DeepSeek Rattles Wall Street With Claims of Cheaper AI Breakthroughs 152
Chinese AI startup DeepSeek is challenging U.S. tech giants with claims it can deliver performance comparable to leading AI models at a fraction of the cost, sparking debate among Wall Street analysts about the industry's massive spending plans. While Jefferies warns that DeepSeek's efficient approach "punctures some of the capex euphoria" following Meta and Microsoft's $60 billion commitments this year, Citi questions whether such results were achieved without advanced GPUs.
Goldman Sachs suggests the development could reshape competition by lowering barriers to entry for startups. Founded in 2023 by former hedge fund executive Liang Wenfeng, DeepSeek's open-source models have gained traction with its mobile app topping charts across major markets. DeepSeek's latest AI model had sparked over $1 trillion rout in US and European technology stocks Monday, before even the U.S. market opened.
Goldman Sachs suggests the development could reshape competition by lowering barriers to entry for startups. Founded in 2023 by former hedge fund executive Liang Wenfeng, DeepSeek's open-source models have gained traction with its mobile app topping charts across major markets. DeepSeek's latest AI model had sparked over $1 trillion rout in US and European technology stocks Monday, before even the U.S. market opened.
Prove it (Score:2)
Re:Prove it (Score:4, Insightful)
The stock markets *are* insanely overpriced and I'm honestly shaken that they have again and again reached these insane, unrealistic peaks over the past decade. What irony it would be if a fresh Chinese startup is the domino piece that triggers the next stock market crash.
Re: (Score:2)
The stock markets *are* insanely overpriced and I'm honestly shaken that they have again and again reached these insane, unrealistic peaks over the past decade. What irony it would be if a fresh Chinese startup is the domino piece that triggers the next stock market crash.
I'm not sure about "insanely" overpriced. The rivers of profits are flowing quite nicely.
But we're probably due for several years of very low, if any, growth in the stock market, as prices normalize.
Re:Prove it (Score:4, Insightful)
You know what a multi-billion breakthrough in AI would look like to me? Natural language coding.
Re:Prove it (Score:5, Insightful)
I think we fundamentally disagree on the value of AI. I think we've barely scratched the surface and its use will keep growing.
I could be wrong! It'll be interesting to see where it goes.
Re: (Score:2)
Re: (Score:3)
It comes from the ability to aim eyeballs, and the fact that companies pay a truly silly amount of money for those aimed eyeballs.
It's one of the shittiest of business models, but also one of the most tried and true.
Re: (Score:2)
I think you're right. I also think that it's likely to cause massive social disruption. Lots of people appear, to me, to be "playing John Henry". (Yeah, he could drive those stakes deeper than the steam engine, but they didn't need to be quite that deep. And it was a lot easier to build more machines than to make equally capable copies of John Henry. Also racing against the machine killed him, but it didn't kill the machine.)
OTOH, I also think that current LLMs don't have the vaguest idea what they're
Re: (Score:2)
I am uncertain here. Nothing currently with LLM AI looks valuable, it does things that are either fluff (chatbots for laughs) or can be done with a good web search. Yes, some other AI is amazing, but gets ignored once the LLM caught people's attention. The Google translate for instance is quite handy, and it would be a good use for LLM as a component (LLM being designed specifically for natural language processing), but it's also very good without it. Image analysis is light years ahead of where it was
Re:Prove it (Score:5, Insightful)
On one hand, you claim that the market is not driven by fundamentals, and on the other hand, you say that "social media and AI" are [almost nothing of value]
. In 2019, Social Media was a $219 billion dollar a year business, expected to hit $251 by this year.
NVidia is expected to sell ~$50 billion USD of H100s.
To summarize, you claim that the market is not driven by fundamentals, implying that the fundamentals of those markets are not healthy, when in truth, they are fucking absurdly so.
