Comment by lsy
6 months ago
I think two things can be true simultaneously:
1. LLMs are a new technology and it's hard to put the genie back in the bottle with that. It's difficult to imagine a future where they don't continue to exist in some form, with all the timesaving benefits and social issues that come with them.
2. Almost three years in, companies investing in LLMs have not yet discovered a business model that justifies the massive expenditure of training and hosting them, the majority of consumer usage is at the free tier, the industry is seeing the first signs of pulling back investments, and model capabilities are plateauing at a level where most people agree that the output is trite and unpleasant to consume.
There are many technologies that have seemed inevitable and seen retreats under the lack of commensurate business return (the supersonic jetliner), and several that seemed poised to displace both old tech and labor but have settled into specific use cases (the microwave oven). Given the lack of a sufficiently profitable business model, it feels as likely as not that LLMs settle somewhere a little less remarkable, and hopefully less annoying, than today's almost universally disliked attempts to cram it everywhere.
> There are many technologies that have seemed inevitable and seen retreats under the lack of commensurate business return (the supersonic jetliner)
I think this is a great analogy, not just to the current state of AI, but maybe even computers and the internet in general.
Supersonic transports must've seemed amazing, inevitable, and maybe even obvious to anyone alive at the time of their debut. But hiding under that amazing tech was a whole host of problems that were just not solvable with the technology of the era, let alone a profitable business model. I wonder if computers and the internet are following a similar trajectory to aerospace. Maybe we've basically peaked, and all that's left are optimizations around cost, efficiency, distribution, or convenience.
If you time traveled back to the 1970s and talked to most adults, they would have witnessed aerospace go from loud, smelly, and dangerous prop planes to the 707, 747 and Concorde. They would've witnessed the moon landings and were seeing the development of the Space Shuttle. I bet they would call you crazy if you told this person that 50 years later, in 2025, there would be no more supersonic commercial airliners, commercial aviation would basically look the same except more annoying, and also that we haven't been back to the moon. In the previous 50 years we went from the Wright Brothers to the 707! So maybe in 2075 we'll all be watching documentaries about LLMs (maybe even on our phones or laptops that look basically the same), and reminiscing about the mid-2020s and wondering why what seemed to be such a promising technology disappeared almost entirely.
I think this is both right and wrong. There was a good book that came out probably 15 years ago about how technology never stops in aggregate, but individual technologies tend to grow quickly and then stall. Airplane jets were one example in the book. The reason why I partially note this as wrong is that even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today.
A better example, also in the book, are skyscrapers. Each year they grew and new ones were taller than the ones last year. The ability to build them and traverse them increased each year with new technologies to support it. There wasn't a general consensus around issues that would stop growth (except at more extremes like air pressure). But the growth did stop. No one even has expectations of taller skyscrapers any more.
LLMs may fail to advance, but not because of any consensus reason that exists today. And it maybe that they serve their purpose to build something on top of them which ends up being far more revolutionary than LLMs. This is more like the path of electricity -- electricity in itself isn't that exciting nowadays, but almost every piece of technology built uses it.
I fundamentally find it odd that people seem so against AI. I get the potential dystopian future, which I also don't want. But the more mundane annoyance seems odd to me.
> even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today
I think they pretty strongly do
The solution seems to be "just lower your standards for acceptable margin of error to whatever the LLM is capable of producing" which should be concerning and absolutely unacceptable to anyone calling themselves an Engineer
7 replies →
> The reason why I partially note this as wrong is that even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today.
The fundamental problem has already been mentioned: Nobody can figure out how to SELL it. Because few people are buying it.
It's useful for aggregation and summarization of large amounts of text, but it's not trustworthy. A good summary decreases noise and amplifies signal. LLMs don't do that. Without the capability to validate the output, it's not really generating output of lasting value. It's just a slightly better search engine.
It feels like, fundamentally, the primary invention here is teaching computers that it's okay to be wrong as long as you're convincing. That's very useful for propaganda or less savory aspects of business, but it's less useful for actual communication.
13 replies →
> even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today
I hate to dogpile on this statement but I can think of two major issues right now:
* Small context windows, and serious degradation when pushing the limits of existing context windows. A human can add large amounts of state to their "context window" every day.
* Realtime learning. My humans get smarter every day, especially in the context of working with a specific codebase.
Maybe the AI companies will figure this out, but they are not "same technique more processor power" kinds of problems.
There are sound math reasons for skyscrapers topping out, mostly due to elevator capacity and the inability to effectively get people in and out of the floorspace as you go past a few hundred ft. There's no construction engineering reason you can't go taller - the Burj Khalifa, for example, is three times taller than a typical Western major city skyscraper - it just doesn't make economic sense unless you're a newly rich nation looking to prove a point.
1 reply →
>I think this is both right and wrong. There was a good book that came out probably 15 years ago about how technology never stops in aggregate, but individual technologies tend to grow quickly and then stall. Airplane jets were one example in the book. The reason why I partially note this as wrong is that even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today.
I don't see any solution to hallucinations, nor do I see any solution in sight. I think that could count as a concrete issue that would stop them.
4 replies →
Yeah, and with LLMs the thing I can't shake, however, is that this time it's pretty strongly (maybe parasitically) latched onto the aggregate progress of Moore's law. Few other technologies have enjoyed such relatively unfettered exponential improvement. It's like if skyscraper materials double in strength every n years, and their elevators approach teleportation speed, the water pumps get twice as powerful, etc., which would change the economics vs the reality that most of the physical world doesn't improve that fast.
Was the problem that supersonic flight was expensive and the amount of customers willing to pay the price was even lower than the number of customers that could even if they wanted to?
3 replies →
From a system optimisation perspective, SSTs solved the wrong problem.
Want to save people time flying? Solve the grotesque inefficiency pit that is airport transit and check-in.
Like, I'm sorry, STILL no high speed, direct to terminal rail at JFK, LAX and a dozen other major international airports? And that's before we get to the absolute joke of "border security" and luggage check-in.
Sure, supersonic afterburning engines are dope. But it's like some 10GHz single-core CPU that pulls 1.2kW out of the wall. Like it or not, an iPhone 16 delivers far more compute utility in far more scenarios.
that's not a technology problem, many airports are super efficient, e.g. Singapore. Public transport in the US is held back by other forces.
