Its more like grapefruit for Ed. Poor kid does not yet understand that banter means listening to the boss's lifestyle flexes and just smiling politely.
I swear it's my best part of the show. I find myself rewinding sometimes. Mans is actually hilarious. Although I know it's a PR nightmare waiting to happen on one of the edginess jokes 😅
It’s the worst part of the show. It’s like the mindless small talk people insist on before meetings. I don’t know you, idgaf what you did on the weekend. Probably another reason why I’m so unsuccessful.
This is my favorite comment I’ve seen about the DeepSeek vs Nvidia news so far: “The one good takeaway from the Deepseek vs Nvidia news, I think, is people might finally realize we don't need to devote more than half the world's energy resources, or set up a Dyson Sphere around the sun, just to help computers solve trillions of multiplication problems to spit out another thousand mediocre web apps.” - Jeff Geerling
That's awesome. I can understand why we would want to keep pushing down the path of robotic automation to produce actual goods. But the "generative AI" interest escapes me completely. Do we really need computers spending huge amounts of resources to quickly create endless mountains of content that could only ever be consumed by other AI?
The winner of the DeepSeek R1 release are the cloud computing providers and the customers. Amazon just released DeepSeek R1 as a Bedrock Foundation model on AWS. AWS has profited enormously from creating services for Open Source software like Postgres, Redis, RabbitMQ, etc. Open Source AI models will be the exact same. If the AI companies cannot maintain a significant performance advantage, in the long run, open source models on cloud infrastructure will be lower cost variations of frontier foundation models that will crowd out proprietary competition because they are more cost effective to operate.
57:48 my accounting teacher's advice was once your stock hits 100% gain, sell half the shares and rebalance -- and keep the rest in that initial stock basically forever. He did this with Microsoft and that stock alone let him retire a millionaire as it rose and split over and over again for the next 20 years. Keep spreading out, you only need one or two big hitters.
Regarding the job market, this is the toughest market I've seen since graduating college in 2023. I've been out of work for 4 months now, having previously worked in wealth management/financial advising. There are minimal, almost non-existent entry-level white collar jobs in finance. Myself, peers, friends, and colleagues can't even land an interview let alone get a job. I've been turned down from minimum wage jobs from Lowe's, Home Depot, Walmart, McDonald's, Target, restaurants/service industry, etc., as well. The market is brutal right now unless you're working in the trades/blue collar. The contrast from 2023 is stark when I was initially seeking jobs post-grad.
It’s not even as bad as it could be, either. May worsen. My peers graduated in 2011, 2012. Dead in the water lol as bad as it feels, it could be worse!
@@Oakbeast I think you might be the fresh fish. Were you in the workforce in 2008-2009? THAT was a bloodbath. It's super soft right now, has been since the COVID blip in 2020.
This is your best episode this year. Great insight & commentary, the guest was great as well. I really enjoyed listening; more of this please! Also bring Claire on the pod more often, she brings good energy.
Here's the reality we need more registered nurses and Physicians assistants. We don't need more MBA students. Also local government is hiring in many areas in the country municipalities, County and State level. Maybe your MBA can be used in place of a MPA. I hate to say it but people need to Pivot in the job market. We need more people in healthcare. It is the fastest growing field, and will continue to be the fastest growing field as the population ages. Many registered nurses can make what Engineers do. Normal Engineers not the California software engineers.
The big thing with deepseek is the open source element of it. That breaks down the whole openAI business model and allows the general population to iterate and improve deepseek
Thank you for the statement about taking some profits off the table. Daddy is not looking to buy a jet but thank you for reinforcing that profit taking is okay.
Deepseek's impact on the markets is driven by its efficiency claims, both in cost and compute. Neither claim is backed by evidence, and they will not provide it. The real story is this - China learned that it can strongly impact the US market with nothing more than a press release. Imagine XYZ company starts selling a Tesla EV equivalent and announces they can produce them for $1,000 and ship them worldwide for $50. The car is real, and nothing special, but would the US market react in the same way? Would no one even ASK for proof? Furthermore, if the claims turn out to be true (unlikely), that means that it's possible to do even MORE with compute, for less cost. That is good for US companies, not bad. Almost all the reporting on this topic is embarrassing and disappointing due to missing these two key points.
My friend, DeepSeek R1 can be downloaded today for free, and ran locally, on relatively inexpensive hardware. It’s real, the proof is readily available. It also doesn’t deliver AGI, so the AI race is still underway. The claimed hardware and dollars required to create the R1 model, is an interesting but immaterial side note.
@@jimgsewellI was gonna say, the scariest thing for nvda is deepseek is open-source. Open ai is in a panic because they no longer have a monopoly on the inner workings of the LLM.
I’ve been saying this for a while due to the way the tech job market has fluctuated in the past 10 years. I genuinely think we’re gonna be talking seriously about unionization in tech in the next few years, especially if people keep losing jobs to AI and if H1B Visa recipients continue to b scapegoated.
