Today's weird possible AI bubble pop - brought to you by an AI breakthrough?

So sometime earlier today, the world started collectively losing its E36 M3 over a new LLM model from a Chinese company called DeepSeek when its client app reached the top of the mobile app stores' most popular downloads after a recent new model release. It's not only a big leap forward from other models, but took about an order of magnitude less computing time in training than other leading models and orders of magnitude less money, and is almost as open as an LLM can be. Since then, NVIDIA's stock value has dropped by over half a trillion dollars, probably due to concerns that not only will the AI industry's endless hunger for graphical computing power become rather satiable, but that the (likely temporary) computing power loss of bypassing Nvidia's "CUDA moat" and using another manufacturer's GPUs or more specialized types of processors for training instead could now be entirely acceptable. Indirectly AI-related industries got 10-20% of their value lopped off today as well. It also gives the appearance that every AI company other than DeepSeek has been doing it all wrong in comparison.

Previously it might have been presumed that an AI bubble pop, caused by running into the hard realities of actually turning a profit from this whole energy-and-cash-guzzling AI business, would make AI use vastly more expensive overnight as investor dollars cease to prop up the affordability of giving customers their first hit, but the fact that this new model is highly open means that anyone could download it and run it on an ordinary gaming PC or anything else with sufficient GPU power, no subscription fees necessary. So the result could be an explosion of AI use instead. Even access to DeepSeek's hosted access plans is vastly cheaper than the existing options.

The fact that Deepseek used relatively little resources to build this and is making so much of it open suggests that they have something much bigger in store.

I don't want to get too much into the STONKS aspects of the issue, but the biggest takeaway may be that the Nvidia AI training hegemony is showing a lot of big cracks.

Some good articles on the subject:

https://www.cbc.ca/news/business/deepseek-ai-startup-1.7442382

https://hackaday.com/2025/01/27/new-open-source-deepseek-v3-language-model-making-waves/

On some questions from the "wild job market" thread:

aircooled said:

In reply to GameboyRMH :

[May] be more accurate to say the US AI bubble has been deflated a bit, by the Chinese demonstrating a cheaper alternative?

So far the US AI bubble has been the world AI bubble. From what I can find all indications are that DeepSeek is a privately held startup, if it were publicly traded today would've been very interesting indeed.

aircooled said:

In reply to GameboyRMH :

I would certainly hope, even if the Chinese are doing it "better" that western countries would not jump in with Chinese based AI! (as in contributing to it's knowledge base)  Of course, it still means it's rather helpful to China.  You might be able to clarify the situation better.

Using it doesn't necessarily mean contributing to its knowledge base, especially if it's used on privately hosted instances, but using DeepSeek's own hosted service certainly would. Just using the Internet basically means you contribute to all of the models out there being trained on anything that's technically downloadable, like this forum we've all made posts on. That said, if this thing takes off, keeping American businesses off of it won't be a whole lot easier than keeping American kids off of TikTok.

A DeepSeek vs. ChatGPT review. No further progress on the hallucination problem it seems:

https://www.wired.com/story/deepseek-chatbot-hands-on-vs-chatgpt/

GameboyRMH said:

That said, if this thing takes off, keeping American businesses off of it won't be a whole lot easier than keeping American kids off of TikTok.

Unlikely. At my company, we are already not allowed to use any "Chatbot" but our own internal LLM, due to fears of proprietary information going to outside servers. China as a whole are masters of stealing IP. When I worked for TWG, we quit doing business with Zoomlion because it was obvious they were trying to steal our IP and reverse engineer our electronic products. 

It's already blocking information about the CCP, Ji Xinping, the situation with the Uyghurs, and Taiwan, etc. 

And given China's history for complete openness and transparency, I'm not buying the $6 million development cost. Just like I don't trust their "tofu dreg" built infrastructure.

Other news account say this was mostly built on open source tools like Meta Llama and is more of a case of "Let's see if a cut rate LLM can do 90% of what a big budget one does for a quarter of the price." A lot of the big names in AI have been pursuing diminishing returns with spending a lot for tiny incremental improvements.

Deep Seek's main accomplishment seem to be demonstrating you can jump off that treadmill and build a workable LLM with mostly off the shelf tools on a budget.

Other possibility: They're lying about how many resources they took to train it.

"Breakthru" events in the computer world are so common I am surprised how strongly the market responded.  It seems to validate my theory that the stock market is emotion driven to a high degree.  A month from now when there is more truth about this app should be interesting.  Speaking of interesting, consider the timing of the release and the maybe coming soon tarif war.

