Artificial Intelligence thread

A

antwerpery

Junior Member
Registered Member
henrik said:
Should investors sell their holding in nvidia, asml and even some energy companies, because the whole AI trade has been displaced by the Chinese model?
Click to expand...
This is a little off topic but
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!

Yeah, the market is not reacting well to Deepseek. It takes sense for Nvidia and ASML and the A.I companies of course, but even other tangibly related stocks like companies that build power lines are being affected. Not a big surprise considering how fucking big the A.I bubble is. US market hasn't opened yet, so we'll have to see.
 
E

european_guy

Junior Member
Registered Member
magmunta said:
Here is my take; of course, I may very well be wrong. Deepseek released the paper, weights and models to be downloaded. In the paper, they say RL has created R1 and explain how in the paper. They have not published the data and codes for implementing training using the data. So, the paper could be deceptive to make American competitors waste their recourses on the futile method that the paper says.
Click to expand...

Deepseek result has already been replicated:

https://twitter.com/i/web/status/1883183099787571519

Actually these unlucky guys from HK university have been working on a very similar Reinforcement Learning (RL) algorithm since 2 months already....and burned on the finishing line by Deepseek release. Bad luck, it happens.

The important point is that they made reasoning capability to "emerge" from a pre-trained model (the other Chinese Qwen model btw) with only 8K examples processed with RL. The so called "aha moment" (when the LLM for the first time and autonomously, reconsiders its previous output) emerged after just 44 iterations. So it seems confirmed that once you have a decent pretrained model, you just need to run the extra mile to make it think. This has huge implications, no surprise NVIDIA is -14% in pre-market.

P.S: Noteworthy the humbleness and modesty of the author that titled "replicated DS results", instead of clarifying that their effort was a parallel, unrelated work.
 
A

antwerpery

Junior Member
Registered Member
european_guy said:
The important point is that they made reasoning capability to "emerge" from a pre-trained model (the other Chinese Qwen model btw) with only 8K examples processed with RL. The so called "aha moment" (when the LLM for the first time and autonomously, reconsiders its previous output) emerged after just 44 iterations. So it seems confirmed that once you have a decent pretrained model, you just need to run the extra mile to make it think. This has huge implications, no surprise NVIDIA is -14% in pre-market.
Click to expand...
The cope about Deepseek having access to massive amounts of GPU is already here.

Please, Log in or Register to view URLs content!
He reportedly built up a store of Nvida A100 chips, now banned from export to China. Experts believe this collection - which some estimates put at 50,000 - led him to launch DeepSeek, by pairing these chips with cheaper, lower-end ones that are still available to import.
Click to expand...

Even the BBC is reporting that Deepseek has 50k A100s. Which is insane, that's roughly the amount of GPUs that a large company like META or Google has, and it will have caused hundreds of millions of dollars. Not to mention that the A100 is banned and smuggling that amount into China will be massive undertaking, let alone having so many given to a lesser known company at the time. Most of the high end chips would be with Alibaba and Bytedance. Oh and their source?

Please, Log in or Register to view URLs content!

Dylan Patel. Who just pulled that figure out of his ass, unless you can explain how he of all people has secret insider information on a chinese company that only became famous a week ago. This is the BBC, I really expected them to be better than this. Dylan Patel is absolutely seething about Deepseek on twitter btw. I think that he has alot of money invested in A.I tech companies and is about to see a massive drop in his stocks.
 
Temstar

Temstar

Brigadier
Registered Member
antwerpery said:
The cope about Deepseek having access to massive amounts of GPU is already here.

Please, Log in or Register to view URLs content!


Even the BBC is reporting that Deepseek has 50k A100s. Which is insane, that's roughly the amount of GPUs that a large company like META or Google has, and it will have caused hundreds of millions of dollars. Not to mention that the A100 is banned and smuggling that amount into China will be massive undertaking, let alone having so many given to a lesser known company at the time. Most of the high end chips would be with Alibaba and Bytedance. Oh and their source?

