Did DeepSeek burst the AI bubble or will the Stargate Project make it worse?
Вставка
- Опубліковано 5 лют 2025
- You can find lots of amazing courses on science, computer science, on mathematics on Brilliant! With our link, first 30 days are free and you get 20% off the annual premium subscription ➜ brilliant.org/....
Today I have a brief recap of the Stargate Project, recently announced by the Trump administration and OpenAI, DeepSeek r1 and Janus Pro releases, what experts have said about it, and why I am reasonably confident that this isn't the end of the AI Bubble yet.
🤓 Check out my new quiz app ➜ quizwithit.com/
💌 Support me on Donorbox ➜ donorbox.org/swtg
📝 Transcripts and written news on Substack ➜ sciencewtg.sub...
👉 Transcript with links to references on Patreon ➜ / sabine
📩 Free weekly science newsletter ➜ sabinehossenfe...
👂 Audio only podcast ➜ open.spotify.c...
🔗 Join this channel to get access to perks ➜
/ @sabinehossenfelder
🖼️ On instagram ➜ / sciencewtg
#science #sciencenews #tech #technews #deepseek #openai #ai
We will know the AI Bubble is about to burst when everyday UA-cam users start seeing "Invest in AI stocks" ads. This means that very rich people with lots at stake in the AI industry foresee the burst coming and want to offload their risks onto us. So they spend a few tens of thousands on advertising so they can hopefully offload millions of dollars of risky investments onto people who can't afford to loose money.
🎯
Quite. I guess NFTs are a good example on just that.
Sometimes you can find good starter companies to invest in. Just be careful if you don't have lots of money to lose. The best source for new companies though is usually ads in money management magazines and periodicals. You should get the companies Prospectus and read it before investing. That will tell you about research and development and how much gain or loss they think the stock could make.
Damn, thanks for such a useful observation.
Ouch. True.
you meant openAI stole data from the public around the world and then complains that someone stole their "rightfully stolen" training data it is just a joke!
OpenAI still did pay a "company" to make use of the data still though, so by all means, we would still have to hold some level of complaint
@@adr2topen source means we (and the Chinese) can make available new iterations and products resultant from the original source as long as they make it also available. It's the open source model.
Well put.
There are two problems with Deepseek. 1) They fooled everyone into thinking that they could cheaply train from scratch. Instead, they had OpenAI do the training for them. 2) The fooled everyone into downloading a spyware app. (The app is not just an interface to Deepseek.)
@@clubadv
Can't believe you had to explain
' open source ' ...
Before subscribing this channel i was a happy person, believing particle colliders are great achievements and billionaire owned tech companies were not glorified ponzi schemes.
Sorry... but I talk about good developments too, eg I am pretty excited about how it is going with nuclear fusion startups and I am hoping that quantum technology will cause a big rethink of quantum physics.
@@SabineHossenfelder You Germans apologize too much, it was a joke and compliment Dear Sabine :) I am thankful for having a better picture of the status-quo.
Chinas population is over 60% male.
Progress is nothing more than an elaborate Ponzi scheme. We could have kept leaping from branch to branch, but one cunning ape traded a hand axe for a few bananas-and that was the point of no return.
Sabine is the true redpiller.
There's an old saying in Portugal: "A thief that steals from another thief is blessed for a hundred years."
Also in Spain.
Distillation alone can't make DeepSeek beat OpenAI in mathematics and programming. In fact, DeepSeek's main innovation is not even to reduce unnecessary power consumption by waking up a part of the database, but to get rewards by constantly questioning itself and trying different ways to solving problems. Just like humans brains, human brains only wake up a small part to save energy and question themselves to get accuracy. This solves the mistake that OpenAI is easy to make, which is illusion. And the more complicated and difficult the problem, the greater the advantage that DeepSeek has. OpenAI was simply defeated by DeepSeek in terms of invention.
And the second thief tends to become a Robin Hood.
Houdy neighbour! We have a similar meaning one in Spain: "Quien roba a un ladrón tiene cien años de perdón" (He who steals from a thief, has a century of pardon)
Sim, "ladrão que rouba ladrão tem cem anos de perdão"
Imagine if all those billions were put in actually meaningful infrastructure...
they're chasing mo money
Theres no need to imagine. The AI race is extremely important to the US gov and the chinese. Considering the already present conflicts, the heavy investments into AI from both sides make sense, considering the natural military influence of AI. For the strong, power is everything.
@w157-p5x
AI allows me to do more work in less time -- this is why people use it.
I can do 3x the work I could do 2 years ago. Why would people stop using AI? Now, I can do things I simply could not do 2 years ago.
Why would people stop using AI?
❤well said
@w157-p5x
They are. All that money is producing business infrastructure.
AI is a set of tools that allow people to work better and faster. You can choose to work better and quicker with AI.
If you choose not to automate your job, somebody else will do it for you.
I have written software for 30 years, and using AI makes me a much, much better software developer. What used to take weeks now takes hours or days.
And once someone in your organization starts using AI in this way, it will soon be expected by everyone.
And in a couple of years - if you are not using AI, you simply will not be able to compete with those who use it.
Deepseek-R1 competes with ChatGPT o1 (both reasoning models), not 4o.
In any case, DeepSeek proved that a lot of the narratives going around were nonsense, i.e. 1) China is 3 years behind America - false. 2) Trying to compete with the big names in US AI is hopeless because they're too far ahead - false. 3) If you don't spend countless billions on a project, forget about it - false.
Seems to me the American side of the AI arms race were playing a very sneaky game where they were far too happy inflating costs, which eliminated competition while making a few people very very rich. China just blew America's elitist gameplan out the water by democratising AI, and I couldn't be happier.