Deepseek is the beginning of PC comparing to Mainframe moment. Pretty much all good news for everyone, more use cases, cheaper and higher quality AI products, except Mainframe people who have NOT made their money back.
This is a very interesting interview. I side with Ed who sees the deepseek phenomenon as rather made-up by China. I also liked Robert Armstrong's views.
Either way…it has opened up a new AI frontier which is open source and not closed and monopolistic…there are further models which will be released by in the coming weeks…those will be refined further…welcome to competition and a free market…rather than rent seeking dominance 🎉
Scott hit on a great point regarding Deepseek. If people want (and can leverage) the Temu of LLMs, then Deepseek is their baby. If not, they are likely going to look to the 'bespoke' AI tools that cost more to develop.
Perhaps a contrarian view but I feel the biggest beneficiary of this would be Google (assuming they’re able to execute). Simply because Google has distribution; LLM and search seem to be such a natural fit. If Google is able to give users AI generated inputs at very low cost along with the search results. It would simply become the default product for a lot of day to day users. Ofc more complex tasks, I guess a bit like Saas companies we will have people developing for specific use cases.
Is it just me or is Ed really bland and pedestrian? He lacks any real insight and it’s painful and boring watching him be pushed forward to offer his opinion and become a star…
China proved it's on the cutting edge of technology. Deepseek is not just cheaper, it's countless times more efficient and when efficiency is all that matters, that just means the product is much better.
If DeepSeek is a big step up in efficiency, that just means your massive AI data centre will get more work done. Getting more work done is the key. You want the enormous super computer. The value isn't in the model, it's in the work. A better, Open Source model can be built upon and we can all get more work done. So Nvidia are still going to sell all their GPUs. It doesn't matter how they did it if they've improved on Chat GPT in either how well it does the work, or how efficiently it does it, if they've made the model open source and everyone else can build on it or duplicate it.
It doesnt matter how much it cost China to train DeepSeek. Its an open source application. The money was spent, and its gone, and now what remains is a product anyone can use. And people have used it. Have ran these models on mac minis instead of supercomputers. That value is immediate and none of us paid a dime for it to exist, and dont need to pay a dime for it going forward--regardless of how much money it took to train.
As a Boomer-I note: My employer donated funds from my paycheck to Social Security "because that's the way it's done." I also contributed to an IRA. How do you think I can receive the Social Security payments and then donate the amount to a charity? Which charities do you suggest? Thanks.
I feel a deflationary song coming on........."Anything you can buy, I can buy cheaper. I can buy anything cheaper than you! Fifty cents?, Forty cents!, Thirty cents?, Twenty cents! No, you can't!, Yes, I can, Yes!, I can...............
32:22 About ORCL shares dropped on Deepseek news, because its share went up quite a bit upon Project Stargate Announcement. It went up from $172 (2025-01-20) to $186 (2025-01-22). Upon the news of Deepseek, it dropped to $158 (2025-01-26). It was an oversold. Now, it bounced back to $170 today (2025-01--30), i.e.the Pre-Stargate Announcement level.
Hi Ed, Deepseek disclosed that they trained Deepseek - V3 model (which is the smaller version model released in Dec 2024) about 6 million dollar. But they did not disclose how much money they spend to train the R1 model (which is the current model with COT and comparable to OpenAI-01). Just to add that note. I think it can be easily verified on their website. They did not lie on this point. I think the company is decent to share all the secret technique in Open Source. FYI
To Ed's point, Geoffrey Hinton the British data scientist widely consider the Godfather of AI stated yesterday that he is dubious of the relative cost of DS stated. He thinks OpenAI probably spent 100 m on their final training run, not billions, opposed to DS's 6 mil. So still a very significant drop in cost, but one that is now more believable. Source and fascinating interview: th-cam.com/video/vxkBE23zDmQ/w-d-xo.html
"Our FT writers have been writing for months that this would happen." That is a very bold statement. Please provide non pay-walled evidence. How many FT bureau staff are based in China these days? Why do you have so many staff focussing on an uninvestable economy?
In his magnum opus The Best and the Brightest David Halberstam included a gloriously funny passage about the British aristocracy and their interaction with the Chinese when they began trading in the mid 1800's. Ed would do well to fire up his chatgpt and find it. Key words include gunpowder, monocles and arrogance.
Guys you are just not getting it. Even if we assume their claims are wrong (which i dont think is the case, people are convulting training costs of the base model with fine tuned model i.e..R1), Open AI behaves like closed AI , not disclosing chain of thought tokens, not disclosing how the models were trained & then on the side you have a chinese company open sourced the model (only the weights though) as well as a paper on all the training details. Even if the chinese are lying, they have helped the open source community & potentially the world as well.
