Sort of. Big fan of 404media usually, but they need to "look through the bubble" to understand this can and likely will be replicated on device when local models and inference catch up, which they will eventually (Apple Intelligence, Hermes, OpenClaw, etc). Slap iCloud and Google Storage sync on the app, and you've got your own character bots in your device ecosystem.
It seems harder to sell on device ai as a subscription, I would guess? although the much lower inference costs mean maybe you could just sell it as a one time purchase
Isn't the majority of AI compute training and not inference? Saying "It doesn't matter if the AI companies fold because I can run models on my computer" is like saying "I don't care if the farm closes because I get my food from the grocery store"
By on device you mean running locally on a mobile device like an iPhone? If so we're a really long way from a chatbot that's engaging to talk to. There's just not enough processing power to do it
No-one's going to want to talk to a 7B model for more than 3 minutes
Yes, iphone, ipad, mac. Maybe your phone makes the inference requests to either Apple servers (requiring iCloud+ and credit based?) or your Mac (Tailscale equivalent).
My assumptions are that inference costs continue to decline (and/or the difficulty level of routing requests declines) and hardware acceleration continues into the future. These assumptions might be wrong, but that is the idea the thesis is based around.
(have built workflows and harnesses that currently target LLM provider APIs under the assumption I'll be able to point "locally-ish" eventually, my career had a strong focus on distributed computing about 10-15 years ago, think managing data and work clusters for LHC detector data processing, and I grew up on hobbyist distributed.net and BOINC projects, all the lenses I see this problem through; queues, jobs, sync vs async performance optimization, etc; recovered sysadmin|infra/devops engineer a lifetime ago)
Yes, current gen hardware was not built to run such models, and are obviously inefficient and bad at it.
Significant investment is being spent on producing hardware that is more suitable for running mid-sized (100B-ish) models. Starting with flash chips with wide enough interfaces to provide the weights (NAND flash can be ~5x more energy-efficient for reads than DRAM), and going into more exotic solutions like compute-in-memory.
I think we're a breakthrough away. I don't think processing power is an issue. A 7B model has a lot of cruft (like knowledge of code syntax) that a character sim doesn't need. It should be possible to extract just the relevant parts of a larger and more powerful model and run it with the same resources you would run a general 7B model. We just don't have the ability to isolate and extract weights like that yet.
So, your contention is that 404media isn't taking into consideration a hypothetical where they wouldn't burn millions of dollars each month running something.
"eventually" is doing a lot of work here. Like sure, yes, eventually. But that could be later this year or it could be 50 years from now. We don't know the answer to that. Assuming it's going to happen in a timeline short enough to matter for this conversation isn't worthwhile. It's like saying eventually we'll have cold fusion, so why is everyone skeptical?
There are a tonne of AI companion apps out there now, lot of them focused on nsfw, it's going to be interesting to see if character.ai is an isolated case, or if it's a domino effect.
I doubt the NSFW-focused ones have taken on venture funding. They're probably leaner shops with a dramatically smaller userbase that is more likely to actually pay.
Before Grok locked down its video generation tools the amount of high-quality and stunningly realistic looking porn you could generate was insane. You’re going to need something like that in order to make billions of dollars in this space. I don’t think anyone is going to take the time/money and deal with all the legal implications that comes from training a video generation model like though.
I can think of an established industry whose major players each have lots of relevant data that they own to use for that purpose and which would have a ready way to profit from use of those models without particularly significant additional legal exposure compared to those in their normal operations.
Oh, and major players from that industry have been publicly reaching out for people with AI image/video generation and training experience on public forums.
Which industry? The porn industry? Because if that’s who you mean even they don’t have the spare billions it’ll take to train a video generation model from scratch that’s actually good. Honestly, if you ask me, they should just make a deal with grok to buy out that portion of their video generation model.
I think it's more nuanced than that. Yes, it's true that the economics are horrible. But it's also true that consumer startups are hard. And then when you tack on legal fees and product changes due to policy concerns, it becomes impossible. Maybe the lesson is that the world isn't ready for a product that targeted minors with fake friends based on existing IP. Maybe it never will be.