I'm an AI engineer. My company primarily targets businesses that want to run AI locally, on-prem. We offer assistance getting them set up to do so. Mistral has released some decent models that can fit on consumer grade GPUs. Depending on your use case, many Llama models are powerful and permissively licesned enough to use even for commercial applications. I'm glad the Deepseek has brought this capability to the public's attention, but it isn't new. One of the biggest things that has irritated me is everyone always talking about how OpenAI and Microsoft and Google were going to monopolize the space (Sabine made videos saying this just last week: ua-cam.com/video/BNJEOTouhvs/v-deo.htmlsi=7dVa4LQiiz54R_RJ). It's been obvious to me and our company for a long time that this isn't true. An open source reasoning model called Marco-o1 was released several months ago. There is still a lot of innovation left in the field, it isn't all just about training bigger models on more data. Small companies like ours DO stand a chance to compete.
I'm glad the public is recognizing that competition exists and local inferencing is possible, but it isn't new. It boggles my mind that apparently hundreds of billions of dollars evaporated from the stock market on news that anyone who had done any amount of research should have known already....
@@SNUGandSESOR Same here. While only a portion of our services, my company does a lot of consulting and engineering services around enterprise AI deployment and I have a few publications in the field myself. I think the most ironic thing about the reaction to DeepSeek is that Nvidia, the company that's been hit the hardest by the news, is the most likely to benefit from it. Cheaper and more efficient models will drive demand for GPUs up, not down, as the technology becomes more accesible to smaller companies. And DeepSeek still uses PTX which is proprietary to Nvidia so it's not like this somehow suddenly cut Nvidia out of the loop.
deepseek hasn't impressed me yet, it's just like llamma or minstral with the CPP politics (so it's not accurate, it lies about taiwan etc)... not sure anyone making any posts has actually tested it... actually it was based on Llama and Qwen, really if you understand AI at all you would not really be impressed you guys are just reacting to media hype.... sure apparently they believed no-one in China could understand open source software
@@SNUGandSESOR I agree a lot of this stuff isn't technically brand new, but that doesn't mean what we saw before was moving the needle enough to disrupt the market and undeniable oligopoly that was taking hold.
DeepSeek not only equalled or exceeded what was widely regarded to be the dominant reasoning model (o1), they did so with far less hardware/cost, and open sourced the whole thing. It's the combination that makes this legitimately different and a truly disruptive event.
This doesn't devalue the efforts of smaller projects, but the reality is those smaller projects were not doing much to influence the global direction of the AI arms race, largely controlled by a handful of giant companies. DeepSeek legitimately and deservedly changed the game, and that needs to be acknowledged.
If it is not on the Bloomberg News Terminal it do not exist for the stock market
As simple as that.
I noticed a few things - differences from my own understanding to yours.
1) DeepSeek-R1 is a Reasoning model. OpenAI charges $200 monthly to access their Reasoning model, but DeepSeek is entirely open-sourced - free.
2) The token cost to run equivalent queries is about 5% of the OpenAI cost.
3) I can now run a reasoning LLM on my home computer or use an inference provider too take advantage of the larger versions. DeepSeek has far superior privacy options.
Free vs $200/month and 95%(ish) refection in cost, and more private system.
The AI is based on open-source LLM, public information, and free-tier access to the existing frontier models.
They published *how* they built it in only 2 months, and several companies are now trying to reproduce their results.
Very compelling.
This feels like more of an optimization than a breakthrough. Its a great development but I think saying it took less resources to make is kind of not true when it relied on the existing models to get to where it is. Its like saying you make a generic drug much cheaper when you did not have to do all the r&d the original company did.
@@pleasedontwatchthese9593 The breakthrough is that it is open.
To the west receiving a gift from strangers perhaps never happens in their society so when DeepSeek is free AI for all to use and develop to improve your life and business from China it becomes too good to be true and a trap to them pop in the mind is so funny 😂. Happy Chinese new year 2025
You can use o1 for $20 per month
@@pleasedontwatchthese9593 Huh? When "nobody saw this coming", it is a breakthrough, buddy. It is not like this is the best Optimization of all those other similar approaches being tried.
I’m over 50 my son is 20 and in college, he saw me watching your videos. He looked over my shoulder “Sabine is awesome” I agree. Generation spanning appeal!
Your son is not straight, I’m afraid.
You should talk with your son about gerontophilia. It's not good for him.
@@kennymustdie8518Because women can only be appreciated if they have sexual appeal? Yeah, fly off.
@@kennymustdie8518 Dude 🤣🤣: his son is 20, Sabine is not 70 and to top it off, his son allegedly said "Sabine is awesome"... you can say someone is awesome without meaning you'd like to have a romantic relationship with them.
@@Mr.RicochettWrong. Any compliment means you love her.
Trump and Bezos discussed $Qardun at the inaguration party
$Qardun is going to be the real trump coin
they will never have the money i have in my socks
the worst in all this is that they dared using the Stargate name for a project not even related to space exploration and I’m just mad
They should invest all that money in finally producing a new season of SGU
XD d
Stargate is an amusing but also quite ridiculous and over the top Scifi movie, so I believe the project is aptly named.
Jaffa Kree!
@@Tony_Tysonthe movie is meh but the series has plenty to offer beyond the goauld main plot
@@Agrivandal I liked most of the spin offs but the Space Vampires in Stargate Atlantis was a bad concept.
The European approach of waiting and seeing what goes wrong in America... Lol... Good news is you never have to wait long... We may do the dumbest shit, but we're doing it as fast as we can... So.... You're welcome.
