The Subprime AI Crisis
wheresyoured.at
wheresyoured.at
Even if all AI investment froze tomorrow, I'd still have my 405B Llama 3.1 model, along with countless other smaller models, and I'd run them to do whatever the heck I felt like doing, with no commitment to any provider.
Writing code with AI? I could swap to a local model. Costs nothing. Provides no revenue to any VC-backed company.
Yes, bigger models will always command a premium for the highest end of reasoning. But you don't always need the best possible reasoning. GPT-3 and early GPT-4 were more than good enough for a ton of use cases last year.
And we've seen the pace of development in the open source world these past two years. The open source community (Meta in particular) has completely obliterated the commercial value of these models.
If there were no "weights available" models, OpenAI would have an incredible, unbeatable moat, and they would be worth truly astounding amounts of money. But as it stands, we all have free, unfettered access to local models far better and cheaper than models that were flagship 18 months ago, and close enough to the performance of current flagship models that it won't make a difference for a ton of use cases.
There is no way to justify the current level of investment, with local models being freely available.
The assumption I'm making, of course, is that this transformer technology won't ever lead to AGI - it will just be another tool in our ever-expanding tool-belt.
Costs nothing? I'd have to buy a desktop computer with a good GPU to run that model. How much would that cost?
As of 2022, Uber had supposedly raised a total of $25.2B in funding over 32 rounds. So they may be ahead in the money-drain business.
"There is no way to justify the current level of investment, with local models being freely available."
The business problem OpenAI faces is that their systems aren't good enough that users can trust the results. Slightly crappier results are also available at much lower cost. Unless OpenAI can definitively fix the "hallucination" problem, and at least return "Don't know" when appropriate, this isn't going to work at OpenAI's price point.
Chatbots are about as good as low-end outsourced call centers. They can sort of help with programming. The systems that generate pictures can do some impressive things. LLMs produce better blithering than most bloggers. It's really impressive. But, absent a major theoretical breakthrough, that's not enough to fund OpenAI.
But every hyperscale data center owner is investing for themselves, major companies capable of the R&D are investing in their own science, and the market for LLMs is slow to materialize relative to the hype. That's money that might have gone to OpenAI if their situation was similar to Uber's.
Hallucinations don't need to be fixed in one go as much as improved progressively. There is some magic threshold where they'll stop being an issue for specific tasks, or alternatively they simply become more reliable than humans. The problem will die with a whimper.
I'm not entirely cynical on the value of LLMs, but I've yet to see one say "I dont know", or "I'm not sure, but here's my best guess".
I've used LLMs to build form autofilling based on unstructured documents. It correctly does not answer fields that it doesn't know, and does not try to guess anything. It has been pretty much error-free.
It's all about your prompting. Without explicitly being given guidance on how not to answer, you're right, they will never say they don't know.
If you've never actually seen that happen, I encourage you to experiment more with LLMs; there's lots that can be achieved with the right prompting.
And your assertion is that when this happens, it's because the human operator is hallucinating?
It depends what you're doing with it. Answering arbitrary questions is very different from document summarization or from mapping arbitrary questions onto a list of canned questions to then answer.
It also doesn't have to be perfect. Instead the cost of an overall system (AI + error reporting & handling) just has to be cheaper than that same overall system using whatever the AI is replacing.
If it makes errors 3% of the time, and people from Mechanical Turk make errors 2% of the time, its usefulness depends on whether or not those 50% more errors cost more than the cost savings from paying a model provider rather than paying humans.
I wonder at what point Grok might become a serious competitor to OpenAI and Anthropic?
The larger models aren't perfect, but worlds apart from Llama.
It's been a while since we've had a major tech company commodify/open source a major dependency (and not just layer some rent seeking pay/subscription service over existing assets). Meta seems to have the best long term plan for LLMs.
AWS only brings in 80B+ a year, must be from everyone self-hosting. I'm curious why people think this is different?
I think the point was that the things LLMs are useful for can increasingly be done with freely available models you can run on your own hardware, hardware you rent from someone else, or via a 3rd party's hosted service. Want to run Meta's models on Google's server? Nothing stopping you.
