I think we’re about to get a crash in 5 hours folks
The companies known as the Magnificent Seven make up over 20% of the global stock market. And a lot of this is based on their perceived advantage when it comes to artificial intelligence (AI).
The big US tech firms hold all the aces when it comes to cash and computing power. But DeepSeek – a Chinese AI lab – seems to be showing this isn’t the advantage investors once thought it was.
DeepSeek doesn’t have access to the most advanced chips from Nvidia (NASDAQ:NVDA). Despite this, it has built a reasoning model that is outperforming its US counterparts – at a fraction of the cost.
Investors might be wondering about how seriously to take this. But Microsoft (NASDAQ:MSFT) CEO Satya Nadella is treating DeepSeek as the real deal at the World Economic Forum in Davos:
“It’s super impressive how effectively they’ve built a compute-efficient, open-source model. Developments like DeepSeek’s should be taken very seriously.”
Whatever happens with share prices, I think investors should take one thing away from the emergence of DeepSeek. When it comes to AI, competitive advantages just aren’t as robust as they might initially look.
I haven’t followed developments in AI all that closely. In a very general sense, it seems to me that the potential benefits of AI (or high-powered automation) are limitless. I think it’s hard to fully comprehend all the practical uses it will have and how it could transform the economy.
HOWEVER, currently the only applications I seem to see are: 1.) ChatGPT-like applications, i.e. ask a question and get a response that at best, writes up some text it would have taken you much longer to write or get a maybe right maybe wrong answer, or 2.) shitty art that can’t even make hands right; and even if it did, it’s just images. Not exactly economically impactful like the spinning jenny.
So… is what I said above really just how AI is being used in the US, and is that the reason for the huge bubble in asset values of companies like Nvidia and Microsoft?
my analysis is roughly the same as yours. I also think 80% of the buzz around AI is just that… every US tech company has been rebranding existing services as somehow incorporating “AI”. I don’t watch a ton of commercials, but when I do see them, tech firms seem to all be investor-oriented hype messaging about how IBM or Apple or Dell Systems or whoever are using “AI” to do/make computer better-good for happy human.
everyone is just saying “AI” about everything.
to me, the market wealth associated is a speculative bubble around a term that has become meaningless.
of course an open source project from China is blowing proprietary performance benchmarking out of the water. the proprietary black box is always smoke and mirrors to attract capital to the secret, inscrutable sauce that is really just watered down ketchup and mayo.
So LLM’s the “AI” that everyone is typically talking about are really good at one statistical thing:
“CLASSIFYING”
What is “CLASSIFYING” you ask? Well it’s basically attempting to take a data and put it into specific boxes. If you want to classify all the dogs you could classify them based on breed for example. LLMs are really good at classifying better than anything we’ve ever made and they adapt very well to new scenarios and create emergent classifications of data fed to them.
However they are not good at basically anything else. The “generation” that these LLMs do is based on the classifier and the model, which basically generates responses based on statistically what the next word is. So for example it’s entirely possible that if you fed an LLM the entirety of Shakespeare and only Shakespeare and you gave it “Two households both alike” as a prompt, it practically may spit out the rest or Romeo and Juliet.
However this means AI’s are not good at the following:
Don’t get me wrong, yes this is a solution in search of problem. But the real reason that there is a bubble in the US for these things is because companies are making that bubble on purpose. The reason isn’t even rooted in any economic reality. The reason is rooted in protectionism. If it takes a small lake of water and 10 data centers to run ChatGPT, that means it’s unlikely you will lose a competitive edge because you are misleading your competition. If every year you need more and more compute to run the models it concentrates who can run them and who ultimately has control of them. This is what the market has been doing for about 3 years now. This is what DeepSeek has undone.
The similarities to BitCoin and crypto bubbles are very obvious in the sense that the mining network is controlled by whoever has the most compute. Etherium specifically decided to cut out the “middle man” of who owns compute and basically says whoever pays into the network’s central bank the most controls the network.
This is what ‘tech as assets’ means practically. Inflate your asset as much as possible regardless of it’s technical usefulness.
