OpenAI Is a Bad Business
wheresyoured.at
wheresyoured.at
But now even regulators are working for them - the more regulations are there, the harder it will be for late comers to join.
OpenAI's spending is mostly buying compute from other people. In other words, OpenAI's growth is paying for Microsoft's data centers. The only real asset OpenAI is building are their models. While they may be the best models available today it is unclear if that those any durable advantage since everyone in the industry is advancing very quickly, and it is unclear if they can really monetize effectively them without the infrastructure to run them.
Which Zuckerberg is doing his best to completely turn them into commodities.
By giving them out for free??
This sort of incentives are what makes unicorns possible.
Also, there are immense lobbying everywhere covertly, if you haven’t seen it, look at the prominent news media, AI is now akin to “productivity gain” and “innovation”, and “adapting AI or we fall behind” attitude.
Look at all the hammers falling on different industries about energy waste, climate change while AI’s usage of massive energy is presented as “innovation and next generation of work and productivity gains to be unlocked”.
The target for OAI(and AI wave in general) is normal people, not tech people. And when people say AI, you conveniently see OAI emerge from news, social media, your colleagues, your boss, your friends.
And what it usually looks like in practice is that regulators will extend an open invite to "industry leaders" to chime in on the proposed regulation. During that process, the interest groups advocating on behalf of the "industry leaders" will utter statements that aim to shape the regulation in a way that is beneficial to the corporation being lobbied for.
The end goal is typically to land on a new bill that will
a) give the regulators new fodder for legislative resume, so they can show their voting constituency that they accomplished something and
b) are not too expensive for the "industry leaders" to comply with. During the regulatory process this will get shrouded in language along the lines of making sure that the regulation won't cost jobs etc. But, if all goes according to plan, the regulation will
c) be insanely expensive for new startups to comply with, thus ensuring that any negative consequences that are caused by the new bill will be "invisible" because they will largely apply to hypotheticals (new startups wanting to enter the market) and small players that aren't on anyone's radar
Market ideologues have both fought regulation and allowed the cozy situations encompassed by regulation capture (allowing the "revolving door" between industry and government based on "who else would know how to regulate? lol")
I mean it is a little.
Amazon didn't extend its losses while growing it's buisness. Yes it made losses (quite large ones for the time) but the underlying principle was pretty solid. They had a market that clearing going to be profitable, because the overheads were much less.
But openAI doesn't have that market. They offer something free that is comparable with the competitors. A paid for option that isn't much better than the free version, and isn't much different in capability from the competition.
Moreover, in order to remain competitive they need to spend _billions_.
Even if we ignore meta kneecaping them, anthorpic et al are yapping at their heels.
The market isn't actually that solid, there isnt anywhere enough cash to maintain the research spend.
"OpenAI (...) has yet to create a truly meaningful product"
Almost everybody I know uses ChatGPT at least semi-regularly for all kinds of things. I'm not sure if it's possible to look at this objectively and claim that ChatGPT isn't "a truly meaningful product."
"OpenAI loses money every single time that somebody uses their product"
I'm not sure if this is true. My understanding is that they (on average, including their free users) make money on usage, but lose money on training. But I could be wrong.
I also think OpenAI isn't yet serious about monetizing its products. You can just go to ChatGPT and use it for free. They have a ton of data about their users that their users give them freely. The free ChatGPT tier is a prime candidate for advertising.
You could be 100% right, and it still could be an unsustainable business. Depending on how much training and R&D costs.
Doesn't OpenAI need to continue training just to remain relevant? Not even just because there is a race between multiple AI companies, but because new information is regularly published. And if that's the case, it's just part of the cost, right?
I have never been quite sure if it is good practice. It does make you vulnerable to a flop.
OpenAI seems to be in a similar position.
The absolute values or the deals performed are irrelevant to the principle. The principle is that revenues and expenditure go up in a way that never results in a profit, but if you broke each generation of development into a separate company it would be a series of profitable investments requiring increasingly larger investment for the next company. If serving ChatGPT stays profitable enough to cover the training of the model that they are currently serving then they are financially secure. It will only become a problem if they cannot recover the cost of training each model by serving it. We don't yet know if that will happen.