Everyone keeps getting this wrong about Nvidia: it's not that their chips are so great, it's the power of CUDA low level programming API's. If it were just hardware then everyone would be using ASICs and FPGAs for training / inference rather than GPUs. RISC-V chips that are simple and just require concurrency are very easy to design compared to Nvidia products but Nvidia has the software development ecosystem everyone wants in tandem with their hardware.
When you talk about MBAs not getting a job within the first 3 months, I would assume a fair majority is the fact that they didn’t get the job offer they wanted or the company they wanted.
I downloaded DeepSeek and asked it what to do if you didn’t vote for Trump and don’t approve of his policies. The response included 12 very valid actions and was entirely in support of democracy.😮😂
American economy : running strong (though Stanford graduates can’t find jobs) - Chinese economy : crashing because their youth can’t find jobs. I think they both lie on some degrees and calling each other lairs. What’s more important is their own people believe their own government and calling the other lairs…
25:58: This is the most convoluted bs I’ve seen in awhile, though ofc protects the elite’s interests rofl: “maybe this is great for Google right maybe this is great for Microsoft who were shoveling money on the assumption that they had to build it themselves at great expense and then they wake up on Tuesday morning and they say maybe we don't have to own it at all it won't be at great expense so that maybe their capex capex budget they're they're they basically investment budget just went down and their revenues did not I that's possible it will be fascinating.”
Aren’t those numbers all very inflated and is there some gouging going along if China did it for the right price? What does it say about our economic way of making money or the greed that the consumer is victim of?
Id like to offer another analogy. To my way of thinking DeepSeek is akin to the cotton gin. Modern wage slaves are unlikely to profit from this technology. Unlike with the cotton gin, the value of wage slaves will likely be lower as more intelligent work can be accomplished by lesser educated/credentialed people. The initial market reaction is accurate. DeepSeek will be deflationary. It reduces costs and has to be used to remain competitive. As some readers know, the inventors of the cotton gin did not die in the lap of luxury. I would suspect this was a plot to attenuate American economic power and so far it is successful. BTW as usual --- luv the show.
You can always do inference on lower-end hardware-just adjust the quantization. Deepseek must use low end hardware because they don't have the ability to leverage high end Nvidia hardware -- after all, openly admitting they used high-end hardware would also mean admitting they got their hands on restricted Nvidia tech. This entire deepseek story is quite suspect and the markets reaction laughable.
The DeepSeek market correction was an overreaction. This LLM, being a distillation of open source LLMs, does not invalidate the AI scaling laws previously propelling the market. Compute capacity still rules in AI, although now we now find that we can get even more out of it. Reduced cost will drive further AI growth and consequently increased demand for chips, servers and energy.
Not sure there is anything extraordinary about 50K H100s in Singapore, it has over 70 data centers itself of 1.4GW. China has their own chips for inference, and everything is cheaper in China, including their energy and 110GW of Solar + Battery.
The international trade theory is going to burn you. 12 years ago, when there were fewer conflicts and Europe hadn't start pulling the plug on its own energy sources, there was an argument to be made that there was a lot of potential growth outside of the U.S. If you invested in VXUS since then, you are not a happy camper. And things are not going to suddenly turn around internationally. Good luck, Scott, with making that trade work. It will have the same outcome as betting against the Yen.
As a person who works with LLMs for a living, this expert knows nothing about the field and it's quite embarrassing. Every open source model can be run in your local machine. The delta here is that the training and inference uses different techniques than in the past and it is substantial. However, deepseek is currently quite bad in many things and the API doesn't work at all. Ed's claims have a lot of merit.
Inference is still expensive and DeepSeek in no way means we will not be pouring massive resources into AI. So many people outside the industry do not understand computing.
None of the AI leaders have hit AGI, so the race isn’t over. All of the AI leaders can incorporate the same techniques that DeepSeek uses, into their own models, so DeepSeek’s current advantage is short lived. The biggest contributor to the market downturn is Trump announcing new 25%-100% tariffs on Taiwanese chips. If Trump actually follows through with tariffs this steep, expect the selloff to grow. We are still at the beginning of AI. The industry should expect occasional leaps forward in ability and efficiency as new training/design techniques are discovered/invented. DeepSeek R1 is a terrific model. If your AI needs can be satisfied with that level of model, congratulations, you have an inexpensive solution available today. If you are waiting for AGI, the race for best foundation model, is still going strong.
Quite frankly, the majority of Americans care about AI as much as they care about bitcoin. Less than nothing. In fact they dread its potential to make their lives even more complicated & dystopian for zero reason at all. The idea of tax dollars pouring into it while they can’t afford healthcare & higher education is enraging. Seeing China knock the legs out from under the industry is actually gratifying. It’s as if there’s an obscene profit margin built in to the American effort. You think?
When someone says "several orders of magnitude" and the metric is most certainly NOT several orders of magnitude different (the training costs were not claimed to be 'several orders of magnitude' lower), then I tend to stop listening to anything that person says. Why? Because he probably doesn't even know that he doesn't understand what he's talking about. He's already made several other either slightly or totally wrong statements / claims. He clearly doesn't really know much about either AI or the technology. Bzzzt. Next.