As a sidenote “putting things in boxes” is the very thing that itself upholds bourgeois democracies and national boundaries as well.
I think this distinction is interesting because if the product of AI is classification it is just a more abstract method of continuing to do the very thing we have been doing for a long time with other statistical methods that fragment data into sets of whatever we wish to draw boundaries to. Essentially it sounds like intensified and far more depersonalized categorizing.
I mean at this raw of an argument you might as well argue for Lysenkoism because unlike Darwinism/Mendelian selection it doesn’t “put things in boxes”. In practice things are put in boxes all the time, it’s how most systems work. The reality is that as communists we need to mediate the negative effects of the fact that things are in boxes, not ignore the reality that things are in boxes.
The failure of capitalism is the fact that it’s systems of meaning making converge into the arbitrage of things in boxes. At the end of the day this is actually the most difficult part of building communism, the Soviet Union throughout it’s history still fell ill with the “things in boxes” disease. It’s how you get addicted to slave labor, it’s how you make political missteps because it’s so easy to put people in a “kulak” in a box that doesn’t even mean anything anymore, it’s how you start disagreements with other communist nations because you really insist that they should put certain things into a certain box.
I’m not arguing for anything and did not suggest that putting things in boxes is good or bad, but that it just is what we do.
And that these models just do the same, but with more abstraction.
No not, actually. The bubble is in the idea that AI requires large amounts of power, cooling, and processing throughput to achieve things like the current OpenAI O1 Reasoning and Logic models. The circle is like this:
The New AI Model Is Bigger --> Needs Bigger Hardware --> Bigger Hardware Needs Better Cooling --> More Cooling and Bigger Hardware Needs More Power --> More Cooling and Bigger Hardware means we can train the next Bigger Model --> Back to Start
So long as the Newest AI model is “bigger” then the last AI model, then everyone in this chain keeps making more money and has higher demand.
However, what Deepseek has done is put out an equivalent to the newest AI model that:
A) Required less up front money to train,
B) Uses considerably less resources than the previous model,
C) Is released on an Open Source MIT License, so anyone can host the model for their use case.
Now the whole snake is unraveling because all this investment that was being dumped into power, cooling, and hardware initiatives are fucked because less power and cooling is required, and older hardware can run the model.
The fact that the US model can eat shit as soon as somebody figures out a way to make it work better and faster instead of forcing a bunch of bloat and ad-infinitum upgrading is hilarious to me.
If that’s not a perfect distillation of the infinitely wasteful US economy I don’t know what is.
That was a great explanation, thanks!
I’d like some LLM features centered around “here is my organized folder of documents for context” but without paying $20/month or buying a $2,000 GPU or giving all of my data to Google/Microsoft. I couldn’t find anything that actually worked even if I did pay money or give my data to Google/Microsoft. Ignoring even the folder thing, I remember asking Claude to summarize a novel I’m writing and it kept mixing character’s with details that unambiguously only applied to other characters.
It doesn’t work in the average case. I’ve seen this tactic from the company that I work for and multiple companies I have contacts at. Bosses think they can simply use “AI” to fix their hollowed out documentation, on-boarding, employee education systems by pushing a bunch of half correct, barely legible “documentation” through an LLM.
It just spits out garbage for 90% of people doing this. It’s a garbage in garbage out process. In order for it to even be useful you need a specific type of LLM (a RAG) and for your documentation to be high quality.
Here’s an example project: https://github.com/snexus/llm-search
The demo works well because it uses a well documented open source library. It’s also not a guarantee that it won’t hallucinate or get mixed up. A RAG works simply by priming the generator with “context” related to your query, if your model weights are strong enough your context won’t outweigh the allure of statistical hallucination.
I’ve seen one or two companies doing exactly this, but specific to certain varieties of bureaucracies (mostly insurance companies, unfortunately). It struck me as one of the few potential real uses for LLMs so long as it can provide a bibliography for any responses that can be confirmed deterministically.
It’s great at digging thru piles of documents looking for specific things. CGP grey (lib) has an old video called “humans need not apply” that touches on this.