produces more PPTs when we want fewer, better presentations. It offers to summarize emails (a summary form to begin with) when the correct response to an email that requires summarization is to delete it unread. It is not without a LOT of high value use cases. But there are a swarm of low value use cases all over the office place now — and it is fine to be annoyed by them.
I agree that many current office uses of AI are annoying: more emails, more slides, more summaries of things that perhaps should not have existed in the first place.
But I am not sure “more low-quality output” is the deepest problem. From an evolutionary perspective, information explosion is not a defect; it is the normal result of a new communication or production layer. Printing, newspapers, radio, TV, the internet, and social media all produced floods of material that existing elites saw as vulgar, useless, or low-quality. Some of it was. Some of it later became new culture, new markets, new institutions, or new coordination mechanisms.
I am also not sure we can define “bad workflow” or “good workflow” from a neutral standpoint. The same workflow may look terrible to a manager, useful to an employee, profitable to a firm, exhausting to a freelancer, or liberating to someone previously excluded from the system. “Quality” and “efficiency” are not fixed categories outside social context.
The real issue is whether people have the right and practical ability to choose, experiment, refuse, exit, and move to better arrangements. Markets and social selection can only work if participants are not locked into systems they cannot leave.
So the question is not how to stop abundance or centrally define quality. It is how to preserve enough openness, competition, mobility, and exit rights for better filters, workflows, and institutions to emerge.
I think Noah and Erik are both pointing at something real here.
Noah is right that dismissing AI as “just slop,” “just autocomplete,” or “just a bubble” is a bad strategy. A technology can be overhyped and still be historically important. But Erik is also right that progressive skepticism is not only about whether the tool works. It is also about who owns the surrounding system, who absorbs the externalities, and who gains power from adoption.
This is why I think there is an important distinction between refusing to use a technology and criticizing the industry around it.
Power plants have environmental, monopoly, regulatory, and political problems. But most progressives do not refuse to use electricity as a matter of ideology. They use electricity while criticizing the ownership structure, externalities, regulation, and political economy of the energy system. The same is true of smartphones, airlines, banks, and the internet.
So when AI use itself is treated as morally suspicious, that may be a sign that AI has not yet become ordinary infrastructure in many people’s daily lives. The anxiety is real, but it is partly pre-adoption anxiety. People are reacting not only to the tool, but to what they fear AI will reorganize: work, authorship, expertise, status, institutional trust, and political power.
I’ve been thinking of this as “AI anxiety”: not simply fear of a new tool, but fear of a new technological layer before society has learned how to live with it, criticize it, and govern it at the same time.
The better position is not “AI is just slop” or “AI is automatically good.” It is: use the technology where it is useful, understand it from inside real workflows, and then criticize much more precisely who owns the nodes, who pays the external costs, who loses bargaining power, and who captures the gains.
But I think the deepest problem inside AI anxiety is not only technological. It is political. Much of the AI debate is still organized around intelligence, capability, alignment, automation, and safety. These are important questions, but they often leave out the harder one: power. Who has the authority to decide how AI is deployed? Through what institutions? With what legitimacy? And can political structures designed around sovereign nation-states govern a technology whose data, supply chains, users, and effects are already global?
That, to me, is the question AI anxiety should push us toward. Not just “should we use AI?” but “what kind of institutional imagination do we need for a civilization in which AI is becoming infrastructure?”
The art historian "insisted that AI was somehow different" because it is. A human working with a technology to create art (photography and film) is different than a human instructing a technology to create art (AI-generated images and video).
That all depends on how much control you abdicate and how much engagement you have with the process. It’s like the difference between just point-and-shoot and considering composition and lighting and all the other things involved in letting a camera create an image by passively processing light that hits its film, with no control from you.
For anyone who’s had to cut pizzas at volume, and for delivery, where getting the pizza cut swiftly to keep temperature high is crucial, the regular pizza cutter disparaged here is an absolute necessity.
It takes a bit of training (maybe 20-30 pizzas in quick succession) to get the feel for it, but once you have, nothing else comes close.
The rocker cutter is second best at roughly 1/2 the speed.