There is insufficient data in this area, Without knowing the capital expenditure on hardware, the costs of running the hardware, and the breakdown between how much of that hardware is running inference and how much is training, we just don't have a reliable way to calculate potential profitability. It's been a long time since GPT4, and all releases since then have been improvements on that generation. Some of those improvements have significantly reduced the cost of inference. There is evidence that they are spending on the next two generations already. We don't yet know how capable the next generation model will be, in the absence of that data point, almost all of the speculation is relatively meaningless.
>but I don't see any indicator that OpenAI will not go the route of enshittifiaction if it gets the opportunity
There is a very significant indicator. We live in a world where it is commonplace to offer free services paid for by advertising. This is the model where the user is not the customer, it is the product. That is the most common path to enshitification, because pleasing your customer takes priority. OpenAI has not yet taken that path (and they have had the opportunity). Both the API and ChatGPT plus work on the old fashioned model of give us some money and we'll give you a service. That is a model that disincentivises screwing over your customers.
It may not stay this way, but they certainly have had opportunities for enshitification that they have chosen not to embrace.
Ed's perspective is from a nonengineering side, not the typical Hacker News user or LLM API consumer.
> My understanding is that they (on average, including their free users) make money on usage, but lose money on training.
The billion-dollar YoY losses can only be possible with negative marginal profits. Training is expensive but not that expensive.
> I also think OpenAI isn't yet serious about monetizing its products. You can just go to ChatGPT and use it for free
That's just modern software business: have the user try it for free (loss-leading) with an incentive to upsell to a paid product, notably ChatGPT+. And if OpenAI didn't do it, others will as is currently the case now. But in a post ZIRP world that doesn't work as well.
> Ed's perspective is from a nonengineering side, not the typical Hacker News user or LLM API consumer.
ChatGPT has 200M weekly active users.
Diminishing returns means that the user gets less marginal benefit with each larger model, and the model collapse phenomenon means that models trained on new training data might be less good than older models. Have straightforward mitigations been put in place such as filtering out from the training data forums where users like to share AI generated content?
That's why OpenAI/Sam Altman has been memeing AGI. None of this will work unless they make God.
Count me in the camp that doesn't and wishes fewer people did. I just asked ChatGPT what it knows about me, and apparently I am known for my work in the field of AI. I've apparently contributed to multiple AI projects. I have a globally-unique name and have done absolutely nothing of the sort.
I don't dispute that there are good uses for it right now, but people taking this stuff at face value for anything important is terrifying.
You're likely to get more value out of it if you try to do something useful with it rather than asking it to answer a question you know ex ante it cannot know the answer for.
I had absolutely zero reason to believe it would mischaracterize me in advance of asking that question. In fact I presumed that would be an easy gimme: simply Googling me is enough to provide a pretty accurate assessment even before following any of the links.
Presumably you already know about yourself, yet that was the example you chose to pick. It comes across as you basically leading the witness just to be able to say say "gotcha! I knew you'd suck".
It's not a search engine, although you can ask some models/chatbots to search the web. Again, you're holding it wrong.
Instead of trying to confirm your bias that LLMs are bad, try to find discomforming evidence.
“The human understanding when it has once adopted an opinion, draws all things else to support and agree with it. And though there be a greater number and weight of instances to be found on the other side, yet these it either neglects and despises, or else by some distinction sets aside and rejects.”
— Francis Bacon
With all due respect, how the hell else am I supposed to evaluate its ability to answer accurately on topics without testing it first on topics I know about?
> Instead of trying to confirm your bias that LLMs are bad, try to find discomforming evidence.
I wasn't trying to confirm my bias. I literally thought that with all the information OpenAI has collected throughout the web they can probably give me a pretty good summary of the publicly available information about me. They probably even have stuff that's a bit more obscure or difficult to find. This seemed like an extremely easy way for me to be impressed.
I was trying to find disconforming evidence. Yet again I was thoroughly disappointed.
