I sincerely appreciate Ed’s in-depth research and financial analysis in this piece. There’s a lot of effort that goes into producing something like this and that’s worth mentioning. I’m an...
I sincerely appreciate Ed’s in-depth research and financial analysis in this piece. There’s a lot of effort that goes into producing something like this and that’s worth mentioning. I’m an AI-hater as much as the next guy here on Tildes, but I just get the sense that this article misses the forest for the trees.
Every executive, venture capitalist, C-suite officer, and financial analyst understands we are in a massive, massive bubble right now. They all agree that this bubble will pop, many people will lose their jobs, and some would probably even agree with you if you said that some of these MAG7 companies will go under. They’re not betting on the growth from AI saving their capex, they’re dumping capex so they can be the last man standing when the bloodbath is over.
One has to recognize these people learned their lessons in the dot-com bubble, and are applying the same lessons here. Yes, companies dumped exorbitant sums on (in hindsight) very useless investments and some went bankrupt because of it. Yet the winners of that bubble were not the ones who played it safe and let the new trend bandwagon pass them by. The winners were the ones who dumped the capex, sometimes ran negative for years, and then came out on top: Amazon, Google, Microsoft, etc.
Every investor believes that only a few of these AI investments will pay off, and just aren’t sure which ones will (or they are all-in on one or two horses). But they fully believe that when all is said and done, the bankruptcies and firesale acquisitions over, the costs falling down, then the new kings of the global economy will have been crowned.
Look, I cannot say if I agree with their analysis. But when I go interact with average folk in their day-to-day life, they seem to love these glorified word predictors. They use them to generate emails, to search for recipes, and sometimes to console them through a tough break-up. I spend time in classrooms and students can’t get enough of “Chat”. I spend time with family and they talk about their new use of AI so that they don’t have to make that stupid presentation at work. I see this tech and I see it being popular, at least in the right way.
Fundamentally, that’s why I hate this tech so much. If Ed Zitron’s market collapse prophecy comes to pass, my gratitude might make me a religious man. The gamble paying off is so, so, so, so, so much worse. Imagine the next Google not just having access to your search history and watch history, but a personalized model that they can use to convince you of things. Imagine a product that allows you to deploy subtle biases into millions of personalized chatbots that consumers are using, changing political opinions and market demands at scale. That’s what the VCs see, and they are damn right to say it’s worth risking bankruptcy for. I just pray that vision is a mirage.
Without even making a judgement about how well these products actually work, there's still an issue here. I'm sure I would love flying around in a private business jet wherever I want a few times...
Exemplary
But when I go interact with average folk in their day-to-day life, they seem to love these glorified word predictors. They use them to generate emails, to search for recipes, and sometimes to console them through a tough break-up.
Without even making a judgement about how well these products actually work, there's still an issue here.
I'm sure I would love flying around in a private business jet wherever I want a few times a month. I can't afford to do that though, and neither can almost anyone else, which is why the charter jet industry is a $25 billion a year industry, not a $5 trillion a year industry. If business jets were hyped beyond any semblance of reality such that venture capitalists, private equity, fund managers, and everyone else subsidized flights so that every single person in the country was taking private jet trips every week, you could definitely make it look like a $5 trillion dollar industry though.
What we're doing with AI isn't sustainable. People use it because generating a funny picture or writing somewhat convincing, but bland, inaccurate text can be fun when you're doing it for free or for 10 bucks a month or whatever.
If people had to pay the true cost of it though? How many boomers would really be willing to shell out $50 bucks a month to churn out pictures of Jesus made of fruit? Probably not many.
There's basically no one I could see being willing to pay this except people who directly use it to help with their jobs, like developers and copywriters. That market is way, way, wayyyy smaller than what the investment in this companies is signaling though. In order for the hype to be justified, we'd have to be living in a completely different reality.
I think this is going to depend a lot on how much that $50 ends up being. Bear in mind that there's a lot of range to these models, from ones that you can run on your laptop for free, to ones that...
How many boomers would really be willing to shell out $50 bucks a month to churn out pictures of Jesus made of fruit?
I think this is going to depend a lot on how much that $50 ends up being. Bear in mind that there's a lot of range to these models, from ones that you can run on your laptop for free, to ones that are tremendously expensive. I suspect the "pictures of Jesus made of fruit" market will be pretty satisfied with cheap models that can be run as a kind of loss leader. And for professionals, $50 a month isn't that much, assuming what you get out of that is genuinely useful.
The bigger question is how much training costs are going to change. The impression I get is that the current prices are fairly sustainable in the sense that, say, Anthropic's $20/mo tier largely covers the cost of running the LLM queries being made. But right now, everyone needs to be spending a lot on research and training for the next model, otherwise they'll be left behind, and that's what's unsustainable. Presumably, at some point, the improvements between generations will not be worth the cost of that research, at which point it will be possible to concentrate only on selling access to a pre-trained model, which is right now the sustainable part of the business.
I agree that we're in a bubble right now, and a lot of AI stuff is being sold just because AI sounds cool, but hype cycles are nothing new. What's interesting this time is that, even outside of the hype cycle, the underlying technology still seems very useful.
Actual cost for running a top-tier model comes in around $0.03 per API request, and that’s running it personally (i.e. accounting for overhead as the container spins up, rather than efficiently...
Actual cost for running a top-tier model comes in around $0.03 per API request, and that’s running it personally (i.e. accounting for overhead as the container spins up, rather than efficiently batching a constant stream of input from multiple users), you can run less capable models for less than 1/10th of that.
The sunk costs into OpenAI et al might be tough to recover, but third-party-hosted Llama and DeepSeek are already viable at reasonable cost.
Are they? When self-hosting Claude 3.7 Sonnet on AWS for use with Cursor, it's extremely easy to run up a $100 bill every day. Sure, we can use worse models, but the cutting edge models are...
Are they? When self-hosting Claude 3.7 Sonnet on AWS for use with Cursor, it's extremely easy to run up a $100 bill every day. Sure, we can use worse models, but the cutting edge models are already barely acceptable.
The most popular use-cases are also the most expensive, and that seems like a problem.
I don’t use coding models (or LLMs in general, really; I just work on the tech they’re built on top of) so I don’t have a great benchmark for what’s reasonable in that workflow - but unless I’m...
I don’t use coding models (or LLMs in general, really; I just work on the tech they’re built on top of) so I don’t have a great benchmark for what’s reasonable in that workflow - but unless I’m missing something it looks like $100 would be multiple millions of tokens even with AWS markup. I’m not saying it’s impossible, but that’s the entirety of War and Peace four or five times over in a single day, which seems like a lot even accounting for the complexities of code generation?
Chat models are easier to eyeball because the input and output is pretty much capped to what an individual can read or write in a reasonable time, so that’s what I had in mind for extrapolating up to reasonable daily use. I’d be surprised if non-code users are regularly generating more than 10k tokens/day (roughly 35 printed A4 pages of text).
[Edit] As a side note, there’s also an upper limit on reasonable cost set by the option of buying dedicated hardware.
You can fit the vast majority of models on a 512GB Mac Studio, which sells for $10k. It may not even be the best choice in a lot of cases, but it gives a good ballpark on what “reasonable” looks like: if a company can support four employees’ Cursor sessions per machine and keeps it for three years, that’s $70/employee/month for effectively unlimited use.
Yes, it's code specific, but that's not an unreasonable sum for an agentic tool like Cline with a big MCP server. Querying multiple repos or project management systems with large projects adds up...
Yes, it's code specific, but that's not an unreasonable sum for an agentic tool like Cline with a big MCP server. Querying multiple repos or project management systems with large projects adds up real quick.
Right, that's the bit I disagree about. As I understand it, the API pricing for most of these services roughly corresponds to the real cost of serving those requests, and the per-month...
Right, that's the bit I disagree about. As I understand it, the API pricing for most of these services roughly corresponds to the real cost of serving those requests, and the per-month subscriptions are similarly profitable as long as you put in sensible rate limits. And currently those prices are fairly reasonable right now.
Like I said, I think the biggest expense right now for these companies is the research and training needed to stay ahead of the curve and bring out their next generation of LLM. And I suspect over time, that expense will go down proportionally as these companies begin to max out what LLMs are reasonably capable of.
Three big thoughts: Microsoft was already an entrenced monopoly. Google was the seach engine and was hurtling towards profitability at record pace. Amazon survived by the skin of its teeth,...
Three big thoughts:
The winners were the ones who dumped the capex, sometimes ran negative for years, and then came out on top: Amazon, Google, Microsoft, etc.
Microsoft was already an entrenced monopoly. Google was the seach engine and was hurtling towards profitability at record pace. Amazon survived by the skin of its teeth, because it had just secured funding before the bubble crashed. If that deal took 3 more months, Amazon would have been as dead as Pets.com, the most infamous. Pets burned $12 million in advertising to attain a peak revenue of less than a million. Which was also sold at a loss.
The stuff that survived had nothing to do with capex, and everything to do with actually having a viable path to profitability.
But when I go interact with average folk in their day-to-day life, they seem to love these glorified word predictors.
I think that's where the Netflix and Spotify comparison comes in. The two largest streaming services have a combined revenue of $67 billion. Do people love these AI chatbots so much that they'll be willing to pay $100 or more a month when the pied piper comes to collect on all that capex expenditure? I think it's a fun novelty to most people that will happily abandon it when faced with the real cost of a sustainable product.
They’re not betting on the growth from AI saving their capex, they’re dumping capex so they can be the last man standing when the bloodbath is over.
And because all of those people are literally burning everyone's retirement funds to do so, I hope they all rot in jail when they are.
I think we should distinguish between money spent on data centers and GPU's and electricity (they are real costs) and the money people think they have due to rising stock prices. The stock market...
I think we should distinguish between money spent on data centers and GPU's and electricity (they are real costs) and the money people think they have due to rising stock prices. The stock market went up quite a bit since 2008. If it goes down again, that means long-term investors are less wealthy than they thought (unless they got out) but they might not have lost money overall.
If you didn't buy during the bubble, it's more like, you thought you were richer for a while, but it was an illusion. It never really existed except in speculator's minds.
Meme stocks and cryptocurrency are purer versions of this, but it's somewhat true of all financial assets in a bubble.
Generally I agree with you. I think Google is the best comparison here. Google was the new king post dot-com. Google won because it became the search engine and became an extraordinarily...
Generally I agree with you.
Google was the seach engine and was hurtling towards profitability at record pace.
I think Google is the best comparison here. Google was the new king post dot-com. Google won because it became the search engine and became an extraordinarily high-margin business.
Do people love these AI chatbots so much that they'll be willing to pay $100 or more a month when the pied piper comes to collect on all that capex expenditure? I think it's a fun novelty to most people that will happily abandon it when faced with the real cost of a sustainable product.
I think most investors want the product to stay free. Let me speak from personal experience. I have had clients walk into the business that I work at, effectively ready to sign a contract without even talking to us. They had asked ChatGPT for a set of services and ChatGPT gave an absolutely glowing review of us. Why? I don’t know. I guess our website got lucky.
People trust the chatbots, especially people that lack tech savvy and don’t understand how the models actually function. Its effectiveness as a convincer is so much better than any modern search engine.
I think the product offering will look a lot more like “pay OpenAI to run some fine-tuning on the latest model so that it recommends our business more often” to users. Subscription services suffer because users aren’t the cash cow, businesses are. In that scenario, whoever has the userbase will win, just like Google did.
I would distinguish between business and home use. At work, that's not an unreasonable price. If you compare $100/month to the monthly salary of an employee who uses the tool, the tool is cheap in...
Do people love these AI chatbots so much that they'll be willing to pay $100 or more a month when the pied piper comes to collect on all that capex expenditure?