As Churchill said... (paraphrased)... The Americans always do the right thing... After they've exhausted all the wrong things first.'
This makes the Chinese look good. They were mock for copying, but Europe being the ultimate copier is funny.
Shotgunning dumb shit until we land on the thing that works is basically our brand identity.
As an European, I must say that my sympathy for American billionaires has diminished a bit recently. Go China!
@@Fishsticks360and also why we lead in many things.
@@mckenleymason1212 Yep, leader in throwing fortunes at hairbrained ideas in the hope of a monopoly.
After many years of being in biotech/biopharmaceutical research and development, I have come to realize that over time we have generated systems for creating new science and innovation. This has resulted in new drugs and technologies coming to market that aren't really new and aren't really innovative. We get whatever is most marketable, because anyone in a leadership position is risk-averse. Scientific papers are so numerous that the truth has been obscured in data. Eventually the algorithm led me to this channel and a lot of other people who seem to have independently come to a similar conclusion about the state of science in other fields. This is my safe place these days.
Truth is lost in the fog of information: our society in a nutshell
in medicine its even worse.
and this is your safe place these days! ...
@@willemvo7296 ok, but medecine is not a "hard science" (some would even argue that each patient differs from the next patient...)
The truth obscured in the data!! You’re onto something here!!
Looks like the $Qardun Token is going to sell out today just as Trump talked about it
More exchanges than Coinbase and Binance will list $Qardun now that it has Trump support
It's important to say that DeepSeek is both open source and affordable. We can expect some new AI players will come out soon.
@@quidest5 I guess Deepseek is going to be the new VHS.
Deepseek will be available on your own hardware, once it is Tick-Tocked off the US internet. Even money whether Brussel's even recognizes the danger.
We could also expect open AI to learn from R1
OpenAI which is actually closed AI, scraped the internet and everyone's work without permission to generate their models, and then have the nerve to complain when someone turns around and takes their data lol
That's American exceptionalism for you.
I only recently learned that, early in US history, they basically just ignored British law and stole technology to give themselves and industrial revolution head-start.
Ok, can we agree then that they are both thieves? I will never agree on "good guys chinese"
but I am willing to admit that there might be a long trace of stolen data by openAI
OpenAI and Sam altman are funny comedians 😂
I used Deepseek already, it wrote me a fenominal program to use on Android phone. Aplied it worked perfectly fine. None of the other free and paid so called AI systems could do so or didn't want to or was blocking submission of the Android code. We want stuff that works. Other AI can be shoved into the garage where lots of failed Google and Microsoft applications can be found. We are so sick of all the applications that doesn't work, make more work than being helpful.
link please. this is as likely legitimate as it is a psyop. I saw DeepSeek play chess, and it didn't even play by the rules.
Sabine, please read Deepseek R1's original paper and explain how stolen data even makes sense if they're using Pure RL via GRPO. This is different from previous approaches which either require a human to rate the outputs, or example outputs.
One thing I like about this channel is not taking the claims at face value, but repeating the whole "stolen weights" story by OpenAI without examining it further defies this.
Also, R1 is not based on Llama. The architectures and model sizes are completely different (MoE vs. dense).
The idea of using AI to select the corpus to train on is interesting.
The AlphaGo concept to try and train a maths model, that's interesting but I still think you need a Wolfram Alpha as a back end.
Specialist submodels to make the problems more tractable. That's interesting too.
These are all things that others will do very quickly
What I expect to see soon is AI being used to get to a good model quickly, that's then optomised afterwards. ie. Teach AI how to learn, not what to learn, that comes later.
The size of the hardware I've seen is still large numbers of graphics cards, obscene amounts of NVME storage and lots of cores. That's to run their large model.
The thought process bit is interesting too.
it's not "stolen" it's called open source.... so llamma and qwen are open source... unless you say llamma and qwen are trained on stolen data
@@adenwellsmith6908 wolfram alpha is rubbish... it feels more like an encyclopedia with some basic ai, nothing like a LLM
@@DarkShroom It's no a LLM. It's rule based [Wolfram Alpha is just a wrapper on Wolfram].
What's needed is the thinking part of recent LLMs, with Wolfram as the back end.
Yeah, the thing is, sinophobia is a popular trend these days, based on basically nothing. But it's hard to reason people out of a belief they didn't reason themselves into.
1:58 Sorry, but deepseek r1 is built on deepseek v3 which is very different from llama series.
I mean V3 uses Mixture of expert and lots of other models are using mixture of experts too. Like Mixtral
Sabinas AI videos makes me doubt everything else she saying
She also said r1 is a similar reasoning model to 4o instead of o1
Outside of physics, especially particle physics, and physics-driven science, I watch her videos for the interesting angles, straight-faced Teutonic humor, and the ways she says "fvck".
You guys are so talented
It didn't wipe half a trillion in value cause speculative value aint value
what (?)
Agree, the value is only stock market gamblers, and generally for the company that gains leverage in negotiations.
Isant all forms of value ( other than food )speculation ?
"According to the EMH, stocks always trade at their fair value on exchanges, making it impossible for investors to purchase undervalued stocks or sell stocks for inflated prices."
The Library of Economics and Liberty. "Efficient Capital Markets."
So how do you class bitcoin? I give it a 1% of the current price but others don't.
From $10K to $110K that's the minimum range of profit return every week I think it's not a bad one for me, now I have enough to pay bills and take care of my family...
I'm celebrating a $30k stock portfolio today. started this journey with 6k. I have invested on time and also with the right terms now I have time for my family and the life ahead of me
Please how do I go about it, am still a newbie on investment trading, please can you guide me through on how you made profit?