If spot-checking AI with softball questions is considered a "gotcha", that is an extremely damning comment on the current state of things.
Ask it some programming question, to draft an email, to create a marketing campaign, to help you dig deeper into a research topic, I don't know.. things an LLM was built for.
Tool-use frameworks with data retrieval tools add can be wrapped around an LLM to make the combined system try to do what you describe, but even the best of those is a lot less reliable than the underlying LLM is at its core functions like summarizing directly provided information.
Yeah, that's something language models are exceptionally bad at, and, yeah, ChatGPT is more than an LLM it is a UI around an LLM core that has some other facilities for data retrieval, but, still way out of its area of strength (unfortunately, that seems to be a lot of what people want LLM-centered products like ChatGPT to be.)
Yeah, I'd be more comfortable if it could say "I don't know" instead of generating bullshit.
Simon Willison's article Think of language models like ChatGPT as a “calculator for words”[1] helped me think about this topic better:
> If you ask an LLM a question, it will answer it—no matter what the question! Using them as an alternative to a search engine such as Google is one of the most obvious applications—and for a lot of queries this works just fine. It’s also going to quickly get you into trouble.
All of this is why I only every use LLMs for things that I can verify.
[1]: https://simonwillison.net/2023/Apr/2/calculator-for-words/
I simply do not believe you. Most people I speak to attempt to use it periodically, find that for any promising use case it fails completely and then give up for another few weeks.
Well, I can't exactly prove it to you. The thing I would offer as evidence, though, is that most of my friends have white-collar jobs. They spend most days sitting in offices writing stuff.
Let me give you an example. One of my friends works in a legal department at a bank. She often gets to review things like ads or promotions the bank runs. These often contain things that are illegal, because the marketing team is not made up of people who have a legal background, and many things that seem like reasonable ideas are illegal. So a lot of her work consists of explaining to people why they can't do the thing they want to do.
She's not particularly good at doing that in an empathetic way, so she's started using ChatGPT to outline these emails for her. "The marketing team at my bank wants to run a competition where people who open a new account get a chance of winning a prize. This is illegal, since it violates sweepstakes laws (see below, where I've pasted the applicable sections). They either have to provide a way to enter the contest freely, without opening an account, or they have to abandon the idea. Can you write an email that explains this in an empathetic, but clear way?"
Obviously, she then has to read the email ChatGPT writes and edit it, but it still saves her time and creates a better end result.
"for any promising use case it fails completely"
I simply do not believe you :-)
Just this paragraph but without the instructions to ChatGPT is sufficient for this case. It's literally more work to get a useful output from ChatGPT than just to do the work yourself. And that's not even accounting for the careful review you need to give every morsel of output from it.
This feels like opinion stated as fact.
I don't write business language well, but I am able to communicate much more effectively with business people if I run it through an LLM to "translate" for me. I do some editing on the output, and I've saved 10 minutes that I would have otherwise agonized over whether or not I've used the right tone while emailing another department.
"I've fixed the scrolling bug in the calendar component, and it should be published after lunch. Can you let Alex know I'll be around until 5 if he wants to go over it?"
becomes
"I have resolved the scrolling issue in the calendar component, and it is scheduled for publication this afternoon. Please let Alex know that I am available until 5 p.m. today if he would like to review it together."
It depends on who you're talking to.
But either way, the point is that people do use LLMs, whether you agree with how they use them or not.
Open AI don't need to report any numbers to see it's incredibly widely used. The site has had over 1.5B visits every month since March 2023.
I am yet to hear of any YC company that succeeded and pointed at Altman as a key factor.
OpenAI only got first mover advantage because it was seen as separate entity that had a higher purpose. Now that the truth is out and others realized that Altman was playing a con the whole time, the competition is catching up and turning them into commodities.
Why do people still put so much stock into him?
And that's before factoring in what other models they are cooking that they're in no hurry to release.
If ChatGPT is so easy to outcompete, where's the competition?
All that feedback has made their product better and they're at a very different scale than everyone else.