I would distinguish between business and home use. At work, that's not an unreasonable price. If you compare $100/month to the monthly salary of an employee who uses the tool, the tool is cheap in comparison - if it's useful, which it definitely is for some developers. (Example.)
For consumers, I expect that many will pay $20/month if they're getting a lot out of it, but costs will go down to the point where it's often free and advertiser-supported. Google's AI search results are an example. (A lot of people hate them because they're being pushed on us, but I see no reason why they won't improve.)
I haven't seen much research actually supporting the supposed productivity increases. It makes devs feel more productive, but all the research I've seen says it actually slows down competent devs.
I haven't seen much research actually supporting the supposed productivity increases. It makes devs feel more productive, but all the research I've seen says it actually slows down competent devs.
Yes, it's anecdotal and likely varies a lot. There is one study that's been shared around showing that some developers were slowed down by it, which seems plausible, but it also might have been a...
Yes, it's anecdotal and likely varies a lot.
There is one study that's been shared around showing that some developers were slowed down by it, which seems plausible, but it also might have been a skill issue that goes away with more experience? There is definitely a learning curve.
My work is paying for ChatGPT and cursor, I want to replace the latter with Claude code starting Monday too. And I think if I reach the limit of the 20€ a month plan, the 100€ a month Claude plan...
My work is paying for ChatGPT and cursor, I want to replace the latter with Claude code starting Monday too. And I think if I reach the limit of the 20€ a month plan, the 100€ a month Claude plan will be a no brainer too.
I am very weary of vibe coding, but if you decide what you want and let the AI do the typing for you and then double check it (seeing it as your junior dev, rather than just accepting it as gospel) the speed up can be immense.
My ADHD brain just has to not get distracted every time I wait for the LLM outputs
I was under the impression this was the explicit interest of incorporating Google AI into search. But it reminds me of this Parks and Rec bit https://youtu.be/K3YNmoV9NsM?t=46
Imagine the next Google not just having access to your search history and watch history, but a personalized model that they can use to convince you of things.
I was under the impression this was the explicit interest of incorporating Google AI into search. But it reminds me of this Parks and Rec bit
-this is Grizzle Vibe it's this new app that we're developing that monitors your mood
-tell them about it
-as you know, the cameras on your phones are always on whether you're using them or not-
-I'm sorry they are?
-this app uses facial recognition software to track your expressions, it's always watching
-well, what do you do with this information?
-well, if the camera senses that you're in a bad mood, then we could geo-match you to say the nearest cup of sweet pick-me-up java but if you're in a good mood, then we could geo-nudge you to like a sweet coffee shop you could just keep the good times rolling
I’m sure I’ve quoted this in relation to online advertising before, but it just keeps getting more apt: Sure, I know the propaganda at scale is working, I can see the impact and it’s honestly...
I’m sure I’ve quoted this in relation to online advertising before, but it just keeps getting more apt:
He had found a Nutri-Matic machine which had provided him with a plastic cup filled with a liquid that was almost, but not quite, entirely unlike tea.
The way it functioned was very interesting. When the Drink button was pressed it made an instant but highly detailed examination of the subject's taste buds, a spectroscopic analysis of the subject's metabolism and then sent tiny experimental signals down the neural pathways to the taste centers of the subject's brain to see what was likely to go down well. However, no one knew quite why it did this because it invariably delivered a cupful of liquid that was almost, but not quite, entirely unlike tea.
Sure, I know the propaganda at scale is working, I can see the impact and it’s honestly scary. But it’s also still a daily occurrence to see companies running astonishing amounts of compute across absurdly granular and invasive data, only to show me ads for a brand that a simple geoIP lookup would’ve told them doesn’t even exist where I am.
I suspect that's coming now,.later, or is already here. Social media content presentation algorithms are already subtly biased and pushing people in directions. The question isn't whether this...
I suspect that's coming now,.later, or is already here. Social media content presentation algorithms are already subtly biased and pushing people in directions.
The question isn't whether this next phase shall come to pass. The question isn't whether a select handful of businesses will become massively rich.
The question is whether people and the entities through which we come together to protect ourselves - governments - will have the will power to do absolutely anything at all about it
The answer is that we will not have the capacity or the power.
Which ones? Tesla seems vulnerable if they stop selling as many cars. The car industry is very competitive and they're less popular than they were. That's not really an AI thing, though. Nvidia...
some would probably even agree with you if you said that some of these MAG7 companies will go under
Which ones?
Tesla seems vulnerable if they stop selling as many cars. The car industry is very competitive and they're less popular than they were. That's not really an AI thing, though.
Nvidia would see a large drop in revenues if the other big tech firms cut back.
This could be good for the profitability of the other firms because it means they stopped spending like mad? On the other hand, their previous capital expenditures go on the balance sheet and don't immediately affect profits. They will be a drag on profits in future years due to depreciation. (Alternatively, it would be a one-time loss if the investment were written off, as companies do sometimes.)
If AI revenue is low for the other big tech companies, that would make them more resilient if the bubble bursts. For example, Google's total revenue is $96 billion (annualized), so losing ~8 billion in AI revenue wouldn't hurt them all that much, and this suggests that they will easily survive.
The stock market might still tank though, because they are pricing in expectations of future AI revenue that would evaporate if the bubble bursts.
Are you sure they aren’t using chat in reference to streamer culture? IE: Chat am I cooked, Chat what do I do? Some of my friends kids say this stuff and it’s not a reference to ChatGPT at all.
I spend time in classrooms and students can’t get enough of “Chat”.
Are you sure they aren’t using chat in reference to streamer culture? IE: Chat am I cooked, Chat what do I do?
Some of my friends kids say this stuff and it’s not a reference to ChatGPT at all.
I’m talking generally about college kids and the people using “Chat” seem to be anthropomorphizing ChatGPT since in the context of the sentence, “Chat” could be replaced with “Google”. “Let me ask...
I’m talking generally about college kids and the people using “Chat” seem to be anthropomorphizing ChatGPT since in the context of the sentence, “Chat” could be replaced with “Google”. “Let me ask Chat.” “Chat told me XYZ.” I’ve heard that use of “Chat” too but honestly I feel like the two uses of “Chat” come from pretty separate communities.
I have not heard any college students refer to chatGPT as "Chat" it always means "present company" referencing like a twitch/live stream chat, but I'm training some RAs soon and I'll ask them...
I have not heard any college students refer to chatGPT as "Chat" it always means "present company" referencing like a twitch/live stream chat, but I'm training some RAs soon and I'll ask them because it's fun to make them cringe at old people using their slang.
Ed Zitron consolidates a whole bunch of economic data surrounding AI, and the results are terrifying. The key snip to remember through the whole thing: And he goes on to explain, in immense...
Ed Zitron consolidates a whole bunch of economic data surrounding AI, and the results are terrifying. The key snip to remember through the whole thing:
In simpler terms, 35% of the US stock market is held up by five or six companies buying GPUs. If NVIDIA's growth story stumbles, it will reverberate through the rest of the Magnificent 7, making them rely on their own AI trade stories.
And he goes on to explain, in immense detail, just how bad those AI stories are.
I read the whole thing, it's well written and breaks down several things that I've been wondering about but haven't been able to articulate to people. The last quote in particular hits on a thing...
I read the whole thing, it's well written and breaks down several things that I've been wondering about but haven't been able to articulate to people.
Yet after that, generative AI feels more like a feature of cloud infrastructure rather than infrastructure itself. AWS and similar megaclouds are versatile, flexible and multifaceted. Generative AI does what generative AI does, and that's about it.
You can run lots of different things on AWS. What are the different things you can run using Large Language Models? What are the different use cases, and, indeed, user requirements that make this the supposed "next big thing"?
The term "agent" is one of the most egregious acts of fraud I've seen in my entire career writing about this crap, and that includes the metaverse.
OpenAI is a terrible business, and the only businesses worse than OpenAI are the companies built on top of it. Large Language Models are too expensive to run, and have limited abilities beyond the ones I've named previously, and because everybody is running models that all, on some level, do the same thing, it's very hard for people to build really innovative products on top of them.
So, really, everything comes down to NVIDIA's ability to sell GPUs, and this industry, if we're really honest, at this point only exists to do so. Generative AI products do not provide significant revenue growth, its products are not useful in the way that unlocks significant business value, and the products that have some adoption run at such a grotesque loss.
If you want to know my true agenda, it's that I see something in generative AI and its boosters something I truly dislike. Large Language Models authoritatively state things that are incorrect because they have no concept of right or wrong. I believe that the writers, managers and executives that find it exciting do so because it gives them the ability to pretend to be intelligent without actually learning anything, to do everything they can to avoid actual work or responsibility for themselves or others.
There is an overwhelming condescension that comes from fans of generative AI — the sense that they know something you don't, something they double down on. We are being forced to use it by bosses, or services we like that now insist it's part of our documents or our search engines, not because it does something, but because those pushing it need us to use it to prove that they know what's going on.
The last quote in particular hits on a thing I'd been feeling strongly the last year or so. While AI has some utility and is interesting, the people around me hyping it up all give off a similar vibe. That vibe is one separate from the usual used-car-salesman ick that accompanies people blindly buying into a trend. It's a discomforting sense that the people who are heavy AI users are commonly bluffing their way through everything. That they're operating solely on AI summaries of everything in their life, and don't actually understand any of it at anything beyond a surface level.
I didn't think I'd read the entire thing, out loud, but it's just too compelling. Ed talks about every single thing I believed and wasn't smart enough to articulate. My feelings on the technology...
I didn't think I'd read the entire thing, out loud, but it's just too compelling. Ed talks about every single thing I believed and wasn't smart enough to articulate. My feelings on the technology haven't changed, nor are they relevant here - but he's right. This particularly cancerous growth of the tech economy is rotten to the core, and as a rotten tree falls it will still shred its neighbours and cause untold damage.
Love to see Ed’s articles. Could someone explain once the data centers and housing are bought, will costs dramatically go down for AI companies? Or will this be an ever accelerating arms race?...
Love to see Ed’s articles. Could someone explain once the data centers and housing are bought, will costs dramatically go down for AI companies? Or will this be an ever accelerating arms race? This is the one explanation for why burning through this much capex is possible if it’s just to build infrastructure.
I can't fully answer your question, or answer it in a detailed enough manner, but data centres are pretty expensive to run/maintain/operate (especially something of such a massive scale) and that...
I can't fully answer your question, or answer it in a detailed enough manner, but data centres are pretty expensive to run/maintain/operate (especially something of such a massive scale) and that equipment will be obsolete in like 5 years and need to be replaced by newer technology anyway, so there'll still be a lot of CapEx left. So IMO the total costs will somewhat go down a little bit but there'll still be plenty of OpEx and CapEx.
The real trouble isn't the infra costs in total IMO. It's the fact that every single LLM -- even the cheapest ones -- are absurdly expensive to operate, even on existing infrastructure. A...
The real trouble isn't the infra costs in total IMO. It's the fact that every single LLM -- even the cheapest ones -- are absurdly expensive to operate, even on existing infrastructure. A subscription you use with any real frequency ought to cost you more than your Internet connection or electric bill. Right now tech companies are giving away the functionality for free to get people hooked, just like carshare apps or delivery food did in the first couple of years. But just like those (now much more expensive) products, eventually LLMs will need to monetize. Injecting ads and product placement will make the product worse, possibly useless and untrustworthy. Charging people hundreds of dollars of compute per month is too expensive even if it is good. Once the free money dries up these products are dead.
That's a really sweeping generalization that can't possibly be true. I can run LLMs on my phone for free. I can run them on my $500 Mac Mini for free. The providers that are selling access to...