AI is sure sounding a lot like BitCoin.

VolvoHeretic said:

AI is sure sounding a lot like BitCoin.

I used to feel the same way until I really started using it. If you know how to properly construct prompts, you'd be surprised at the amount of useful information you can get. 

In reply to z31maniac :

My internet search experience has been going downhill for years but with the new buzz about AI it has taken a nosedive.

VolvoHeretic said:

AI is sure sounding a lot like BitCoin.

I assume you meant that negatively, but have you seen the price of Bitcoin over the past couple of years?

MadScientistMatt said:

Other news account say this was mostly built on open source tools like Meta Llama and is more of a case of "Let's see if a cut rate LLM can do 90% of what a big budget one does for a quarter of the price." A lot of the big names in AI have been pursuing diminishing returns with spending a lot for tiny incremental improvements.

Deep Seek's main accomplishment seem to be demonstrating you can jump off that treadmill and build a workable LLM with mostly off the shelf tools on a budget.

 

Yep. I think that the DeepSink news timed into a skittish AI market that needed an excuse to adjust.

It seems a bit strange to me that anything open source or freely useable would come out off China(?) Seems kind of suspicious to me, but, as noted, if it sucks someone into using their servers, that could be a tactic.

Also as noted, the company I work for is very paranoid about using any AI.  No actual company data or processes can be put through one.  Even those that are (at least claimed to be) closed systems.

In reply to VolvoHeretic :

Unrelated to the topic at hand, but I pay to search with Kagi. Every once in a while I use Google out of curiosity, and I don't think I could ever go back. Kagi is decent at getting "real" results from forum posts, some dude's blog whom happens to be an expert on what I'm searching for, etc. Google just returns blogspam and SEOd fake review sites. Highly recommend checking out Kagi.

porschenut said:

"Breakthru" events in the computer world are so common I am surprised how strongly the market responded.  It seems to validate my theory that the stock market is emotion driven to a high degree.  A month from now when there is more truth about this app should be interesting.  Speaking of interesting, consider the timing of the release and the maybe coming soon tarif war.

Exactly what I was thinking reading about this last night.

The news is saying there is an app for this that is being downloaded heavily, so it seems like they are getting a lot of input / data into their servers from those that download the app at least.

In reply to aircooled :

The timing is suspect.

Huge amount of downloads (by Americans) at nearly the exact moment a different Chinese app was set to be outlawed?  

It seems they (Chinese gov.) wanted to figure out a way to stay on American phones.

What are the chances that DeepSeek is actually just a bunch of underpaid Chinese workers in a nondescript office tower answering questions?

spedracer said:

In reply to VolvoHeretic :

Unrelated to the topic at hand, but I pay to search with Kagi. Every once in a while I use Google out of curiosity, and I don't think I could ever go back. Kagi is decent at getting "real" results from forum posts, some dude's blog whom happens to be an expert on what I'm searching for, etc. Google just returns blogspam and SEOd fake review sites. Highly recommend checking out Kagi.

I'll give them a try. The traditional search engines seem more gummed up than ever with useless results.

So, is the stock market really run by a bunch of computers running algorithms and when one starts selling, the rest follow suit and also sell causing the market to crash until someone pulls the plug which brings a brief pause and then resets and the cycle continues?

Is AI already involved with all of those decisions?

We are doomed.

In reply to Indy - Guy :

I am sure it will be fine... just relax.. go with it...

JG Pasterjak said:

What are the chances that DeepSeek is actually just a bunch of underpaid Chinese workers in a nondescript office tower answering questions?

I don't think that would possibly work... It's never been done before?

VolvoHeretic said:

So, is the stock market really run by a bunch of computers running algorithms and when one starts selling, the rest follow suit and also sell causing the market to crash until someone pulls the plug which brings a brief pause and then resets and the cycle continues?

Is AI already involved with all of those decisions?

We are doomed.

Computer aided trading has been a thing for decades. In 2009 Spread Networks, built a fiber optic connection from Chicago to to New Jersey to drop trading times from 17 milliseconds to 13 milliseconds. 

A little intro:

https://www.forbes.com/forbes/2010/0927/outfront-netscape-jim-barksdale-daniel-spivey-wall-street-speed-war.html

JG Pasterjak said:

What are the chances that DeepSeek is actually just a bunch of underpaid Chinese workers in a nondescript office tower answering questions?

Get out of my brain!  I was thinking that exact same thing when I heard about it!

There is no such thing as a "privately held startup" in China.