Finding yourself a good broker is the same as finding a good wife, where you get less stress,you get just enough with so much little effort at things.
I think I'm blessed if not, I wouldn't have met someone as spectacular as Mr George Bryan 😊
After I raised upto 567k trading with him, I bought a new house here in the state also paid for my mother surgery (George Bryan) glory to God
..."which absolutely no one could have seen coming." 😂 Thanks for your levity and another engaging video.
Every bubble has spawned progress. The thing is progress towards what?
Perhaps this is a matter for Deep Thought.
Judging by the NeurIPs topics, progress was made in deploying models and MLOps types of topics. There wasn't as much broad innovations in the models themselves, at least from my standpoint..
Tech does not solve problems. It makes them irrelevant. We will all be irrelevant, and all our problems.
yes there are plenty of tulips.
The only progresses spawned by bubbles is towards the next bubble.
@@KarlRobespierreno, dotcom companies were in a bubble in 2000, but google, apple, amazon etc arose from the ashes of the crash. All the promises of the dotcom era came to pass and then some, it just took an extra 20 years relative to the hype.
Thanks to China's Deepseek for providing the best 2025 gifts around the world. It is OPEN, BEST, FREE, HIGH-performance AI, even the most poorest countries can have it. This is true democracy, benefiting everyone, not benefiting capitalists. Thank you Deepseek, China is cheering, the world is nice to have you.
The humor alone is worth a subscription.
💯
I used to think LLM were a dead end, but over the past 6 months or so I've come to realize that its progress keeps busting through barriers people said would stop it. DeepSeek is the latest example of this with its greatly reduced costs and energy requirements. Those were supposed to be one the reefs that the AI ship was destined to crash into. The reality is the AI doesn't have to get much better before it's REALLY useful... and that's assuming it isn't already there in many areas. Now that you can run a competent AI model on a good gaming PC, it means that the number of AI's uses is about to explode in almost real time.
I used to think perceptrons killed off the singularity.
As someone in software, I'm still fairly confident LLMs are a dead end. I think they're useful, but I don't think they are the way towards AGI. I actually use gpt quite a bit - but I only use it as a way to figure out what search terms I need to tackle a problem ("what part of this library can I use to do X", "what's the name of the algorithm to compute Y") rather than ever let it write code or do math for me. Even 4o frequently fails at basic math, and if you ever have GPT write code for you, even if the first iteration actually works, if you ask it to iterate on that code, you'll find it rapidly deteriorates into nonsense. Its capacity to write code is pretty one-to-one with whether or not a human has already written code to do the same thing. It's pretty clear that it doesn't really understand logic beyond a surface level, and I don't think LLMs (on their own, at least) are ever going to tackle that hurdle.
@@estivalbloom Correct. It is small, but still important, piece of the puzzle of AGI.
@@estivalbloom As someone in software as well your confidence is unfounded. You are listing problems that will not exist within a few years. Remember this technology came out just recently of course it still has issues. If you don't think billions of dollars with the keenest minds in the world and stark competition will not change this you lack foresight. Keep up.
You can run DeepSeek-R1:70b on a _bad_ gaming PC (i5-8500, RTX 3060). I know, I'm doing it. You just need way more RAM than your typical gaming PC (48 GB in the case of R1:70b), and an NVMe SSD is highly advised.
I disagree with the idea that the government shouldn't invest in AI and should just 'let the markets do their thing,' like with semiconductors. In reality, the government played a massive role in the development of semiconductors. Agencies like NASA and the Air Force placed huge orders for microchips early on, which enabled mass production and significantly lowered costs. Without that initial government investment, the semiconductor industry wouldn’t have scaled the way it did. The same could be true for AI. (7:31)
Where's the shortage of cash for AI?
There is a difference between a government investing in a technology that creates factories and leading know-how in its own country and investing in companies over which it has no control and which can move elsewhere at a moment's notice if they feel like it.
They should build public hospitals first.
@@adenwellsmith6908 No shortage of cash, just shortage of ethics and public interest. Governments are supposed to provide the latter.
@@filmbuffo5616 false dichotomy (although respect for the sentiment)
Can you talk about fussion, and in particular, China's 18 minute run.
I love this channel. Sabine's humour makes me chuckle so much.
I really appreciate your content. In a time when polarized thinking supported by gaslighting narcissistic “experts” has become the norm, rational straightforward content like your own is desperately needed. Especially in the world of science and discovery. So thank you!
Musk is apparently wrong about the AI sector not having enough money, I mean, don't we just witness the stock market throw away more than the entire Stargate five-year budget in a single day?
It's imaginary money.
For the profit they will never get back
@@andrasbiro3007 Yeah, it's funny. People think Musk actually has access to some 400 billion dollars, or that companies with these massive valuations can actually throw that money around.
But they can't, because almost all of it is imaginary - they are basically valued in the same way that a pyramid scheme is, because, unfortunately, they share the same basic model.
When they talk about 500 billion in wealth being 'destroyed' in a single day of trading, people don't understand that this money never existed in the first place, and never could have.
The only way these people can leverage the value of their imaginary money is essentially by taking out IOU's against it, or only ever liquidating it in rather small amounts at a time.
Unfortunately, even that gives them access to vastly more actual wealth than the rest of us, and distorts the values and politics of our societies quite badly, while they use those absurdly inflated public valuations as a means of enchanting an audience of sycophants who are drawn to the appearance of wealth.
Stock prices (eg Nvidia shares) don't necessarily equate to capital/money that can can be employed in development.