Now, you may disagree with the products on some philosophical level- and I certainly agree with you on the matter of Social Media, but to somehow deny that they're not making money hand over fucking fist, and that that isn't a market fundamental, is pure idiocy.
In short, future earnings potential is a market fundamental. It is the market fundamental.
You are otherwise correct that this obscene-scale money funneling isn't innovative, but innovation is not a market fundamental.
Re:Prove it (Score:4, Insightful)
The stock market is overvalued in that, to get to a reasonable P/E of 20, earnings would have to increase by about 50% or prices would have to drop by about 33%.
Re: (Score:2)
Re: (Score:2)
And you left out an absolutely absurd claim by the parent poster: "The last actual innovation that came from tech was a smartphone."
Re: (Score:3)
Gotta pick and choose your battles.
Re: (Score:2)
In short, future earnings potential is a market fundamental. It is the market fundamental.
By the same logic, pets.com [wikipedia.org] is now the world's most valuable company, because future valuation is the same thing as earnings.
Market fundamentals is value that company produces - goods and services. It is typically reflected in P/E ratio, where 9-10 is an established company that produces good value. Look up P/E for any of the social media companies or AI companies or even NVIDIA. It is 100s to 1000s (hint: more is not better in P/E).
Re: (Score:2)
By the same logic, pets.com [wikipedia.org] is now the world's most valuable company, because future valuation is the same thing as earnings.
I'll take a non-sequitur for $100, Alex, lol.
Market fundamentals is value that company produces - goods and services.
Incorrect.
That is one tiny bit of market fundamentals.
By your reasoning, an entity that, for example, statutorily must be paid by people in return for nothing has no value.
This is of course nonsense, since such an entity has a fucking amazing P/E.
It is typically reflected in P/E ratio
What an absurd statement.
In the P/E ratio, the P is 100% in the control of the company.
They set the P/E by adjusting P (via splits). This makes it more or less attractive to particular investor types.
If you're an e
Re: Prove it (Score:2)
P/E is price per share to ernings per share so splitting shares has no effect. https://www.investopedia.com/t... [investopedia.com]
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
Natural language coding.
Albeit a flawed experience, this is essentially what experiences like Copilot are already doing. You use natural language to tell it what you want, then it spits out changes that you can accept or decline. Yes the output is a traditional computer language, but that's because 'natural language' programming is fundamentally imprecise and therefore unusable. And even if you could get a computer to reliably compile it while filling in the blanks in ways that make sense to the authors, you still have the challen
Re:Prove it (Score:5, Insightful)
Current prices aren't about profits, but about future earnings potential. This is very visible in everything from Tesla to Meta to Google, to tech in general. It is why much most stable companies that offer greater return on investment aren't as expensive. No such future potential.
What opensourcing DeepSeek models did is change the view of future potential. And with it comes re-evaluation of the companies who's value comes from future potential.
Re: (Score:2)
Future potential = fantasy. One of our great strengths as a species was the ability to believe in fantasy to the point that we could manifest it into reality. But that only works for things where the collective buys in *AND* the fantasy is something that is actually achievable. Will the prognosticators of AI be proven right and most of humanity will be thrown out of the "must work to provide for self" norm of today? I have my doubts it'll happen on anything resembling their timelines. It will be disruptive
Re: (Score:2)
What you call "fantasy" is actually "ability to comprehend future". It's something that humanity only gained recently, out of the Ancient Egypt > Ancient Greece > Ancient Rome, where Egyptians realized that sacrifice need not be like for like (alchemical concept of equal value trade, commonly expressed as human sacrifice for future success in most culture, i.e. trading human life for future human life), Greeks created a theory out of this finding, and Romans put it to test against reality and made pra
Re: (Score:2)
What you call "fantasy" is actually "ability to comprehend future". It's something that humanity only gained recently, out of the Ancient Egypt > Ancient Greece > Ancient Rome, where Egyptians realized that sacrifice need not be like for like (alchemical concept of equal value trade, commonly expressed as human sacrifice for future success in most culture, i.e. trading human life for future human life), Greeks created a theory out of this finding, and Romans put it to test against reality and made practical applications.