> Character.ai’s bankruptcy is the one of the first dominos to fall in the long-foreseen AI bubble.
Kind of speculative, and a very bold speculation at that. I don’t have anything else to say, but one company that was obviously mismanaged doesn’t really get my attention.
News is less about inspection and more about speculation these days. Ironically, journalists are largely paid the same way Ai is, per word. The incentives might be better if both were "salaried"
Groups desperate to prove an increasingly absurd suggestion are latching onto this because actual evidence of overinvestment or of broader correlated issues is thin on the ground.
Character.ai is hardly representative of the AI market. It is at best a tiny sliver. There's coding tools, business process applications and more. Now the AI market likely has some challenges around financing build outs given recent inflation driven interest rates etc. But that's very different from what character.ai
I have been seeing this more and more with “top” tier publications but “Character.ai’s bankruptcy is the one of the first dominos to fall in the long-foreseen AI bubble.”
A simple grammatical mistakes like this should have been caught on review/edit phase.
Either that or writers are relying too much on AI.
The economics of AI are terrible. The exact inverse of what VCs normally say they strive for.
It’s capital intensive, doesn’t easily scale without linear costs, is basically a commodity, seemingly has negative cashflow forever, and customers aren’t willing to pay what it actually costs. It’s basically the worst tech business model and yet somehow VCs caught a spell and threw every available penny into the fire.
Watching this all unravel is going to be interesting.
I think the closest analogy we have of the AI boom is the mid 90s browser wars.
First all the money went into the browsers like Netscape. Then the browser got commoditized, but that didn't mean the value went away. A better business model came along, and lo and behold, the Chrome browser went open source. The browser became free, and the scale of adoption meant a lot of value could still be extracted.
A free browser product didn't mean the value of Chrome went away. Rather the opposite. It's a gateway for buying stuff through Google search.
Marc Andressen was the "Mosaic browser" guy before he was the investor guy, so the folks in the early days did not lose all their money. Quite the opposite. Many are now billionaire oligarchs.
LLMs will likely go the way of the browser... commoditized, then free, then appropriated to shove ads into your face. I predict a lot of safety and interpretability research will be leveraged to one day push ads, just like "Golden Gate Claude" could be manipulated to think of itself as the Golden Gate Bridge.
On the models yes. All indications are models will basically be a free commodity soon.
On the datacenter side the closest analogy is the fiber buildout of the .com bust. While that fiber ended up being useful decades later it’s unclear what value a bunch of half-completed datacenters will bring if their owners go bust.
Half of the cost of the datacenter is the building, power, and cooling infra, and half is the silicon. Datacenter investments will depreciate faster than fiber, but the value won't disappear.
Holding value is contingent on whether a company can manipulate the regulators to create a lasting monopoly or expand the market dramatically. My suspicion is that at some point the investors will gravitate towards the latter by turning LLMs into more active participants in the economy
Am I understanding this correctly as an idiomatic reference to an enchantment or curse? If so, I'd be interested to know where the expression "catching a spell" comes from.
(I'm really interested in the phrasing and I hope you don't take this the wrong way. For me personally, the metaphorical idea that someone with inexplicable behavior has been bewitched is a normal thing to express, but I'd be more likely to say they "fell under" a spell than that they "caught" it. I believe in Mandarin the usual idiom uses 中 "be struck [by a projectile]".)
Sure, but that's normal for moonshots. Maybe LLMs really will stall and spend 10 years getting better at anything and everything except taste, and then there'll be a completely new approach that has nothing to do with massive GPU farms. Maybe not.
I suspect this one wasn't the "economics of AI" though. This one was bait and switch (intentional or not). They generated a huge user base, then clamped down on many of the things that drew the user base to them in the first place, then acted surprised when those users left because they couldn't get what they wanted there.
And yes, I'm talking about the landslide of NSFW stuff that had to be culled to turn it into an acquisition-friendly target.
Well said. The anti-AI sentiment is strong in a few camps on HN. Praying that AI will go away so as a software engineer, only familiar with the SDLC, you become relevant again is not a winning strategy. AI is like a soup with many bubbles, some will pop while new ones begin.