1 reply →
They don't optimize for our convenience, they optimize for their profit.
1 reply →
SST came and went in an era when none of that security theater existed to begin with
1 reply →
The crucial point is that we simply do not know yet if there is an inherent limitation in the reasoning capabilities of LLMs, and if so whether we are currently near to pushing up against them. It seems clear that American firms are still going to increase the amount of compute by a lot more (with projects like the Stargate factory), so time will tell if that is the only bottleneck to further progress. There might also still be methodological innovations that can push capabilities further.
What reasoning capabilities?
2 replies →
I don't think we're anywhere near peak capability for LLMs yet. It won't take 50 years but still it's been just 4 years.
> So maybe in 2075 we'll all be watching documentaries about LLMs (maybe even on our phones or laptops that look basically the same), and reminiscing about the mid-2020s and wondering why what seemed to be such a promising technology disappeared almost entirely.
It's hard for me to believe that anyone who works with technology in general, and LLMs in particular, could think this.
Technology has been stagnant for over 50 years, so it is easy to believe if you live in reality.
But sure if you exclusively look at the one field that has been advancing until about now I can see how one would end up where you are.
But this is exactly right, we are incrementally moving into the lamest cyberpunk future one can imagine, and we have been for some time.
Oh no, LLMs won't disappear but they will be a lot less loud.
Progress is often an S shaped curve and we are nearing saturation.
slower, no fast option, no smoking in the cabins, less leg room, but with TVs plastered on the back of every chair, sometimes
its actually kind of scary to think of a world where generative AI in the cloud goes away due to costs, in favor of some other lesser chimera version that can't currently be predicted
but good news is that locally run generative AI is still getting better and better with fewer and fewer resources consumed to use
The problem with supersonic commercial jets was mainly one of marketing/politics. The so called "sonic boom" problem was vastly overhyped, as anyone who lives near an air force base can tell you.
The conspiracy theorist tells me the American aerospace manufacturers at the time (Boening, McDonnell-Douglas, etc.), did everything they could to kill the Concorde. With limited flyable routes (NYC and DC to Paris and London I think were the only ones), the financials didn't make sense. If overland routes were available, especially opening up LA, San Francisco and Chicago, it might have been a different story.
>as anyone who lives near an air force base can tell you.
In the US, the Air Force is simply not allowed to fly supersonic anywhere near a city or a suburb with only a few exceptions.
One exception is Edwards Air Force Base in the California desert: there are houses nearby, but the base (and supersonic warplanes) preceded the construction of the homes, so the reasoning is that the home builders and home buyers knew what they were buying into.
Another exception (quoting Google Gemini):
>From 1964 to 1966, the FAA and U.S. Air Force conducted supersonic flights over St. Louis and other cities like Oklahoma City to gauge public reaction to daily sonic booms. The goal was to understand public tolerance for commercial supersonic transport (SST) operations. Reactions in St. Louis, as elsewhere, were largely negative, contributing to the eventual ban on commercial supersonic flight over land in the U.S.
Have you have experienced sonic booms? I have (when my family visited West Germany in 1970) and I certainly would not want to be subjected to them regularly.
Seems... wrong. Booms broke windows and drove zillions of complaints. Supersonic flight near airbases is controlled and happens on specific traffic corridors, right?
> The so called "sonic boom" problem was vastly overhyped, as anyone who lives near an air force base can tell you.
The pilots don't shit where they eat. Ask some farmer a bit further away how many sheep die a year from panic instead.
> “most people agree that the output is trite and unpleasant to consume”
That is a such a wild claim. People like the output of LLMs so much that ChatGPT is the fastest growing app ever. It and other AI apps like Perplexity are now beginning to challenge Google’s search dominance.
Sure, probably not a lot of people would go out and buy a novel or collection of poetry written by ChatGPT. But that doesn’t mean the output is unpleasant to consume. It pretty undeniably produces clear and readable summaries and explanations.
> People like the output of LLMs so much that ChatGPT is the fastest growing app ever
While people seem to love the output of their own queries they seem to hate the output of other people's queries, so maybe what people actually love is to interact with chatbots.
If people loved LLM outputs in general then Google, OpenAI and Anthropic would be in the business of producing and selling content.
> While people seem to love the output of their own queries they seem to hate the output of other people's queries
Listening or trying to read other peoples chats with these things is like listening to somebody describe a dream. It’s just not that interesting most of the time. It’s remarkable for the person experiencing it but it is deeply personal.
Low effort Youtube shorts with AI voice annoy the crap out of me.
After all this hype, they still can't do text to speech properly. Pause at the wrong part of the sentence all the time.
Google does put AI output at the top of every search now, and sometimes it's helpful and sometimes it's crap. They have been trying since long before LLMs to not just provide the links for a search but also the content.
Google used to be interested in making sure you clicked either the paid link or the top link in the results, but for a few years now they'd prefer that a user doesn't even click a link after a search (at least to a non-Google site)
1 reply →
I think the thing people hate about that is the lack of effort and attention to detail. It’s an incredible enabler for laziness if misused.
If somebody writes a design or a report, you expect that they’ve put in the time and effort to make sure it is correct and well thought out.
If you then find the person actually just had ChatGPT generate it and didn’t put any effort into editing it and checking for correctness, then that is very infuriating.
They are essentially farming out the process of creating the document to AI and farming out the process of reviewing it to their colleagues. So what is their job then, exactly?
These are tools, not a replacement for human thought and work. Maybe someday we can just have ChatGPT serve as an engineer or a lawyer, but certainly not today.
1 reply →
If I cared about the output from other people's queries then wouldn't they be my queries? I don't care about ChatGPTs response to your queries is because I don't care about your queries. I don't care if they came from ChatGPT or the world's foremost expert in whatever your query was about.
> That is a such a wild claim. People like the output of LLMs so much that ChatGPT is the fastest growing app ever.
The people using ChatGPT like its output enough when they're the ones reading it.
The people reading ChatGPT output that other people asked for generally don't like it. Especially if it's not disclosed up front.
Had someone put up a project plan for something that was not disclosed as LLM assisted output.
While technically correct it came to the wrong conclusions about the best path forward and inevitably hamstrung the project.