It underlines all human success today, and is the primary explaining factor why people in Africa fail to do even most basic infrastructure maintenance left to them by others, while Chinese built a successful civilization out of rubble of Cultural Revolution and Great Leap Forward.
So you're massively underselling "fantasy".
I said we have the ability to make the fantasy real. If you think that's underselling it, that's you not really getting what I said, not me underselling it.
We literally create our reality by building on the fantasy until it's real. But it doesn't always work. And sometimes the grifters grab hold of that innate tendency to build our fantasies and ride the wave to large sums of cash for no actual progress. That's the universe we exist in now, and I'm of the belief that this is what the AI push is. A giant gri
Re: (Score:2)
We don't. We are not using our minds to bend reality. That is a complete and utter mystical misrepresentation of reality, that is common to cultures that weren't touched by the innovation I described. This is how you get places like Rhodesia with excellent infrastructure just degrade as locals tear infrastructure up to sell it as raw materials, sell books from school libraries and so on. It is the ancient understanding of how "future" works, that is utterly unsuitable for modern world.
Which is my point. You
Re:Prove [the price is right] (Score:2)
Interesting branch, but my take is that the stock prices are just opinions. One question is whose opinion, though these days it may mostly be a computer that is just faking the having of opinion. But more interesting and dangerous are two other aspects. One is that the the prices are (normally) determined by sales involving only a tiny number of the existing shares. The second is probably more dangerous. That's the lack of any transaction charge. The shares are traded in an imaginary universe of zero cost p
Re: (Score:2)
What you miss in this understanding is that fundamentally everything humans do is a matter of opinion. Because we're not rational, but emotional beings. Which is why irrational opinion widely held by populace shapes reality more than purely rational belief held by few.
Re: (Score:2)
The stock markets *are* insanely overpriced and I'm honestly shaken that they have again and again reached these insane, unrealistic peaks over the past decade. What irony it would be if a fresh Chinese startup is the domino piece that triggers the next stock market crash.
The tech sector is wayyyy past due for a correction, and it's going to be nasty when it happens. But it has to happen.
Re: (Score:2)
You mean like Warren Buffett who is sitting on $325 billion in cash [apnews.com]? How horrible it must be for him, and others, to buy low and sell high. What is the world coming to?
Re:Prove it (Score:5, Informative)
sure, download ollama and run deepseek locally: https://ollama.com/library/dee... [ollama.com]
deepseek r1 competes with openai's o1, except it's actually open and you can run it locally for free. i have an old GPU and it can run a smaller version, but if you have beefy hardware you can run the full model.
and this is where we're at: companies getting hundreds of billions in capital to deploy monstrous datacenters, to charge companies $$$ a month to run on those models and spy on ALL of their data even if they claim not to...
or spend a few thousand bucks on a beefy server and hit the r1 model all day every day with paying just for the cost of electricity. I'm currently working for a company that wants to offer AI clients but every client "wants AI" but also "don't want to give up all of their data to openai for fun".
if deepseek's approach is to be truly open and let you run things locally, this could change the whole AI business which is basically "we sell you access to this black box that spits out content in exchange for money".
Re:Prove it (Score:5, Interesting)
also, the local code LLMs have gotten so good, that Microsoft is now giving you Copilot for free. It used to be a $10/mo subscription but there is now a free tier. This is because pretty much any computer with a GPU from the last decade, that has enough VRAM, can use a code LLM.
Now why does Microsoft try so hard to keep you locked in? Running that free tier, especially at that scale, costs a LOT of money. They really, really want your content. Maybe not your code, but oh boy do they want to learn how you use it in order to train their next generation coding LLM to "code like real humans code".
Re:Prove it (Score:5, Interesting)
Given the DeepSeek selloff news today... Not looking so good for "Stargate" or whatever it's called.... stock up on popcorn.