That's a really sweeping generalization that can't possibly be true. I can run LLMs on my phone for free. I can run them on my $500 Mac Mini for free. The providers that are selling access to baremetal for running LLMs on are almost certainly selling their services for a profit and you can see that it's not expensive. LLMs range too widely to make a claim that "every single LLM are absurdly expensive to operate". There's tons of useful models out there that run on cheap hardware.
What are these costs though? How much does it really cost to run an llm once the model is made and the servers are running? Does anyone have actual figures for this?
What are these costs though? How much does it really cost to run an llm once the model is made and the servers are running? Does anyone have actual figures for this?
Electricity, GPU, etc adds up. Maybe they'll get more efficient eventually, but every new model consumes an order of magnitude more power so far. Most people want the latest models, not the cheap...
Electricity, GPU, etc adds up. Maybe they'll get more efficient eventually, but every new model consumes an order of magnitude more power so far. Most people want the latest models, not the cheap stuff.
But more importantly, this very article makes a separate point: NVIDIA's stock price is predicated on everyone else increasing GPU spend every quarter. If models get way more efficient, NVIDIA's numbers get worse... which ought to impact the stock market a lot since NVIDIA accounts for a massive chunk of the market!
This website might help? https://simonwillison.net/2025/May/7/llm-prices/ For open source LLM’s, the model is free so whatever they’re charging to run it just includes inference costs and their...
Every new technology has its haters, even the printing press had its own. A Benedictine monk and scribe once wrote: "the pen is a virgin, but the printing press is a whore." Others thought it...
Every new technology has its haters, even the printing press had its own. A Benedictine monk and scribe once wrote: "the pen is a virgin, but the printing press is a whore." Others thought it debased, not democratized, knowledge, as printers rushed to print out cheap crap, like love poetry or bibles that the masses could read without the help of some priest. Some thought that printers were out to make a quick buck, not do good for society.
Anyway, the printing press led to the Reformation and many wars and completely changed society and the world — eventually for the better in the very long run, most would agree, although it caused much bloodshed and pain for a few centuries.
I think everyone knows the current generation of AI is limited, but everyone knows this is a Gutenberg moment and everything will be forever transformed, whether with the current gen models or the next gen models, so everyone's positioning themselves for the coming transformation. What's a few hundred billion to burn when we're on the verge of the next big human revolution and there are trillions of dollars of opportunity at stake?
I feel like a comparison to the printing press is exactly the type of thing this author is railing against. The reason the printing press was revolutionary is because it allowed the dissemination...
Exemplary
I feel like a comparison to the printing press is exactly the type of thing this author is railing against.
The reason the printing press was revolutionary is because it allowed the dissemination of new ideas, something that LLMs are structurally incapable of doing. LLMs can produce text that is strikingly similar to human writing. That's it. It cannot understand the ramifications of what it is writing, it cannot intentionally elicit an emotion in a reader, it cannot empathize with a reader, and it cannot contain a lived experience of the world in which we all live. It generates text that is statistically consistent with text a human could write, that's it.
Also, printers in that day printed cheap screed because it made them money; the main thrust of this article is that these companies are losing billions to produce generally useless text. The analogy would be apt if the printing press could only print cheap trash.
If you're strictly using LLMs to produce text that resembles human writing, then it feels like you're only using them for 5% of their potential. They're capable of collating, organizing,...
Exemplary
If you're strictly using LLMs to produce text that resembles human writing, then it feels like you're only using them for 5% of their potential. They're capable of collating, organizing, proofreading, editing, translating, filtering, and processing inputs in hundreds of different formats. And they're getting more powerful every day.
Need to parse two giant PDFs and find common points between them? An LLM can do that in seconds. Convert your RTF document to Markdown? Sure thing. Review a translation for any errors? Of course. Have a word on the tip of your tongue? Describe it, the LLM will find the answer.
And that's just their tool use. They're often just as applicable to other domains. For instance, I've found they're very useful for self-motivated learners. If you're studying a new topic, try asking it to create a lesson plan, or to quiz you.
I recently spent 15 minutes with an LLM and came away with a stronger understanding of quaternions than I've ever had in my life. Being able to explain my understanding, then have a personalized explanation correct my misunderstandings was more impactful than even hours of videos or lectures would have been.
These are omni-tools on a level never seen before. They use natural language processing, making them accessible to almost anyone, and are surprisingly capable in ways we're still exploring. I implore you not to write them off as simple story generators. In my mind, that is their least interesting property.
At the same time, I caution you to not just understand what they can do, but what they can't. A token generator is a statistical machine, and should not be understood as a source of truth. This is particularly the case when asking questions about niche or recent events. I see far too many people making this mistake.
Less-common but just as important is understanding that context plays a major role in using LLMs effectively, and both your inputs and the LLM outputs will bias future responses. If you lead it towards a local minima, it will follow. Use tools like conversation branching, or just starting new chats frequently if you feel you've introduced too much bias.
AI tools are a bit like driving a car. They're extremely useful, but potentially dangerous if you don't know how to operate them. Read the manual, spend a few hours getting familiar with the controls, and stay vigilant when using one. If you follow the rules, you'll be far more productive with them than you would be without.
Or it's a Beanie Baby moment and the rest of us will be glad when the fad is over. This is about as transformative as spellcheck, and about as intelligent as Zoltar.
everyone knows this is a Gutenberg moment and everything will be forever transformed
Or it's a Beanie Baby moment and the rest of us will be glad when the fad is over.
This is about as transformative as spellcheck, and about as intelligent as Zoltar.
We’ve got the annoying difficulty where everyone’s got a different meaning in mind when they say “AI”, but you just need to look at the latest Nobel prize awards to see what non-LLM applications...
We’ve got the annoying difficulty where everyone’s got a different meaning in mind when they say “AI”, but you just need to look at the latest Nobel prize awards to see what non-LLM applications of the technology are achieving.
We could turn off all the chatbots tomorrow and the improvements in medicine, climate science, mathematics, etc. that stem from the same neural net advancements would still be utterly transformative. Arguably more so, because then we wouldn’t be fighting OpenAI for GPU availability!
Even on the LLM side, even if you hate the things, I think it’s hard to deny the impact. I actually agree with a lot of what @Wes said just above - these are the most capable data parsers ever created, and the genuine utility of that is not to be underestimated - but even at maximum pessimism I think it’s hard to deny that the way LLMs are shaping and will shape online discourse has a pretty big (and largely negative) impact. Transformation sadly doesn’t have to be good…
I read and write professionally. No, it isn't. LLMs generate 'good enough' text for a lot of use cases. But most of those use cases are situations where we don't really need text anyway. Sort of...
I read and write professionally. No, it isn't. LLMs generate 'good enough' text for a lot of use cases. But most of those use cases are situations where we don't really need text anyway.
Sort of like AI art: yes, I can ask an LLM to generate an image of a stained glass window containing a woman reading a book with a raven on her shoulder, holding a key in its mouth. But why? What does that image accomplish?
Similarly, I can ask an LLM to generate text for a newsletter. But why? So I can not care about the text? If I want to do a good job writing the newsletter, I then need to edit and fix all of the text.
LLMs can definitely help people hastily create PoC code, unit tests, crappy text, bloated prose, and a whole lot more. But I don't want to consume any of that. If you give a single flying fuck about what you're making and you put it together by hand, I'd rather consume that.
I know I'm being a bit of a grumpy Luddite here. But I also know that, decades after CGI took over cinema, I still prefer movies that use practical effects. Dark Crystal is a great example -- the recent show used some CGI, but grounded it in practical effects. Maybe someday we'll get there with LLMs. But I think it'll take about as long as it's taken autonomous vehicles or even good quality CGI (indistinguishable from practical effects).
And let's not kid ourselves that LLMs will automate away most human labor. To put it kindly: that's astronomically fucking stupid and shortsighted and I cannot respect anyone with that stance. Maybe once we develop real artificial intelligence with sentience, instead of fake can-sorta-pass-the-Turing-test LLMs. But at that point we also need to deal with human rights issues around UBI and guaranteed food, along with the Black Mirror ethical implications of goddamn machine slavery.
Professionally, I can't really speak to the other things you mentioned, but in regards to this: I can't disagree with you more here. A lot of software engineering is knowing when you make software...
LLMs can definitely help people hastily create PoC code, unit tests, crappy text, bloated prose, and a whole lot more. But I don't want to consume any of that. If you give a single flying fuck about what you're making and you put it together by hand, I'd rather consume that.
Professionally, I can't really speak to the other things you mentioned, but in regards to this: I can't disagree with you more here. A lot of software engineering is knowing when you make software that's maintainable in the long term, and when to make crappy PoCs that will only run a few times, and a lot of code is the latter. Getting something that works and demonstrates that what you want is achievable is a really useful skill, and being able to automate that is an incredibly powerful tool.
More importantly, it can represent the democratisation of computing. I watched a Steven Mould video the other day where he got hold of a device that could only be used by programming a microcontroller - a skill that he did not have at all. With the help of some coding AI, he was able to get it working just fine, and deliberate a handful of really cool ways the device could be used. He didn't need high-quality, artisanal code. He needed something that worked and made his computer do what he wanted. That is incredible! That has been the basic goal of HCI since Douglas Englebert sat down and gave the mother of all demos.
I agree that there are all sorts of ethical and social issues that need to be dealt with. But I don't get this argument that LLMs are not incredibly useful.
It's cool that Steven was able to use LLMs to help him program a microcontroller. If LLMs can help provide an additional 'learning path' to make different kinds of programming accessible to more...
It's cool that Steven was able to use LLMs to help him program a microcontroller. If LLMs can help provide an additional 'learning path' to make different kinds of programming accessible to more people, that's absolutely helpful.
What line of software engineering work are you in? In my experience, I need to make a crappy PoC maybe once or twice a year, tops. Maybe an LLM can help me do that twice as fast. But I also kind of enjoy building a project from scratch, and it helps me understand how all of the various pieces fit together. So using an LLM to abstract away most of that sounds like a hack that will likely slow down my learning experience.
As far as the "democratisation of computing" goes, I couldn't disagree more. Using a term like that suggests to me that you're a little deep in the kool-aid. You know what would actually 'democratise computing'? Opening up hardware platforms for people to run their own code! Regulating powerful tech monopolies like Apple and Google and Facebook who prevent people from hacking improvements into their workflows! Teaching kids how to program a computer in school instead of showing them how to customize the appearance of a Google Doc on their Chromebook! Adopting Linux and open source software in government so individuals don't get forced into running proprietary software for interop!
Sure, LLMs might democratize hacky, unmaintainable solutions to niche problems. But when I've tried to use LLMs to solve those kinds of problems, they generally fail because they don't have exact Stack Overflow solutions or documentation to cheat off of.
Thanks for articulating your opinion in a polite way, it's one of the things I really love here on Tildes. Hopefully I haven't been too disrespectful in my response :)
I work in web development, and building proofs of concept to show how a particular feature might work or to explore whether something is feasible is definitely part of that. It's not something I...
I work in web development, and building proofs of concept to show how a particular feature might work or to explore whether something is feasible is definitely part of that. It's not something I do every day, but it something that happens regularly.
That said, it's not something I'm good at: like you, I like doing things "properly", taking my time, and learning something in the process. I see this particularly in little side projects: I'll spend a couple of hours setting up type-checking and proper minification for a twenty line JS script for my blog. It's interesting to me, but beyond that it's rarely productive, especially as I'm going to write that script once and then almost certainly never change it for the rest of its life.
This sort of stuff seems to be one of the places that LLM coding systems shine: hacky solutions that don't need to be maintained, because they wouldn't be maintained in the first place. And that's a lot of code. I think those of us with careers in software development further just how much code can get way with being bad or mediocre code.