I only discovered this later when attempting to fix the mess and having my own chat with an LLM and getting mysteriously similar responses.
The problem was that the assumptions made when asking the LLM were incorrect.
LLMs do not think independently and do not have the ability to challenge your assumptions or think laterally. (yet, possibly ever, one that does may be a different thing).
Unfortunately, this still makes them as good as or better than a very large portion of the population.
I get pissed off not because of the new technology or the use of the LLM, but the lack of understanding of the technology and the laziness with which many choose to deliver the results of these services.
I am more often mad at the person for not doing their job than I am at the use of a model, the model merely makes it easier to hide the lack of competence.
7 replies →
Especially if it's not disclosed up front, and especially when it supplants higher-value content. I've been shocked how little time it's taken for AI slop SEO optimized blogs to overtake the articles written by genuine human experts, especially in niche product reviews and technical discussions.
However, whether or not people like it is almost irrelevant. The thing that matters is not whether economics likes it.
At least so far, it looks like economics absolutely loves LLMs: Why hire expensive human customer support when you can just offload 90% of the work to a computer? Why pay expensive journalists when you can just have the AI summarize it? Why hire expensive technical writers to document your code when you can just give it to the AI and check the regulatory box with docs that are good enough?
1 reply →
I'm not really countering that ChatGPT is popular, it certainly is, but it's also sort of like "fastest growing tire brand" that came along with the adoption of vehicles. The amount of smartphone users is also growing at the fastest rate ever so whatever the new most popular app is has a good chance of being the fastest growing app ever.
No… dude… it’s a new household name. We haven’t had those in software for a long time, maybe since TikTok and Fortnite.
5 replies →
> That is a such a wild claim.
Some people who hate LLMs are absolutely convinced everyone else hates them. I've talked with a few of them.
I think it's a form of filter bubble.
This isn't some niche outcry: https://www.forbes.com/sites/bernardmarr/2024/03/19/is-the-p...
And that was 18 months ago.
Yes, believe it or not, people eventually wake up and realize slop is slop. But like everything else with LLM development, tech is trying to brute force it on people anyway.
6 replies →
Maybe he's referencing how people don't like when other humans post LLM responses in the comments.
"Here's what chatGPT said about..."
I don't like that, either.
I love the LLM for answering my own questions, though.
"Here's what chatGPT said about..." Is the new lmgtfy
4 replies →
> AI apps like Perplexity are now beginning to challenge Google’s search dominance
Now that is a wild claim. ChatGPT might be challenging Google's dominance, but Perplexity is nothing.
It’s not a wild claim, though maybe your interpretation is wild.
I never said Perplexity individually is challenging Google, but rather as part of a group of apps including ChatGPT, which you conveniently left out of your quote.
At some point, Groupon was the fastest growing company ever.
People "like" or people "suffice" with the output? This "rise of whatever" as one blog put it gives me feelings that people are instead lowering their standards and cutting corners. Letting them cut through to stuff they actually want to do.
> People like the output of LLMs so much that ChatGPT is the fastest growing app ever
And how much of that is free usage, like the parent said? Even when users are paying, ChatGPT's costs are larger than their revenue.
> That is a such a wild claim. People like the output of LLMs so much that ChatGPT is the fastest growing app ever.
And this kind of meaningless factoid was immediately usurped by the Threads app release, which IMO is kind of a pointless app. Maybe let's find a more meaningful metric before saying someone else's claim is wild.
Asking your Instagram Users to hop on to your ready made TikTok Clone is hardly in the same sphere as spinning up that much users from nothing.
And while Threads growth and usage stalled, ChatGPT is very much still growing and has *far* more monthly visits than threads.
There's really nothing meaningless about ChatGPT being the 5th most visited site on the planet, not even 3 years after release. Threads doesn't make the top 50.
4 replies →
I would pay $5000 to never have to read another LLM-authored piece of text ever again.
...I do wonder what percent of ChatGPT usage is just students cheating on their homework, though.
Neal Stephenson has a recent post that covers some of this. Also links to teachers talking about many students just putting all their work into chatgpt and turning it in.
https://nealstephenson.substack.com/p/emerson-ai-and-the-for...
24 replies →
> > “most people agree that the output is trite and unpleasant to consume”
> That is a such a wild claim.
I think when he said "consume" he meant in terms of content consumption. You know, media - the thing that makes Western society go round. Movies, TV, music, books.
Would I watch an AI generated movie? No. What about a TV show? Uh... no. What about AI music? I mean, Spotify is trying to be tricky with that one, but no. I'd rather listen to Remi Wolf's 2024 Album "Big Ideas", which I thought was, ironically, less inspired than "Juno" but easily one of the best albums of the year.
ChatGPT is a useful interface, sure, but it's not entertaining. It's not high-quality. It doesn't provoke thought or offer us some solace in times of sadness. It doesn't spark joy or make me want to get up and dance.
I'm confused with your second point. LLM companies are not making any money from current models? Openai generates 10b USD ARR and has 100M MAUs. Yes they are running at a loss right now but that's because they are racing to improve models. If they stopped today to focus on optimization of their current models to minimize operating cost and monetizing their massive user base you think they don't have a successful business model? People use this tools daily, this is inevitable.
They might generate 10b ARR, but they lose a lot more than that. Their paid users are a fraction of the free riders.
https://www.wheresyoured.at/openai-is-a-systemic-risk-to-the...
This echoes a lot of the rhetoric around "but how will facebook/twitter/etc make money?" back in the mid 2000s. LLMs might shake out differently from the social web, but I don't think that speculating about the flexibility of demand curves is a particularly useful exercise in an industry where the marginal cost of inference capacity is measured in microcents per token. Plus, the question at hand is "will LLMs be relevant?" and not "will LLMs be massively profitable to model providers?"
58 replies →
That's fixable, a gradual adjusting of the free tier will happen soon enough once they stop pumping money into it. Part of this is also a war of attrition though, who has the most money to keep a free tier the longest and attract the most people. Very familiar strategy for companies trying to gain market share.
15 replies →
Then cut off the free riders. Problem solved overnight.
Are you saying they'd be profitable if they didn't pour all the winnings into research?