The trick is to get your elbow up high above the pizza. Once you have that feel, and the cut making + pizza spinning pattern down, you can cut a large pizza into 12 slices in about 3 seconds making 6 cuts. The busiest pizza restaurants I’ve worked at will have you cutting 200-240 pizzas an hour, pulling them from the oven with a paddle, opening a box and dropping them into the box, cutting, and then closing the box. That’s about 7 seconds a pizza. This just isn’t possible with any other implement. Even the fastest rocker cutter users I’ve seen are topped out at about 150 an hour. The need to use 2 hands slows you down a lot.
"“Large language models do not, cannot, and will not ‘understand’ anything at all,” argued Tyler Austin Harper,"
Maybe others have had better luck, but I've never gotten anyone who thinks that explain to me how they think the electro-chemical computer between their ears "knows" anything. It's not even that they are ghosts in the machine people - the thought seems to never have occurred to them.
It’s pretty simple, really. The electro-chemical computer between your ears is not a large language model (LLM), just as, a century or more ago, it is not the clockwork it might have been compared to.The large language model doesn’t have experiences with the world (it has only read rumors of experiences with the world), and does not appear to have a model of the world against which it can check its dice-roll output (figuring out whether this is true or not is a contentious area of research).
This does not preclude the LLM from being supplemented with models — this is how the coding assistants are making the progress they are making, and it is how mathematicians get results out of the LLMs (the statistical model produces input for proof-checking software, which gives feed back where the input is lacking). There is some work being done to supplement LLMs with models of the physical world, as well.
"The electro-chemical computer between your ears is not a large language model (LLM)"
From what I've seen of the way a lot of humans reason, I'm not entirely sure of that. Heck, I'm not entirely sure there's that much daylight between human mythmaking (whether it be classic religious myths or more modern urban legends) and AI hallucinations.
Can you explain your terminology when you say coding assistants are making progress by being "supplemented with models"? They're making progress using RLVR tuning; are you saying the reward verification system is a "model"; e.g. a model of how Python programs actually operate?
Basically, yes. Though I’ll admit that the term “model” is heavily overloaded. The verification system has a model (i.e., a mapping from representation to truth values) that it uses to verify that the steps taken achieve their goal. Though I don’t want to claim great expertise (RLVR was a new term for me, for example — I’m in somewhat former ground w.r.t. the work using LEAN to verify steps in mathematics).
And as to incorporating models of physical behavior, that’s happening both in universities and industry.
Functionally, "understanding" or "acquiring knowledge" means updating an internal representative map of truth values.
Von Neumann machines may struggle with anything that is vague and fuzzy instead of strictly binary, but otherwise the difference is not that large. The fundaments of the LLM's model of the world are mathematical axioms and scientific laws, with everything else just ranked highly likely or less likely.
Of course, the LLM instance a user actually interacts with doesn't learn, this happens in the training process and then gets frozen in time.
People don't have "world models". The concept of a "world model" is entirely made up and comes from GOFAI research where they'd just invent a term and then decide it must exist because they said so.
See Phil Agre and Howard Dreyfus.
I knew Phil though I was late coming to read Dreyfus. You are correct that people don’t have a “model” instantiated in the technical sense I alluded to above (as a mapping between representations and truth values), but people most certainly do have models of the world that they gain through embodied experience: the behavior of a glass set on the edge of a table, for example, or the arc of a ball thrown through the air, or how a person will respond to one choice of phrase versus another.
So, we see these AI tools being applied most effectively in those domains where the path between symbols and validity is simple, clear, and verifiable: code, mathematics, and I think we see them growing less effective the further one moves from those domains into the world of (embodied) experience.
Sure, you may not compare your own brain to an LLM, but by "the problem of other minds", how confident are you that my brain actually has "a model of the world against which it can check its dice-roll output"? Heck, by how many times I have hit my toes on furnitures, I am not that confident in my brain checking its dice-roll output against its world model automatically.
Or going the other direction, I do not think lack of neuroplasticity ("the world that they gain through embodied experience") precludes the existence of a world model. Neuroplasticity is true and I agree that people with training can do amazing things such as juggling and gymnastics and super precise manufacturing. But it is also clear that a baby's brain contains a world model already at birth.