Re: (Score:3, Interesting)
I, for one, hope NVIDIA stock comes crashing down, and all of that money is reinvested more diversely.
Re: (Score:3)
I, for one, hope NVIDIA stock comes crashing down, and all of that money is reinvested more diversely.
How does NVIDIA stock crashing produce money that can be "reinvested more diversely"?
Re: (Score:2)
1. Stock prices go down when more people are _stupid_ than _stupider_.
2. When a person sells a stock, their cash assets _shart_ by the amount of the sale.
3. This means that money can be "_given_ more _to my wife_".
Re: (Score:2)
When you buy something, you trade your money in exchange for something. In this case, "all that money" already went in to other peoples' pockets when the stock changed hands.
If the stock drops, the "value" goes poof. If sold, the seller gets less money back than they spent in the first place. It is as if you bought ice that melted away.
If you mean tanking NVidia would make future investment flows look different, sure, most likely. But that'
Re:Prove it (Score:5, Interesting)
running a model and training a model are two different things. Their claim is they _trained_ their model for millions of dollars not billions like their competitors.
Re:Prove it (Score:4, Insightful)
For investors, this detail isn't a huge factor. The fact remains a party they weren't even looking at produced something that endusers can just run offline. So it highlights a huge risk they might not have considered to the relevance of the massive money they have poured into folks like Anthropic and OpenAI.
Re: (Score:3)
or spend a few thousand bucks on a beefy server and hit the r1 model all day every day with paying just for the cost of electricity. I'm currently working for a company that wants to offer AI clients but every client "wants AI" but also "don't want to give up all of their data to openai for fun".
if deepseek's approach is to be truly open and let you run things locally, this could change the whole AI business which is basically "we sell you access to this black box that spits out content in exchange for money".
Good points. I am currently working on a project to run an AI model locally on a Mac for the same reasons. We can control the data and ensure it remains private.
Re: Prove it (Score:2)
Re: (Score:2)
The local Cursor only has the ability to talk back to Cursor.
That made it a flat out "no." for our organization.
Re: (Score:2)
It's such a shame, as Cursor does such a great job, and has been improving fast.
Re: (Score:2)
Re: (Score:2)
Just sayin', the street is irrational.
Re: (Score:2)
Nobody is listening to you. There's a saying "Buy on rumour, sell on news". Just sayin', the street is irrational.
Yea, that's been the big argument overtime. Nobels have been won arguing both sides of it; Chicago even has the Fama / Thaler mixed doubles team.
Re: (Score:3)
Not really. This is just more of Jensen's Law [techradar.com]. Yeah, the (several) techniques they used helped save compute. But everything with AI has been getting more compute efficient for years - hardware, inference, and models simultaneously (and their impacts are multiplicative). And they're going to continue to. This fact is not news.
The thing that's more disturbing people in the field, rather, is the realization that Chinese companies can be quite competitive. It's also concerning that it basically has the Great
Re: (Score:3)
Their claim is that the model that anyone can already download and run is compelling, at least for scenarios where ChatGPT-o1 are compelling. So far I haven't found anyone trying it call it as misrepresented, but have seen a large amount of praise for it as it runs in pretty modest environments.
Beyond the purported accomplishment, for investors it raises the possibility that even if there's a massive reshaping of the world to "AI", it won't necessarily mean that the models are captive exclusively to places
Re: (Score:3)
Their claim is that the model that anyone can already download and run is compelling, at least for scenarios where ChatGPT-o1 are compelling. So far I haven't found anyone trying it call it as misrepresented, but have seen a large amount of praise for it as it runs in pretty modest environments.
Actually, not everyone can train this model. The details are sparse, but the claim is training on $5.6 million, compared to the expected $50-100 million. That efficiency, if true, is indeed impressive.
But there are some things to read in the numbers. First is that hyperscalers still have an advantage. We're still talking a huge number of GPUs for a long time, even if that long time is shorter.