From where I'm standing, the models are useful as is. If Claude stopped improving today, I would still find use for it. Well worth 4 figures a year IMO.
They'd be profitable if they showed ads to their free tier users. They wouldn't even need to be particularly competent at targeting or aggressive with the amount of ads they show, they'd be profitable with 1/10th the ARPU of Meta or Google.
And they would not be incompetent at targeting. If they were to use the chat history for targeting, they might have the most valuable ad targeting data sets ever built.
40 replies →
That's calculating value against not having LLMs and current competitors. If they stopped improving but their competitors didn't, then the question would be the incremental cost of Claude (financial, adjusted for switching costs, etc) against the incremental advantage against the next best competitor that did continue improving. Lock in is going to be hard to accomplish around a product that has success defined by its generalizability and adaptability.
Basically, they can stop investing in research either when 1) the tech matures and everyone is out of ideas or 2) they have monopoly power from either market power or oracle style enterprise lock in or something. Otherwise they'll fall behind and you won't have any reason to pay for it anymore. Fun thing about "perfect" competition is that everyone competes their profits to zero
But if Claude stopped pouring their money into research and others didn't, Claude wouldn't be useful a year from now, as you could get a better model for the same price.
This is why AI companies must lose money short term. The moment improvements plateau or the economic environment changes, everyone will cut back on research.
For me, if Anthropic stopped now, and given access to all alternative models, they still would be worth exactly $240 which is the amount I'm paying now. I guess Anthropic and OpenAI can see the real demand by clearly seeing what are their free:basic:expensive plan ratios.
1 reply →
> Well worth 4 figures a year IMO
only because software engineering pay hasn't adjusted down for the new reality . You don't know what its worth yet.
13 replies →
> If they stopped today to focus on optimization of their current models to minimize operating cost and monetizing their user base you think they don't have a successful business model?
Actually, I'd be very curious to know this. Because we already have a few relatively capable models that I can run on my MBP with 128 GB of RAM (and a few less capable models I can run much faster on my 5090).
In order to break even they would have to minimize the operating costs (by throttling, maiming models etc.) and/or increase prices. This would be the reality check.
But the cynic in me feels they prefer to avoid this reality check and use the tried and tested Uber model of permanent money influx with the "profitability is just around the corner" justification but at an even bigger scale.
> In order to break even they would have to minimize the operating costs (by throttling, maiming models etc.) and/or increase prices. This would be the reality check.
Is that true? Are they operating inference at a loss or are they incurring losses entirely on R&D? I guess we'll probably never know, but I wouldn't take as a given that inference is operating at a loss.
I found this: https://semianalysis.com/2023/02/09/the-inference-cost-of-se...
which estimates that it costs $250M/year to operate ChatGPT. If even remotely true $10B in revenue on $250M of COGS would be a great business.
9 replies →
Revenue is _NOT_ Profit
And ARR is not revenue. It's "annualized recurring revenue": take one month's worth of revenue, multiply it by 12--and you get to pick which month makes the figures look most impressive.
14 replies →
It's a good point. Any business can get revenue by selling Twenty dollar bills for $19. But in the history of tech, many winners have been dismissed for lack of an apparent business model. Amazon went years losing money, and when the business stabilized, went years re-investing and never showed a profit. Analysts complained as Amazon expanded into non-retail activities. And then there's Uber.
The money is there. Investors believe this is the next big thing, and is a once in a lifetime opportunity. Bigger than the social media boom which made a bunch of billionaires, bigger than the dot com boom, bigger maybe than the invention of the microchip itself.
It's going to be years before any of these companies care about profit. Ad revenue is unlikely to fund the engineering and research they need. So the only question is, does the investor money dry up? I don't think so. Investor money will be chasing AGI until we get it or there's another AI winter.
How many of those MAUs are crappy startups building a janky layer on top of the OpenAI API which will cease to exist in 2 years?
Last year, ChatGPT was 75% of OpenAI's revenue[1], not the API.
[1]: https://www.businessofapps.com/data/chatgpt-statistics/
> that's because they are racing improve models. If they stopped today to focus on optimization of their current models to minimize operating cost and monetizing their user base you think they don't have a successful business model?
I imagine they would’ve flicked that switch if they thought it would generate a profit, but as it is it seems like all AI companies are still happy to burn investor money trying to improve their models while I guess waiting for everyone else to stop first.
I also imagine it’s hard to go to investors with “while all of our competitors are improving their models and either closing the gap or surpassing us, we’re just going to stabilize and see if people will pay for our current product.”
> I also imagine it’s hard to go to investors with “while all of our competitors are improving their models and either closing the gap or surpassing us, we’re just going to stabilize and see if people will pay for our current product.”
Yeah, no one wants to be the first to stop improving models. As long as investor money keeps flowing in there's no reason to - just keep burning it and try to outlast your competitors, figure out the business model later. We'll only start to see heavy monetization once the money dries up, if it ever does.
2 replies →
It’s just the natural counterpart to dogmatic inevitabilism — dogmatic denialism. One denies the present, the other the (recent) past. It’s honestly an understandable PoV though when you consider A) most people understand “AI” and “chatbot” to be synonyms, and B) the blockchain hype cycle(s) bred some deep cynicism about software innovation.
Funny seeing that comment on this post in particular, tho. When OP says “I’m not sure it’s a world I want”, I really don’t think they’re thinking about corporate revenue opportunities… More like Rehoboam, if not Skynet.
> most people understand “AI” and “chatbot” to be synonyms
This might be true (or not), but for sure not on this site.
1 reply →
Making money and operating at a loss contradict each other. Maybe someday they’ll make money —but not just yet. As many have said they’re hoping capturing market will position them nicely once things settle. Obviously we’re not there yet.
It is absolutely possible for the unit economics of a product to be profitable and for the parent company to be losing money. In fact, it's extremely common when the company is bullish on their own future and thus they invest heavily in marketing and R&D to continue their growth. This is what I understood GP to mean.
Whether it's true for any of the mainstream LLM companies or not is anyone's guess, since their financials are either private or don't separate out LLM inference as a line item.
No, because if they stop to focus on optimizing and minimizing operating costs, the next competitor over will leapfrog them with a better model in 6-12 months, making all those margin improvements an NPV negative endeavor.