That's what I mean when I talk about democratisation. People cannot run their own code unless they can write their own code, and most people (a) would benefit from being able to write little bits of hacky code to link things in their computer, and (b) cannot write code. There have been a bunch of solutions to that (visual coding, YAML configuration files, etc), but the benefit of LLMs is that you can now essentially program in natural language: tell your computer exactly what you want it to do, and let it figure out how to make that happen.
FWIW, I don't use LLMs to program at all myself. I use them occasionally as a replacement for Google for things that I know Google won't be good at ("what word am I thinking of", etc), and for copy editing, and I find it very useful for these sorts of things. On the other hand, the rise of agentic programming has all happened while I've been on parental leave, so I've not really had much chance to try that out. I know some developers that I respect a lot have gotten a lot out of their LLM usage, but I know others who have found it more hindrance than help, so I've not really made up my mind on how useful it is for professional programming. (In other words: I certainly don't think I've drunk any kool-aid!)
But I do strongly believe that for lay computer users, natural-language has been an HCI holy grail for decades, and it finally feels like something realistic, as opposed to a feature of science fiction. That's not to say that everything's perfect yet, but stories like Steven Mould's really demonstrate what the potential of LLMs can be for non-developers.
Low-code solutions have been around forever, and they mostly work. Visual Basic debuted in 1991. The problem, just as with the LLM code, is once you needed to dig in and maintain or expand an...
Low-code solutions have been around forever, and they mostly work. Visual Basic debuted in 1991. The problem, just as with the LLM code, is once you needed to dig in and maintain or expand an existing thing, it falls apart rapidly and now you need a specialist to dive deep and reverse-engineer a codebase with 0 commments.
There has never been a real barrier to computing literacy any more than there is to reading literacy and is less difficult than learning a second language IMO. Coding can be taught at a first-grade reading level with Scratch, another low-code solution.
The only genuine barrier has been learned helplessness toward technology and a tolerance of it. In part because it was much more lucrative to sell solutions to a public that can't make their own.
Sure, but most of the time you don't need to maintain or expand your code, at least in the cases I'm talking about. In fact, I would guess that most code outside of professional projects is...
The problem, just as with the LLM code, is once you needed to dig in and maintain or expand an existing thing, it falls apart rapidly and now you need a specialist to dive deep and reverse-engineer a codebase with 0 commments.
Sure, but most of the time you don't need to maintain or expand your code, at least in the cases I'm talking about. In fact, I would guess that most code outside of professional projects is written once, and then either appended to or rewritten entirely if changes are needed. I'm thinking about research software, bash scripts to manage your local environment, IFTTT-style glue scripts, hell even my mechanical keyboard needs to be programmed to get the most out of it.
And you're right, low-code solutions have existed for a long time, but they tend to be very limited because they can only do what has been programmed in by the designer. The advantage of LLMs here is that their output is standard code, which means if it can be programmed, you can (in theory) get an LLM to write it. This is also why tools like Scratch aren't really the answer: they're too basic if you're someone who enjoys programming, and they're too limiting to be useful if you're not interested in programming and just want to get stuff done. And someone needs to have preprogrammed all the behaviours you're interested in: there's no Scratch module for programming your keyboard, for example.
I agree that in theory, anyone can learn how to program. Most people, however, don't want to. So the question becomes: how can someone who has no desire to program at all still be fully in control of their own computer? And not just in the "you own all the software (but you've got no idea how it works and you can't change it)" FOSS sense, but in the "you own a device that can truly do anything you think of, and you can come up with more things for it to do" sense. I think LLMs represent a serious step forwards in the regard.
Fundamentally, I think this is an HCI thing. Go back to the 1970s and '80s and show researchers back then a way of telling your computer to do stuff that used natural language, was iterative, and used your own workspace as its context. They'd truly have thought you'd solved personal computing. I think that's the most underappreciated part of LLMs, and I suspect once the hype dies down, it's going to be one of the parts that most changes how we use computers long-term.
I think the proof is in the pudding on this one. https://twitter.com/Thom_Wolf/status/1924399746447269963 I've spent 20 years learning to code well. If new technologies allow kids to do the same...
As far as the "democratisation of computing" goes, I couldn't disagree more.
I've seen just as much happening out of a weekly coding program for kids starting from the age of 7, run by a handful of teenagers in their spare time, without the AI involvement. At what point is...
I've seen just as much happening out of a weekly coding program for kids starting from the age of 7, run by a handful of teenagers in their spare time, without the AI involvement.
At what point is the technology no longer making a tool accessible, but becoming a crutch that renders the skill useless?
Despite all of our enhanced technology, reading comprehension is at the lowest point it's been in years, and dropping. All of the aid in the world can't compensate for the mental skills that need developed to learn to read.
It's like the extension of helicopter parenting: By removing all obstacles, nobody learns to do anything.
We're going to end up with a generation that doesn't make art because they found picking up a crayon too daunting.
Honestly, is it such a bad thing if programming does become useless? I say that as somebody that loves to write code, and has probably built some part of my own identity around it. But I recognize...
At what point is the technology no longer making a tool accessible, but becoming a crutch that renders the skill useless?
Honestly, is it such a bad thing if programming does become useless? I say that as somebody that loves to write code, and has probably built some part of my own identity around it. But I recognize that programming is a means to an end - the ability to make a computer perform a task for me - and that the methods and tools have always undergone change. The code I write today looks nothing like it did 20 years ago, or even 10 years ago.
I understand the temptation to argue that new learners need to do it "the hard way", just as we did, but that strikes me as fallacy. I've never had to write raw assembly, deal with pointers, or carefully manage my own memory. When I program today, I don't have to write my own quick sort algorithm because every standard library already includes one. Higher-level languages smooth out the implementation details so we can focus on the important stuff.
The way I see it, "vibe coding" - though I don't love the term - is a natural extension to that concept. If English becomes the hot new programming language, isn't that a good thing, in the end? It means that almost anyone can create applications, websites, tools and neat interactions. And if they enjoy that, it's a natural onramp to start learning more about the underlying technologies.
I would argue that we have some obligation to not pull the ladder up behind us. Vibe coding isn't killing reading comprehension, and it's not making kids afraid of crayons. It is helping to remove the friction that's traditionally made programming an unapproachable subject for most people.
Counterpoint: it's 100% guaranteed that by now you have read something that was either entirely or partially LLM-written or written with the assistance of an LLM and you hadn't realized it. It's...
Counterpoint: it's 100% guaranteed that by now you have read something that was either entirely or partially LLM-written or written with the assistance of an LLM and you hadn't realized it.
It's like what they say about cosmetic surgery: people only notice bad cosmetic surgery. Good cosmetic surgery is invisible, and it's quite common in affluent, urban areas. I know lots of people who've gotten work done.
Sort of like AI art: yes, I can ask an LLM to generate an image of a stained glass window containing a woman reading a book with a raven on her shoulder, holding a key in its mouth. But why? What does that image accomplish?
I know architects, landscape architects, and designers who are using image AI generators to assist them with design explorations and comparing design options in high fidelity. For them, it's like having an office of infinite interns at their disposal.
I know artists who are embracing AI and technology to push the boundaries of art in radical new directions and to new heights. Model and asset generators are enabling indie game dev teams to go faster and farther with smaller teams.
I know people working on tactical urban improvement tools that can help people visualize different urban interventions and infrastructure improvements in their neighborhoods. Many people can't think or communicate visually, so these tools help them understand and communicate their own ideas.
The bar for art, design, and creativity is so much higher now, and as an insider in the art and design worlds, I find the current generation of work much more interesting.
And let's not kid ourselves that LLMs will automate away most human labor. To put it kindly: that's astronomically fucking stupid and shortsighted and I cannot respect anyone with that stance. Maybe once we develop real artificial intelligence with sentience, instead of fake can-sorta-pass-the-Turing-test LLMs.
LLMs have already 100% passed the Turing test.
Two centuries ago, ~75% of Americans worked in agriculture. The Malthusians thought such was an inescapable human condition, that we were bound by the mouths we had to feed. Now, less than 2% work in agriculture, and we drown in food.
along with the Black Mirror ethical implications of goddamn machine slavery.
Anatomically modern humans have been around for 200–300,000 years. We have existed in hellish conditions for much of that, our numbers have been tiny because death beset us from all directions. Only in the past 200 years, through mechanization and automation, have we finally been freed.
All of our abundance and comfort we have, we owe to machines.
Of course. But I also read a lot of crap on the internet. I guarantee that I haven't ready anything written by an LLM that I actually thought was good. Fair point! Cosmetic surgery, CGI, LLMs: if...
Counterpoint: it's 100% guaranteed that by now you have read something that was either entirely or partially LLM-written or written with the assistance of an LLM and you hadn't realized it.
Of course. But I also read a lot of crap on the internet. I guarantee that I haven't ready anything written by an LLM that I actually thought was good.
It's like what they say about cosmetic surgery: people only notice bad cosmetic surgery. Good cosmetic surgery is invisible, and it's quite common in affluent, urban areas. I know lots of people who've gotten work done.
Fair point! Cosmetic surgery, CGI, LLMs: if they're used sparingly, they can be useful.
For them, it's like having an office of infinite interns at their disposal.
It feels like you didn't really engage with my point here: if the work is so unimportant that they'd previously toss it to an intern, and they can replace those interns (or just never hire interns at all) with an LLM, did they even need to do that work at all? I would argue that the point of this kind of low-stakes internship work is to give young, inexperienced people a chance to learn about a space without risking anything too important, so they can get a foothold in an industry and start working their way up. It's great that your architect friends can save money and effort on interns, but aren't you concerned that this is going to erode the talent pipeline? Was the point of that art actually HAVING the art, or was it just an excuse to create a talent pipeline, and we're essentially removing that Chesterton's Fence without understanding it?
I know people working on tactical urban improvement tools that can help people visualize different urban interventions and infrastructure improvements in their neighborhoods. Many people can't think or communicate visually, so these tools help them understand and communicate their own ideas.
Did you use AI to write this? What on earth does 'tactical urban improvement' or 'visualize different urban interventions' mean? Are you serious when you say that 'many people can't think or communicate visually'? Because anyone can draw a diagram on a chalkboard. You don't need AI for any of that. If you can't do it, hire an intern to do it who has some art skills. Your argument here seems to boil down to 'lets's stop paying anyone to ever do art again', which... seems like a pretty shitty thing to do to the art and graphics design industries?
LLMs have already 100% passed the Turing test.
I am aware. LLMs create sequences of likely text. But the Turing test was kind of a crappy test anyway, because it doesn't evaluate intelligence or consistency or ability, it just evaluates if something sounds convincingly human given certain inputs. Let me put it this way: the 'Turing test' of driving would ask if you could differentiate a machine-driven car from a human-driven car. But humans are terrible at driving! I don't want a machine-driven car that occasionally rolls coal at cyclists, or stops in the crosswalk. The Turing test tells you one thing: can I tell if this thing is not a human in a short text conversation? Don't get me wrong, it's very very cool that LLMs can do that. But it doesn't make them good at everything!
Anatomically modern humans have been around for 200–300,000 years. We have existed in hellish conditions for much of that, our numbers have been tiny because death beset us from all directions. Only in the past 200 years, through mechanization and automation, have we finally been freed.
All of our abundance and comfort we have, we owe to machines.
