Analysts said the Monday sell-off underscores anxieties about whether the massive spending on artificial intelligence ― and the specialized chips, data centers and related power infrastructure ― are justified....
DeepSeek is a China-based start-up that last week launched a free AI assistant that it says can operate at a lower cost than American AI models like ChatGPT.... DeepSeek has shaken the market because it purports to need fewer and less advanced chips than other AI models, while still performing as well as U.S. rivals — challenging the premise that only big, well-capitalized companies can make breakthroughs in the sector.
২৭ জানুয়ারী, ২০২৫
"China’s DeepSeek AI app sends U.S. tech stocks reeling/The tech-heavy Nasdaq index lost nearly 4 percent in early trading Monday, with chipmaker Nvidia down nearly 12 percent.."
WaPo reports (free-access link).
এতে সদস্যতা:
মন্তব্যগুলি পোস্ট করুন (Atom)
পোস্টগুলি
৫৩টি মন্তব্য:
This is great news for consumers.
The Singularity Is Near has turned into The Singularity Is Fear.
The market will bounce back. DeepSeek has the same flaws that inhibit other Chinese tech, particularly the built-in bias towards omitting any subject matter that reflects poorly on the CCP. And the CCP being paranoid authoritarians see a LOT of subjects as being off-limits. But if you just want a cheap Chinese knockoff of a "thinking" machine, sure go ahead.
It could be an issue...if you take both the Chinese and WaPo at face value. Don't recommend that.
I'm just going to say that this is another attempt by the CCP to mine technical expertise from the west. Rather odd that this comes on the heels of our deregulating TikTok.
US technology stocks get bubbly and then crash every generation or two (e.g., Great Depression; early 1970s; late 1980s; early 2000s). International stocks / emerging markets bubble and then crash every generation or two as well.
International investing strikes me as overdue for a simple momentum boom, and US tech overdue for a correction. China, as sometimes an emerging market, may bubble up for 10 or 20 years. Compare this to Japan's Nikkei bubble of the 1980s. Everything in investing is half fad and half fact.
China has flooded our universities with their top students, and they are getting trained in all our most advanced technologies, including AI and quantum computing. In a few more years they won't need to steal our technology secrets, they'll be developing the latest advancements themselves.
The good news though is that they will be shocked by how much more communist our universities are compared to theirs, so they'll start withdrawing their students from them.
We can't trust the ChiComs with anything.
They are the enemy, they are implacable, and their advantage is headcount - they can swarm any productive effort with people that are trained observers. The educational institutions are their happy hunting ground, as there is a lot of grant-type collaboration on the cutting edge between industry, government, and academy in those places.
They have improved their skill at stealing Intellectual Property, turning it around, and issuing it as a final product out in front of the original I.P. developers. The defense industry isn't very different. I've already been reading that this is probably one of those efforts.
Just my uninformed opinion, but I think anyone who is making financial decisions based on this news is an idiot. We need at least two weeks, and probably more to see if there is anything to this.
Or
If it sounds too good to be true it probably isn't. Although I admit "probably" is doing a lot of work in that sentence.
...challenging the premise that only big, well-capitalized companies can make breakthroughs in the sector.
It's hard to discern Biden policy. Marc Andreessen alleged the covert Biden policy was for "government control AI to the point of being market makers, allowing only a couple of companies who cooperated with the government to thrive." Meanwhile, at Politico: "Biden official worried AI development is too reliant on a few tech giants."
After famed investor Marc Andreessen met with government officials about the future of tech last May, he was “very scared” and described the meetings as “absolutely horrifying.” These meetings played a key role on why he endorsed Trump, he told journalist Bari Weiss this week on her podcast.
What scared him most was what some said about the government’s role in AI, and what he described as a young staff who were “radicalized” and “out for blood” and whose policy ideas would be “damaging” to his and Silicon Valley’s interests.
He walked away believing they endorsed having the government control AI to the point of being market makers, allowing only a couple of companies who cooperated with the government to thrive. He felt they discouraged his investments in AI. “They actually said flat out to us, ‘don’t do AI startups like, don’t fund AI startups,” he said.
Open source. Cheaper chips and running, so maybe you don't need Nvidia chips and nuclear plants to implement AI. Just a licensing fee to DeepSeek. If you're developing an accounting or legal or medical app, do you really care that it won't let you ask about Xi or the CCP?
Markets do what they're going to do. Lately, they've been overvalued; in particular the big tech stocks.
"DeepSeek has shaken the market because it purports to need fewer and less advanced chips than other AI models, while still performing as well as U.S. rivals — challenging the premise that only big, well-capitalized companies can make breakthroughs in the sector."
If true, how, in the long run, is this not a good thing? However, we only have China's word for it at the moment; in particular, their claim of lower training costs.
I just want to thank the second commenter for pointing out that Jews are to blame for this. That is the type of deep thinking that makes this site so worthwhile.
