Anthropic CEO Reveals New Details About DeepSeek R1

  Рет қаралды 178,512

Matthew Berman

Matthew Berman

Күн бұрын

Try Recraft, a premium image generation and editing tool for creators and teams, and get $12 off with my promo code BERMAN12: go.recraft.ai/...
Join My Newsletter for Regular AI Updates 👇🏼
forwardfuture.ai
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
👉🏻 Instagram: / matthewberman_ai
👉🏻 Threads: www.threads.ne...
👉🏻 LinkedIn: / forward-future-ai
Media/Sponsorship Inquiries ✅
bit.ly/44TC45V
Links:
x.com/DarioAmo...
darioamodei.co...
Disclosure: I am a small investor in Groq.

Пікірлер: 2 100
@pb5109
@pb5109 7 күн бұрын
The most important thing about DeepSeek is the fact that they open-sourced it. This is gonna be a good foundation for future models that are likely to be much much better than anything private
@matthew_berman
@matthew_berman 7 күн бұрын
100%
@niyazzmoithu20
@niyazzmoithu20 7 күн бұрын
Yeah. 100%.
@maxziebell4013
@maxziebell4013 7 күн бұрын
200%
@TheReferrer72
@TheReferrer72 7 күн бұрын
That is wishful thinking.
@maxziebell4013
@maxziebell4013 7 күн бұрын
@@TheReferrer72 Yeah, probably it's wishful thinking - the real aspect of it is that it's much better than anything most other countries than the U.S. had prior to this release. So it does in a way level the playing field (For now, at least.).
@DailyHomerClips
@DailyHomerClips 7 күн бұрын
OpenAI literally distills the whole internet and that's okay, but when DeepSeek allegedly does it, now it's a public freakout
@Synthalog
@Synthalog 7 күн бұрын
Well said!
@matthew_berman
@matthew_berman 7 күн бұрын
Exactly.
@justinmeader
@justinmeader 7 күн бұрын
It's more about the claims that they started from scratch and made this model for $6m, that's the part that's disingenuous.
@wadecrawford8494
@wadecrawford8494 7 күн бұрын
You seem to have glossed over the accusation that DeepSeek distilled from stolen Openai data that was created by crawling the internet and compiling it for years. it is not even close to the same thing. It like a farmer claiming they can grow corn in a week and the only thing they did was go to another farmer and empty their grain silo.
@MoeShlomo
@MoeShlomo 7 күн бұрын
In other words, DeepSeek did to OpenAI what OpenAI did to countless other sources. The irony of OpenAI complaining about it could not be any richer.
@shonnspencer1162
@shonnspencer1162 7 күн бұрын
Mathew, I appreciate 99.9% of your reporting; however, to do this topic justice, you should start by emphasizing that OpenAI was originally intended to be an open-source project. Every single model today is being trained on redundant data-we all know this. It is critical to recognize that these models are being trained on data that we, as a community, own. Many organizations and companies have co-opted this data, and now they claim that the very data they took from us is somehow being stolen from them. This is an outrageous stance that is damaging to the AI community. We all understand that these organizations will go to great lengths to discredit open-source AI due to the enormous financial interests at stake. This will not be the last attempt to undermine open models, and they will continue to look for ways to attack and discredit them. However, those of us in the AI community remain committed to open source and will not fall for these tactics. In reality, they are guilty of the very same-if not worse-practices with our data.
@seanpitcher1102
@seanpitcher1102 7 күн бұрын
I agree 100% there should be a lawsuit if open AI isn't going to open its models then they stole data from sources on the web. If it's open we were fine with accepting that they used public data for public benefit. However, it's NOT for public benefit anymore!
@edwardduda4222
@edwardduda4222 7 күн бұрын
Most people know that Open AI, at least Mathew’s audience. We’re all AI professionals/enthusiasts here and he’s talked about in other videos.
@normanlove222
@normanlove222 7 күн бұрын
Couldn't say it any better.
@Dan-e4h7b
@Dan-e4h7b 7 күн бұрын
I knew not reading the fine print and clicking the "I Agree" button would bite us in the ass.
@b3nz0
@b3nz0 7 күн бұрын
I concur.
@LakesouthTiger-tw6es
@LakesouthTiger-tw6es 6 күн бұрын