What are you even talking about? ASI and AGI require some degree of independent thought and introspection: that's sentience. My point was that if we actually achieve machine sentience, we have automatically achieved machine slavery. Fortunately LLMs are not sentient. But a lot of AI enthusiasts have been hyping 'AGI' and 'ASI' 'soon' for a couple of years now. If we achieve that (and I don't think there's a path from LLMs to that, personally, largely because ASI/AGI would require some degree of sentience, which LLMs lack), we need to immediately deal with the implications! It is not ethical to imprison millions or billions of sentient machines to do our busywork. Fortunately this isn't a today issue or likely even a this-decade or this-century issue. But it is something we need to discuss and regulate as a society before it happens, especially since machine intelligences might experience the passage of time differently from ourselves.
It's important labor-intensive work. Having an intern make six different versions of a front reception would take days, maybe weeks. Now you can make a hundred versions and iterations within...
It feels like you didn't really engage with my point here: if the work is so unimportant that they'd previously toss it to an intern, and they can replace those interns (or just never hire interns at all) with an LLM, did they even need to do that work at all? I would argue that the point of this kind of low-stakes internship work is to give young, inexperienced people a chance to learn about a space without risking anything too important, so they can get a foothold in an industry and start working their way up. It's great that your architect friends can save money and effort on interns, but aren't you concerned that this is going to erode the talent pipeline? Was the point of that art actually HAVING the art, or was it just an excuse to create a talent pipeline, and we're essentially removing that Chesterton's Fence without understanding it?
It's important labor-intensive work. Having an intern make six different versions of a front reception would take days, maybe weeks. Now you can make a hundred versions and iterations within minutes.
I'm not worried. Art and technology have always been intimately connected, and the advancement of technology reduces the labor of art and makes it more abundant and therefore accessible.
In the distant past, art was for the very elite: paintings and statues were extremely expensive and labor-intensive to produce. They were commissioned and produced by studios of artists and their apprentices. The masses had to go to church to see any artisanal art; the average person would have never laid their eyes on a real painting.
Now, we are awash in rich media in every style. But the boundary keeps getting pushed. I think young people will be fine and find their niche. There will be a future where, through automation, it'll be easier for architects to have solo practices. Architectural design services, so far reserved for the wealthy—tell someone that you had an architect design your home and they'll see you as a fancy pants rich person—will become commonplace.
What on earth does 'tactical urban improvement' or 'visualize different urban interventions' mean?
Are you serious when you say that 'many people can't think or communicate visually'? Because anyone can draw a diagram on a chalkboard. You don't need AI for any of that... seems like a pretty shitty thing to do to the art and graphics design industries?
I am serious. Why do you think architects spend so much time and money on high-fidelity renderings and building models (which can $100k's in labor for the fancy skyscraper ones) to communicate their vision to their clients? If one wants to convince anyone to spend real money on a capital improvement project, they have to convince stakeholders of their vision.
The whole point of design school is to develop advanced visual and spatial intelligence. So, relatively few people can do it at a level where they can sell their ideas.
But much like how Canva made basic graphic design more accessible to everyday people who'd previously have to pay several hundred dollars to a local graphic designer to make a simple restaurant menu or event flyer. People are now using image AIs to help them reimagine their home interior designs, outfits, yards, and so on.
People are using image AIs to revisualize neighborhood streets with bike infrastructure, more trees and benches, with different pavement material, and so on. High-fidelity images, not crude sketches, get stakeholders actually excited and talking.
There's an incredible explosion in creativity.
But the Turing test was kind of a crappy test anyway, because it doesn't evaluate intelligence or consistency or ability, it just evaluates if something sounds convincingly human given certain inputs.
LLMs have won medals in the International Math Olympiad. The vast majority of people probably couldn't even solve the easiest problem in the IMO.
What are you even talking about? ASI and AGI require some degree of independent thought and introspection: that's sentience.
Many humans exhibit a low degree of independent thought and introspection — the Nazi regime exploited this fault to induce many ordinary humans to do extraordinarily evil things. In Adolf Eichmann, Hannah Arendt saw not an evil monster but a very ordinary human who simply followed orders, went home to dinner with his family, and never questioned what he was told. From that she quipped that evil is banal.
Humans are also capable of reasoning in an unconscious state. People sleep talk to their partners or sleep walk and perform coherent, complex tasks (washing dishes, even driving a car).
The connection between intelligence and sentience or consciousness is much more tenuous than it seems.
I'm sorry, but: I feel deeply passionate about the subject of human-centric infrastructure, so I can't not respond to this. High-fidelity images are not the thing that holds back infrastructure...
I'm sorry, but:
People are using image AIs to revisualize neighborhood streets with bike infrastructure, more trees and benches, with different pavement material, and so on. High-fidelity images, not crude sketches, get stakeholders actually excited and talking.
There's an incredible explosion in creativity.
I feel deeply passionate about the subject of human-centric infrastructure, so I can't not respond to this. High-fidelity images are not the thing that holds back infrastructure improvements. NIMBY attitudes, stubbornness, and car-dependent lifestyles do that. Similarly, let's not pretend that automating away certain low-level tasks will 'democratize computing' or 'change the world' or anything so dramatic. All I've seen proof of is LLMs' ability to knock out the bottom rungs of the employment ladder. By ignoring that point, I can't help but wonder what your agenda really is. Is it really so hard to understand that a lot of us don't find LLMs that useful?
Hear me out: Canva isn't any more accessible than Microsoft Publisher from circa 1991. You could click through some wizards, change fonts and layouts, and add some clipart. All in about 10 clicks....
But much like how Canva made basic graphic design more accessible to everyday people who'd previously have to pay several hundred dollars to a local graphic designer to make a simple restaurant menu or event flyer.
Hear me out: Canva isn't any more accessible than Microsoft Publisher from circa 1991. You could click through some wizards, change fonts and layouts, and add some clipart. All in about 10 clicks.
Many humans exhibit a low degree of independent thought and introspection
We know the answer to this: Because of a lack of education. From Hitler's #2: "Education is dangerous - every educated person is a future enemy."
If you put in a 2nd graders effort into something, you should get a 2nd grader's result.
We shouldn't be focus on optimizing the results of 2nd graders, to make it look like a 4th grader's. We should be focused on how to better teach them to be 4th graders.
I guess this is somewhat off topic: Although I agree that the AI bubble is damaging in a lot of ways, there are other ways that AI is going to be even more negatively disruptive, and there's no...
I guess this is somewhat off topic: Although I agree that the AI bubble is damaging in a lot of ways, there are other ways that AI is going to be even more negatively disruptive, and there's no way for us to be sure how.
I've been thinking about how we will see increased use of AI for scams.
Currently there are a lot of "manual" scams which involve cold calling or spam emails that misrepresent a relationship or an opportunity. Phishing attacks and other ways to get you to click a link and give up secret info. Then there are other types of fraud and blackmail, like "I've got pictures of you from your webcam, so pay me money or I'm sending them to people you know". These types are going to become way worse soon as various chatbot technologies have conversations with people to talk them into giving up something valuable. So instead of getting a text message or a recording, you'll get a full interactive shakedown. Then people will start selling AI counter agents to these AI scammers, but the AI counter agents will also invade your privacy and waste your time and money.
I’d put that down to filter bubbles. If you spend a lot of time talking to VC investors and MBAs you get a very different idea of the prevailing opinion compared to talking to heavily online...
I’d put that down to filter bubbles. If you spend a lot of time talking to VC investors and MBAs you get a very different idea of the prevailing opinion compared to talking to heavily online techies, and if you talk to average end users you get a whole different view again. Possibly one expressed with a few too many em-dashes.
I sincerely appreciate Ed’s in-depth research and financial analysis in this piece. There’s a lot of effort that goes into producing something like this and that’s worth mentioning. I’m an AI-hater as much as the next guy here on Tildes, but I just get the sense that this article misses the forest for the trees.
Every executive, venture capitalist, C-suite officer, and financial analyst understands we are in a massive, massive bubble right now. They all agree that this bubble will pop, many people will lose their jobs, and some would probably even agree with you if you said that some of these MAG7 companies will go under. They’re not betting on the growth from AI saving their capex, they’re dumping capex so they can be the last man standing when the bloodbath is over.
One has to recognize these people learned their lessons in the dot-com bubble, and are applying the same lessons here. Yes, companies dumped exorbitant sums on (in hindsight) very useless investments and some went bankrupt because of it. Yet the winners of that bubble were not the ones who played it safe and let the new trend bandwagon pass them by. The winners were the ones who dumped the capex, sometimes ran negative for years, and then came out on top: Amazon, Google, Microsoft, etc.
Every investor believes that only a few of these AI investments will pay off, and just aren’t sure which ones will (or they are all-in on one or two horses). But they fully believe that when all is said and done, the bankruptcies and firesale acquisitions over, the costs falling down, then the new kings of the global economy will have been crowned.
Look, I cannot say if I agree with their analysis. But when I go interact with average folk in their day-to-day life, they seem to love these glorified word predictors. They use them to generate emails, to search for recipes, and sometimes to console them through a tough break-up. I spend time in classrooms and students can’t get enough of “Chat”. I spend time with family and they talk about their new use of AI so that they don’t have to make that stupid presentation at work. I see this tech and I see it being popular, at least in the right way.
Fundamentally, that’s why I hate this tech so much. If Ed Zitron’s market collapse prophecy comes to pass, my gratitude might make me a religious man. The gamble paying off is so, so, so, so, so much worse. Imagine the next Google not just having access to your search history and watch history, but a personalized model that they can use to convince you of things. Imagine a product that allows you to deploy subtle biases into millions of personalized chatbots that consumers are using, changing political opinions and market demands at scale. That’s what the VCs see, and they are damn right to say it’s worth risking bankruptcy for. I just pray that vision is a mirage.
Without even making a judgement about how well these products actually work, there's still an issue here.
I'm sure I would love flying around in a private business jet wherever I want a few times a month. I can't afford to do that though, and neither can almost anyone else, which is why the charter jet industry is a $25 billion a year industry, not a $5 trillion a year industry. If business jets were hyped beyond any semblance of reality such that venture capitalists, private equity, fund managers, and everyone else subsidized flights so that every single person in the country was taking private jet trips every week, you could definitely make it look like a $5 trillion dollar industry though.
What we're doing with AI isn't sustainable. People use it because generating a funny picture or writing somewhat convincing, but bland, inaccurate text can be fun when you're doing it for free or for 10 bucks a month or whatever.
If people had to pay the true cost of it though? How many boomers would really be willing to shell out $50 bucks a month to churn out pictures of Jesus made of fruit? Probably not many.
There's basically no one I could see being willing to pay this except people who directly use it to help with their jobs, like developers and copywriters. That market is way, way, wayyyy smaller than what the investment in this companies is signaling though. In order for the hype to be justified, we'd have to be living in a completely different reality.
I think this is going to depend a lot on how much that $50 ends up being. Bear in mind that there's a lot of range to these models, from ones that you can run on your laptop for free, to ones that are tremendously expensive. I suspect the "pictures of Jesus made of fruit" market will be pretty satisfied with cheap models that can be run as a kind of loss leader. And for professionals, $50 a month isn't that much, assuming what you get out of that is genuinely useful.
The bigger question is how much training costs are going to change. The impression I get is that the current prices are fairly sustainable in the sense that, say, Anthropic's $20/mo tier largely covers the cost of running the LLM queries being made. But right now, everyone needs to be spending a lot on research and training for the next model, otherwise they'll be left behind, and that's what's unsustainable. Presumably, at some point, the improvements between generations will not be worth the cost of that research, at which point it will be possible to concentrate only on selling access to a pre-trained model, which is right now the sustainable part of the business.
I agree that we're in a bubble right now, and a lot of AI stuff is being sold just because AI sounds cool, but hype cycles are nothing new. What's interesting this time is that, even outside of the hype cycle, the underlying technology still seems very useful.
$50 if everyone paid. If only professionals did, $1000+.