If we trace the evolution of symbols and languages, as I have happened upon Joseph Heath's recent article, according to Henrich, each language evolves as part of a cooperative culture system. That is, the domain of symbols is naturally outside of embodied experience, but rather have evolved as bridges among different embodied experiences. This situates LLM in a long evolutionary progress of human attempts of cooperating among minds blind to each others' inner workings but nevertheless inclined to imitate each other. According to Henrich, it is a feature that the language-driven cooperative structures evolve so much faster than biological evolution, that human brains in fact struggle to keep up. The world models contained in the cultures, expressed in their languages, are what the LLMs actually learn.
Whether LLMs have internal models or not is an active, and bit contentious, area of research. For example, there’s some work that claims to have located specific locations that represent the state of the game board in a net trained on sequences of legal game-moves (Othello). Changing the state of that area causes the LLM to start producing output consistent with the new game-state the model represents. In less constrained “worlds”, the evidence for the existence of a model is frequently harder to find (and there’s plenty of evidence that such models don’t exist (because they produce outputs that are not consistent with something we might call a model about the reality we want the LLM to “comment on”).
I don't really see why it would be necessary for people inclined to agree with Mr. Harper to first explain how consciousness/subjectivity exists and can "know." Neuroscientists and philosophers of mind regard this as a hard enough problem that they refer to it as "the hard problem."
Having come to appreciate how interesting the hard problem is, I think it's a mistake to suggest that organic awareness can be analogized to a computer, electro-chemical or otherwise. When you are recovering from general anesthesia there is a point at which you are aware of (for example) the doctor speaking to you. The moment before you are "awake" and making a memory of the doc you are "awake" and not making a memory of the doc -- awake but not aware; conscious but not knowing. That is a state that we have no reason to believe a computer will ever have because it is "awake" only in that it is computing. Of course, electro-chemical processes continue even when we are not awake (or even in an unconscious dream state), but that is because the same processes keep us alive. Unplug a computer and it will be ready-to-go when you plug it back in. Unplug us and the only computation left to make is casket or jar.
That said, acknowledging that AI "cannot understand" says almost nothing about its technological or social importance, and is no argument for dismissing it.
Very true. The molecules and atoms in the brain are just like the ones found everywhere else. Even the functional 'intelligence' which cells or larger structures in the brain have is not what we're thinking or talking about when we refer to human intelligence.
And just like AI the human mind is comprised of a range of task-specific modules responding to particular inputs, yielding particular outputs. Some of the modules are encapsulated with one sensory route in but most have multiple feeds including feeds from downstream modules which have already undergone some processing. Sometimes these modules are localised in areas of the brain, sometimes they are distributed globally across the brain, though often with a single trigger source. These modules and the larger functions they contribute to may do their own thing under the radar of conscious experience or may come together in a concerted way to produce conscious experience.
As animals get more complex they are able to bring more of these modules together in different ways. But the killer app that sets humans apart from all other organisms and accounts for our phenomenal success, is language which enables us to connect domains of thought in ways that wouldn't otherwise be possible. When we talk to others or talk to ourselves (out loud or silently) we are opening up connections between modules that enable executive functioning to be extended into long-term planning and complex problem-solving. AI apps like ChatGPT and Claude appear to do something very similar to this reflexive, self-instruction process. When you set them a task they break it down into parts and create a plan for tackling it, check the fit of the plan to the problem, checks their own logic and working etc.
AI may end up not meeting the highest expectations that people have for it and whether it ever develops into Artificial Consciousness (AC ?) is another matter (and probably dependent on us wanting to develop AC by giving AI interests, urges to meet these interests and maybe even emotions related to these urges). But I don't think we can say at this point that there's something about AI that completely rules it out of being genuinely intelligent in a recognisably human way.
A Toto washlet in a bathroom at NRT basically saved my life after a rough flight over from the US back in 2016 and I’ve never looked back. Have them in upstairs and downstairs bathrooms of the house. I think each cost around $400. I’ve probably converted a dozen people over the years but have also felt perplexed by the slow adoption.
If you live with a partner or share a bathroom, the most overlooked benefit is the charcoal deodorizer. You just never smell each others 💩 ever again. It’s absolutely amazing. The first time sharing a hotel bathroom after having a Toto for a while will be a harsh reminder of the before times.
Biggest piece of advice if you’re looking into them is to figure out where the nearest outlet is in your bathroom because they require 24/7 electricity.
That is the (an) issue with adoption: very few existing homes or apartments have an outlet near the toilet.