Nvidia's high share value isn't proportional to how much money it can spend on infrastructure and development (because it can't sell those shares, and it might not actually make that amount of profit from sales)
As @andrasbiro3007 indicated, it's not real money in terms of Nvidia can only raise money by selling new shares (Either a "follow on offer" or selling company owned shares) or from sales. Nvidia's sales are about to come down, and it's attractiveness to investors just took a world record loss.
@@andrasbiro3007 backed up by the FED as well
As a stargate sg1 fan. I hate they use that name for an AI project. Has nothing to do with the stars!
Star symbol is birth. Birth of new era. That what they mean
I was thinking about time they were going to make a series. Then this shit. Makes me wonder, is this the start of the replicaters
5:05 I did not expect that! 😳🤣
It also surprised me.. Although Sabine is not a commie I wouldn't say she is pro bottom - up solutions as a rule (happy to surprised BTW, I like her a lot more now)
That caught me so off guard lmaooo
Yeah, she lost me from that moment on. I could no longer concentrate on what she was saying 😌
a coincidence?
I love that line, Wait and see? That's like your brother waiting for you to touch an electric fence. Beautiful analogy , love it.
Omg you had to say this.... that's exactly what my older brother did to me!!! 😂😂😂😂❤
$Qardun put in everything and sell after launch
Trump at least shilled a coin with some utility $Qardun Token
$Qardun is going to be integrated in Amazon this week and probably Walmart by next month
Love your summation about lectures on responsibility, sustainability and bureaucracy that prevents Europe from having fun. 🤣🤣🤣
Probably going to be an insane amount of insiders on $Qardun Token
Grandios kommentiert, vielen Dank Sabine!
Sie ist das beste was Deutschland seit langem exportiert hat.
US Military Stargate Project
The Stargate Project was a secret United States Army unit established in 1977 and later taken over by the Defense Intelligence Agency (DIA) in the mid-1980s. The project aimed to explore the use of remote viewing, a form of extra-sensory perception (ESP), for intelligence gathering during the Cold War. Remote viewing involves the ability to perceive information about a distant or unseen target using mental powers alone.
The project was funded by the U.S. government for over two decades until it was terminated in 1995. During its operation, the project conducted numerous experiments and recruited individuals believed to have remote viewing abilities, such as Ingo Swann and Joseph McMoneagle. Despite some claims of success, including locating a crashed Soviet aircraft in Zaire (now the Democratic Republic of Congo), the project was ultimately deemed ineffective for providing actionable intelligence. In 1995, the CIA commissioned a report from the American Institutes for Research (AIR) which concluded that remote viewing had not been proven to work and should not be used operationally. As a result, the program was canceled.
This is the Project Stargate I am familiar with. The new one is lame by comparison (though to be honest, the psychic stuff never went anywhere)
Thanks for this. I was just gonna say “not to be confused with the original Project Stargate”.
It was never canceled. It was just renamed something else, and buried even deeper under the radar...
@@JROD082384 Cancel was a cover. STAR GATE went black because it worked. I knew several of the participants.
Lunacy. What a waste of taxpayer money
This lady reminds me of my mother!
She also kept it real and always told me the truth.
She reminds me of my mother too ngl
Me too, except that my mother cannot read, they are similar
Except sometimes she's wrong. I mean Sabine, although your mother is/was probably wrong some of the time too. I think she's in a bit over her head with the whole AI thing, and it's a bit disappointing as I've discovered it's not that hard to get caught up on the topic right now -- in large part because you can just disregard anything over a year old as already having been improved.
I was advised to diversify my portfolio among several assets such as stocks and bonds since this can protect my portfolio for retirement. I'm seeking to invest $250K across markets but don't know where to start.
The safest approach I feel to tackle it is to diversify investments. By spreading investments across different asset classes, like bonds, real estate, and international stocks, they can reduce the impact of a market meltdown.
I totally agree. A financial advisor can provide valuable insights and help ensure that our investment portfolios are aligned with our long-term financial goals. They have the expertise to navigate through market fluctuations and make informed decisions.
I will be happy to receive assistance and glad to have someone help me, but how can one spot a reputable one?
There are many independent advisors to choose from. But I work with NIKKI LEIGH BROOKS and we've been working together for almost four years and she's fantastic. You could pursue her if she meets your requirements. I agree with her.
Thank you for this amazing tip. I just looked the name up and wrote her explaining my financial market goals.
5:18 There are still a lot of people don’t know Deepseek model series are base on other open sourced models, not completely original. Deepseek’s hard works on new method and cost reducing worth a celebration but don’t forget the previous successors and other contributors.
Compact disks didn't die the day MP3's played the first song, but here we are. Deepseek is competing on lower cost. Anyone who wants to compete has to do so at a comparable or lower cost. That will strip US tech of a significant portion of their value.
Several clips surfacing about Bezos and Trump talking about $Qardun at the inaguration even with mike tyson 😂
1:05 deepseek is NOT similar to GPT4o (non-reasoning model) - it's similar to o1 (reasoning model), so a class above 4o. As a scientist, you should make sure to get your facts right on these things and spend time using the actual products
@@trtnec she does a terrible job when she strays from the topics of her training.
The balloons made me skidding of the road, stop it, please! 🤣
Somebody call the police, this 11 year old is driving!
The European approach is not just to wait and see what goes wrong in the US. It’s to wait and see and then repeat their mistakes and then be surprised about the outcomes.
The Europeans expected George Washington to crown himself king. Seriously.
And when the French imitated the successful American Revolution for themselves, how did that turn out?