Open AI charges $60/m token output for their 01 model. DeepSeek R1 is estimated to cost $0.03 - $0.05/m token output. That’s a few orders of magnitude difference.
MBA students are graduating too young and never networked in college. It's who you know not what you know. That's the whole point of the HARVARD degree, networking.
@@tonyjones3514 It makes the videos unshareable. There is a lot of good content in his podcasts that people would like to share, and they can't reasonably say ignore the first 30 seconds of dirty jokes while you are at the office.
30:15 - I am still amazed that people do not understand what it takes to run AI compute (just inference) for a thousand, hundred thousand, millions, billions of customers 24/7. People and businesses without expertise are going to run the models? How, in which app and collaboration suites? They are going to build it themselves? No? ISVs are going to? OK, then where do they run the AI compute? You still need to incredibly large scale infrastructure all around the world. The model gets optimized 10x ,100x ,1000x ... Great you can sell your products and services to 10x 100x 1000x more customer than $AMZN $GOOG $META $MSFT had anticipated because the compute capacity was the bottleneck. $NVDA going down? It is still a made race for the biggest AI cloud providers to scale up globally.
How much of the hype around DeepSeek is just a result of the novelty phase of the product? Additionally, since the explosion of users will increase the cost/demand for more advanced chips, do you think the conversation around how cheap it was to make the model is misplaced?
Whoever edits these episodes...needs to show us the faces of the people after Scott finishes his corny jokes😂 The intro ruins the fun especially since we get a preloaded ad🤦.
Supposedly lower cost, lower cost GPU's, lower number of GPU's. Supposedly. From the functional point of view I think Deepseek as an Opensource product can be downloaded and run off line or online. Thats dramatic. The process it uses to create results is far better than the others.
"Should interest rates come down." > This is like a person walking into a bank and saying "I want to borrow money...and i want to set the interest rate i borrow at." Its a ridiculous notion. The US Govt is the largest borrower of US Dollars in the world. Of course the leader of that government wants lower interest rates. All borrowers do. But lenders have the power, and until Trump is the one lending money, he has no say in the matter.
Great guest, Robert Armstrong. Ian Bremmer is another great guest.
Two minutes of banter (no matter how mundane) between Scott and Ed is like drinking freshly squeezed orange juice.
Its more like grapefruit for Ed. Poor kid does not yet understand that banter means listening to the boss's lifestyle flexes and just smiling politely.
I swear it's my best part of the show. I find myself rewinding sometimes. Mans is actually hilarious. Although I know it's a PR nightmare waiting to happen on one of the edginess jokes 😅
It’s the worst part of the show. It’s like the mindless small talk people insist on before meetings. I don’t know you, idgaf what you did on the weekend. Probably another reason why I’m so unsuccessful.
11:43 that’s the smile that tells how proud is Prof G of Ed’s analysis
This is my favorite comment I’ve seen about the DeepSeek vs Nvidia news so far:
“The one good takeaway from the Deepseek vs Nvidia news, I think, is people might finally realize we don't need to devote more than half the world's energy resources, or set up a Dyson Sphere around the sun, just to help computers solve trillions of multiplication problems to spit out another thousand mediocre web apps.” - Jeff Geerling
Well, I just wrote "Them chizzles ate. our. lizzunch." So, ya might push that one aside now?
That's awesome. I can understand why we would want to keep pushing down the path of robotic automation to produce actual goods. But the "generative AI" interest escapes me completely. Do we really need computers spending huge amounts of resources to quickly create endless mountains of content that could only ever be consumed by other AI?
Lovin’ Robert Armstrong this morning 👍🏼
The winner of the DeepSeek R1 release are the cloud computing providers and the customers. Amazon just released DeepSeek R1 as a Bedrock Foundation model on AWS. AWS has profited enormously from creating services for Open Source software like Postgres, Redis, RabbitMQ, etc. Open Source AI models will be the exact same. If the AI companies cannot maintain a significant performance advantage, in the long run, open source models on cloud infrastructure will be lower cost variations of frontier foundation models that will crowd out proprietary competition because they are more cost effective to operate.
I like that you put links to the headlines articles in the description :)
I haven’t noticed them before and wanted to say thank you 😊
Great guest! Really brought insight and humor
57:48 my accounting teacher's advice was once your stock hits 100% gain, sell half the shares and rebalance -- and keep the rest in that initial stock basically forever. He did this with Microsoft and that stock alone let him retire a millionaire as it rose and split over and over again for the next 20 years. Keep spreading out, you only need one or two big hitters.