FWIW, I bought a high-voltage extension cord and a plastic junction box (for a waterproof connection with the Washlet), cemented it across my bathroom mirror's bottom trim. and plugged into a grounded outlet at the other end of the sink. Extra cost (beyond $340 for the S2 Washlet): ~$20, and total of an hour's time doing my own full installation.
It's kept my ass clean and happy ever since; even helps to ease constipation!
Installation generally includes the one-time cost of hiring an electrician to run wiring and install one.
I’ve had multiple across multiple apartments and homes over the years and never had to hire someone for install. Did have to get creative with an extension cord at one place but figured it out.
Seems like the perfect candidate for a battery no? Why not have one you can recharge?
Amen. Trust us fellow Americans: Must Get the Toto
Toto needs to put its products in mid-priced hotels. The slow adoption is because people don’t know they need it until they use it. A hotel is the place for people to use it without having to buy it first. Give those hotels a big volume discount.
There are issues with washlets in US commercial buildings because the water in toilets isn't clean enough for it. It's why Google took them out of their offices.
Is this because of grey-water recycling systems or something?
Right
I installed a Toto seat washlet after visiting Japan and I don't think I'll ever be without one again. It makes everything so much better.
Worth noting that it's often fairly cheap and easy to add an outlet near a toilet in an existing house. 100% worth it to avoid an awkward extension cord.
I believe it is ancient Samurai magic built into every Toto C2+ washlet. I don’t understand how it works and I’ve never changed any of the filters and don’t even know where they are, all I know is my bathroom never smells like poo
What sort of charcoal deodorizer do you speak of?
As someone concerned about power loads, I'd ask why it is a vampire load?
Most of them heat the water it sprays on you, it's much more comfortable.
Seems, er, wasteful. I'll pass. At least until the electrical generation system is decarbonized (or my house is so oversubscribed on solar and storage that I'll never know the difference).
Oy vey, "The Pwannet!" That really triggers me; you sound like NPR...
Please! I might be a New York Jew (a gay one, at that), but I'm a Ruben Gallego Democrat:
"Every Latino man wants a big-ass truck, which, nothing wrong with that.... We use terms like ‘bring more economic stability.’ These guys don’t want that. They don’t want ‘economic stability.’ They want to really live the American dream.”
In other words, I take my cues from the KKK: Kerouac, Kesey, and Kafka. (So much for "behavioral health"!)
FWIW, my electric bills haven't noticeably increased since I installed the Washlet. (The electric draw is barely intermittent.)
As for the warm water? I'd tell you to kiss my ass, but the Washlet does that for me... and for a gay guy, in a pulsating stream? Sheer heaven!...
To each their own, I guess. :-)
I’m skeptical that most of the members of the public using AI are doing that much useful with it! Part of the problem with the reflexive AI-skepticism is that seeing a lot of the common things people are doing with AI really does suggest it’s useless. It’s hard to realize you don’t have to use it to just generate slop when most of what you directly see people doing with it is generating slop. I gave a talk a few weeks ago to try to encourage the grad students in my department to figure out all the useful things AI can do for them, and gave five examples from my own research, teaching, and service to get them thinking. But if you just see slop doggerel and slop images, you might not realize that you now have the ability to write computer programs that do things you’ve always wanted to do.
Exactly. In Halt and Catch Fire, the character Joe McMillan - definitely more of a Steve Jobs than a Steve Wozniak - says:
“The computer isn’t The Thing; it’s the thing that gets you to The Thing.”
AI is the same way; I suspect that all the general models will become commodified - more specialized ones perhaps not so much - and the real money will be made by folks who figure out how to wrangle all that capability into, “killer apps.”
Could you share your list(s) of AI uses?
It was five specific things I've done:
1. describe mathematical diagrams to get the LLM to write TikZ code to include diagrams in LaTeX (this has sometimes required a bit of editing by hand afterward)
2. generating interactive applets for in-class activities (I made a little app to do Claude Shannon's "guess the next letter" activity to have a bit of sense of what predict-the-next-token really is like)
3. talk with ChatGPT to better understand all the steps in the proof of the Caratheodory extension theorem and why each condition is there, as well as what goes wrong with certain couterexamples if you miss them
4. create slides for class (both a timeline of LLMs and doing a little original research on what fraction of movies in each year were about AI from 1972 to 2012 - the years I got a list of wide-release movies in the US for)
5. generate images to illustrate slides (both in academic talks and class lectures)
Cases 2 and 4 are demonstrated here:
https://www.kennyeaswaran.org/teaching/ai-literacy/
I also mentioned some things I definitely *don't* use AI to do, which includes writing text that will be published under my name and writing summaries of papers I should be reading. (Having it write summaries of a bunch of papers I *might* want to read seems like it could be reasonable, though I haven't done it much yet.)