Second, Nvidia GPUs were used. The PR didn't say that, but it also didn't mention Chinese chips. If Chinese chips ha
Re: (Score:2)
I've seen Chinese AI firms state that even if they could get Chinese chips for free, it'd still not be worth it to train on them over NVidia - that's how much of an advantage NVidia has.
Re: (Score:2)
It feels like the underlying model is more capable than the finetune lets it be, IMHO. For some things I use DeepSeek, but for a lot, Claude 3,5 Sonnet is still king.
DeepSeek's work definitely deserves attention, but not the hyperbolic reaction that most people who couldn't name 5 LLMs if their life depended on it have been giving it.
Re: (Score:2)
V2 isn't all that impressive.
R1 though, I can see definitely... shaking the foundations a good bit. It shook mine.
Watching its reasoning in its <thought> tags is.... eery. And it works a lot of the time to give answers far, far better than Llama does on its own.
In complex questions- better than any of the bigs I've used (ChatCPT/Claude/Gemini)
The downside, is I don't find it useful for local inference for any tools because i
Re: (Score:2)
Yeah, that's why you need more than one reasoning network. One to handle situations you're already considered, that need FAST response. Another to handle new situations. Perhaps a third that's more a "reflex arc". And a fourth for "let me get back to you on that one".
I really doubt that one reasoning method can handle all four of those situations.
Re:Prove it (Score:5, Insightful)
Companies like Tesla are meme stocks, insanely overpriced driven by a cult of personality. Autopilot, then Full Self Driving, now Full Self Driving (Supervised) has been "next year" for over a decade now. Their robots can barely shuffle walk, at the latest even the dancing ones were secured to the floor while the servers couldn't lift their own trays, and the ones in the crowd were surrounded by handlers while being driven remotely. I doubt we'll see a roadster with "cold gas thrusters" unless we figure out a break the laws of energy conservation.
You get the the idea. Still "reactive" Wall Street prices them higher than every other automaker, most of whom deliver far more cars yearly.
Now if Wall Street were truly reactive wouldn't you think they'd react to that abject failure? Actually not just failure when you add in revelations of staged videos pronouncing it as working all by itself.
Re:Prove it (Score:5, Informative)
If they can prove it, then yes, many AI companies are way overvalued and a bubble is bursting. I have no idea if they can back up their claim, but Wall Street will react to everything so it remains to beset if this is a rout or buy opportunity...
Fortunately, the Chinese company released their code. So, at least theoretically, the required amount of hardware to train can be verified.
NVDA is getting pummeled today. Down over 11%. The big question is whether this development (if true for the hardware training claims) helps or hurts Nvidia. Model advancement is something that usually leads to futher model advancement and more importantly more uses cases, and those two things are good for Nvidia. The bottom line is ... well, we'll see what Nvidia's bottom line is in the next few quarters. For NVDA, that's the only thing that's important.
Re: (Score:3)
Nvidia was always in for a huge fall. The export ban ensured it. Of course Chinese companies were going to work hard to remove their dependence on foreign hardware.
Re: (Score:2)
Nvidia was always in for a huge fall. The export ban ensured it. Of course Chinese companies were going to work hard to remove their dependence on foreign hardware.
Was the training done on Chinese or Nvidia chips? It was almost certainly Nvidia chips because the PR didn't mention Chinese chips, and using Chinese chips would have been far bigger news.
The Chinese market for Nvidia is currently 10-15% of total revenue. That compares to an estimate of total worldwide demand for Nvidia systems of 3x to 15x current supply. If all Chinese went away, it wouldn't make a blip on Nvidia's sales.
Re: (Score:2)
The chips used are only relevant to the time taken to do the training. Nvidia ones are the fastest, but since this AI model needs dramatically less work to train it can be done on cheaper and slower chips.
So even if they did use Nvidia chips this time, the future direction is pretty clear. That's why Nvidia stock is down so hard, they are no longer essential to AI training.