I laughed at the blind climber/dog joke 😅
Regarding the job market, this is the toughest market I've seen since graduating college in 2023. I've been out of work for 4 months now, having previously worked in wealth management/financial advising. There are minimal, almost non-existent entry-level white collar jobs in finance. Myself, peers, friends, and colleagues can't even land an interview let alone get a job. I've been turned down from minimum wage jobs from Lowe's, Home Depot, Walmart, McDonald's, Target, restaurants/service industry, etc., as well. The market is brutal right now unless you're working in the trades/blue collar. The contrast from 2023 is stark when I was initially seeking jobs post-grad.
It’s not even as bad as it could be, either. May worsen. My peers graduated in 2011, 2012. Dead in the water lol as bad as it feels, it could be worse!
Very true.
Wow, you guys are fresh fish Good luck out there. It’s a bloodbath.
Get used to the tradesmen living in the best houses and driving the best cars and stealing girls away from suits.
@@Oakbeast I think you might be the fresh fish. Were you in the workforce in 2008-2009? THAT was a bloodbath. It's super soft right now, has been since the COVID blip in 2020.
This is your best episode this year. Great insight & commentary, the guest was great as well. I really enjoyed listening; more of this please! Also bring Claire on the pod more often, she brings good energy.
52:43 😂😂 Greatest. Reaction. Ever. Especially, to a question i didn't understand.
Here's the reality we need more registered nurses and Physicians assistants. We don't need more MBA students. Also local government is hiring in many areas in the country municipalities, County and State level. Maybe your MBA can be used in place of a MPA. I hate to say it but people need to Pivot in the job market. We need more people in healthcare. It is the fastest growing field, and will continue to be the fastest growing field as the population ages. Many registered nurses can make what Engineers do. Normal Engineers not the California software engineers.
I love this podcast so much!!@🎉
Do these hosts disclose their portfolios?
Scott generally does. He seems quite transparent
We’re told Ed’s 401k is mostly domestic index like VTI or VOO
@@Delrin3030*occasionally does.
55:00- Crunchy Tune! Loves it
The big thing with deepseek is the open source element of it. That breaks down the whole openAI business model and allows the general population to iterate and improve deepseek
loved the opening
Thank you for the statement about taking some profits off the table. Daddy is not looking to buy a jet but thank you for reinforcing that profit taking is okay.
Oh look the banter joke.... its so great to hear it again, and again, and again, and ....... again.
Deepseek's impact on the markets is driven by its efficiency claims, both in cost and compute. Neither claim is backed by evidence, and they will not provide it. The real story is this - China learned that it can strongly impact the US market with nothing more than a press release. Imagine XYZ company starts selling a Tesla EV equivalent and announces they can produce them for $1,000 and ship them worldwide for $50. The car is real, and nothing special, but would the US market react in the same way? Would no one even ASK for proof?
Furthermore, if the claims turn out to be true (unlikely), that means that it's possible to do even MORE with compute, for less cost. That is good for US companies, not bad.
Almost all the reporting on this topic is embarrassing and disappointing due to missing these two key points.
My friend, DeepSeek R1 can be downloaded today for free, and ran locally, on relatively inexpensive hardware. It’s real, the proof is readily available. It also doesn’t deliver AGI, so the AI race is still underway. The claimed hardware and dollars required to create the R1 model, is an interesting but immaterial side note.
@@jimgsewellI was gonna say, the scariest thing for nvda is deepseek is open-source. Open ai is in a panic because they no longer have a monopoly on the inner workings of the LLM.
I’ve been saying this for a while due to the way the tech job market has fluctuated in the past 10 years. I genuinely think we’re gonna be talking seriously about unionization in tech in the next few years, especially if people keep losing jobs to AI and if H1B Visa recipients continue to b scapegoated.
Deepseek is the beginning of PC comparing to Mainframe moment. Pretty much all good news for everyone, more use cases, cheaper and higher quality AI products, except Mainframe people who have NOT made their money back.
“AI is going to enable 1 worker to do the work of 10”. Feel like we hear this every year…
This is a very interesting interview. I side with Ed who sees the deepseek phenomenon as rather made-up by China. I also liked Robert Armstrong's views.
Either way…it has opened up a new AI frontier which is open source and not closed and monopolistic…there are further models which will be released by in the coming weeks…those will be refined further…welcome to competition and a free market…rather than rent seeking dominance 🎉
Robert is fantastic. Thanks for having him on!!
0:47 now THAT’S GOOD!!! 😂😂😂
My favorite workout podcast, thanks for existing!
When I workout I don't listen when I get into it. I don't seem to be able to walk and chew gum at the same time.
I do lyft for a side hustle, I listen to your show and people always ask about it. I think I’ve turned on 30 to 40 people in the last year.
Scott hit on a great point regarding Deepseek. If people want (and can leverage) the Temu of LLMs, then Deepseek is their baby. If not, they are likely going to look to the 'bespoke' AI tools that cost more to develop.
Perhaps a contrarian view but I feel the biggest beneficiary of this would be Google (assuming they’re able to execute).
Simply because Google has distribution; LLM and search seem to be such a natural fit. If Google is able to give users AI generated inputs at very low cost along with the search results.