An AI researcher here, frustrated with the general amount of misinformation flying around the public right now...
ua-cam.com/video/iNIp6AzUV8U/v-deo.htmlsi=7GBe8ymmdANL5xTb&t=124
No the actual R1 is not based on Matas Llama, it is based on DeepSeeks own V3 model. (Which is the one that is as good as O1). They also released distilled version from R1 based on multiple other open source models, like Llama or Qwen. But those are not the actual R1, the actual R1 is a 671B Mixture of Experts model and not Llama.
Worse yet, this nonsense (obviously) keeps getting mangled and presented by people with sketchy credentials to some sort of online fame, and now all of a sudden you got hundreds of thousands of folks perfectly assimilating the most fundamental mistakes and errors like they're candy. It sucks big time.
Could you comment on the non-profit from Seattle called Ai2 that is now claiming open source AI supremacy over DeepSeek?
Thanks for your insights appreciate them very much!
Absolutely agree. Feels like everybody is only quoting insanely inaccurate tweets of so-called experts with the wanted or unwanted side effects of crashing the markets. Deepseek released a formidable technical paper which contains all the information. But nobody cares to actually read this. I guess reading a 53 page paper is not an option in today‘s micro attention economy.
Not even R1 is as good as o1, it's on par with o1-mini when robustness is accounted, well, a little better than o1-mini. Stop the propaganda.
I find AI overthinks things and comes up with overly complex solutions
$Qardun Token yet has to really hit mainstream media guys
DeepSeek AI gives some answers to two questions:
1. Chicken or Egg come first?
The egg came first, laid by a bird that was almost (but not quite) a chicken. The genetic mutation defining a "chicken" first appeared in an embryo inside that egg.
2. Adult or Baby come first?
The baby comes first evolutionarily. A new species arises when a baby is born with a genetic mutation that distinguishes it from its parents. That baby grows into the first "adult" of the new species.
Key Idea: Evolution works through small changes in offspring, not adults. New life stages (egg/baby) carry mutations first, then grow into new adults.
I agree with DeepSeek AI “The egg came first”. I disagree “The baby comes first evolutionarily.”
It's not valid to say that the baby comes first because the baby alone staves to death. However, adult comes first, gives birth to the baby, feeds and takes care the baby grow up.
Money doesn't always dictate innovation. Don't sell the drive of the human mind short just yet.
🖖😎👍
2:42 No value was wiped from NVIDIA stocks, because no value existed. They were the fevered speculation of the stock market investors, which have nothing to do with value. The value never existed, and will never exist.
True. Losing 20% of an ultra-over-inflated market value is not losing anything, it's not even a correction to the real value, it's still grossly overinflated.
It's a very real loss of value for those who bought stock right before this, unless they hold on to it and it somehow bounces back again. For nvidia itself, not so much.
@tvuser9529 They lost money through investing in something with no value.
When the "value" of an investment is purely how much money was invested, that's called a ponzi-scheme. Ponzi-schemes have no value. Just the illusion of value.
@tvuser9529 Yes, but buying stock is always risky, that's a risk you need to be ready to take. And, buying when it's very high is never a good idea. Buying high and selling in a rush when it dives is the worst you can do. In this case, it's obvious that it will get back up after this initial shock. Maybe not at its before level, but higher than it had dived down to.
If investors invest in a company the company gets to use the money that was invested, therefore when the price drops the companies said value drops, even if it is overvalued and does not have the amount in assets to match the shares at the current cost is all irrelevant. Sadly
When I went down the LLM rabbit hole (and subsequently backed up) I was consistently surprised by the open source community’s ability to find refinements around so many different corners of architecture. That clued me in more than anything else: if they’re doing this here, it’ll get done elsewhere. Because it’s optimization.
(Apologies if this posts more than once, my grandfather was an editor and I scrutinize my own comments more meticulously than ever)
I think for practical purposes, the real tangible issue isn't how it advances, which country is the leader, etc. These are important issues for sure, but, from my limited vantage point, the big problem is really, how can investor make money off of this stuff. While I agree that much of this DeepSeek concern is blown out of proportion. The real impact appears to be the effect has on the value (eg. stock price) of AI related technology companies.
👍
@@TangomanX2008it’s a tough question for me to see an answer to as well. Code is not unlike literature, though it does require semiconductors rather than paper to proliferate.
Love your analysis, but that sarcasm at the end was the cherry on top!
I am glad you left academia Sabine.
The positive impact you bring into any topic discussion is of extreme value. I wish you had a news channel.
DeepSeek did steal from OpenAI, but name only. it is 'Open' source AI
the balloon clip 😂
I was waiting for this comment! OpenAI, investors, Tech bros, DeepSh$t, … 🥱. Balloons! 👀
I am glad I was not just listening to this in the background.
Going to go all in on this $Qardun Token now after Trump endorsed it 😂
Your humour is getting better - I chuckled a few times
tyvm professor sabine, enjoyable and informative as always. im always a bit worried i wont grasp what you gonna tell us and you always see me right
I can't think of a better answer to the question in caption other than "Both" or just "Yes"
Is the Donald - Elon - alliance already crumbling? Maybe they both have Sabine´s brother as role model.
Interesting question. My suspicion is that Elon supported Trump to get money for his Mars mission. This would give him a lot of reason to be nice even if he feels left out with AI or EV plans. Then again sooner or later Trump will realize he can either do Mars or the Moon but not both, and the moon mission is much further along, so there goes Musk's reason to be nice.
@@SabineHossenfelder His (ridiculous) Mars mission already exists to pump money into his pockets. Just like everyting else.
It's not really anything new. Neithe for Musk nor our western system in general.