One thing we're seeing in the software engineering agent space right now is how many people are angry with Cursor [1], and now Claude Code [2] (just picked a couple examples; you can browse around these subreddits and see tons of complaints).
What's happening here is pretty clear to me: Its a form of enshittification. These companies are struggling to find a price point that supports both broad market adoption ($20? $30?) and the intelligence/scale to deliver good results ($200? $300?). So, they're nerfing cheap plans, prioritizing expensive ones, and pissing off customers in the process. Cursor even had to apologize for it [3].
There's a broad sense in the LLM industry right now that if we can't get to "it" (AGI, etc) by the end of this decade, it won't happen during this "AI Summer". The reason for that is two-fold: Intelligence scaling is logarithmic w.r.t compute. We simply cannot scale compute quick enough. And, interest in funding to pay for that exponential compute need will dry up, and previous super-cycles tell us that will happen on the order of ~5 years.
So here's my thesis: We have a deadline that even evangelists agree is a deadline. I would argue that we're further along in this supercycle than many people realize, because these companies have already reached the early enshitification phase for some niche use-cases (software development). We're also seeing Grok 4 Heavy release with a 50% price increase ($300/mo) yet offer single-digit percent improvement in capability. This is hallmark enshitification.
Enshitification is the final, terminal phase of hyperscale technology companies. Companies remain in that phase potentially forever, but its not a phase where significant research, innovation, and optimization can happen; instead, it is a phase of extraction. AI hyperscalers genuinely speedran this cycle thanks to their incredible funding and costs; but they're now showcasing very early signals of enshitifications.
(Google might actually escape this enshitification supercycle, to be clear, and that's why I'm so bullish on them and them alone. Their deep, multi-decade investment into TPUs, Cloud Infra, and high margin product deployments of AI might help them escape it).
[1] https://www.reddit.com/r/cursor/comments/1m0i6o3/cursor_qual...
[2] https://www.reddit.com/r/ClaudeAI/comments/1lzuy0j/claude_co...
[3] https://techcrunch.com/2025/07/07/cursor-apologizes-for-uncl...
Exactly. This is basically the argument of “AI as Normal Technology”.
https://news.ycombinator.com/item?id=43697717
Thanks for the link. The comparison to electricity is a good one, and this is a nice reflection on why it took time for electricity’s usefulness to show up in productivity stats:
> What eventually allowed gains to be realized was redesigning the entire layout of factories around the logic of production lines. In addition to changes to factory architecture, diffusion also required changes to workplace organization and process control, which could only be developed through experimentation across industries.
This seems like one the only sane arguments in this whole sea of articles.
My take since day one:
(1) Model capabilities will plateau as training data is exhausted. Some additional gains will be possible by better training, better architectures, more compute, longer context windows or "infinite" context architectures, etc., but there are limits here.
(2) Training on synthetic data beyond a very limited amount will result in overfitting because there is no new information. To some extent you could train models on each other, but that's just an indirect way to consolidate models. Beyond consolidation you'll plateau.
(3) There will be no "takeoff" scenario -- this is sci-fi (in the pejorative sense) because you can't exceed available information. There is no magic way that a brain in a vat can innovate beyond available training data. This includes for humans -- a brain in a vat would quickly go mad and then spiral into a coma-like state. The idea of AI running away is the information-theoretic equivalent of a perpetual motion machine and is impossible. Yudkowski and the rest of the people afraid of this are crackpots, and so are the hype-mongers betting on it.
So I agree that LLMs are real and useful, but the hype and bubble are starting to plateau. The bubble is predicated on the idea that you can just keep going forever.
The next step is clearly improved vision and everyday-physics models. These can also solve hallucinations.
clearly
>> There are many technologies that have seemed inevitable and seen retreats under the lack of commensurate business return
120+ Cable TV channels must have seemed like a good idea at the time, but like LLMs the vast majority of the content was not something people were interested in.
I think the difference between all previous technologies is scope. If you make a super sonic jet that gets people from place A to place B faster for more money, but the target consumer is like "yeah, I don't care that much about that at that price point", then your tech sort is of dead. You are also fully innovated on that product, like maybe you can make it more fuel efficient, sure, but your scope is narrow.
AI is the opposite. There are numerous things it can do and numerous ways to improve it (currently). There is lower upfront investment than say a supersonic jet and many more ways it can pivot if something doesn't work out.
The number of things it can actually do is significantly lower than the number of things the hype men are claiming it can do.
It's not a great analogy. The only parallel with Concorde is energy consumption. I think a better analogy would have been VR.
I mean, thats the point, they aren't the same. Concorde was one dimensional, AI is not.
Most of the comments here feel like cope about AI TBH. There's never been an innovation like this ever, and it makes sense to get on board rather than be left behind.
> There's never been an innovation like this ever
There have been plenty of innovations like this. In fact, much of the hype around LLMs is a rehash of the hype around "expert systems" back in the '80s. LLMs are marginally more effective than those systems, but only marginally.
3 replies →
The difference is that the future is now with LLMs. There is a microwave (some multiple) in almost every kitchen in the world. The Concord served a few hundred people a day. LLMs are already ingrained into hundreds of millions if not billions of people’s lives, directly and indirectly. My dad directly uses LLMs multiple times a week if not daily in an industry that still makes you rotate your password every 3 months. It’s not a question of whether the future will have them, it’s a question of whether the future will get tired of them.
The huge leap that is getting pushback is the sentiment that LLMs will consume every use case and replace human labor. I don't think many are arguing LLMs will die off entirely.
Developers haven't even started extracting the value of LLMs with agent architectures yet. Using an LLM UI like open ai is like we just figured fire and you use it to warm you hands (still impressive when you think about it, but not worth the burns), while LLM development is about building car engines (here is you return on investment).
> Developers haven't even started extracting the value of LLMs with agent architectures yet
There are thousands of startups doing exactly that right now, why do you think this will work when all evidence points towards it not working? Or why else would it not already have revolutionized everything a year or two ago when everyone started doing this?
Most of them are a bunch of prompts and don't even have actual developers. For the good reason that there is no training system yet and the wording of how you call the people that build these system isn't even there or clearly defined. Local companies haven't even setup a proper internal LLM or at least a contract with a provider. I am in France so probably lagging behind USA a bit especially NY/SF but the word "LLM developer" is just arriving now and mostly under the pressure of isolated developers and companies like me. This feel really really early stage.