Re: (Score:2)
So even if they did use Nvidia chips this time, the future direction is pretty clear. That's why Nvidia stock is down so hard, they are no longer essential to AI training.
That sounds a little optimistic. China's capacity to design and manufacture advanced chips to modern Western standards remains severely limited.
Still, suffice it to say that even if Nvidia chips remain the standard in this space, Nvidia's market opportunity may level out as customers start not to need the most advanced (and expensive) chips or as many of them.
Re: (Score:3)
Deepseek v3 changes the necessary balance between memory, memory bandwidth and compute. Both in training and inference. NVIDIA makes money from their big slabs of tensor cores, which is exactly what has become less relevant.
Of course no one else is making hardware particularly suited for low precision MoE models either, so that won't have any immediate impact.
Re: (Score:2)
Re: (Score:2)
They released their weights and a paper, not their code.
That said, there is an open project to replicate it, which would give us an open codebase.
Re: (Score:2)
It's real, or the whole company is a massive fraud. Not just their product, but somehow they have been posting profits and paying their bills while offering their AI models at a price which would require them to have developed a much cheaper way of training it. They had to have paid for that GPU time somehow, so unless it's massive financial fraud... We should probably be worried. Again.
Re: (Score:2)
It's real, or the whole company is a massive fraud. Not just their product, but somehow they have been posting profits and paying their bills while offering their AI models at a price which would require them to have developed a much cheaper way of training it. They had to have paid for that GPU time somehow, so unless it's massive financial fraud... We should probably be worried. Again.
Fraud? In a startup? I'm shocked...
Not sayin it is, but I've learned to be skeptical of game changing breakthrough claims. We should be worried in either case.
Re: (Score:2)
They had to have paid for that GPU time somehow, so unless it's massive financial fraud...
I'm not saying its fraud, but I will note that there have been a number of historic instances of the Chinese government "dumping" subsidized product at a loss, in order to gain entree into a desired industry and drive away competitors. It wouldn't be terribly surprising to find out the Chinese government has been supplying this company with free electrical power and/or helping them pay for gear.
Re: (Score:2)
Several other Chinese companies have been forced to lower prices and take losses on their AI offerings because of this.
Re: (Score:2)
Kind of like cold fusion!
But it doesn't take proof to rattle Wall Street. Just convincing marketing.
Re: (Score:2)
The issue is that it's already popular, and that's what rattles Wall Street. Wall Street gets rattled every time they see a mouse, before getting any details or doing analysis, etc. Wall Street run on emotion, not logic. If Wall Street did run on logic, then no way would AI stocks be vaued that high to begin with. Hype is a balloon that is easily poked with a pin.
sounds like bullshit (Score:2)
but apparently that's what the world runs on
Let's assume for a moment that it's true (Score:5, Insightful)
I'm reminded of a panel at Comic-Con, many years ago, by Lloyd Kaufman of Troma Studios. He was talking about why multi-hundred million dollar mega-budget productions are so popular with the big studios, when it would be easier to make a profit from movies with a budget of a few million (and make a lot more of them). His bottom line was
"You can't embezzle $20 million from a $5 million budget."
Since China's goals are political, not financial, they aren't concerned with how much they can extract from investors. They just want to disrupt the west as much as they can. (And for that, it doesn't actually matter if their claims are true or not.)
Re: (Score:2)
Big bingo! Mod taustins comment up.
Re: (Score:3)
But they are doing it by disrupting embezzlers, so how bad is that really?
Re: (Score:2)
Given the effect it had on stock prices just from the announcement, I can't help but wonder who among the obscenely wealthy investors in the US sold short before the stocks tanked, and how much they'll make selling long now as those prices come back up.
China's interests are definitely served by actively helping US billionaires get even richer, thus pissing off people like you. Maybe they actively conspired, maybe they didn't, but their interests are definitely served by helping make that happen.