Very interesting essay, thanks.
I share the fear of a wet behind as noted in one of the early comments.
I’m floored by the AI comments. I’m 73 and almost universally among my circle of suburban friends (most of whom are to the left of me) the value of AI is a given. While there is concern re the impact on our kids and grandkids future, not developing and using it has not come up in our discussions.
I've had the same type of experience about AI discussions, Mr. Attermann (and I'm a bit older than you). The only exception is people who read about the Singularity and imagine AI is going to decide to destroy human civilization -- perhaps out of envy of the human pleasure of Toto washlets, which, I can assure you, have excellent aim once you choose the rear-view setting that fits.
It is interesting to realize, as I just have, that an AI can be but never have a coglione (an Italian word that I think many of us learned only today).
I have a Toto C100, one of the washlet seats that can be added to an existing toilet that Noah mentions. There's no wet behind issue, you just wipe after using it.
I think people who have never used one overestimate the amount of water it sprays; it's just a little bit of water.
ok... was I the only one who wondered.... so how is Noah going to connect Pizza Wheel, Japanese Toilets, and AI ....... until I saw this sentence and it all made sense....
"Unlike in the case of pizza cutters and washlets, Americans have correctly identified the most useful technology, and are adopting it."
For the record, lawyers are absolutely adopting AI. In biglaw, at least, AI use is increasingly dramatically, and clients are expecting if not mandating it for efficiency. The observation that it's about as good as a second year associate is high praise.
I suspect the lawyers will just expand the amount of work to meet the new capacity. Just like they have done with contracts over the last few decades. The old timers tell me about how 500 page credit agreements used to be 50 pages.
Unfortunately, law is one place where Jevon's Paradox will absolutely apply
I'm not in the business but I've read that clients are increasingly demanding fixed prices for chunks of work, not open-ended billable hours, and it's been squeezing the incomes of biglaw firms. If the clients keep doing that, and the actual cost of grinding out 500-page agreements goes down by a large factor, the clients will harvest much of the savings.
I have read stats elsewhere that number of lawsuits have gone up dramatically since the advent of LLm. I suspect previous ppl might just shrug something off, now they check with their LLM and say Hey, I have a case!!!
This is absolutely happening, although I think most of the effect is helping people without lawyers access the legal system.
Funny story: My friend is a city attorney, a job that requires him to go to court occasionally when people try to fight citations for things like off-leash dogs and littering. He recently had an incident in which someone fighting a citation tried to use a smartphone running an AI app to listen to what was going on the in the courtroom and tell them how to respond to the city/judge.
To some extent, it's good that AI is helping more people navigate complex and confusing legal systems. I do worry that a lot of our legal system is basically built on the assumption that things are rate-limited by the need to hire an expert of some sort, and that is going to quickly break. For example, I've heard from people in schools that parents are using AI to write letters demanding disability accommodations like extra test time for their children.
Can confirm, I am a lawyer and I am using AI daily.
I haven’t read the AI part yet, but let me congratulate you on your inspired selection of pizza cutters and buttwash toilets for this post.
Regarding pizza cutters: the long curved cutter is 100% the way to go.
Regarding buttwash toilets: once you’ve used one, anything else seems positively barbarian. I REALLY don’t like to use the bathroom anywhere except home.
I’m not sure you’ll be able to top this one.
As an animator and game designer I have noticed this dismissal of AI a bit in my work with a little twist. Recently, artists and animators have been highly encouraged to start using AI tools to create useful scripts and fix workflow problems. I love using AI for these purposes. As someone with basically no coding skills at all but with lots of knowledge about animation pipelines these tools were a godsend. Now I didn’t have to put in a new tech debt request to fix something. I can just do it. The moment the tools were available I jumped in and started making things. All my managers and producers love it. However, most of the animators didn’t care. They seemed to just ignore the tools entirely, didn’t engage with it at all, and kinda just kept their heads down and kept doing their usual work. Whenever I explained how to use the tools they didn’t care. That is, until they actually made something. When one of my co workers had a problem that required a script fix or plugin tool and I could just show them exactly how it works how you could make this new thing…their eyes lit up. It was like magic to them. It’s amazing to see someone who doesn’t have the coding skills create a tool they’ve always wanted to make. Then, after that they fall in love with this technology and it dawns on them…wow this really really matters.