18 replies →
>Or why else would it not already have revolutionized everything a year or two ago when everyone started doing this?
The internet needed 20 years to take over the world. All of the companies of the first dot com bust are in the past. The tech is solid.
Every booster argument is like this one. $trite_analogy triumphant smile
[flagged]
1 reply →
3 years into automating all white collar labor in 6 months.
It's 3 years away.
lol you must not be looking for a white collar job right now then outside of IT.
The only thing that is over hyped is there is no white collar bloodbath but a white collar slow bleed out.
Not mass firing events but transition by attrition over time. A bleed out in jobs that don't get back filled and absolutely nothing in terms of hiring reserve capacity for the future.
My current company is a sinking ship, I suspect it will go under in the next two years so I have been trying to get off but there is absolutely no place to go.
In 2-3 years I expect to be unemployed and unemployable, needing to retrain to do something I have never done before.
What is on display in this thread is that human's are largely denial machines. We have to be otherwise we would be paralyzed by our own inevitable demise.
It is more comforting to believe everything is fine and the language models are just some kind of doge coin tech hype bullshit.
>> Developers haven't even started extracting the value of LLMs with agent architectures yet.
What does this EVEN mean? Do words have any value still, or are we all just starting to treat them as the byproduct of probabilistic tokens?
"Agent architectures". Last time I checked an architecture needs predictability and constraints. Even in software engineering, a field for which the word "engineering" is already quite a stretch in comparison to construction, electronics, mechanics.
Yet we just spew the non-speak "Agentic architectures" as if the innate inability of LLMs in managing predictable quantitative operations is not an unsolved issue. As if putting more and more of these things together automagically will solves their fundamental and existential issue (hallucinations) and suddenly makes them viable for unchecked and automated integration.
This means I believe we currently underuse LLM capabilities and their empirical nature makes it difficult to assess their limitations without trying. I've been studying LLMs from various angles during a few months before coming to this conclusion, as an experienced software engineer and consultant. I must admit it is however biased towards my experience as an SME and in my local ecosystem.
Hallucinations might get solved by faster, cheaper and more accurate, vision and commonsense-physics models. Hypothesis: Hallucinations are a problem only because physical reality isn't text. Once people switch to models that predict physical states instead of missing text, then we'll have domestic robots and lower hallucination rates.
1 reply →
Theyre doing it so much it's practically a cliche.
There are underserved areas of the economy but agentic startups is not one.
> Developers haven't even started extracting the value of LLMs with agent architectures yet.
For sure there is a portion of developers who don't care about the future, are not interested in current developements and just live as before hoping nothing will change. But the rest already gave it a try and realized tools like Claude Code can give excellent results for small codebases to fail miserably at more complex tasks with the net result being negative as you get a codebase you don't understand, with many subtle bugs and inconsistencies created over a few days you will need weeks to discover and fix.
This is a bit developer centric, I am much more impressed by the opportunities I see in consulting rather than applying LLMs to dev tasks. And I am still impressed by the code it can output eventhough we are still in the funny intern stage in this area.
1 reply →
>evelopers haven't even started extracting the value of LLMs with agent architectures yet.
Which is basically what? The infinite monkey theorem? Brute forcing solutions for problems at huge costs? Somehow people have been tricked to actually embrace and accept that now they have to pay subscriptions from 20$ to 300$ to freaking code? How insane is that, something that was a very low entry point and something that anyone could do, is now being turned into some sort of classist system where the future of code is subscriptions you pay for companies ran by sociopaths who don't care that the world burns around them, as long as their pockets are full.
I cannot emphasize how much I agree with this comment. Thank you for writing it, I would never have had written it as well.
I don't have a subscription not even an Open AI account (mostly cause they messed up their google account system). You can't extract value of an LLM by just using the official UI, you just scratch the surface of how they work. And yet there aren't much developers able to actually build an actual agent architecture that does deliver some value. I don't include the "thousands" of startups that are clearly suffer from a signaling bias: they don't exist in the economy and I don't care about them like at all in my reasonning. I am talking about actual LLM developers that you can recruit locally the same way you recruit a web developer today, and that can make sense out of "frontier" LLM garbage talk by using proper architectures. These devs are not there yet.
I pay $300 to fly from SF to LA when I could've just walked for free. Its true. How classist!
Let's not ignore the technical aspects as well: LLMs are probably a local minima that we've gotten stuck in because of their rapid rise. Other areas in AI are being starved of investment because all of the capital is pouring into LLMs. We might have been better off in the long run if LLMs hadn't been so successful so fast.
> first signs of pulling back investments
I agree with you, but I’m curious; do you have link to one or two concrete examples of companies pulling back investments, or rolling back an AI push?
(Yes it’s just to fuel my confirmation bias, but it’s still feels nice:-) )
Most prominent example was this one: https://www.reuters.com/technology/microsoft-pulls-back-more...
I think that's more reflective of the deteriorating relationship between OpenAI and Microsoft than an true lack of demand for datacenters. If a major model provider (OpenAI, Anthropic, Google, xAI) were to see a dip in available funding or stop focusing on training more powerful models, that would convince me we may be in a bubble about to pop, but there are no signs of that as far as I can see.
There are pretty hidden assumption in this comment. First of all, not every business in the AI space is _training_ models, and the difference between training and inference is massive - i.e. most businesses can easily afford inference, perhaps depending on model, but they definitely can.
Another several unfounded claims were made here, but I just wanted to say LLMs with MCP are definitely good enough for almost every use case you can come up with as long as you can provide them with high quality context. LLMs are absolutely the future and they will take over massive parts of our workflow in many industries. Try MCP for yourself and see. There's just no going back.
LLMs with tools*
MCP isn’t inherently special. A Claude Code with Bash() tool can do nearly anything a MCP server will give you - much more efficiently.
Computer Use agents are here and are only going to get better.
The conversation shouldn’t be about LLMs any longer. Providers will be providing agents.
correct and companies will be exposing their data via mcp instead of standard rest apis.
3 replies →
> I just wanted to say LLMs with MCP are definitely good enough for almost every use case you can come up with as long as you can provide them with high quality context.