It would simply become the default product for a lot of day to day users.
Ofc more complex tasks, I guess a bit like Saas companies we will have people developing for specific use cases.
Feels like Ed went big on Nvida
Hahah I caught the same vibe
Is it just me or is Ed really bland and pedestrian? He lacks any real insight and it’s painful and boring watching him be pushed forward to offer his opinion and become a star…
@@19shall820 0 insight other than regurgitating CNN talking points.
Not many know: DeepSeek was also trained with Huawei's GPU. (Ascend 910C)
@@19shall820 I like Ed.
China proved it's on the cutting edge of technology. Deepseek is not just cheaper, it's countless times more efficient and when efficiency is all that matters, that just means the product is much better.
If DeepSeek is a big step up in efficiency, that just means your massive AI data centre will get more work done. Getting more work done is the key. You want the enormous super computer. The value isn't in the model, it's in the work. A better, Open Source model can be built upon and we can all get more work done. So Nvidia are still going to sell all their GPUs. It doesn't matter how they did it if they've improved on Chat GPT in either how well it does the work, or how efficiently it does it, if they've made the model open source and everyone else can build on it or duplicate it.
construction sites in my part of town are now empty of workers....for 3 or 4 days now.
Guess they will have to pay real American wages and have those people pay taxes
It doesnt matter how much it cost China to train DeepSeek. Its an open source application. The money was spent, and its gone, and now what remains is a product anyone can use. And people have used it. Have ran these models on mac minis instead of supercomputers. That value is immediate and none of us paid a dime for it to exist, and dont need to pay a dime for it going forward--regardless of how much money it took to train.
As a Boomer-I note: My employer donated funds from my paycheck to Social Security "because that's the way it's done." I also contributed to an IRA. How do you think I can receive the Social Security payments and then donate the amount to a charity? Which charities do you suggest? Thanks.
I like the addition of Clara and hope she’s on more. 👌🏾
I think you need an AI expert to chat about Deepseek, cause it's not that simple.
I feel a deflationary song coming on........."Anything you can buy, I can buy cheaper. I can buy anything cheaper than you! Fifty cents?, Forty cents!, Thirty cents?, Twenty cents! No, you can't!, Yes, I can, Yes!, I can...............
I don't see it. Nothing is cheaper than before
@@mrwilliamwonder On Bloombergs today....'Market Minutes';- US Bond Yields hint at stagflation risk.
32:22 About ORCL shares dropped on Deepseek news, because its share went up quite a bit upon Project Stargate Announcement. It went up from $172 (2025-01-20) to $186 (2025-01-22). Upon the news of Deepseek, it dropped to $158 (2025-01-26). It was an oversold. Now, it bounced back to $170 today (2025-01--30), i.e.the Pre-Stargate Announcement level.
27:44 yea that went over my head too Scott 😅
What is Magnificent 10 at 07:52? Never heard of it
Top 10 stocks by market cap in the US
@Beangoeszoom thanks
Hi Ed, Deepseek disclosed that they trained Deepseek - V3 model (which is the smaller version model released in Dec 2024) about 6 million dollar. But they did not disclose how much money they spend to train the R1 model (which is the current model with COT and comparable to OpenAI-01). Just to add that note. I think it can be easily verified on their website. They did not lie on this point. I think the company is decent to share all the secret technique in Open Source. FYI
To Ed's point, Geoffrey Hinton the British data scientist widely consider the Godfather of AI stated yesterday that he is dubious of the relative cost of DS stated. He thinks OpenAI probably spent 100 m on their final training run, not billions, opposed to DS's 6 mil. So still a very significant drop in cost, but one that is now more believable. Source and fascinating interview: th-cam.com/video/vxkBE23zDmQ/w-d-xo.html
"Our FT writers have been writing for months that this would happen."
That is a very bold statement. Please provide non pay-walled evidence. How many FT bureau staff are based in China these days? Why do you have so many staff focussing on an uninvestable economy?
In his magnum opus The Best and the Brightest David Halberstam included a gloriously funny passage about the British aristocracy and their interaction with the Chinese when they began trading in the mid 1800's. Ed would do well to fire up his chatgpt and find it. Key words include gunpowder, monocles and arrogance.
Guys you are just not getting it. Even if we assume their claims are wrong (which i dont think is the case, people are convulting training costs of the base model with fine tuned model i.e..R1), Open AI behaves like closed AI , not disclosing chain of thought tokens, not disclosing how the models were trained & then on the side you have a chinese company open sourced the model (only the weights though) as well as a paper on all the training details. Even if the chinese are lying, they have helped the open source community & potentially the world as well.
Everyone keeps getting this wrong about Nvidia: it's not that their chips are so great, it's the power of CUDA low level programming API's. If it were just hardware then everyone would be using ASICs and FPGAs for training / inference rather than GPUs. RISC-V chips that are simple and just require concurrency are very easy to design compared to Nvidia products but Nvidia has the software development ecosystem everyone wants in tandem with their hardware.