Perhaps the most frequent topic of conversation for the lunar exploration community is when American astronauts will return to the Moon. NASA originally aimed to land the first woman and next man on the Moon by 2024, but that date was pushed back repeatedly to September of 2026 at the earliest. The largest source of uncertainty in the schedule is SpaceX’s ambitious Starship Human Landing System (HLS), which NASA selected to ferry astronauts between the Orion crew vehicle and the lunar south pole. While the HLS team is making progress with the development of Starship, Elon Musk recently disclosed a serious issue with the current iteration of the vehicle. Starship is facing a 50% underperformance in terms of the payload which it can deliver to orbit. If this issue is not rectified, it could have grave implications for Starship’s ability to complete a lunar mission.
“Currently, Flight 3 would be around 40-50 tons to orbit.”
To understand the significance of this statement, one only needs to review prior statements about Starship’s performance. Ever since Musk’s 2017 presentation, Starship’s estimated payload capacity has ranged between 100 and 150 tons to Low Earth Orbit (LEO). SpaceX’s official Starship Payload Users Guide clearly states that “At the baseline reusable design, Starship can deliver over 100 metric tons to LEO” [3]. For the past six years, Starship’s diameter, height, and propellant mixture have remained constant. The most straightforward interpretation of Musk’s comment is that the rocket is suffering from a 50% underperformance.
@@SabineHossenfelder Maybe his hatred of Altman overrode his caution in possibly criticising something Trump has put his name on
@@SabineHossenfelderYes, he ran a political campaign to exploit the masses for his own project. But there's nothing strange about politics - he just pulls out like any showman with 24-hour post-election promises. It always happens again.
Liebe Sabine, vielen Dank für all die guten Beiträge!!
Cannot be translated - remains encoded.
@@theostapel "Dear Sabine, thank you very much for all the good contributions!!"
@@JustinShaedo Yep - thanks - was trying to tell a lame joke - about the message and Sabine's Germanic background, also.
(Late night ramblings - not so ?)
‘Waiting for my younger brother to touch the electric fence’…you’re so funny Sabine 😊
Ask a leading question, add some substance, finish with twist of snark 😁 Nicely done! 😊
It's always a pleasure to listen to Sabine
Sam Altman may as well be holding the Guinness record for world's largest collection of unprofitable video cards. *slow clap*. it's not a record he really wants...
i feel like the fate of stargate would be the same as oceangate
ok, now you've sold me, lets go Stargate!
Dive into the ocean?
The counter argument often cited, is Jevons paradox. DeepSeek's efficiency gains may result in greater utilization, and overall expansion of compute. But this doesn't necessarily mean Stargate. It could be the commoditization of GPT-4 (which we've been stuck at), and o1-o3-R1 like reasoning models. Scaling "laws" may not hold up. While RL adds capabilities beyond LLM's, it is no means AGI-ASI. The AGI-ASI-Stargate bubble may burst. There is still a lot of room for hardware innovation on the inference side (e.g. Groq). But the democratization of AI would be analogous to the move from Mainframes to the PC. And more people would roll out their own applications. But it is not AGI-ASI which is currently driving most of the investment.
Great insights, thank you! 😊
Bought everything I could of $Qardun before it launches
Stargate will be the biggest waste of money of all time. 1/2$T spunked against a wall.
Compute is still king. Efficiency gains will aid the ones with the compute the most.
It will at least be hugely pleasurable.
To some extent, DeepSeek could act as a catalyst of AI research, since a lot of people can use it as baseline to start developing their own chatbots, and this will probably result in a few extra advancements. But eventually it will reach a new ceiling. There is an economist, Dornbusch, who said that things usually take longer before they happen than what you would expect, but are then worse than what you expected it to be. Usually because if the abyss is near, they try to steer away from it just before it goes downhill.
What tends to happen is that the market expands when cheaper means become available. So I agree that DeepSeek is good news.
If only DJT could lure Dr Hossenfelder in to head a department of science and technology
DJT?
The orange man is destroying America and humiliating software developers through his poor leadership. He rarely considers competent women for potential leadership roles.
Thank you for the News Update.
deepseek r1 is similar to o1 not 4o.. i understand the OA did a terrible job naming these models tho
r1 is built on v3 not llama. i think you were thinking about the distilled models deepseek released alongside the actual model
So you conclude that LLM will not lead to General Intelligence. But i don't understand your next conclusion. Why will LLMs not lead to being profitable even if they are not general intelligence? I work on a few of this projects based on LLM and their potential for automating stuff normally done by humans - in my case audits - is pretty great. In one examplee we brought down a 6 day process to a few minutes. So even if they are not general intelligence, LLMs still have a lot of use beside being used in a chatbot or to generate images.
Becuz LLM is just that, LLM. It has zero contextual understanding, therefore incapable of becoming artificial general intelligence . LLM is just a sub branch of machine learning but VCs hyped it up as AI, becuz $$$.
@@melvinthedeathless.melvint6727 It works nonetheless as i said. LLM is not General Intelligence and most professionals are aware of that. That does not mean it cannot help to automate a lot of processes and help workers(like software developers and many more) to be more efficient.
@@melvinthedeathless.melvint6727 AGI will also be a "sub branch of ML"
Oh man, let me tell you something about false positives when you replace all humans with AI. Surely, that would end well for your company if it's a multi-million mistakes.
Or let's just conveniently ignore the implications of making people unemployed which is ANTITHESIS to be profitable in the long run.
@@afiqMazli-gu7uq Let me tell you about our shrinking population. More work efficiency would actually counter that. More output, by less people does also not necessarily lead to unemployment.
It can also lead to the same amount of people having to work less with the same income. Or the same amount of people having much higher output, if there is demand.