The more interesting thing is when the user tests the AI ideas. Like running some code or applying advice for your hobbies. The user returns again and again communicating the outcomes in order to get new advice. The model can collect all chats across many days by topic, and judge them in retrospective. Which answers were good or bad?
The user base is so large, 200M users means lots of perspectives, lots of ideas, chatGPT puts trillions of tokens in human brains per month. They influence on a grand scale. After a while some article will describe the outcomes, and that data percolates back in the next training set. A feedback loop. And an indirect agent, working through people. An experience flywheel, collecting from hundreds of millions and serving back.
Only few use cases need the best models. Most of them work ok on mini models and LLaMAs. The problem is that it costs very much to develop this shrinking segment of high end. With each new release in open / smaller models, the island of superiority shrinks for OpenAI. But people's task don't get any more difficult on average. So eventually your phone model will suffice for 99% of use cases, and be free and private. What will OpenAI do with 1% of the market?
These "catch all" use cases sound cool, but in practice are nearly worthless. Because you'll spend more time verifying than if you just did it yourself. The compelling use cases are the human-less use cases. Imagine ridding your entire accounting department because it can be handled by highly specific LLMs and calculators.
It takes like 10 seconds.
You’re just using it to point you in a direction, not solve the entire problem.
That's just not a compelling use case. It makes far more sense to just hire junior engineers. Because then they're actual engineers, and their knowledge base will grow and transfer over time.
If you have proprietary systems, then it may make sense to train an LLM on those. But I would consider that a more "focused" use case. I just don't see the main value of AI being a glorified, generalized google search.
There's potential here to automate tasks with MUCH less friction. Now business folk and people who understand processes (but not algorithms) can potentially automate business processes.
Yesterday, I implemented a recursive PostgreSQL function to crunch some pretty complex tables to produce a new output. As a senior engineer, I knew the output I wanted and understood the trade offs, but I didn’t want to spend 3 whole days trying to figure out how to build the recursive function myself.
Claude did it correctly in about 5 minutes, and then I spent another hour or two just tweaking and exploring alternate options.
I think it also goes without saying that I could not handle this to a junior engineer. We’re quite clearly past the “GPT is only as good as a junior engineer” phase.
It’s so different from anything else that’s ever existed. You could not do what I did yesterday without spending a LOT more time. Google would not have solved that problem, nor would a junior engineer.
This hasn't been the case in my experience.
I guess the moral of that story is that OpenAI may be profitable once they start injecting Temu ads into ChatGPT responses. Profitability, but at what cost?
They appear to already be working on something along those lines: https://news.ycombinator.com/item?id=41658837
Throwing money away isn't the clever bit in the Google story.
It’s called Survivorship bias
People mocking that waste of money were 99% right, and Google is one of the very few exceptions.
Looking at this arbitrary chart from a quick search[1], OpenAI is likely still at the "young growth" stage
[1] https://vlp.teju-finance.com/courses/pluginfile.php/1844/mod...
Both of those two points alone emphasise the fact that comparing the accelerated growth of a tech company in 2024 to the likes of a Sega Megadrive in the 90s, for example, isn’t a like for like comparison.
I don't know how to go more in depth now, but I use Claude, 4o, and o1, both mini and preview in parallel and o1 both mini and preview succeed in many tasks where the others fail.
Most things that succeed have a very clear value proposition that justifies their cost.
But I watch, and if your envisioned future arrives, that will be interesting.
LLMs are still prone to hallucinations, and we're only adding more data and workarounds to make it happen less often. Prompt injections limit their usefulness on untrusted inputs. They can't do precise logic and reasoning, and are too likely to follow memorized patterns instead.
We've got something amazing, way better than what we've had before, but the current architecture is still based on a fuzzy translator. It's hard to say whether this is it, and it's going to plateau at this level for a while, or whether there are more breakthroughs around the corner.
This is the only thing that matters imo and the biggest open question to me. All the math surrounding this is irrelevant if being the premier genAI API/service is something you can charge good money for (to businesses who likely won't scrimp). Winning that market share battle is worth burning billions for most likely and they are ahead of everyone else.
All that feedback has made their product better and they're at a very different scale than everyone else.