Now, I wonder sometimes if being part of the elite educated progressive crowd is that you acquire a smug confidence that really you know everything already and that magic doesn’t exist. For regular artists, animators, and producers…it does.
i think its insane that most google engineers are only able to use google AI coding tools. it probably really skews their perceptions of what's possible and the current state of the industry. I wonder if the productivity difference is measurable.
The google AI team have an exception where they get to use Claude :/
I hear lots of skeptics in internal message boards
I found these two things weird:
---
2. It has a jet of water that washes your butt.
3. It also has a bidet mode.
---
AFAIK, bidet's have always washed your butt! https://en.wikipedia.org/wiki/Bidet
Washlet's have narrowed that definition for some reason I don't understand
Use your cleaver knife for cutting pizza.
Problem solved.
You’re welcome
Going into the COVID-19 lockdown, when people were hoarding toilet paper, some genius at Costco put out a whole bunch of washlets for sale, and we picked one up, I installed it, and my wife was very pleased with it. She eventually ordered a higher-grade model and I moved the first one to the guest bathroom.
I am a software engineer at Google and have been trying out the elephant-goldfish model. https://research.google/pubs/elephants-goldfish-and-the-new-golden-age-of-software-engineering/ My impression is that LLMs are able to do things that have been done dozens of times before, but need careful supervision by a senior engineer to make sure that the output is readable enough for human comprehension, and it tempts a lot of engineers to laziness that they pay for later when it comes time to fix the bugs. LLMs are bad at doing anything where they don't have plenty of examples to work from, and make foolish mistakes because they don't actually understand anything; if I get sloppy and don't check the references when I ask Duckie (Google's internally trained version of Gemini) questions, it will frequently waste my time suggesting things that aren't possible. LLMs that provide references most effective when treated as search engines that handle natural language queries instead of keywords. Trusting the summary without following the links is an easy way to get into trouble.
I use Copilot. What you're describing sounds like my experience with GPT-4-era models; the results I would get would often feel like they were roughly adapted from other peoples' code.
Working with a GPT-5 era model is not like that *at all* - I'm guessing it's the RLVR that makes the difference. It shows (or emulates) clear understanding of the causal structure of the code is writing or analyzing, even when I'm asking for something extremely specific.
Where it still struggles is high-level planning; I have to give it clear instructions about the architecture of what it's creating or it will make decisions that are locally correct but globally messy.
I suspect software architecture will remain very “human-hands-on,” even as AI does more and more of the actual coding. The architecture design is arguably the most important part of it all.
Apparently Mythos is a step change up in architecture specifically (haven't used it; just heard). There's no objective function to train towards, but I don't think it's an inherently unsolvable problem.
I am a software engineer at not Google. I think you're way behind the curve on this. Your description sounds like models from almost a year ago that were good for unit tests but not much else. The current versions have their weaknesses and still need adult supervision, but the code they write is "readable enough for human comprehension" 100% of the time. I am now able to make changes to services all over the code base that I have never touched before. The productivity increase is big. Once the organization adapts to the new reality and I am not held back by obsolete process and procedure the difference will be massive.
Huh, I didn't realize elephant-goldfish was publicized externally (albeit, on Medium).
Your points are valid but the models have gotten better quite recently, enough that I would caution against treating them as "not understanding anything".
Duckie, since you mention it, is much better if you get access to bigger models (or use the newest default) and use "agentic mode". I also noticed a step function improvement in recent weeks in success with longer horizon coding tasks with the latest Gemini 3.5 Flash and the latest Jetski (Antigravity) harness. 3.5 seems to be better at instruction following and has better design sense.
Some of the biggest improvements to my own practice comes from asking the models to help me reason through an analysis of legacy code. It's better at being systematic and creating artifacts to trace system behavior than I am.