When you talk about MBAs not getting a job within the first 3 months, I would assume a fair majority is the fact that they didn’t get the job offer they wanted or the company they wanted.
Prof G may know lot about Liquor stocks, can you please cover about beaten down spirit and beer stocks like Diageo. It is non US as well.
Breaking News! World leaders are congratulating CHINA , DEEPSEEK RI for helping the world for FREE ! Thank you CHINA!😊😊😊
I downloaded DeepSeek and asked it what to do if you didn’t vote for Trump and don’t approve of his policies. The response included 12 very valid actions and was entirely in support of democracy.😮😂
American economy : running strong (though Stanford graduates can’t find jobs) - Chinese economy : crashing because their youth can’t find jobs. I think they both lie on some degrees and calling each other lairs. What’s more important is their own people believe their own government and calling the other lairs…
Editors might want to make a cut at 28:30
Tragedy took place between filming and going live
Robert didn’t once mention the refinancing wall of US govt debt. That’s why “interest rates must come down”.
Do people realise AI just went from a US technology to a global technology? Bullish AF for AI infra
don't rebalance by selling your long term winners, just add into new under-allocated positions to round out the portfolio pie.
China lies alot. Rofl. Well said Ed.
Where do the assets that Ed is saying trade at a discount to value trade? I thought he was talking about private investments
25:58: This is the most convoluted bs I’ve seen in awhile, though ofc protects the elite’s interests rofl: “maybe this is great for Google right maybe this is great for Microsoft who were shoveling money on the assumption that they had to build it themselves at great expense and then they wake up on Tuesday morning and they say maybe we don't have to own it at all it won't be at great expense so that maybe their capex capex budget they're they're they basically investment budget just went down and their revenues did not I that's possible it will be fascinating.”
Paris literally sucks…
I was told almost a half century ago. (So does it remain true?) That Paris was just another dirty city. (or something like that)
Too many Arabs. They’ve ruined the place, as per usual.
Ok Prof G.
Why isn't Claire's photo included in the channel's thumbnail?
Aren’t those numbers all very inflated and is there some gouging going along if China did it for the right price? What does it say about our economic way of making money or the greed that the consumer is victim of?
Robert Armstrong is Awesome. Please bring him in more often.
Id like to offer another analogy. To my way of thinking DeepSeek is akin to the cotton gin. Modern wage slaves are unlikely to profit from this technology. Unlike with the cotton gin, the value of wage slaves will likely be lower as more intelligent work can be accomplished by lesser educated/credentialed people. The initial market reaction is accurate. DeepSeek will be deflationary. It reduces costs and has to be used to remain competitive. As some readers know, the inventors of the cotton gin did not die in the lap of luxury. I would suspect this was a plot to attenuate American economic power and so far it is successful. BTW as usual --- luv the show.
6:47 - these first steps from a 2 year old market release! (decades of research) What’s it looking like in 2035+😮
We are in the 14K download speed point in AI development.
I can't wait for 28.8 speed 😉
You can always do inference on lower-end hardware-just adjust the quantization. Deepseek must use low end hardware because they don't have the ability to leverage high end Nvidia hardware -- after all, openly admitting they used high-end hardware would also mean admitting they got their hands on restricted Nvidia tech. This entire deepseek story is quite suspect and the markets reaction laughable.
The DeepSeek market correction was an overreaction. This LLM, being a distillation of open source LLMs, does not invalidate the AI scaling laws previously propelling the market. Compute capacity still rules in AI, although now we now find that we can get even more out of it. Reduced cost will drive further AI growth and consequently increased demand for chips, servers and energy.
This sounds like Seth Rogan if he didn't smoke weed hahaha
Not sure there is anything extraordinary about 50K H100s in Singapore, it has over 70 data centers itself of 1.4GW. China has their own chips for inference, and everything is cheaper in China, including their energy and 110GW of Solar + Battery.
China isn't the only place that lies...
OpenAI is an Open, Transparent, Non-Profit, Sam Altman owns none of it and will never profit from it and Sam Altman never lies.
There's a glut in the MBA market, is that the take away?
There are a number of issues with DeepSeek (e.g., data collection from users; back-end data leakage; model theft?) so be careful in using.
Don’t trust anything coming out of China even if it’s open source. I simply won’t use it.
The international trade theory is going to burn you. 12 years ago, when there were fewer conflicts and Europe hadn't start pulling the plug on its own energy sources, there was an argument to be made that there was a lot of potential growth outside of the U.S. If you invested in VXUS since then, you are not a happy camper. And things are not going to suddenly turn around internationally. Good luck, Scott, with making that trade work. It will have the same outcome as betting against the Yen.
take what you want, give what you can
Do you guys know about the ICP blockchain - Onicai project currently running Deepseek R1 models on 32-bit canisters on fully on chain.
Marc Andreessen would be a good guest
He lacks integrity. He thinks Zuck is a ‘good guy’. Laughable.
