Oracle's mega moves in the market (cpl hundred $B in market cap),due to their claim that OpenAI was doing a multi-year commit to move much of their workloads to their cloud ... likely a heavy revenue play rather than profit... with an aspiration to push as much CapEx deprecation to out years as possible (btw Oracle where is all the capex?) AKA Financial aengineering... just shows how overly leveraged this bubble has become.
The Economist recently featured a piece pointing out that it's no longer risk that drives the market but a balance of fear of loss and fear of missing out (https://www.economist.com/finance-and-economics/2025/08/06/w...). FOMO is out of control right now
This Oracle surge and revenue predictions really feels like jumping the shark. I mean, it's Oracle.... I've never felt confident enough to bet against a company, but a short position on Oracle may well be too tempting for me.
This technology demos incredibly well and you can just see how everyone gets giddy with excitement around using it. I watch my colleagues and executives proud to show what they could do or make endless jokes about it. It reminds me of when people first got their phones and couldn't stop showing everyone how cool they were.
This leads to an over rotation in the perceived value.. the value is significant just as the mobile phone was, but not going to live up to the hype in the near term.
It's definitely interesting how in anonymous forums there's a lot more people pointing out that they think this is hype whereas when we wear our professional hats many of us join in. It's like we all want you to party going no we all know what's going to happen
This “article” is clickbait. Controversial title with no substance, asking “why are companies investing heavily in a technology that works for some (limited but valuable) use cases, when they could invest in pure R&D for something that might be better someday”.
Even if every major company in the US spends $100,000 a year on subscriptions and every household spends $20/month, it still doesn't seem like enough return on investment when you factor in inference costs and all the other overhead.
New medical discoveries, maybe? I saw OpenAI's announcement about gpt-bio and iPSCs which was pretty amazing, but there's a very long gap between that and commercialization.
I think what we have right now is an excellent interface for low-friction, non-specialised interaction by humans (work or personal use) with a vast array of highly specialised and highly-complex systems.
What it isn't is the actual final "thing" itself. It's just the thin veneer right now.
I'm not convinced that that revolution was worth whatever trillions we'll end up spending, but fortunately that's not on my shoulders to be worried about.
Apparently the total market capitalisation of the US stock market is $62.8 trillion. Shiller's CAPE ratio for the S & P index is currently about 38 -- CAPE is defined as current price / (earnings, averaged over the trailing 10 years)
That suggests that over the last 10 years, the average earnings of the US stock market is about $1.7 trillion annually.
So $344B of spending is about 1/5 of the average earnings of the total US stock market.
Still hard to interpret that, but 1/5 is an easier number to think about.
They MUST. It doesn't matter if it looks fragile or how much money it is.
LLMs risk most of those companies business, they can't afford to not be ahead. If they aren't ahead, there's a risk that the entire US's economy would be in a terrible shape.
American Big Tech companies that make plenty of INTERNATIONAL revenue from Ads (Meta, Google), can quickly become a shell of its former self.
How? Countries and economic blocks could quickly substitute their American products counterparts if they have nothing to offer and could roll out their own.
The US's economy has become very dependant on FAANG cashflow, it's what gets other parts of the economy moving.
No wonder they had a dinner with Trump. If this fades away, US will look very weak and with a terrible economic outlook.
It's obviously a lot of money, and of course it's too much money, but I think we can still get LLMs much further and I think they're probably the currently most interesting approach.
I don't even care about multimodality etc. I think pure text models are a very appealing idea.
One can call into question the paucity of AI-critical posts & comments here on HN. Not much is being said about the economics, everybody's living off the hope that "they'll figure it out." And AGI is just a few hundred-billion-dollar loans away, so why quit while we're ahead?
If they do "figure it out" (both AGI and a viable business model), a lot of people here will likely be out of a job. If they don't, the whole thing will come crashing down, taking our invested savings with it.