Actual cost for running a top-tier model comes in around $0.03 per API request, and that’s running it personally (i.e. accounting for overhead as the container spins up, rather than efficiently batching a constant stream of input from multiple users), you can run less capable models for less than 1/10th of that.
The sunk costs into OpenAI et al might be tough to recover, but third-party-hosted Llama and DeepSeek are already viable at reasonable cost.
Are they? When self-hosting Claude 3.7 Sonnet on AWS for use with Cursor, it's extremely easy to run up a $100 bill every day. Sure, we can use worse models, but the cutting edge models are already barely acceptable.
The most popular use-cases are also the most expensive, and that seems like a problem.
I don’t use coding models (or LLMs in general, really; I just work on the tech they’re built on top of) so I don’t have a great benchmark for what’s reasonable in that workflow - but unless I’m missing something it looks like $100 would be multiple millions of tokens even with AWS markup. I’m not saying it’s impossible, but that’s the entirety of War and Peace four or five times over in a single day, which seems like a lot even accounting for the complexities of code generation?
Chat models are easier to eyeball because the input and output is pretty much capped to what an individual can read or write in a reasonable time, so that’s what I had in mind for extrapolating up to reasonable daily use. I’d be surprised if non-code users are regularly generating more than 10k tokens/day (roughly 35 printed A4 pages of text).
[Edit] As a side note, there’s also an upper limit on reasonable cost set by the option of buying dedicated hardware.
You can fit the vast majority of models on a 512GB Mac Studio, which sells for $10k. It may not even be the best choice in a lot of cases, but it gives a good ballpark on what “reasonable” looks like: if a company can support four employees’ Cursor sessions per machine and keeps it for three years, that’s $70/employee/month for effectively unlimited use.
Yes, it's code specific, but that's not an unreasonable sum for an agentic tool like Cline with a big MCP server. Querying multiple repos or project management systems with large projects adds up real quick.
Claude 3.7 Sonnet is also like 6x more expensive than Deepseek R1 0528, due to licensing costs.
Right, that's the bit I disagree about. As I understand it, the API pricing for most of these services roughly corresponds to the real cost of serving those requests, and the per-month subscriptions are similarly profitable as long as you put in sensible rate limits. And currently those prices are fairly reasonable right now.
Like I said, I think the biggest expense right now for these companies is the research and training needed to stay ahead of the curve and bring out their next generation of LLM. And I suspect over time, that expense will go down proportionally as these companies begin to max out what LLMs are reasonably capable of.
Three big thoughts:
Microsoft was already an entrenced monopoly. Google was the seach engine and was hurtling towards profitability at record pace. Amazon survived by the skin of its teeth, because it had just secured funding before the bubble crashed. If that deal took 3 more months, Amazon would have been as dead as Pets.com, the most infamous. Pets burned $12 million in advertising to attain a peak revenue of less than a million. Which was also sold at a loss.
The stuff that survived had nothing to do with capex, and everything to do with actually having a viable path to profitability.
I think that's where the Netflix and Spotify comparison comes in. The two largest streaming services have a combined revenue of $67 billion. Do people love these AI chatbots so much that they'll be willing to pay $100 or more a month when the pied piper comes to collect on all that capex expenditure? I think it's a fun novelty to most people that will happily abandon it when faced with the real cost of a sustainable product.
And because all of those people are literally burning everyone's retirement funds to do so, I hope they all rot in jail when they are.
I think we should distinguish between money spent on data centers and GPU's and electricity (they are real costs) and the money people think they have due to rising stock prices. The stock market went up quite a bit since 2008. If it goes down again, that means long-term investors are less wealthy than they thought (unless they got out) but they might not have lost money overall.
If you didn't buy during the bubble, it's more like, you thought you were richer for a while, but it was an illusion. It never really existed except in speculator's minds.
Meme stocks and cryptocurrency are purer versions of this, but it's somewhat true of all financial assets in a bubble.
Generally I agree with you.
I think Google is the best comparison here. Google was the new king post dot-com. Google won because it became the search engine and became an extraordinarily high-margin business.
I think most investors want the product to stay free. Let me speak from personal experience. I have had clients walk into the business that I work at, effectively ready to sign a contract without even talking to us. They had asked ChatGPT for a set of services and ChatGPT gave an absolutely glowing review of us. Why? I don’t know. I guess our website got lucky.
People trust the chatbots, especially people that lack tech savvy and don’t understand how the models actually function. Its effectiveness as a convincer is so much better than any modern search engine.
I think the product offering will look a lot more like “pay OpenAI to run some fine-tuning on the latest model so that it recommends our business more often” to users. Subscription services suffer because users aren’t the cash cow, businesses are. In that scenario, whoever has the userbase will win, just like Google did.
I would distinguish between business and home use. At work, that's not an unreasonable price. If you compare $100/month to the monthly salary of an employee who uses the tool, the tool is cheap in comparison - if it's useful, which it definitely is for some developers. (Example.)
For consumers, I expect that many will pay $20/month if they're getting a lot out of it, but costs will go down to the point where it's often free and advertiser-supported. Google's AI search results are an example. (A lot of people hate them because they're being pushed on us, but I see no reason why they won't improve.)
I haven't seen much research actually supporting the supposed productivity increases. It makes devs feel more productive, but all the research I've seen says it actually slows down competent devs.
Yes, it's anecdotal and likely varies a lot.
There is one study that's been shared around showing that some developers were slowed down by it, which seems plausible, but it also might have been a skill issue that goes away with more experience? There is definitely a learning curve.
My work is paying for ChatGPT and cursor, I want to replace the latter with Claude code starting Monday too. And I think if I reach the limit of the 20€ a month plan, the 100€ a month Claude plan will be a no brainer too.
I am very weary of vibe coding, but if you decide what you want and let the AI do the typing for you and then double check it (seeing it as your junior dev, rather than just accepting it as gospel) the speed up can be immense.
My ADHD brain just has to not get distracted every time I wait for the LLM outputs
I was under the impression this was the explicit interest of incorporating Google AI into search. But it reminds me of this Parks and Rec bit
https://youtu.be/K3YNmoV9NsM?t=46
I’m sure I’ve quoted this in relation to online advertising before, but it just keeps getting more apt:
Sure, I know the propaganda at scale is working, I can see the impact and it’s honestly scary. But it’s also still a daily occurrence to see companies running astonishing amounts of compute across absurdly granular and invasive data, only to show me ads for a brand that a simple geoIP lookup would’ve told them doesn’t even exist where I am.
I suspect that's coming now,.later, or is already here. Social media content presentation algorithms are already subtly biased and pushing people in directions.
The question isn't whether this next phase shall come to pass. The question isn't whether a select handful of businesses will become massively rich.
The question is whether people and the entities through which we come together to protect ourselves - governments - will have the will power to do absolutely anything at all about it
The answer is that we will not have the capacity or the power.
Which ones?
Tesla seems vulnerable if they stop selling as many cars. The car industry is very competitive and they're less popular than they were. That's not really an AI thing, though.
Nvidia would see a large drop in revenues if the other big tech firms cut back.
This could be good for the profitability of the other firms because it means they stopped spending like mad? On the other hand, their previous capital expenditures go on the balance sheet and don't immediately affect profits. They will be a drag on profits in future years due to depreciation. (Alternatively, it would be a one-time loss if the investment were written off, as companies do sometimes.)
If AI revenue is low for the other big tech companies, that would make them more resilient if the bubble bursts. For example, Google's total revenue is $96 billion (annualized), so losing ~8 billion in AI revenue wouldn't hurt them all that much, and this suggests that they will easily survive.
The stock market might still tank though, because they are pricing in expectations of future AI revenue that would evaporate if the bubble bursts.
Are you sure they aren’t using chat in reference to streamer culture? IE: Chat am I cooked, Chat what do I do?
Some of my friends kids say this stuff and it’s not a reference to ChatGPT at all.
I’m talking generally about college kids and the people using “Chat” seem to be anthropomorphizing ChatGPT since in the context of the sentence, “Chat” could be replaced with “Google”. “Let me ask Chat.” “Chat told me XYZ.” I’ve heard that use of “Chat” too but honestly I feel like the two uses of “Chat” come from pretty separate communities.
Honestly if I heard that, I'd assume it means they're asking a group chat rather than chatgpt
I have not heard any college students refer to chatGPT as "Chat" it always means "present company" referencing like a twitch/live stream chat, but I'm training some RAs soon and I'll ask them because it's fun to make them cringe at old people using their slang.
Ed Zitron consolidates a whole bunch of economic data surrounding AI, and the results are terrifying. The key snip to remember through the whole thing:
And he goes on to explain, in immense detail, just how bad those AI stories are.
I read the whole thing, it's well written and breaks down several things that I've been wondering about but haven't been able to articulate to people.
The last quote in particular hits on a thing I'd been feeling strongly the last year or so. While AI has some utility and is interesting, the people around me hyping it up all give off a similar vibe. That vibe is one separate from the usual used-car-salesman ick that accompanies people blindly buying into a trend. It's a discomforting sense that the people who are heavy AI users are commonly bluffing their way through everything. That they're operating solely on AI summaries of everything in their life, and don't actually understand any of it at anything beyond a surface level.
I didn't think I'd read the entire thing, out loud, but it's just too compelling. Ed talks about every single thing I believed and wasn't smart enough to articulate. My feelings on the technology haven't changed, nor are they relevant here - but he's right. This particularly cancerous growth of the tech economy is rotten to the core, and as a rotten tree falls it will still shred its neighbours and cause untold damage.
Love to see Ed’s articles. Could someone explain once the data centers and housing are bought, will costs dramatically go down for AI companies? Or will this be an ever accelerating arms race? This is the one explanation for why burning through this much capex is possible if it’s just to build infrastructure.
I can't fully answer your question, or answer it in a detailed enough manner, but data centres are pretty expensive to run/maintain/operate (especially something of such a massive scale) and that equipment will be obsolete in like 5 years and need to be replaced by newer technology anyway, so there'll still be a lot of CapEx left. So IMO the total costs will somewhat go down a little bit but there'll still be plenty of OpEx and CapEx.
The real trouble isn't the infra costs in total IMO. It's the fact that every single LLM -- even the cheapest ones -- are absurdly expensive to operate, even on existing infrastructure. A subscription you use with any real frequency ought to cost you more than your Internet connection or electric bill. Right now tech companies are giving away the functionality for free to get people hooked, just like carshare apps or delivery food did in the first couple of years. But just like those (now much more expensive) products, eventually LLMs will need to monetize. Injecting ads and product placement will make the product worse, possibly useless and untrustworthy. Charging people hundreds of dollars of compute per month is too expensive even if it is good. Once the free money dries up these products are dead.
That's a really sweeping generalization that can't possibly be true. I can run LLMs on my phone for free. I can run them on my $500 Mac Mini for free. The providers that are selling access to baremetal for running LLMs on are almost certainly selling their services for a profit and you can see that it's not expensive. LLMs range too widely to make a claim that "every single LLM are absurdly expensive to operate". There's tons of useful models out there that run on cheap hardware.
What are these costs though? How much does it really cost to run an llm once the model is made and the servers are running? Does anyone have actual figures for this?
Electricity, GPU, etc adds up. Maybe they'll get more efficient eventually, but every new model consumes an order of magnitude more power so far. Most people want the latest models, not the cheap stuff.
But more importantly, this very article makes a separate point: NVIDIA's stock price is predicated on everyone else increasing GPU spend every quarter. If models get way more efficient, NVIDIA's numbers get worse... which ought to impact the stock market a lot since NVIDIA accounts for a massive chunk of the market!
This website might help?
https://simonwillison.net/2025/May/7/llm-prices/
For open source LLM’s, the model is free so whatever they’re charging to run it just includes inference costs and their markup.