But perhaps
Re: (Score:2)
Re: (Score:2)
What China wants is to disrupt the west in any way they can. Increasing the disparity in wealth between the richest and poorest, and emphasizing that disparity, does that. And they can spin it as making communism look better than capitalism (despite the fact that China does not practice communism in any meaningful sense, and their wealth disparity is even greater than in the US).
I didn't forget anything. Benefitting billionaires is, in fact, part of the plan for world domination. The more they concentrate w
Re: (Score:2)
The announcement was Monday, the herd started running on Friday.
Anyone wanting to make the bet could have, no insider knowledge necessary.
Re: (Score:2)
They just want to disrupt the west as much as they can.
They absolutely are, but in this particular case they are saving us from much bigger disaster down the line. Having a bubble pop now, while it is still in Billions is much more manageable than down the line, when it gets into Trillions. It is not yet Too Big To Fail. Unless your view is that Big Tech would have somehow stumbled into an AI product people want.
Re: (Score:2)
You're assuming the bubble will pop, and the stock prices will never recover.
I'm skeptical that is the case. AI is the pony to ride right now, and still has some life left in it.
Re: (Score:2)
Remember, China is not just the government. The companies want to make money.
That said, if the government told a company to act in a particular way, it would act in that way. (OTOH, there are many examples of US companies doing precisely the same thing.)
I don't think we can depend on announcements. And the code isn't open. (Weights aren't sufficient.) But it's available enough that folks can test the claim, so it's probably reasonable to believe the claim. Early reports (seen on Slashdot) seem to in
GPU count (Score:2)
Still no real AI and no idea how to get there (Score:4, Insightful)
And now all AI researchers can think of is to scale auto-complete further with billions more dollars on GPUs. Bigger, better results, but NOT SMARTER.
AI researchers still cannot ( and hardly seem to be trying to ) come up with A REAL THEORY OF AI. Of how humans do intelligence with milliwatts not Gigawatts.
It is as pathetic how much time they have wasted for so little actual progress, just like how little the actual *science* of psychology itself has advanced.
Re: (Score:2)
And fucking MAGNETS, man, how do they work? And I don't wanna talk to a scientist - y'all motherfuckers lying, and getting me pissed.
Re: (Score:2)
AI researchers still cannot ( and hardly seem to be trying to ) come up with A REAL THEORY OF AI. Of how humans do intelligence with milliwatts not Gigawatts.
It reminds me of the joke about the guy whose brother thinks he's a chicken. They ask him "why don't you send your brother to a psychiatrist?", and he answers "we would, but we need the eggs!"
i.e. it doesn't matter all that much if researchers come up with a REAL THEORY OF AI or not; if the AIs are providing people with useful functionality, that'll be good enough.
Yes, you can argue to what extent AIs are being useful; in my limited experience, though, there is definitely some benefit there.
Re: (Score:2)
Utility and intelligence aren't the same, and I think I sort of agree that LLMs aren't intelligent. But at the same time LLMs aren't all of AI, and some AIs seem to be intelligent, i.e. they respond to expecting states of the real world in an attempt at a reasonable manner.
To me, LLMs are an extremely useful PART of an AI. We don't yet have a real AI, except for very simple demonstrators, that are missing crucial pieces. But we don't know how many crucial pieces. One piece is a world model, which every
Re: (Score:2)
People don't use milliwatts. Your brain is the most costly organ in your body and requires about 20 watts. A typical large AI model might use 100 times that but it also works much more quickly. The gigawatts are either for serving millions of requests at the same time (lets see how well your brain does with that) or training, which again is accomplished many times faster than human brains do it.
The training is cerainly very inefficient and the underlying computation isn't great because it's not optimized fo
Re: (Score:2)
Technically speaking, you are correct.
But practically speaking, we do indeed have AI. The fancy auto-complete is so good, that it is nearly indistinguishable from actual AI. It's certainly good enough to defeat all of today's Turing tests!