If they aren't making massive profits right now then there may well be a problem.
I would imagine lobbying efforts and scaremongering is already underway. This is a powerful technology that the general public (and small businesses) might not have access to in the near future.
Not really, as Claude is competitive both on quality and price.
(agree with the rest of your comment)
I do think this moat can still exist though, as providing a product is different than running a model somewhere, even if you have the compute power to allow prompting by a large amount of people. And making a model ready for the business user is quite a bit of work.
Although true, the capabilities of commercial products will probably be eclipsed sooner rather than later. We saw that in image generation. Crowd sourcing the problem did produce better results. Generating videos is probably still locked behind enormous compute power (well, the 400+B model of Llama is too...)
I think the focus of these companies isn't the performance of the model itself, it is providing interfaces to all kinds of systems and maybe solve a specific task.
It takes time for new advancements to get proliferated through the economy.
Where was this claimed? The author quotes OpenAI’s multi-billion dollar revenues.
The article does mention that OpenAI has huge revenue.
> While The Information reported that OpenAI's revenue is $3.5 to $4.5 billion a year in July, The New York Times reported last week that OpenAI's annual revenues have "now topped $2 billion," which would mean that the end-of-year numbers will likely trend toward the lower end of the estimate.
But then the author claims that the business value is questionable.
> And even if they did, it isn't clear whether generative AI actually provides much business value at all. The Information reported last week that customers of Microsoft's 365 suite [snip]
I would have appreciated a deeper discussion of why OpenAI's revenue isn't a data point toward generative AI having some business value. Presumably if nobody was using generative AI in a way that gives them value, OpenAI wouldn't be using all those GPU hours. That's what I was missing from the article personally.
> Based on how unprofitable they are, I hypothesize that if OpenAI or Anthropic charged prices closer to their actual costs, there would be a ten-to-a-hundred-times increase in the price of API calls, though it's impossible to say how much without the actual numbers.
Lots of revenue is flowing into generative AI but would that trend continue if they started needing to actually cover costs? And how much of that revenue is from AI companies that would pop right alongside them?
Firstly, the revenue numbers are rumors. I have no doubt OpenAI has significant revenue, but both reported numbers are imprecise suggesting they likely are estimates at best.
Furthermore, a non-trivial amount of OpenAI's revenue is certainly coming from other AI startups. They in turn are likely burning investor cash, which isnt an indication that OpenAI is providing business value, its an indication that they provide a tool to speculate on future value. Or, more cynically, the provide a tool for companies to convince investors to give them more money.
>I believe that a lot of businesses are "trying" AI at the moment, and once those trials end (Gartner predicts that 30% of generative AI projects will be abandoned after their proof of concepts by end of 2025), they'll likely stop paying for the extra features, or stop integrating generative AI into their companies' products.
Making AI pictures of Pee Wee Herman riding a shark blindfolded is indeed fun, but not profitable for Microsoft because it's too hardware intensive for ads to cover. I gotta make more goofy pics before the bottom falls out...
If one looks at most the bubbles of the past, the writing is on the wall. AI won't go away, but will probably take longer to make profitable than anticipated, just like dot-coms and smart-speakers. Force feeding it is causing indigestion, and it's likely to PukeGPT.
The job losses have already happened. Companies have laid off quite a bit of employees because they wanted to get ahead of the AI wave. They thought they could replace most of their engineers, and turn 1x into 10x. The only field that has benefited is the parasitical companies that have sprung up around these AI services trying to rentseek their way to profitability. So when the bubble bursts, laid off talent will be able to demand a premium to come back and fix the smoldering remains.
That said, it's still going to cause reasonably bad damage as a whole because so much of the tech industry is dependent on angel investors which behaves in almost cult-like ways when it comes to trying to find something to fund.
Is that really a good example of LLM capabilities? LLMs don't even see those letters because of tokenization.
It's a bit like asking a Chinese speaker questions about imaginary Latin alphabet letters in Han characters. Sure, it demonstrates a limitation, but it's a bit of an edge case.
Here's another example of a simple question that a state of the art model (Claude 3.5) gets wrong, as tested just now:
Prompt: How many words are in this sentence?
Response: This sentence contains 7 words.