I hold a neutral position on this, I will sit and watch how this plays out.
I had predicted that Uber would not become profitable. But it did eventually.
You can keep the bus rolling for a surprisingly long time before the wheels fall off.
We need to realize that, the target fir current AI wave is NOT tech folks(claude/mistral/cursor et al. serve that group), but business people who want the “gain in productivity”, “fire the expensive employees and get that pat on the back when margins go high and same work get done by AI and some cheap junior mini team”, “quickly prepare that keynote/email/report for random topic”, “want to learn language X by building a wrapper over API” etc.
That part, and the points about subscriptions and losses on the dollar comes in later though. Maybe there is something I don’t understand, but I almost stopped reading the article when it touched on the SoftBank investment. How is that even remotely relevant?
Nowhere does the author explain why on earth OpenAI would be desperate to accept what they call “dumb money”. Maybe if it was enough money to buy leadership, but it obviously isn’t. So what is the issue here exactly?
There's no need to compete directly. Yet.
> "If OpenAl disappeared tomorrow, we have all the IP rights and all the capability. We have the people, we have the compute, we have the data, we have everything. We are below them, above them, around them."
~ Satya Nadella
By allowing their search algorithm to atrophy they have given up this moat to the point that others can breach it with LLMs that give similar results without the usage data.
Then they expanded into an ecosystem, e.g. Gmail. That was a moat.
All that feedback has made their product better and they're at a very different scale than everyone else.
> OpenAI has not had anything truly important since the launch of GPT-3.5
> [OpenAI] has yet to create a truly meaningful product outside of Sam Altman's marketing expertise
So 4, 4o, o-1, DALL-E, ChatGPT, and their API platform aren't meaningful products?
> To be abundantly clear, as it stands, OpenAI currently spends $2.35 to make $1.
Uh, that's how venture capital works? Uber burnt cash like nobody's business, took SoftBank money, and is now a $175B public profitable company.
> It’s extremely worrying that the biggest player in the game only makes $1 billion (less than 30% of its revenue) from providing access to their models.
Name a single API that hit $1b in revenue in 2-4 years...
The macroeconomics of venture capital are a bit different now than in the 2010's. OpenAI may be the one case too finanically extreme for conventional VC logic.
If you are growing revenue to $B in <5 years, and creating the fastest growing consumer app in world history, then 2010's VC rules still very much apply today.
That said, I reckon the vast majority of their subscribers neither read nor care about the true philosophical implications of their terms of use, which gives them millions of people paying to get “brain raped.” Imitating users, stealing users’ jobs, and banning users who attempt to compete with them, can keep OpenAI around for a long time, because that’s the big tech equivalent of taking souls. Not something anyone with morals wants to take part in, but surely possible for that to evolve into a profitable capitalist business model long term.
"A bunch of monopolistic companies" is an oxymoron. The entire race-to-the-bottom with pricing is because of the lack of moat and open-source pressure.
Their terms of use words it as:
> Use Output to develop models that compete with OpenAI.
I think this is acceptable for 99.99% of users.
> in an underhanded way
Could you elaborate on why you see that as underhanded? It's not secret, as it's a bullet point in their TOS, and I don't see how it's dishonest. Billions were put into this model. Not letting the competitors save a good chunk of that seems fairly reasonable, from a "we need to make the billions back eventually" business perspective.
I would claim there's a much better chance that Llama is the more underhanded model, existing partly to wipe out/impede competition, including OpenAI, before going behind a pay wall.
At least if Meta makes a great Llama model (since 3.1) both sides can benefit and it doesn’t have a ridiculous customer noncompete, and I can use the model on my own computer as much as I like.
The difference is, one side has a noncompete, and the other is competitive.
I think Llama is closer to, "Good luck trying to make a profit with us giving this stuff away for free. Oh look, there aren't any real competitors left! Introducing Llama 5.0, with 10x the fee and vendor lock in! Btw, thanks for all the research that made it possible!"
This is also expected behavior for a large tech company trying to fight for market dominance. To see similar examples that match this template, see the history of Intel and Microsoft.