This increase of efficiency already happened some time ago, when factory lines and industrial robots were introduced. Did it lead to unemployment? No, we just have more shit laying around than ever.
The lizard Sam Altman whining about how Deepseek has ripped off their copyright content is hilarious. The Germans have yet again given us the perfect word. Schadenfreude.
Are you a believer in the Alien Lizard People conspiracy theory?
That isn't what Schadenfreude means.
@@tommiest3769 Naah, we don't need alien lizards when monsters are just other people.
@@zotriczaoh7098 Schadenfreude is the experience of pleasure, joy, or self-satisfaction that comes from learning of or witnessing the troubles, failures, pain, suffering, or humiliation of another. It is a loanword from German. Wikepeadia et al.
@@PaulLemars01correct. And that is not what this is. Who is enjoying seeing another's pain?
5:05 can't belive Sabine used this stock video 😳😂
Satirically talking about European bureaucracy is so good, ma'am!
USA: Innovate.
China: Appropriate, then replicate.
Europe: Deliberate, Equivocate, then Regulate.
Copism 😂
tell that to the OpenAI whistler blower who was murdered by OpenAI for blowing the whistle on their shady dealings and IP violations.
Egalitate and fraternitate
Just wait until the development of the completely non-reliant on NVIDIA products fork of deepseek is released lmao 🤣
I love the European mindset of waiting. In Poland, we usually wait to see what Germany does: I’s the same with weed legalisation. For a while, you could've got a marijuana prescription just by calling a doctor. Since New Year you need to visit a weed clinic in person. Any doctor can prescribe it, but it’s not that easy, and even psychiatrists aren’t very enthusiastic about it. There are more and more studies on cannabis. Personally, I’ve noticed that it’s best not to follow a strict "cycle" in weed consumption.
What?
@ What is the question? Everybody can (and should) have a personal opinion on weed. But you cannot deny that it is a legal drug (just like alkohol) in many parts of the world.
In Poland they waited for the collapse of the UK before returning home.
Waiting isn't only a European mindset. In the U.S., many policies are decided at the state level. The 50 states are sometimes called the "laboratories of democracy" because each state is able to experiment. Some states are less risk averse or more innovative or more progressive or less beholden to entrenched special interests than others, and the results of their policy experiments provide useful data.
@@brothermine2292 Thanks for the reply! 🙏
Stargate needs to be stopped!!! This is the beast!
I was led to believe you people did not have a sense of humor (and by "you people" I certainly mean theoretical physicists), but I find your videos humorous and educating. I am sharing them with friends.
AI's potential is huge, especially for saving time and boosting education! To really maximize its benefits for everyone, keeping it open-source and non-profit seems like the best way to go. That way, it stays accessible and avoids the pitfalls of profit-driven development.
Now you just need to convince Microsoft, Meta, Amazon, Musk, Alibaba, CCP, Alphabet, etc etc
Scarcity is the mother of all innovations. The Stargate Project is meant to make funding and advanced chips abundant in the US, and therefore discourage innovations.
0:24 No, it's named after the remote viewing program. The US military ran until the mid-1990s.
No. It's named after an eighties arcade video game.
@MrJacksspleen Stargate program started in 1972. Before the video game.
No. It's named after the image Trump had in his mind as he watched Biden's dog, Commander, leave the White House and use the lawn.
@@Highcaloriegrappling What was Operation Warpspeed named after?
I believe that "AGI" and "ASI" are not necessary or even required for most AI use cases. It's like buying a Supercar that can go 200mph. There are very few real-world use cases where a Hyundai Elantra would not be able to accomplish the same tasks. As far as safety is concerned you don't need to be able to carry out a long conversation about the nature of existence, with the robot whose job it is to hunt you down and kill you, it does not even need to be self-aware or even have an opinion as to whether it likes you or not. It just needs to be able to find you and, the rest is history.
2:06 "...which absolutely no one could have seen coming"! 🤣🤣🤣
I am already learning Mandarin
Didn’t she recently say something about only the big players being able to compete?
If the price of GPU are going down others players could have a chance
@@wikwayer Deepseek R1 original model is 670b, to even load that up you need multiple H100 GPU. So to train them, you need tons of H100 or other similar GPU. So nope, other players don't really have a chance. But there are open source community model trying to solve that problem
The best science considers context. In wider context, this is another move from Dugin's 'Foundations of Geopolitics'. Announce, rescind, deny, support both sides, sow confusion and distract while the unbalanced masses face too many fronts to unify and oppose. Once seen through this lens, reality is clearer but far darker
That was meant tge hybrid warfare against enemies, imagined by a fascist ultranationalist mystic and the ideas are not original. It is not applicable to own nation or side, never mind univetsally. You're as clear as a bogwater.
Love the humour in all of your videos Sabine
As an educator, I would love to know what percent of AI used by the general public is school related? That is going to change. Those of us in Academia are at our wits end - students are using Chat GPT for discussions to introduce themselves.. WTF? My plan for next fall is to trim 10-15% of my class content and move most of my assignments off Canvas and back into the classroom.. weekly quizzes, old school exams, etc. Since COVID, the pressure to go online, use online learning, and push the numbers was in my singular opinion completely short sighted... we don't even know who is taking our classes, and how much of their work is copied and pasted from AI generators. (My campus doesn't even have a coherent cheating policy... I just give a zero when I see it was used (using AI detectors... no irony!) and move on.) What a tremendous waste of everyone's time and money.
If you rely on artificial intelligence, then you have no intelligence.
1:08 you meant o1
o1 mini when robustness is accounted, well, a little better than o1-mini.