Every new technology has its haters, even the printing press had its own. A Benedictine monk and scribe once wrote: "the pen is a virgin, but the printing press is a whore." Others thought it debased, not democratized, knowledge, as printers rushed to print out cheap crap, like love poetry or bibles that the masses could read without the help of some priest. Some thought that printers were out to make a quick buck, not do good for society.
Anyway, the printing press led to the Reformation and many wars and completely changed society and the world — eventually for the better in the very long run, most would agree, although it caused much bloodshed and pain for a few centuries.
I think everyone knows the current generation of AI is limited, but everyone knows this is a Gutenberg moment and everything will be forever transformed, whether with the current gen models or the next gen models, so everyone's positioning themselves for the coming transformation. What's a few hundred billion to burn when we're on the verge of the next big human revolution and there are trillions of dollars of opportunity at stake?
I feel like a comparison to the printing press is exactly the type of thing this author is railing against.
The reason the printing press was revolutionary is because it allowed the dissemination of new ideas, something that LLMs are structurally incapable of doing. LLMs can produce text that is strikingly similar to human writing. That's it. It cannot understand the ramifications of what it is writing, it cannot intentionally elicit an emotion in a reader, it cannot empathize with a reader, and it cannot contain a lived experience of the world in which we all live. It generates text that is statistically consistent with text a human could write, that's it.
Also, printers in that day printed cheap screed because it made them money; the main thrust of this article is that these companies are losing billions to produce generally useless text. The analogy would be apt if the printing press could only print cheap trash.
If you're strictly using LLMs to produce text that resembles human writing, then it feels like you're only using them for 5% of their potential. They're capable of collating, organizing, proofreading, editing, translating, filtering, and processing inputs in hundreds of different formats. And they're getting more powerful every day.
Need to parse two giant PDFs and find common points between them? An LLM can do that in seconds. Convert your RTF document to Markdown? Sure thing. Review a translation for any errors? Of course. Have a word on the tip of your tongue? Describe it, the LLM will find the answer.
And that's just their tool use. They're often just as applicable to other domains. For instance, I've found they're very useful for self-motivated learners. If you're studying a new topic, try asking it to create a lesson plan, or to quiz you.
I recently spent 15 minutes with an LLM and came away with a stronger understanding of quaternions than I've ever had in my life. Being able to explain my understanding, then have a personalized explanation correct my misunderstandings was more impactful than even hours of videos or lectures would have been.
These are omni-tools on a level never seen before. They use natural language processing, making them accessible to almost anyone, and are surprisingly capable in ways we're still exploring. I implore you not to write them off as simple story generators. In my mind, that is their least interesting property.
At the same time, I caution you to not just understand what they can do, but what they can't. A token generator is a statistical machine, and should not be understood as a source of truth. This is particularly the case when asking questions about niche or recent events. I see far too many people making this mistake.
Less-common but just as important is understanding that context plays a major role in using LLMs effectively, and both your inputs and the LLM outputs will bias future responses. If you lead it towards a local minima, it will follow. Use tools like conversation branching, or just starting new chats frequently if you feel you've introduced too much bias.
AI tools are a bit like driving a car. They're extremely useful, but potentially dangerous if you don't know how to operate them. Read the manual, spend a few hours getting familiar with the controls, and stay vigilant when using one. If you follow the rules, you'll be far more productive with them than you would be without.
Or it's a Beanie Baby moment and the rest of us will be glad when the fad is over.
This is about as transformative as spellcheck, and about as intelligent as Zoltar.
We’ve got the annoying difficulty where everyone’s got a different meaning in mind when they say “AI”, but you just need to look at the latest Nobel prize awards to see what non-LLM applications of the technology are achieving.
We could turn off all the chatbots tomorrow and the improvements in medicine, climate science, mathematics, etc. that stem from the same neural net advancements would still be utterly transformative. Arguably more so, because then we wouldn’t be fighting OpenAI for GPU availability!
Even on the LLM side, even if you hate the things, I think it’s hard to deny the impact. I actually agree with a lot of what @Wes said just above - these are the most capable data parsers ever created, and the genuine utility of that is not to be underestimated - but even at maximum pessimism I think it’s hard to deny that the way LLMs are shaping and will shape online discourse has a pretty big (and largely negative) impact. Transformation sadly doesn’t have to be good…
I read and write professionally. No, it isn't. LLMs generate 'good enough' text for a lot of use cases. But most of those use cases are situations where we don't really need text anyway.
Sort of like AI art: yes, I can ask an LLM to generate an image of a stained glass window containing a woman reading a book with a raven on her shoulder, holding a key in its mouth. But why? What does that image accomplish?
Similarly, I can ask an LLM to generate text for a newsletter. But why? So I can not care about the text? If I want to do a good job writing the newsletter, I then need to edit and fix all of the text.
LLMs can definitely help people hastily create PoC code, unit tests, crappy text, bloated prose, and a whole lot more. But I don't want to consume any of that. If you give a single flying fuck about what you're making and you put it together by hand, I'd rather consume that.
I know I'm being a bit of a grumpy Luddite here. But I also know that, decades after CGI took over cinema, I still prefer movies that use practical effects. Dark Crystal is a great example -- the recent show used some CGI, but grounded it in practical effects. Maybe someday we'll get there with LLMs. But I think it'll take about as long as it's taken autonomous vehicles or even good quality CGI (indistinguishable from practical effects).
And let's not kid ourselves that LLMs will automate away most human labor. To put it kindly: that's astronomically fucking stupid and shortsighted and I cannot respect anyone with that stance. Maybe once we develop real artificial intelligence with sentience, instead of fake can-sorta-pass-the-Turing-test LLMs. But at that point we also need to deal with human rights issues around UBI and guaranteed food, along with the Black Mirror ethical implications of goddamn machine slavery.
Professionally, I can't really speak to the other things you mentioned, but in regards to this: I can't disagree with you more here. A lot of software engineering is knowing when you make software that's maintainable in the long term, and when to make crappy PoCs that will only run a few times, and a lot of code is the latter. Getting something that works and demonstrates that what you want is achievable is a really useful skill, and being able to automate that is an incredibly powerful tool.
More importantly, it can represent the democratisation of computing. I watched a Steven Mould video the other day where he got hold of a device that could only be used by programming a microcontroller - a skill that he did not have at all. With the help of some coding AI, he was able to get it working just fine, and deliberate a handful of really cool ways the device could be used. He didn't need high-quality, artisanal code. He needed something that worked and made his computer do what he wanted. That is incredible! That has been the basic goal of HCI since Douglas Englebert sat down and gave the mother of all demos.
I agree that there are all sorts of ethical and social issues that need to be dealt with. But I don't get this argument that LLMs are not incredibly useful.
It's cool that Steven was able to use LLMs to help him program a microcontroller. If LLMs can help provide an additional 'learning path' to make different kinds of programming accessible to more people, that's absolutely helpful.
What line of software engineering work are you in? In my experience, I need to make a crappy PoC maybe once or twice a year, tops. Maybe an LLM can help me do that twice as fast. But I also kind of enjoy building a project from scratch, and it helps me understand how all of the various pieces fit together. So using an LLM to abstract away most of that sounds like a hack that will likely slow down my learning experience.
As far as the "democratisation of computing" goes, I couldn't disagree more. Using a term like that suggests to me that you're a little deep in the kool-aid. You know what would actually 'democratise computing'? Opening up hardware platforms for people to run their own code! Regulating powerful tech monopolies like Apple and Google and Facebook who prevent people from hacking improvements into their workflows! Teaching kids how to program a computer in school instead of showing them how to customize the appearance of a Google Doc on their Chromebook! Adopting Linux and open source software in government so individuals don't get forced into running proprietary software for interop!
Sure, LLMs might democratize hacky, unmaintainable solutions to niche problems. But when I've tried to use LLMs to solve those kinds of problems, they generally fail because they don't have exact Stack Overflow solutions or documentation to cheat off of.
Thanks for articulating your opinion in a polite way, it's one of the things I really love here on Tildes. Hopefully I haven't been too disrespectful in my response :)
I work in web development, and building proofs of concept to show how a particular feature might work or to explore whether something is feasible is definitely part of that. It's not something I do every day, but it something that happens regularly.
That said, it's not something I'm good at: like you, I like doing things "properly", taking my time, and learning something in the process. I see this particularly in little side projects: I'll spend a couple of hours setting up type-checking and proper minification for a twenty line JS script for my blog. It's interesting to me, but beyond that it's rarely productive, especially as I'm going to write that script once and then almost certainly never change it for the rest of its life.
This sort of stuff seems to be one of the places that LLM coding systems shine: hacky solutions that don't need to be maintained, because they wouldn't be maintained in the first place. And that's a lot of code. I think those of us with careers in software development further just how much code can get way with being bad or mediocre code.
That's what I mean when I talk about democratisation. People cannot run their own code unless they can write their own code, and most people (a) would benefit from being able to write little bits of hacky code to link things in their computer, and (b) cannot write code. There have been a bunch of solutions to that (visual coding, YAML configuration files, etc), but the benefit of LLMs is that you can now essentially program in natural language: tell your computer exactly what you want it to do, and let it figure out how to make that happen.
FWIW, I don't use LLMs to program at all myself. I use them occasionally as a replacement for Google for things that I know Google won't be good at ("what word am I thinking of", etc), and for copy editing, and I find it very useful for these sorts of things. On the other hand, the rise of agentic programming has all happened while I've been on parental leave, so I've not really had much chance to try that out. I know some developers that I respect a lot have gotten a lot out of their LLM usage, but I know others who have found it more hindrance than help, so I've not really made up my mind on how useful it is for professional programming. (In other words: I certainly don't think I've drunk any kool-aid!)
But I do strongly believe that for lay computer users, natural-language has been an HCI holy grail for decades, and it finally feels like something realistic, as opposed to a feature of science fiction. That's not to say that everything's perfect yet, but stories like Steven Mould's really demonstrate what the potential of LLMs can be for non-developers.
Low-code solutions have been around forever, and they mostly work. Visual Basic debuted in 1991. The problem, just as with the LLM code, is once you needed to dig in and maintain or expand an existing thing, it falls apart rapidly and now you need a specialist to dive deep and reverse-engineer a codebase with 0 commments.
There has never been a real barrier to computing literacy any more than there is to reading literacy and is less difficult than learning a second language IMO. Coding can be taught at a first-grade reading level with Scratch, another low-code solution.
The only genuine barrier has been learned helplessness toward technology and a tolerance of it. In part because it was much more lucrative to sell solutions to a public that can't make their own.
Sure, but most of the time you don't need to maintain or expand your code, at least in the cases I'm talking about. In fact, I would guess that most code outside of professional projects is written once, and then either appended to or rewritten entirely if changes are needed. I'm thinking about research software, bash scripts to manage your local environment, IFTTT-style glue scripts, hell even my mechanical keyboard needs to be programmed to get the most out of it.
And you're right, low-code solutions have existed for a long time, but they tend to be very limited because they can only do what has been programmed in by the designer. The advantage of LLMs here is that their output is standard code, which means if it can be programmed, you can (in theory) get an LLM to write it. This is also why tools like Scratch aren't really the answer: they're too basic if you're someone who enjoys programming, and they're too limiting to be useful if you're not interested in programming and just want to get stuff done. And someone needs to have preprogrammed all the behaviours you're interested in: there's no Scratch module for programming your keyboard, for example.
I agree that in theory, anyone can learn how to program. Most people, however, don't want to. So the question becomes: how can someone who has no desire to program at all still be fully in control of their own computer? And not just in the "you own all the software (but you've got no idea how it works and you can't change it)" FOSS sense, but in the "you own a device that can truly do anything you think of, and you can come up with more things for it to do" sense. I think LLMs represent a serious step forwards in the regard.