As a person who works with LLMs for a living, this expert knows nothing about the field and it's quite embarrassing.
Every open source model can be run in your local machine. The delta here is that the training and inference uses different techniques than in the past and it is substantial. However, deepseek is currently quite bad in many things and the API doesn't work at all. Ed's claims have a lot of merit.
Inference is still expensive and DeepSeek in no way means we will not be pouring massive resources into AI. So many people outside the industry do not understand computing.
52:44 that's me watching the nightly news with trump in office
None of the AI leaders have hit AGI, so the race isn’t over. All of the AI leaders can incorporate the same techniques that DeepSeek uses, into their own models, so DeepSeek’s current advantage is short lived. The biggest contributor to the market downturn is Trump announcing new 25%-100% tariffs on Taiwanese chips. If Trump actually follows through with tariffs this steep, expect the selloff to grow. We are still at the beginning of AI. The industry should expect occasional leaps forward in ability and efficiency as new training/design techniques are discovered/invented. DeepSeek R1 is a terrific model. If your AI needs can be satisfied with that level of model, congratulations, you have an inexpensive solution available today. If you are waiting for AGI, the race for best foundation model, is still going strong.
Quite frankly, the majority of Americans care about AI as much as they care about bitcoin. Less than nothing. In fact they dread its potential to make their lives even more complicated & dystopian for zero reason at all. The idea of tax dollars pouring into it while they can’t afford healthcare & higher education is enraging. Seeing China knock the legs out from under the industry is actually gratifying. It’s as if there’s an obscene profit margin built in to the American effort. You think?
What tax dollars are being spent on AI?
When someone says "several orders of magnitude" and the metric is most certainly NOT several orders of magnitude different (the training costs were not claimed to be 'several orders of magnitude' lower), then I tend to stop listening to anything that person says. Why? Because he probably doesn't even know that he doesn't understand what he's talking about. He's already made several other either slightly or totally wrong statements / claims. He clearly doesn't really know much about either AI or the technology. Bzzzt. Next.
Open AI charges $60/m token output for their 01 model. DeepSeek R1 is estimated to cost $0.03 - $0.05/m token output. That’s a few orders of magnitude difference.
Love RA
MBA students are graduating too young and never networked in college. It's who you know not what you know. That's the whole point of the HARVARD degree, networking.
Chat and DeepSeek will replace search. I don’t even use google anymore.
Thanks for not doing porn jokes
you don't like them?
@@tonyjones3514 It makes the videos unshareable. There is a lot of good content in his podcasts that people would like to share, and they can't reasonably say ignore the first 30 seconds of dirty jokes while you are at the office.
@@donaldjohnson-y6nnot only that, the porn jokes are aweful. They are not thoughtful or funny.
@@alexk729if you didn't know Prof G you'd think "why the hell is my friend or family sending me a video of a dirty old man??"
I love his dad jokes ☹️
30:15 - I am still amazed that people do not understand what it takes to run AI compute (just inference) for a thousand, hundred thousand, millions, billions of customers 24/7.
People and businesses without expertise are going to run the models? How, in which app and collaboration suites? They are going to build it themselves? No? ISVs are going to? OK, then where do they run the AI compute?
You still need to incredibly large scale infrastructure all around the world.
The model gets optimized 10x ,100x ,1000x ... Great you can sell your products and services to 10x 100x 1000x more customer than $AMZN $GOOG $META $MSFT had anticipated because the compute capacity was the bottleneck.
$NVDA going down? It is still a made race for the biggest AI cloud providers to scale up globally.
How much of the hype around DeepSeek is just a result of the novelty phase of the product? Additionally, since the explosion of users will increase the cost/demand for more advanced chips, do you think the conversation around how cheap it was to make the model is misplaced?
16:56 struggling to find work? Or too proud to start at the bottom?
Whoever edits these episodes...needs to show us the faces of the people after Scott finishes his corny jokes😂
The intro ruins the fun especially since we get a preloaded ad🤦.
install the plugin that skips ads.
Let Ed tell a joke. DO IT!
he is the joke though, so cynical so incurious
How did the deepseek story even break? Was it from the scaleAI guy? Is he credible?
Supposedly lower cost, lower cost GPU's, lower number of GPU's. Supposedly.
From the functional point of view I think Deepseek as an Opensource product can be downloaded and run off line or online. Thats dramatic. The process it uses to create results is far better than the others.
So we're the ai companies here misrepresenting the cost to build to pump up their stock values believing they held all the cards?
"Should interest rates come down." > This is like a person walking into a bank and saying "I want to borrow money...and i want to set the interest rate i borrow at." Its a ridiculous notion. The US Govt is the largest borrower of US Dollars in the world. Of course the leader of that government wants lower interest rates. All borrowers do. But lenders have the power, and until Trump is the one lending money, he has no say in the matter.
learn about the passive bid, that's all that matters