The conclusion of the article is wrong - that the Economics of AI Don't Work.
A better report would have been how the Economics of AI Wrappers Don't Work.
Character.ai was likely paying frontier LLM API rates (which are known to be ridiculously expensive at 90% margins) and then had their own R&D plus lawsuit overhead. Most people can use a $20 Claude or ChatGPT sub to build a completely custom Character.ai. Once Character.ai employed MBA Strategy #3 (aka enshittification), churn killed the company (as it should)
---
What doesn't work is attempting to build an AI Wrapper. Those companies had a good run 2022-2025, but with how good both open weights and cheap frontier models have become, AI Wrappers can no longer demand fat margins.
I'm sure people will now argue what exactly AI Wrappers are in the comments below, but for now I'm more interested in understanding what the HN hivemind's take is on the topic as a whole.
It's a bit silly to write an article like this without mentioning the company was hollowed out by Google in 2024.
It is difficult to make a company in a fast-growing sector work when all of your technical staff and leadership are working for Google, and your remaining staff can see the writing on the wall and are checking out.
The headline doesn’t match the story. It is TBD whether the economics of AI actually work or not. A small company (that wasn’t vertically-integrated) going bankrupt doesn’t demonstrate a whole lot, other than that some paths aren’t viable for an AI-based business.
For what it's worth Michael Burry is pretty much short on the entire thing. At some point(I think mid 2027) those data centers bankrolled by loans will have to start paying rent to investors who bought those loans.
If companies that are basing X% of their future earnings on new AI based products don't pay up a bunch of those data centers are not gonna be re-leased. It's gonna be a problem. An seasoned analyst can take educated guesses as to what that X is: investments made in pure chips by Nvidia/AmD vs how much capital investment— not VC — Faang(i.e ones that are just using the chips for building software) are putting into them. The difference between investment devoted to pure DC build out minus what they are spending is a good proxy to understand the overinvestment.
Not worth much. Hasn’t Burry called a number of bubbles since the great financial crisis, none of which have come to pass? Seems like a broken clock to me (he’ll still be right twice a day.)
Does it though? It arguably illustrates how the economics of building custom specialized small sub-frontier LLMs and building an service around them that can easily be replicated by open tools wrapping either frontier LLMs for equal or better quality or personally-controlled open LLMs for privacy doesn't work, especially in a market where privacy is a selling point.
But I think that it is dubious whether that is a good proxy for “the economics of AI” broadly.
Which I am not saying are good—I just don't think Character.ai has much probative value on the broad question on way or the other.
46 comments
[ 0.27 ms ] story [ 12.3 ms ] threadNo-one's going to want to talk to a 7B model for more than 3 minutes
My assumptions are that inference costs continue to decline (and/or the difficulty level of routing requests declines) and hardware acceleration continues into the future. These assumptions might be wrong, but that is the idea the thesis is based around.
(have built workflows and harnesses that currently target LLM provider APIs under the assumption I'll be able to point "locally-ish" eventually, my career had a strong focus on distributed computing about 10-15 years ago, think managing data and work clusters for LHC detector data processing, and I grew up on hobbyist distributed.net and BOINC projects, all the lenses I see this problem through; queues, jobs, sync vs async performance optimization, etc; recovered sysadmin|infra/devops engineer a lifetime ago)
Significant investment is being spent on producing hardware that is more suitable for running mid-sized (100B-ish) models. Starting with flash chips with wide enough interfaces to provide the weights (NAND flash can be ~5x more energy-efficient for reads than DRAM), and going into more exotic solutions like compute-in-memory.
"eventually" is doing a lot of work here. Like sure, yes, eventually. But that could be later this year or it could be 50 years from now. We don't know the answer to that. Assuming it's going to happen in a timeline short enough to matter for this conversation isn't worthwhile. It's like saying eventually we'll have cold fusion, so why is everyone skeptical?
https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo (7B)
Oh, and major players from that industry have been publicly reaching out for people with AI image/video generation and training experience on public forums.
Kind of speculative, and a very bold speculation at that. I don’t have anything else to say, but one company that was obviously mismanaged doesn’t really get my attention.