Re: (Score:2)
Ah but it's useful.
I'm not sure if you've actually tried AI, given your response. If you had, you'd know that AI is pretty decent at generating prose, instructions, code, and images, and does a better job of searching the web than Google. AI has already saved me a ton of time. You might say I'm a very happy customer! You might find it useful too, if you can get past your preconceptions and try it out.
Re: (Score:2)
I think I need AI to parse your words.
Re: (Score:2)
All the decades of time and billions of dollars on AI research, and the best the researchers could come up with is the discovery that stupid auto-complete can be scaled to provide impressive results.
This is unfalsifiable rhetoric. If I entered "Tonight's winning lotto numbers are "... or "The solution to the alpha sticking problem is..." and the technology were to accurately complete the sentence it would still be an "autocomplete".
And now all AI researchers can think of is to scale auto-complete further with billions more dollars on GPUs. Bigger, better results, but NOT SMARTER.
Are recent ARC-AGI and FrontierMath o3 benchmark results smarter or merely bigger and better? How does one differentiate between "bigger, better" and "SMARTER" and why does it even matter? Is there an objective criteria or test?
AI researchers still cannot ( and hardly seem to be trying to ) come up with A REAL THEORY OF AI. Of how humans do intelligence with milliwatts not Gigawatts.
The only metric anyone cares about is cost.
POP - ppssssh (Score:2)
Brute force approach (Score:2)
It's clear the brute force approach works and is sometimes the best way to advance. Once you see the path to getting the job done, take it. I see projects like ITER and ChatGPT4 in this category. Can you learn to do these things more efficiently given a few extra years of R&D? Obviously. But you are also a few years behind, and you have probably spent a lot of effort prematurely optimizing.
BFMI has it's points (Score:2)
While BFMI (brute force and massive ignorance) has it's points, it's usually not the best approach unless speed it your only criterion, and often not then.
That said, that's not the approach taken by ANY AI company. They're all based on optimizing something or other. LLMs are based off of optimizing cheap training material. This is a focused approach, and it produced results. But the results only optimized for "Producing reasonable sentences.", and the training material contained lots of garbage of vario
A buying opportunity (Score:4, Funny)
I follow Nvidia stock closely and I've made a few $thousands of easy money over the past 2-3 months by just buying the dips and selling the highs. A couple of times I've botched the timing but you can just hold on to the stock and eventually it will recover. So I welcome this drop, and when the price stabilizes or starts rising I'll buy another batch of shares.
As for DeepSeek, their main innovation appears to be that they figured out how to significantly reduce costs associated with training and compacting models. They appear to have used Nvidia chips they had stockpiled prior to the embargo (one source said they have 10,000). Probably their techniques will be adopted by the big AI companies, and the result will simply be more powerful AI that is less expensive to produce. Still there will be a bottomless appetite for Nvidia devices, and all this sudden alarm is misguided.
Re: (Score:2)
But this model is so efficient, it’s running on MacBook Pros. No need for Nvidia at all!
Re: (Score:2)
They used Nvidia chips to train their models.
Will they or won't they ban it (Score:3)
Instead of competing directly with Deepseek, the large American companies may ask Trump ban it in the United States on data leakage concerns like they did with Tik Tok.
See:
https://www.newsweek.com/deepseek-tiktok-ban-donald-trump-2021406
Re: (Score:2)
AI isn't necessarily evil, but it is safe to assume it is being implemented in ways which threaten your data security.
Because it's almost always opaque - they don't tell you how or where the processing is done and what telemetry is involved - I would always choose to disable any AI feature I can.
Re: (Score:2)
AI isn't necessarily evil, but it is safe to assume it is being implemented in ways which threaten your data security.
In my view, "implemented in ways which threaten your data security" qualifies.
But is it actually disabled? (Score:2)
Some things that say they are disabled keep running, and just disable the user interface.
Re: (Score:3)
Re: phones (Score:2)