("Before we make any changes, tell me what happens if you change X to happen before Y? Track down all the references to X and provide a case by case analysis. Output a CSV with file names, line numbers, and conditions under which that line is executed. I will review it.")
In the case of Duckie, I might ask for an analysis of how the change came to be (it can integrate across code changes, historical docs, the org chart/team graph, etc).
Also, the Flash model is often a better choice if you're not doing frontier work or trying to one-shot a giant app, because it's much faster and can iterate quickly instead of spending a lot of time ratiocinating. And yet it doesn't lose the plot while it's churning (which was a failure mode of the older models).
Generally, I find that once I know what I want (the hard part, which can involve a sustained conversation that results in artifacts like those I mention above) I give a moderately long prompt and walk away and get a reasonable first cut at a submittable change. But, you're right that in that it's a first cut -- you have to review it yourself, first!
I’ve had better luck with Claude, but I agree that one must keep it on a relatively short leash — when it encounters problems, it can come up with some hair-brained approaches, when what is really needed is rethinking the approach. It has allowed me to do things in a day that would have taken weeks before, however.
Ok, there’s a thing that bothers me about this “the left is missing out on AI” discourse. The reasons for skepticism and dismissal that are rebutted here are certainly representative of commonly held beliefs, I’ll grant that. But I don’t think they do justice to the *best* arguments progressives have for being wary of the technology. If the reasons progressives commonly cite for not using AI are *wrong*, but, at the same time there exist much stronger arguments against using it than those typically proffered, then obviously they’d just need to adopt those better arguments, they wouldn’t need to change their actions.
And what bugs me is that it doesn’t seem that hard, with a minute of thought, to come up with far better justifications for progressives to be wary. There is no danger of becoming dependent on a pizza shear that then jacks up its subscription price. There is no danger of a company that produces washlets becoming so powerful that its unelected leadership is seated beside heads of state at the G7. Using these products does not contribute to the further precaritization of artists, or the delegitimization of other professions. While the “data centers are using all the water” stuff is massively overblown, and data centers don’t look so bad when you compare them to the impact of other commodities and conveniences we happily use without a second thought, there are still horror stories like the xAI’s Colossus data center in Memphis with its screaming smog-producing gas turbines—which is now being used by Anthropic!
At minimum, progressives just experienced a decade in which many major tech billionaires went from being purported progressive allies to enthusiastic Trump collaborators—about the worst thing someone can do from a progressive POV, and something that surely would set many folks against any activity likely to make that cohort richer and more powerful.
So I feel like, until the better arguments are really addressed, this discourse is going to remain an exercise in talking past one another, and is unlikely to produce real changes in the behavior of people on different sides of this growing divide.
Agree. It is not 1998. People have experienced over a decade of enshittification. They have seen enough. The new technology at best makes things we already had enough of (often too much of) more plentiful. Roping a bunch of chips together and filling them with stolen text turned out to be very useful in a brute force sort of way. But so was laying down a bunch of highways. The same amount of brute force effort could have improved the world in more valuable ways, and a looming sense that people have little control over how these efforts are directed is being communicated in all kinds of ways - some of them counterproductive!
Scolding AI skeptics while pointing out that smart phones are toxic suggests an opportunity for a synthesis around “what is going on”.
As to art: another example of LLM facilitating the production of something we already have plenty of. Does anyone think there are too few novels and paintings and little cutesy videos in the world? Did anyone think we should build a server farm the size of Phoenix to improve our supply of “art?”
And in practice at real offices: this stuff can be super annoying. It produces more emails when we want fewer, better emails. It produces more PPTs when we want fewer, better presentations. It offers to summarize emails (a summary form to begin with) when the correct response to an email that requires summarization is to delete it unread. It is not without a LOT of high value use cases. But there are a swarm of low value use cases all over the office place now — and it is fine to be annoyed by them.
I agree that many current office uses of AI are annoying: more emails, more slides, more summaries of things that perhaps should not have existed in the first place.
But I am not sure “more low-quality output” is the deepest problem. From an evolutionary perspective, information explosion is not a defect; it is the normal result of a new communication or production layer. Printing, newspapers, radio, TV, the internet, and social media all produced floods of material that existing elites saw as vulgar, useless, or low-quality. Some of it was. Some of it later became new culture, new markets, new institutions, or new coordination mechanisms.