This just shows you lack imagination.
I have a lot of use cases that they are not good enough for.
I mean, it all depends on how one defines "high quality context".
I do wonder where in the cycle this all is given that we've now seen yet another LLM/"Agentic" VSCode fork.
I'm genuinely surprised that Code forks and LLM cli things are seemingly the only use case that's approached viability. Even a year ago, I figured there'd be something else that's emerged by now.
But there are a ton of LLM powered products in the market.
I have a friend in finance that uses LLM powered products for financial analysis, he works in a big bank. Just now anthropic released a product to compete in this space.
Another friend in real estate uses LLM powered lead qualifications products, he runs marketing campaigns and the AI handles the initial interaction via email or phone and then ranks the lead in their crm.
I have a few friends that run small businesses and use LLM powered assistants to manage all their email comms and agendas.
I've also talked with startups in legal and marketing doing very well.
Coding is the theme that's talked about the most in HN but there are a ton of startups and big companies creating value with LLMs
Yup. Lots of products in the education space. Even doctors are using LLMs, while talking with patients. All sorts of teams are using the adjacent products for image and (increasingly) video generation. Translation freelancers have been hit somewhat hard because LLMs do "good enough" quite a bit better than old google translate.
Coding is relevant to the HN bubble, and as tech is the biggest driver of the economy it's no surprise that tech-related AI usages will also be the biggest causes of investment, but it really is used in quite a lot of places out there already that aren't coding related at all.
LLMs are amazing at anything requiring text analysis (go figure). Everyone I know doing equity or economic research in finance is using it extensively for that, and from what I hear from doctors the LLMs are as good as that in their space if not better
It feels like there's a lot of shifting goalposts. A year ago, the hype was that knowledge work would cease to exist by 2027.
Now we are trying to hype up enhanced email autocomplete and data analysis as revolutionary?
I agree that those things are useful. But it's not really addressing the criticism. I would have zero criticisms of AI marketing if it was "hey, look at this new technology that can assist your employees and make them 20% more productive".
I think there's also a healthy dose of skepticism after the internet and social media age. Those were also society altering technologies that purported to democratize the political and economic system. I don't think those goals were accomplished, although without a doubt many workers and industries were made more productive. That effect is definitely real and I'm not denying that.
But in other areas, the last 3 decades of technological advancement have been a resounding failure. We haven't made a dent in educational outcomes or intergenerational poverty, for instance.
> most people agree that the output is trite and unpleasant to consume
This is likely a selection bias: you only notice the obviously bad outputs. I have created plenty of outputs myself that are good/passable -- you are likely surrounded by these types of outputs without noticing.
Not a panacea, but can be useful.
> 2. Almost three years in, companies investing in LLMs have not yet discovered a business model that justifies the massive expenditure of training and hosting them,
I always think back to how Bezos and Amazon were railed against for losing money for years. People thought that would never work. And then when he started selling stuff other than books? People I know were like: please, he's desperate.
Someone, somewhere will figure out how to make money off it - just not most people.
To use the Internet as a comparison:
Phase 1 - mid to late 1990s:
- "The Internet is going to change EVERYTHING!!!"
Phase 2 - late 1990s to early 2000s:
- "It's amazing and we are all making SO much money!"
- "Oh no! The bubble burst"
- "Of course everyone could see this coming: who is going to buy 40 lb bags of dogfood or their groceries over the Internet?!?!?"
Phase 3 - mid 2000s to 2020:
- "It is astounding the amount of money being by tech companies"
- "Who could have predicted that social media would change the ENTIRE landscape??"
My guess is that LLM's are bridge technology, the equivalent of cassette tapes. A big step forward, allowing things that we couldn't before. But before long they'll be surpassed by much better technology, and future generations will look back on them as primitive.
You have top scientists like LeCun arguing this position. I'd imagine all of these companies are desperately searching for the next big paradigm shift, but no one knows when that will be, and until then they need to squeeze everything they can out of LLMs.
ML models have the good property of only requiring investment once and can then be used till the end of history or until something better replaces them.
Granted the initial investment is immense, and the results are not guaranteed which makes it risky, but it's like building a dam or a bridge. Being in the age where bridge technology evolves massively on a weekly basis is a recipe for being wasteful if you keep starting a new megaproject every other month though. The R&D phase for just about anything always results in a lot of waste. The Apollo programme wasn't profitable either, but without it we wouldn't have the knowledge for modern launch vehicles to be either. Or to even exist.
I'm pretty sure one day we'll have an LLM/LMM/VLA/etc. that's so good that pretraining a new one will seem pointless, and that'll finally be the time we get to (as a society) reap the benefits of our collective investment in the tech. The profitability of a single technology demonstrator model (which is what all current models are) is immaterial from that standpoint.
Nah, if TSMC got exploded and there was a world war, in 20 years all the LLMs would bit rot.
Eh, I doubt it, tech only got massively better in each world war so far, through unlimited reckless strategic spending. We'd probably get a TSMC-like fab on every continent by the end of it. Maybe even optical computers. Quadrotor UAV are the future of warfare after all, and they require lots of compute.
Adjusted for inflation it took over 120 billion to build the fleet of liberty ships during WW2, that's like at least 10 TSMC fabs.
2 replies →
> model capabilities are plateauing at a level where most people agree that the output is trite and unpleasant to consume.
What are you basing this on? Personal feelings?
LLMs need significant optimization or we get significant improvement on computing power while keeping the energy cost the same. It's similar with smartphone, when at the start it's not feasible because of computing power, and now we have one that can rival 2000s notebooks.
LLMs is too trivial to be expensive
EDIT: I presented the statement wrongly. What I mean is the use case for LLM are trivial things, it shouldn't be expensive to operate
LLM can give you thousands of lines of perfectly working code for less than 1 dollar. How is that trivial or expensive?
Looking up a project on github, downloading it and using it can give you 10000 lines of perfectly working code for free.
Also, when I use Cursor I have to watch it like a hawk or it deletes random bits of code that are needed or adds in extra code to repair imaginary issues. A good example was that I used it to write a function that inverted the axis on some data that I wanted to present differently, and then added that call into one of the functions generating the data I needed.