The comparison to crypto keeps coming up. Not everyone's savings went into crypto, but a lot of people's savings and retirement funds are being invested in funds tied to the stock market. And right now its growth depending on pumping cash into the AI bubble.
there is clearly overhype. Given the influx of people building here (and i am not aware if it happened previously too), in a bid to differentiate from other startups building the same thing, many just stripped the nuance out of any technical idea, and made it a simple marketing term. As an exec where ten startups are promising you "training on your data" to provide the best chatbot, it's hard to tell the difference between who would be actually training and who would be just tweaking prompts. This has happened to other concepts too (anecdotally, most famous is how everyone is offering deep research). yes, it helps growth, but comes at the cost of trust. There is this startup which promised "experience based learning" when all they were doing is adding memory to the prompt to get it to perform better. (you can look it up, recenty raised series A).
This does not mean ideas are not working. I personally think pretraining has done its job. We did not know what the job previously was, but now we do given the way RL works. Pretraining and test time compute enables models to develop a generalized prior they can use to solve any given problem (much like how humans solve such problems). Sometimes priors are lacking so you need to train more using RLVR, and still early days, but directionally I think we have another scaling curve here.
18 comments
[ 1601 ms ] story [ 1484 ms ] threadThe Economist recently featured a piece pointing out that it's no longer risk that drives the market but a balance of fear of loss and fear of missing out (https://www.economist.com/finance-and-economics/2025/08/06/w...). FOMO is out of control right now
Exactly. The whole stock market is currently behaving like the crypto bubbles.
This leads to an over rotation in the perceived value.. the value is significant just as the mobile phone was, but not going to live up to the hype in the near term.
It's definitely interesting how in anonymous forums there's a lot more people pointing out that they think this is hype whereas when we wear our professional hats many of us join in. It's like we all want you to party going no we all know what's going to happen
Even if every major company in the US spends $100,000 a year on subscriptions and every household spends $20/month, it still doesn't seem like enough return on investment when you factor in inference costs and all the other overhead.
New medical discoveries, maybe? I saw OpenAI's announcement about gpt-bio and iPSCs which was pretty amazing, but there's a very long gap between that and commercialization.
I'm just wondering what the plan is.
What it isn't is the actual final "thing" itself. It's just the thin veneer right now.
I'm not convinced that that revolution was worth whatever trillions we'll end up spending, but fortunately that's not on my shoulders to be worried about.
Apparently the total market capitalisation of the US stock market is $62.8 trillion. Shiller's CAPE ratio for the S & P index is currently about 38 -- CAPE is defined as current price / (earnings, averaged over the trailing 10 years)
That suggests that over the last 10 years, the average earnings of the US stock market is about $1.7 trillion annually.
So $344B of spending is about 1/5 of the average earnings of the total US stock market.
Still hard to interpret that, but 1/5 is an easier number to think about.
LLMs risk most of those companies business, they can't afford to not be ahead. If they aren't ahead, there's a risk that the entire US's economy would be in a terrible shape.
American Big Tech companies that make plenty of INTERNATIONAL revenue from Ads (Meta, Google), can quickly become a shell of its former self.
How? Countries and economic blocks could quickly substitute their American products counterparts if they have nothing to offer and could roll out their own.
The US's economy has become very dependant on FAANG cashflow, it's what gets other parts of the economy moving.
No wonder they had a dinner with Trump. If this fades away, US will look very weak and with a terrible economic outlook.
LLMs are a million times better than Crypto currencies.
I don't even care about multimodality etc. I think pure text models are a very appealing idea.
... I mean, of course they haven't. They are a natural consequence of how the things work!
If they do "figure it out" (both AGI and a viable business model), a lot of people here will likely be out of a job. If they don't, the whole thing will come crashing down, taking our invested savings with it.
The comparison to crypto keeps coming up. Not everyone's savings went into crypto, but a lot of people's savings and retirement funds are being invested in funds tied to the stock market. And right now its growth depending on pumping cash into the AI bubble.
This does not mean ideas are not working. I personally think pretraining has done its job. We did not know what the job previously was, but now we do given the way RL works. Pretraining and test time compute enables models to develop a generalized prior they can use to solve any given problem (much like how humans solve such problems). Sometimes priors are lacking so you need to train more using RLVR, and still early days, but directionally I think we have another scaling curve here.