I am also not sure we can define “bad workflow” or “good workflow” from a neutral standpoint. The same workflow may look terrible to a manager, useful to an employee, profitable to a firm, exhausting to a freelancer, or liberating to someone previously excluded from the system. “Quality” and “efficiency” are not fixed categories outside social context.
The real issue is whether people have the right and practical ability to choose, experiment, refuse, exit, and move to better arrangements. Markets and social selection can only work if participants are not locked into systems they cannot leave.
So the question is not how to stop abundance or centrally define quality. It is how to preserve enough openness, competition, mobility, and exit rights for better filters, workflows, and institutions to emerge.
I think Noah and Erik are both pointing at something real here.
Noah is right that dismissing AI as “just slop,” “just autocomplete,” or “just a bubble” is a bad strategy. A technology can be overhyped and still be historically important. But Erik is also right that progressive skepticism is not only about whether the tool works. It is also about who owns the surrounding system, who absorbs the externalities, and who gains power from adoption.
This is why I think there is an important distinction between refusing to use a technology and criticizing the industry around it.
Power plants have environmental, monopoly, regulatory, and political problems. But most progressives do not refuse to use electricity as a matter of ideology. They use electricity while criticizing the ownership structure, externalities, regulation, and political economy of the energy system. The same is true of smartphones, airlines, banks, and the internet.
So when AI use itself is treated as morally suspicious, that may be a sign that AI has not yet become ordinary infrastructure in many people’s daily lives. The anxiety is real, but it is partly pre-adoption anxiety. People are reacting not only to the tool, but to what they fear AI will reorganize: work, authorship, expertise, status, institutional trust, and political power.
I’ve been thinking of this as “AI anxiety”: not simply fear of a new tool, but fear of a new technological layer before society has learned how to live with it, criticize it, and govern it at the same time.
The better position is not “AI is just slop” or “AI is automatically good.” It is: use the technology where it is useful, understand it from inside real workflows, and then criticize much more precisely who owns the nodes, who pays the external costs, who loses bargaining power, and who captures the gains.
But I think the deepest problem inside AI anxiety is not only technological. It is political. Much of the AI debate is still organized around intelligence, capability, alignment, automation, and safety. These are important questions, but they often leave out the harder one: power. Who has the authority to decide how AI is deployed? Through what institutions? With what legitimacy? And can political structures designed around sovereign nation-states govern a technology whose data, supply chains, users, and effects are already global?
That, to me, is the question AI anxiety should push us toward. Not just “should we use AI?” but “what kind of institutional imagination do we need for a civilization in which AI is becoming infrastructure?”
I wrote more about this here: https://substack.com/@deepbitcheesebrew/p-204107626
The art historian "insisted that AI was somehow different" because it is. A human working with a technology to create art (photography and film) is different than a human instructing a technology to create art (AI-generated images and video).
Try to find the clear defining line between various tools an artist may use and an artist using AI.
That all depends on how much control you abdicate and how much engagement you have with the process. It’s like the difference between just point-and-shoot and considering composition and lighting and all the other things involved in letting a camera create an image by passively processing light that hits its film, with no control from you.
The comments about pizza wheels are just silly. One such as this:
https://www.amazon.com/Kitchy-Pizza-Cutter-Wheel-Protective/dp/B019S3W8AO/134-6755186-1193922
doesn't have any of the drawbacks you mention.
For anyone who’s had to cut pizzas at volume, and for delivery, where getting the pizza cut swiftly to keep temperature high is crucial, the regular pizza cutter disparaged here is an absolute necessity.
It takes a bit of training (maybe 20-30 pizzas in quick succession) to get the feel for it, but once you have, nothing else comes close.
The rocker cutter is second best at roughly 1/2 the speed.
The trick is to get your elbow up high above the pizza. Once you have that feel, and the cut making + pizza spinning pattern down, you can cut a large pizza into 12 slices in about 3 seconds making 6 cuts. The busiest pizza restaurants I’ve worked at will have you cutting 200-240 pizzas an hour, pulling them from the oven with a paddle, opening a box and dropping them into the box, cutting, and then closing the box. That’s about 7 seconds a pizza. This just isn’t possible with any other implement. Even the fastest rocker cutter users I’ve seen are topped out at about 150 an hour. The need to use 2 hands slows you down a lot.