Of course, somewhere in the pipeline it added the call into every data generating function. Cue a very confused 20 minutes a week later when I was re-running some experiments.
4 replies →
well I presented the statement wrongly. What I mean is the use case for LLM are trivial things, it shouldn't be expensive to operate
and the 1 dollar cost for your case is heavily subsidized, that price won't hold up long assuming the computing power stays the same.
1 reply →
Thousands of lines of perfectly working code? Did you verify that yourself? Last time I tried it produced slop, and I've been extremely detailed in my prompt.
3 replies →
But the thing is, LLMs are already incredibly cheap to operate compared to the alternatives. Both for trivial things and for complex things.
Well recently cursor got a heat for rising price and having opaque usage, while anthropic's claude reported to be worse due to optimization. IMO the current LLMs are not sustainable, and prices are expected to increase sooner or later.
Personally, until models comparable with sonnet 3.5 can be run locally on mid range setup, people need to wary that the price of LLM can skyrocket
1 reply →
Imagine telling a person from five years ago that the programs that would basically solve NLP, perform better than experts at many tasks and are hard not to anthropomorphize accidentally are actually "trivial". Good luck with that.
>programs that would basically solve NLP
There is a load-bearing “basically” in this statement about the chat bots that just told me that the number of dogs granted forklift certification in 2023 is 8,472.
1 reply →
"hard not to anthropomorphize accidentally' is a you problem.
I'm unhappy every time I look in my inbox, as it's a constant reminder there are people (increasingly, scripts and LLMs!) prepared to straight-up lie to me if it means they can take my money or get me to click on a link that's a trap.
Are you anthropomorphizing that, too? You're not gonna last a day.
1 reply →
Yeah it solved NLP about 50% of the time, and also mangles data badly and in often hard-to-detect ways.
It still doesn't pass the Turing test, and is not close. Five years ago me would be impressed but still adamant that this is not AI, nor is it on the path to AI.
Calling LLMs trivial is a new one. Yea just consume all of the information on the internet and encode it into a statistical model, trivial, child could do it /s
> all of the information on the internet
Total exaggeration—especially given Cloudflare providing free tools to block AI and now tools to charge bots for access to information.
well I presented the statement wrongly. What I mean is the use case for LLM are trivial things, it shouldn't be expensive to operate
Oh wow I forgot that the microwave oven was once marketed as the total replacement of cooking chores and in futuristic life people can just press a button and have a delicious good meal ( well you can now but microwave meals are often seen as worse than fastfood ).
Investments are mostly in model training. We have trained models now, we'll see a pullback in that regard as businesses will need to optimize to get the best model without spending billions in order to compete on price, but LLMs are here to stay.
> 2. Almost three years in, companies investing in LLMs have not yet discovered a business model that justifies the massive expenditure of training and hosting them, the majority of consumer usage is at the free tier, the industry is seeing the first signs of pulling back investments, and model capabilities are plateauing at a level where most people agree that the output is trite and unpleasant to consume.
You hit the nail on why I say to much hatred from "AI Bros" as I call them, when I say it will not take off truly until it runs on your phone effortlessly, because nobody wants to foot a trillion dollar cloud bill.
Give me a fully offline LLM that fits in 2GB of VRAM and lets refine that so it can plug into external APIs and see how much farther we can take things without resorting to burning billions of dollars' worth of GPU compute. I don't care that my answer arrives instantly, if I'm doing the research myself, I want to take my time to get the correct answer anyway.
We actually aren't too far off from that reality. There are several models you can run fully offline on your phone (phi-3, Gemma-3n-E2b-it, Qwen2.5-1.5b-instruct all run quite well on my Samsung S24 ultra). There are a few offline apps that also have tool calling (mostly for web search but I suspect this is extendable).
If you want to play around a bit and are on android there is PocketPal,ChatterUI, MyDeviceAI, SmolChat are good multi-model apps and Google's Edge gallery won't keep your chats but is a fun tech demo.
All are on github and can be installed using Obtainium if you don't want to
You aren’t extrapolating enough. Nearly the entire history of computing has been one that isolates between shared computing and personal computing. Give it time. These massive cloud bills are building the case for accelerators in phones. It’s going to happen just needs time.
That's fine, that's what I want ;) I just grow tired of people hating on me for thinking that we really need to localize the models for them to take off.
1 reply →
> (the supersonic jetliner) ... (the microwave oven)
But have we ever had a general purpose technology (steam engine, electricity) that failed to change society?
It wouldn't be general purpose if it fails to bring change. I'd take every previous iteration of "AI" as example, IBM Watson, that stuff
They didn't really need the cloud either and yet...
I don't really buy your point 2. Just the other day Meta announced hundreds of billions of dollars investment into more AI datacenters. Companies are bringing back nuclear power plants to support this stuff. Earlier this year OpenAI and Oracle announced their $500bn AI datacenter project, but admittedly in favor of your point have run into funding snags, though that's supposedly from tariff fears with foreign investors, not lack of confidence in AI. Meta can just finance everything from their own capital and Zuck's decree, like they did with VR (and it may very well turn out similarly).
Since you brought up supersonic jetliners you're probably aware of the startup Boom in Colorado trying to bring it back. We'll see if they succeed. But yes, it would be a strange path, but a possible one, that LLMs kind of go away for a while and try to come back later.
You're going to have to cite some surveys for the "most people agree that the output is trite and unpleasant" and "almost universally disliked attempts to cram it everywhere" claims. There are some very vocal people against LLM flavors of AI, but I don't think they even represent the biggest minority, let alone a majority or near universal opinions. (I personally was bugged by earlier attempts at cramming non-LLM AI into a lot of places, e.g. Salesforce Einstein appeared I think in 2016, and that was mostly just being put off by the cutesy Einstein characterization. I generally don't have the same feelings with LLMs in particular, in some cases they're small improvements to an already annoying process, e.g. non-human customer support that was previously done by a crude chatbot front-end to an expert system or knowledge base, the LLM version of that tends to be slightly less annoying.)
Sort of a followup to myself if I come back searching this comment or someone sees this thread later... here's a study that just came out on AI attitudes: https://report2025.seismic.org/
I don't think it supports the bits I quoted, but it does include more negativity than I would have predicted before seeing it.