Please, Log in or Register to view URLs content!

Dylan Patel. Who just pulled that figure out of his ass, unless you can explain how he of all people has secret insider information on a chinese company that only became famous a week ago. This is the BBC, I really expected them to be better than this. Dylan Patel is absolutely seething about Deepseek on twitter btw. I think that he has alot of money invested in A.I tech companies and is about to see a massive drop in his stocks.
Click to expand...
So either DeepSeek is just that much superior and bursting the US AI bubble is well deserved, or the chip sanction did jack all and China could buy GPUs with no problem.

Not sure which is worse.
 
A

antwerpery

Junior Member
Registered Member
GulfLander said:
Its kinda similar to the narrative used by the indian youtuber, youtube video was posted earlier by a member..
Click to expand...
I was the one that posted him. He was having a meltdown about Deepseek on twitter, so I was reading up on some his older articles and videos for schadenfreude. You would think that people would learn to not underestimate China by now.
Temstar said:
or the chip sanction did jack all and China could buy GPUs with no problem.
Click to expand...
There's no way that 50k A100s got smuggled to China and all got allocated to Deepseek. It's more than possible that there's a few tens of thousands of A100s in China, but they would be spread out among many companies. And a new and inexperienced company like Deepseek wouldn't have access to fucking 50k of them.

And the sanctions forced chinese companies to be more efficient, so I guess the sanctions worked out well in the end
 
C

Clark Gap

Junior Member
Registered Member
tphuang said:
The DeepSeek is pretty sticky and they are very quick with updates. I think longer term, this might be where DS makes their money.

https://twitter.com/i/web/status/1883863740799680922
Click to expand...

Their computing power pool has reached capacity limits, with multiple system crashes occurring today. Given that their ultra-low-cost infrastructure is fundamentally built upon hardware optimizations, they can't simply lease computational resources from third-party providers to meet demand. They'll need to prioritize building new in-house clusters post-New Year, which will likely dominate their operational agenda. Considering they've already made v3 and r1 services freely available, potential monetization strategies could include premium features in APP/web clients such as enhanced function calls, real-time web search integration during AI inference, and multimodal capabilities.
 
E

Eventine

Junior Member
Registered Member
antwerpery said:
This is a little off topic but
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!

Yeah, the market is not reacting well to Deepseek. It takes sense for Nvidia and ASML and the A.I companies of course, but even other tangibly related stocks like companies that build power lines are being affected. Not a big surprise considering how fucking big the A.I bubble is. US market hasn't opened yet, so we'll have to see.
Click to expand...
Investors are panic dumping Nvidia, which is justified given the insane valuation, but in my opinion, premature.

What Deep Seek showed is that - as Google claimed almost a year ago - there is no moat. This is important because the valuation of many AI companies was based on them being able to create a moat and jack up the prices until they're profitable. A race to the bottom on API prices will not be kind to the industry in the short term.

But in the long term, as model sizes increase and algorithms become more sophisticated, a moat will emerge. Not off of technology, but off of infrastructure and incumbent advantage. After all, does Windows have a moat? Does Android? Does Google Search? Does AWS? The answer on the technical side is, not really - others can replicate their results; but on the industry side, absolutely - once consumers and companies have gotten used to a platform and have built an ecosystem around it, they are surprisingly sticky. A new challenger will have to move mountains to convince them to switch - just ask Microsoft on search.

So there is - will be - a moat. It's just that Open AI doesn't have it, because they along with Anthropic are by far the worst priced solutions right now, and must beat out the competition by an order of magnitude to justify their prices.

All the while, Google is rising, and leading the Western race to the bottom on API prices. This is why Sam is on X right now saying o3-mini is going to be free with generous requests / day (though still not nearly as generous as Google). The nature of the competition has just changed - it's now a free for all grab for adoption.
 
You must log in or register to reply here.
Top