Assuming anybody's going to be consuming anything after the massive layoffs.
My understanding is one of the issues with teaching an AI system is access to data. There's a lot of free data on the Internet, but there is also a lot of data protected from AI learning. The developers have to get permission and pay license fees to use it. Long process and expensive. The Chinese don't give two fs and use it freely. That's why they were able to train DeepSeek so quickly and cheaply.
There's some talk about whether the information is really true or another Chinese psy-op.
"We need at least two weeks, and probably more to see if there is anything to this."
Sane words Duty of Inquiry.
The people making money on this aren't interested in the long-term impact. They're just playing with other people's short-term reactions. Buy on the rumor, sell on the news. Or in this case, short on the report, clean up when the real story comes out.
Really? They can tank the stock markets and keep on killing kids during a ceasefire? Those Jews are like a mighty super race or something!
I read a "Twitter" post at 2am last night that I can no longer find. In it a researcher who had reviewed Deepseek, indicated that the Chinese had done two things differently , that improved their results significantly. One as I recall was their approach to reinforcement learning, the second improvement I can not presently recollect. He went onto say that the US labs have made hundreds of small, meaningful improvements over the past couple of years.
He predicted major gains would happen quickly when the US labs combined the Chinese improvements with those of their own. Basic thrust was the Chinese are very smart, but it is still early innings.
Probably the Chinese are writing in assembler, which is typically 3 times faster than compiled code.
Also many commentators predict that cheaper AI will only increase demand.
Western economies will not be using any Chinese-developed AI products due to substantial security issues.
No businesses will willingly feed in corporate or personal information to a Chinese-backed AI software product.
Nice journalistic hyperbole today but nothing to see here, aside from a little 1.5% correction to the S&P all time high.
I can’t believe Trump's inner circle didn't see this coming. After all they are the tech titans. Or maybe they're just rich. Oh well.
Looks like that comment was removed.
China was the world leader in technology prior to the European Rennaisance of the 1500s. Do not underestimate their potential, nor that of Japan and Korea. The main force working against East Asia is aging and its population crash. But still, China has >1B people, so even with a crash they match the US and Europe.
A certain troll is being aggressively purged by Althouse today.
I would expect this is an excuse to push tech stocks lower from someone. This seems completely overblown, plus china is a**hole.
With the large drop in nVidia stock today it's a good thing Nancy Pelosi sold a bunch of her shares recently.
But I'm sure she had no advance knowledge of what was going on.
"DeepSeek has the same flaws that inhibit other Chinese tech, particularly the built-in bias towards omitting any subject matter..."
It's open source. Anybody can run it, anybody with the technical knowledge can understand what it is doing. I have a Mac workstation with four physical CPUs, many more "virtual" ones, that could probably run it pretty easily. Just not on a massive scale, but maybe that's not necessary for most purposes.
It's open source. I get a feeling that you don't really understand the stuff you write about.
Somehow worrying that there will be all of this spare computing capacity reminds me of the worries of "dark fiber" and the days when IBM said that there would only be a need for six computers in the world.
I still think that quantum computing is largely hype, I am not sure that one exists that could balance your checkbook, even as they can solve problems in quantum physics that can't be solved by conventional computers due to the extensive calculations required, but if they ever reached the level of their hype, well that will be a world shattering event.
It is? Exactly how much AI have you been consuming?
As I said it will be used to mine more advanced technology by the CCP. Nothing coming out of China is used to benefit anybody outside the CCP.
If we were bothered about TikTok's data gathering activities, just wait til we heat what DeepSeek is doing. Literally every question you ask is being added to your profile and the overall knowledge stored by the CCP. You can learn an awful lot and make a bunch of logical connections from the questions someone asks.
Do you remember the "Internet of Things"? Everything was going to talk to everything else, over the internet. Your refrigerator would tell your television to remind you that you hadn't closed the door. The data from the sensors in your car's engine would be sent to your mechanic for real-time analysis. The reasons Big Tech hyped IoT were
a) If you want to be a CEO, you have to do the "vision thing". You can't just say, "It's a great company, with great engineers, and I'm sure they'll think of something to sell". You need a shiny object to show the board.
b) IoT was concocted so as to make use of exactly those new features which big tech had figured out how to deploy - massive bandwidth and edge computing.
AI is the new IoT. It uses exactly what Nvidia has figured out how to produce; vast amounts of low-precision floating-point arithmetic. What makes AI different, is that it does things the general public thinks are cool. Nobody was really very excited about having his refrigerator talk to his television. But AI makes pretty pictures and tells pretty lies. A lot of rather non-technical people have somehow concluded that AI is "smart". They treat it like some kind of oracle. Look at Althouse. At least one AI post a day.