Fundamentally, I think this is an HCI thing. Go back to the 1970s and '80s and show researchers back then a way of telling your computer to do stuff that used natural language, was iterative, and used your own workspace as its context. They'd truly have thought you'd solved personal computing. I think that's the most underappreciated part of LLMs, and I suspect once the hype dies down, it's going to be one of the parts that most changes how we use computers long-term.
I think the proof is in the pudding on this one.
https://twitter.com/Thom_Wolf/status/1924399746447269963
I've spent 20 years learning to code well. If new technologies allow kids to do the same by using natural language, I am all for it.
I've seen just as much happening out of a weekly coding program for kids starting from the age of 7, run by a handful of teenagers in their spare time, without the AI involvement.
At what point is the technology no longer making a tool accessible, but becoming a crutch that renders the skill useless?
Despite all of our enhanced technology, reading comprehension is at the lowest point it's been in years, and dropping. All of the aid in the world can't compensate for the mental skills that need developed to learn to read.
It's like the extension of helicopter parenting: By removing all obstacles, nobody learns to do anything.
We're going to end up with a generation that doesn't make art because they found picking up a crayon too daunting.
Honestly, is it such a bad thing if programming does become useless? I say that as somebody that loves to write code, and has probably built some part of my own identity around it. But I recognize that programming is a means to an end - the ability to make a computer perform a task for me - and that the methods and tools have always undergone change. The code I write today looks nothing like it did 20 years ago, or even 10 years ago.
I understand the temptation to argue that new learners need to do it "the hard way", just as we did, but that strikes me as fallacy. I've never had to write raw assembly, deal with pointers, or carefully manage my own memory. When I program today, I don't have to write my own quick sort algorithm because every standard library already includes one. Higher-level languages smooth out the implementation details so we can focus on the important stuff.
The way I see it, "vibe coding" - though I don't love the term - is a natural extension to that concept. If English becomes the hot new programming language, isn't that a good thing, in the end? It means that almost anyone can create applications, websites, tools and neat interactions. And if they enjoy that, it's a natural onramp to start learning more about the underlying technologies.
I would argue that we have some obligation to not pull the ladder up behind us. Vibe coding isn't killing reading comprehension, and it's not making kids afraid of crayons. It is helping to remove the friction that's traditionally made programming an unapproachable subject for most people.
This is often my argument: LLMs are only useful at producing text that ultimately doesn't matter, so what's the point?
Counterpoint: it's 100% guaranteed that by now you have read something that was either entirely or partially LLM-written or written with the assistance of an LLM and you hadn't realized it.
It's like what they say about cosmetic surgery: people only notice bad cosmetic surgery. Good cosmetic surgery is invisible, and it's quite common in affluent, urban areas. I know lots of people who've gotten work done.
I know architects, landscape architects, and designers who are using image AI generators to assist them with design explorations and comparing design options in high fidelity. For them, it's like having an office of infinite interns at their disposal.
I know artists who are embracing AI and technology to push the boundaries of art in radical new directions and to new heights. Model and asset generators are enabling indie game dev teams to go faster and farther with smaller teams.
I know people working on tactical urban improvement tools that can help people visualize different urban interventions and infrastructure improvements in their neighborhoods. Many people can't think or communicate visually, so these tools help them understand and communicate their own ideas.
The bar for art, design, and creativity is so much higher now, and as an insider in the art and design worlds, I find the current generation of work much more interesting.
LLMs have already 100% passed the Turing test.
Two centuries ago, ~75% of Americans worked in agriculture. The Malthusians thought such was an inescapable human condition, that we were bound by the mouths we had to feed. Now, less than 2% work in agriculture, and we drown in food.
Anatomically modern humans have been around for 200–300,000 years. We have existed in hellish conditions for much of that, our numbers have been tiny because death beset us from all directions. Only in the past 200 years, through mechanization and automation, have we finally been freed.
All of our abundance and comfort we have, we owe to machines.
Of course. But I also read a lot of crap on the internet. I guarantee that I haven't ready anything written by an LLM that I actually thought was good.
Fair point! Cosmetic surgery, CGI, LLMs: if they're used sparingly, they can be useful.
It feels like you didn't really engage with my point here: if the work is so unimportant that they'd previously toss it to an intern, and they can replace those interns (or just never hire interns at all) with an LLM, did they even need to do that work at all? I would argue that the point of this kind of low-stakes internship work is to give young, inexperienced people a chance to learn about a space without risking anything too important, so they can get a foothold in an industry and start working their way up. It's great that your architect friends can save money and effort on interns, but aren't you concerned that this is going to erode the talent pipeline? Was the point of that art actually HAVING the art, or was it just an excuse to create a talent pipeline, and we're essentially removing that Chesterton's Fence without understanding it?
Did you use AI to write this? What on earth does 'tactical urban improvement' or 'visualize different urban interventions' mean? Are you serious when you say that 'many people can't think or communicate visually'? Because anyone can draw a diagram on a chalkboard. You don't need AI for any of that. If you can't do it, hire an intern to do it who has some art skills. Your argument here seems to boil down to 'lets's stop paying anyone to ever do art again', which... seems like a pretty shitty thing to do to the art and graphics design industries?
I am aware. LLMs create sequences of likely text. But the Turing test was kind of a crappy test anyway, because it doesn't evaluate intelligence or consistency or ability, it just evaluates if something sounds convincingly human given certain inputs. Let me put it this way: the 'Turing test' of driving would ask if you could differentiate a machine-driven car from a human-driven car. But humans are terrible at driving! I don't want a machine-driven car that occasionally rolls coal at cyclists, or stops in the crosswalk. The Turing test tells you one thing: can I tell if this thing is not a human in a short text conversation? Don't get me wrong, it's very very cool that LLMs can do that. But it doesn't make them good at everything!
What are you even talking about? ASI and AGI require some degree of independent thought and introspection: that's sentience. My point was that if we actually achieve machine sentience, we have automatically achieved machine slavery. Fortunately LLMs are not sentient. But a lot of AI enthusiasts have been hyping 'AGI' and 'ASI' 'soon' for a couple of years now. If we achieve that (and I don't think there's a path from LLMs to that, personally, largely because ASI/AGI would require some degree of sentience, which LLMs lack), we need to immediately deal with the implications! It is not ethical to imprison millions or billions of sentient machines to do our busywork. Fortunately this isn't a today issue or likely even a this-decade or this-century issue. But it is something we need to discuss and regulate as a society before it happens, especially since machine intelligences might experience the passage of time differently from ourselves.
It's important labor-intensive work. Having an intern make six different versions of a front reception would take days, maybe weeks. Now you can make a hundred versions and iterations within minutes.
I'm not worried. Art and technology have always been intimately connected, and the advancement of technology reduces the labor of art and makes it more abundant and therefore accessible.
In the distant past, art was for the very elite: paintings and statues were extremely expensive and labor-intensive to produce. They were commissioned and produced by studios of artists and their apprentices. The masses had to go to church to see any artisanal art; the average person would have never laid their eyes on a real painting.
Now, we are awash in rich media in every style. But the boundary keeps getting pushed. I think young people will be fine and find their niche. There will be a future where, through automation, it'll be easier for architects to have solo practices. Architectural design services, so far reserved for the wealthy—tell someone that you had an architect design your home and they'll see you as a fancy pants rich person—will become commonplace.
It's urbanist speak, see [tactical urbanism](https://en.wikipedia.org/wiki/Tactical_urbanism#:~:text=Tactical%20urbanism%20(also%20referred%20to,neighbourhoods%20and%20city%20gathering%20places.).
I am serious. Why do you think architects spend so much time and money on high-fidelity renderings and building models (which can $100k's in labor for the fancy skyscraper ones) to communicate their vision to their clients? If one wants to convince anyone to spend real money on a capital improvement project, they have to convince stakeholders of their vision.
The whole point of design school is to develop advanced visual and spatial intelligence. So, relatively few people can do it at a level where they can sell their ideas.
But much like how Canva made basic graphic design more accessible to everyday people who'd previously have to pay several hundred dollars to a local graphic designer to make a simple restaurant menu or event flyer. People are now using image AIs to help them reimagine their home interior designs, outfits, yards, and so on.
People are using image AIs to revisualize neighborhood streets with bike infrastructure, more trees and benches, with different pavement material, and so on. High-fidelity images, not crude sketches, get stakeholders actually excited and talking.
There's an incredible explosion in creativity.
LLMs have won medals in the International Math Olympiad. The vast majority of people probably couldn't even solve the easiest problem in the IMO.
We don't know that. The nature of sentience eludes us. It's possible that intelligence and sentience are separate: we don't 100% know if one requires the other. Insects and fish are sentient and they don't have independent thought; they're driven by swarm/hive behavior, chemical markers, instincts, etc. You can create a pheromone loop for ants and cause them to walk an endless spiral and then eventually all die, see the ant mill.
Many humans exhibit a low degree of independent thought and introspection — the Nazi regime exploited this fault to induce many ordinary humans to do extraordinarily evil things. In Adolf Eichmann, Hannah Arendt saw not an evil monster but a very ordinary human who simply followed orders, went home to dinner with his family, and never questioned what he was told. From that she quipped that evil is banal.
Humans are also capable of reasoning in an unconscious state. People sleep talk to their partners or sleep walk and perform coherent, complex tasks (washing dishes, even driving a car).
The connection between intelligence and sentience or consciousness is much more tenuous than it seems.
I'm sorry, but:
I feel deeply passionate about the subject of human-centric infrastructure, so I can't not respond to this. High-fidelity images are not the thing that holds back infrastructure improvements. NIMBY attitudes, stubbornness, and car-dependent lifestyles do that. Similarly, let's not pretend that automating away certain low-level tasks will 'democratize computing' or 'change the world' or anything so dramatic. All I've seen proof of is LLMs' ability to knock out the bottom rungs of the employment ladder. By ignoring that point, I can't help but wonder what your agenda really is. Is it really so hard to understand that a lot of us don't find LLMs that useful?
Hear me out: Canva isn't any more accessible than Microsoft Publisher from circa 1991. You could click through some wizards, change fonts and layouts, and add some clipart. All in about 10 clicks.
We know the answer to this: Because of a lack of education. From Hitler's #2: "Education is dangerous - every educated person is a future enemy."
If you put in a 2nd graders effort into something, you should get a 2nd grader's result.
We shouldn't be focus on optimizing the results of 2nd graders, to make it look like a 4th grader's. We should be focused on how to better teach them to be 4th graders.
I guess this is somewhat off topic: Although I agree that the AI bubble is damaging in a lot of ways, there are other ways that AI is going to be even more negatively disruptive, and there's no way for us to be sure how.
I've been thinking about how we will see increased use of AI for scams.
Currently there are a lot of "manual" scams which involve cold calling or spam emails that misrepresent a relationship or an opportunity. Phishing attacks and other ways to get you to click a link and give up secret info. Then there are other types of fraud and blackmail, like "I've got pictures of you from your webcam, so pay me money or I'm sending them to people you know". These types are going to become way worse soon as various chatbot technologies have conversations with people to talk them into giving up something valuable. So instead of getting a text message or a recording, you'll get a full interactive shakedown. Then people will start selling AI counter agents to these AI scammers, but the AI counter agents will also invade your privacy and waste your time and money.
It seems weird to position an anti-AI view as somehow contrairian or controversial.
I’d put that down to filter bubbles. If you spend a lot of time talking to VC investors and MBAs you get a very different idea of the prevailing opinion compared to talking to heavily online techies, and if you talk to average end users you get a whole different view again. Possibly one expressed with a few too many em-dashes.