A simple grammatical mistakes like this should have been caught on review/edit phase.
Either that or writers are relying too much on AI.
(Edit: fixed a grammatical mistake. See?)
It’s capital intensive, doesn’t easily scale without linear costs, is basically a commodity, seemingly has negative cashflow forever, and customers aren’t willing to pay what it actually costs. It’s basically the worst tech business model and yet somehow VCs caught a spell and threw every available penny into the fire.
Watching this all unravel is going to be interesting.
First all the money went into the browsers like Netscape. Then the browser got commoditized, but that didn't mean the value went away. A better business model came along, and lo and behold, the Chrome browser went open source. The browser became free, and the scale of adoption meant a lot of value could still be extracted.
A free browser product didn't mean the value of Chrome went away. Rather the opposite. It's a gateway for buying stuff through Google search.
Marc Andressen was the "Mosaic browser" guy before he was the investor guy, so the folks in the early days did not lose all their money. Quite the opposite. Many are now billionaire oligarchs.
LLMs will likely go the way of the browser... commoditized, then free, then appropriated to shove ads into your face. I predict a lot of safety and interpretability research will be leveraged to one day push ads, just like "Golden Gate Claude" could be manipulated to think of itself as the Golden Gate Bridge.
On the datacenter side the closest analogy is the fiber buildout of the .com bust. While that fiber ended up being useful decades later it’s unclear what value a bunch of half-completed datacenters will bring if their owners go bust.
Holding value is contingent on whether a company can manipulate the regulators to create a lasting monopoly or expand the market dramatically. My suspicion is that at some point the investors will gravitate towards the latter by turning LLMs into more active participants in the economy
Am I understanding this correctly as an idiomatic reference to an enchantment or curse? If so, I'd be interested to know where the expression "catching a spell" comes from.
(I'm really interested in the phrasing and I hope you don't take this the wrong way. For me personally, the metaphorical idea that someone with inexplicable behavior has been bewitched is a normal thing to express, but I'd be more likely to say they "fell under" a spell than that they "caught" it. I believe in Mandarin the usual idiom uses 中 "be struck [by a projectile]".)
That's a big "if".
Even for coding these things are not ready to take away real decision making and it's not clear they will ever get there.
And yes, I'm talking about the landslide of NSFW stuff that had to be culled to turn it into an acquisition-friendly target.
Don't have data... I'm guessing.
A better report would have been how the Economics of AI Wrappers Don't Work.
Character.ai was likely paying frontier LLM API rates (which are known to be ridiculously expensive at 90% margins) and then had their own R&D plus lawsuit overhead. Most people can use a $20 Claude or ChatGPT sub to build a completely custom Character.ai. Once Character.ai employed MBA Strategy #3 (aka enshittification), churn killed the company (as it should)
---
What doesn't work is attempting to build an AI Wrapper. Those companies had a good run 2022-2025, but with how good both open weights and cheap frontier models have become, AI Wrappers can no longer demand fat margins.
I'm sure people will now argue what exactly AI Wrappers are in the comments below, but for now I'm more interested in understanding what the HN hivemind's take is on the topic as a whole.
It is difficult to make a company in a fast-growing sector work when all of your technical staff and leadership are working for Google, and your remaining staff can see the writing on the wall and are checking out.
Character.ai has initiated voluntary Chapter 11 bankruptcy
https://news.ycombinator.com/item?id=50068036
If companies that are basing X% of their future earnings on new AI based products don't pay up a bunch of those data centers are not gonna be re-leased. It's gonna be a problem. An seasoned analyst can take educated guesses as to what that X is: investments made in pure chips by Nvidia/AmD vs how much capital investment— not VC — Faang(i.e ones that are just using the chips for building software) are putting into them. The difference between investment devoted to pure DC build out minus what they are spending is a good proxy to understand the overinvestment.
Not worth much. Hasn’t Burry called a number of bubbles since the great financial crisis, none of which have come to pass? Seems like a broken clock to me (he’ll still be right twice a day.)
But I think that it is dubious whether that is a good proxy for “the economics of AI” broadly.
Which I am not saying are good—I just don't think Character.ai has much probative value on the broad question on way or the other.