Interestingly, it seemed to count the number of words in my sentence correctly, but still answered incorrectly.
It only indicates that tokenization is used. The LLM architecture can be used without tokenization. It would just mean that the available space in the context window would be used less efficiently. For example, a 10 letter word which could be represented by one token would instead take 10 slots.
That's not the relevant figure. It's more about how much capital has been wagered by large tech cos and the major startups & VCs on a near-complete takeover of white-collar work by AI (as that's the only level of payoff that could possibly justify the mind-boggling level of investment).
The implication is if AI is a bust, Google, Microsoft and Meta alone would have to run leaner organisations to make sense. Each of them is massively betting on LLMs are their core growth engine.
Management will try to hold on to the capital for as long as shareholders let them. But if the growth prospects from AI significantly diminish, 20,000 jobs in tech is a pretty conservative estimate for lay-offs.
Maybe OpenAI doesn't make it and maybe it turns out way too early for AGI ... but there is way too much stuff is working right now in multiple domains to say the industry will flop.
The second issue I see is that data is becoming widely unavailable-everything is getting blocked, crawlers are getting cut off immediately, 1000 API calls cost as much as a mid-sized flat in a European capital - the complete opposite of what the internet was supposed to be.
The third issue is the fact that people are gladly using chatgpt to answer questions on stackoverflow for example. We know how badly llms start performing when you train them on their own data... Not to mention the considerable spike in critical bugs in open source projects over the last two years-there's a good chance it a lot to do with it.
The fourth is social media-there are already tons of examples where troll farms are no longer paid workers in some dump in siberia but are in fact powered by chatgpt out some other service. And to think that I laughed at the dead internet theory when I first heard about it...
My problem with AI boom seems is that the insane valuations of these AI companies seems to be based on nothing more than the power of the better funded AI companies to outbid possible competitors for limited amounts of chips available from chip foundries, ie TSMC et al.
If there is a glut of chips or the models become more refined and efficient then what power do these companies then have?
Need to write a report to present at your company but nobody will care about it? Use AI.
Blog posts just for SEO? Use AI.
Illustration image as a header for a post where you need an image just to share the post on social? Use AI.
I am not saying people is in the right to do this, but this is where I saw AI being really useful at.
I use ai-chat quite a bit and the better it gets, I feel more threatened but that is because it leaves me some free time to think like that. And every other month when ai-chat starts spewing garbage answers, I feel pissed at the AI for making me do my research, but it gives me heart warm knowing that this shit cannot replace me.
I have also come to realise that AI needs to be trained to give you correct answers and cannot simply innovate on its own, which is what it needs to be "revolutionary". Also, our entire tech industry is based on products that do deterministic information retrieval. Whether that is getting accurate numbers from a bank account, or computing medical parameters or the velocity of incoming missiles from a bunch of formulae. AI on the other hand seems like tech that will give out answers like "the sum of 1 and 2 is 3 with 99.9% probability".
In any case, these are all just feelings and though I find myself nodding along with the article, there is no information here that is concrete.
Meta and others have released open weights with the claim that they compare to GPT-4; I imagine these are good enough for many of the similar tasks. There are bound to be at least a few more improvements in open weights before bust.
Apple is already building laptops with a mind towards local AI. As NVidia's and AMD's et al AI chips drop in price, they will be included in regular desktops and laptops.
While local-AI becomes more practical, the prices on remote-AI will go up, further driving local-AI improvements. Perhaps at some point will will have a subscription based weights service, where you get updated proprietary weights for your local model for $X/yr.
Local-AI will be fine for Microsoft. And for Google. So, I don't think AI is going to disappear; If anything, it will become more ubiquitous. Weights may start being released less frequently and the SaaS model may go, but that would likely be a net gain.
AGI to end the humanity will need to be financed by the Chinese Communist Party, though.
This will further alter the personality, intent, and product of Silicon Valley; refuseniks will be winnowed, and many people may find themselves working on projects they may find disturbing.
I couldn't agree more.
But my hunch is that NVIDIA is overpriced, with a P/E ratio of 55. It is not a growth company.
Maybe it was counting the seconds _mississipily_.