The problem is, AI does not know anything. AI tries to make something that is like the other things it was "trained" on. But "like" does not include any truth function. 2 + 2 = 5 is "like" 2 + 2 = 4. AI is great for making pretty pictures, but you don't want it managing your bank account.
It can also be used to mine more advanced tech by the US. It has no master. It is open source. You seem to exist to repeat propaganda tropes designed to keep us in a fighting frame of mind, and to not think about peaceful coexistence.
Only the Chinese do psy-ops, BTW. The Chinese and the Russians. We never do them.
DeepSeek is open source, or rather, open weights, and cheap. Things are changing so fast that we have leader of the month. How things play out long term, who knows. I don't expect the same sort of monopolies as happened with Google and Microsoft, but of course, I could be wrong.
The ability for AI to reason is coming along nicely. Current AI isn't that different from everyday conversation, "2 + 2 = 5" isn't unusual on the internet.
"Compare this to Japan's Nikkei bubble of the 1980s."
What crushed Japan's "bubble" was that we pressured them to implement "self imposed" restrictions on exports to US markets, which amounted to the equivalent of a massive tariff on Japanese goods. This is how ChatGPT puts it, and it comports with an article I had read:
Voluntary Export Restraints (VERs) on Japanese cars in the early 1980s functioned much like a hidden tariff: although no formal duty was imposed, the export quota drove up U.S. prices on Japanese autos. Most estimates put the “tariff equivalent” of these self‐imposed export limits in the range of about 20–25 percent. - ChatGPT
I get the feeling you don't understand that open source is not a magical privacy wand.
What amazes me is that people must believe that the ai data centers full of powerful energy hungry nvidia chips must be somehow useless and unnecessary. Oh, Chinese people are so smart they can replicate using simple chips, and we believe. Have the technical papers been published with prediction accuracy and loss computations? I checked, and no. So we believe that a trillion dollar industry is stupid because China makes a press release!!
Training models is what the big deal is, inference is the easy cheap part. Make sure you watch the pea when the hucksters make the play.
The greatest problem is that everyone is looking at the general models trained on all the modeler can grab from the internet. The ip provenance of that data is being questioned by the owners, it’s now largely being used for free, especially by China and its expert hackers.
Until the press release is turned into some reported facts, I will just consider this an insider pump and dump allowing the market kings to shuck the rubes.
Chat gpt and the like are still toys, entertainment. Real working ai will use expert data, not the sweepings of the web. Ai for professions like accounting, law, engineering etc. will be trained on expert data, with a lot of recommender and expert system ad ins.
It is the training that is power intensive. There is a trend to use already trained AI to train it's own replacements, and that seems to yield good results.
And I really have no doubt that ai for the professions will work very well, engineers will have ai assistants to design bridges, and it will not just be an llm. It will be many things, operations research optimization, Physics, and economics bundled as a work flow from site evaluation to cutting the ribbon.
@Jaq: I suspect market participants that are overreacting today will look back on this as a good buying opportunity -- similar to how some bought NVDA under $100 during the Summer
DeepSeek may have found a workaround to export restrictions (as other Chinese companies have), but cannot admit to that in its white paper. Even if it didn't, the $6M training figure excluded a lot of other expenses (e.g. purchase of its chips, R&D, data wrangling, etc.) that the general media are not reporting on.
Nevertheless, if everything being reported were 100% accurate, the reduction in costs might accelerate demand for inference as companies holding off previously due to ROI concerns may be able to justify the experimentation with these tools now.
"that open source is not a magical privacy wand"
Is it that Chinese code is "inscrutable"?
Being an open-weight model, which you can run in your own open-source runtime, or integrate into your app using an open source library, it can’t phone China with your data unless you build in that capability yourself. So, no, not “substantial security issues”. Not even being code, it simply doesn’t have the ability.
Of course, being a product of China means that it will have specific blinders and biases, that you have to account for if you want to use it. Though, you can also fine tune it with your own data to get rid of those biases if that’s important to you, and I’m sure you’ll have options to choose from in the not too distant future as others do just that.
And, as with any other model, you have to validate it for your own use cases.
I’ve been playing with the 70B and 32B versions of the model myself, and the results are remarkably good for the model size, comparable quality-wise to the published frontier models in many respects.
that it says can operate at a lower cost than American AI models like ChatGPT.... DeepSeek has shaken the market because it purports to need fewer and less advanced chips than other AI models, while still performing as well as U.S. rivals
So they say. Any evidence to back it or are we just trusting China to tell us the truth?
"open source" is not a panacea and the root programming still goes back to the originator who will let you see what they want you to see.
If it comes out of China, especially now that they are preparing the battle space for war with Taiwan and by extension us, then this is part of that preparation. China is the CCP.
Deepseek has accomplished its goals of hurting the US economy, and providing the CCP another door into US personal data streams. And it won't even say anything about 1989 Chinese history or CCP mass genocide!
একটি মন্তব্য পোস্ট করুন