No.1 No one said the total cost of Deepseek project is $6 million. They only said the cost of training model is $6 million; No 2. Distillation is legal, it is the standard technique in the AI industry. Everyone does the same. So did Open AI GPT. There is no ground to accuse Deepseek of “stealing”. No. 3 Deepseek made it opensource and available for everyone to use. This is a big deal as it will spread the use of AI in the world which in return will lead to even greater innovation in AI. NO. 4 The main innovation is how Deepseek structure the GPUs for computing which made its efficiency ten time better than OpenAI. No one point that out. Recognizing its state of the art reasoning capabilities, Nvidia now offers Deepseek R1 model in its NIM microservices.
@michaelwoythaler
@michaelwoythaler 6 күн бұрын
Deepseek is a warning shot from China.
@1963tjc
@1963tjc 6 күн бұрын
1000% 🙏🙏
@thecount25
@thecount25 6 күн бұрын
This is more about calming investors than anything else
@ajithboralugoda8906
@ajithboralugoda8906 6 күн бұрын
hey you might be correct but the Gas guzzler Model guys need the Gravvy train keep going!! ;-/)
@mjk9388
@mjk9388 6 күн бұрын
If you steal something from someone else with the intent of giving it to everyone...it's still stealing. I'm not saying they stole it or didn't steal it...just pointing out the moral.
@chrismanchin
@chrismanchin 6 күн бұрын
Stealing from a thief and give it back to the owners is a great deed. A modern Robin Hood.
@rambrasil
@rambrasil 6 күн бұрын
Really CCP is Robin Hood? Education has failed us
@jimoconnor7491
@jimoconnor7491 5 күн бұрын
They didn't give the data "back to the owners" though.
@ttemp2631
@ttemp2631 5 күн бұрын
@@jimoconnor7491 They make the data available "back to the owners" ok
@celxoirealyx
@celxoirealyx 5 күн бұрын
​@@jimoconnor7491 technically they do. OpenAI whole foundation is plagiarized and monopolized. Deepseek broke the monopoly and made it open source. It should've been like this to begin with because the whole tech wouldn't exist without the combined data stolen from said owners.
@celxoirealyx
@celxoirealyx 5 күн бұрын
​@@jimoconnor7491 technically they do by making plagiarized and monopolized tech open to everyone.
@medialcanthus9681
@medialcanthus9681 7 күн бұрын
The Indian whistle blower guy who accused openAi of stolen data was found dead. How about that?
@litewavve
@litewavve 7 күн бұрын
More background and detail about this case: - Suchir Balaji worked at OpenAI for nearly four years, including 1.5 years on the development of ChatGPT. - He left OpenAI in August 2024 due to concerns about the company's practices. - In October 2024, Balaji publicly accused OpenAI of violating US copyright law in developing ChatGPT. - He argued that generative AI models could produce substitutes for the data they were trained on, making fair use defense unlikely in many cases. - Balaji was found dead in his San Francisco apartment on November 26, 2024. - The San Francisco medical examiner's office initially ruled his death a suicide. - Police stated they found no evidence of foul play during the initial investigation. - Balaji's family, particularly his mother Poornima Ramarao, has disputed the suicide ruling and called for a more thorough investigation. - They claim that Balaji's apartment appeared ransacked and that some items, including a pin drive, were missing. - Congressman Ro Khanna has called for a "full and transparent" FBI investigation into Balaji's death. - New crime scene photos have emerged, casting doubt on the suicide theory. - Elon Musk has described the allegations surrounding Balaji's death as "extremely concerning". - The San Francisco Police Department has reportedly reopened the case as an "active investigation". The case continues to be a subject of intense debate and speculation, with many questioning the circumstances surrounding Balaji's death and its potential connection to his role as a whistleblower against OpenAI.
@freetorobandloot
@freetorobandloot 6 күн бұрын
the media won't talk about that because they are paid by these tech-bros.
@caleb9231
@caleb9231 6 күн бұрын
he was brown so no one cares
@alvarny77
@alvarny77 6 күн бұрын
Seriously? I know he was an employee in Open AI
@fredwarsaw5782
@fredwarsaw5782 6 күн бұрын
Yeah, that's Suchir Balaji. He was in his mid-20s
@sdnnvs
@sdnnvs 7 күн бұрын
Our leadership in AI... Americans are so self-centered that Narcissus would feel envy.
@shadeypotion
@shadeypotion 7 күн бұрын
I don't want US oligarchs getting AGI or ASI first. I feel much better with open source getting it first.
@tomirkpl
@tomirkpl 6 күн бұрын
Give money, you will have. No money, no AGI. :D You need oligarchs, even Chinese need oligarchs to train open source.
@PACotnoir1
@PACotnoir1 6 күн бұрын
@@tomirkpl We don't need plutocrats; instead, we need wise individuals and open-source research. It's fundamental in science.
@amdenis
@amdenis 6 күн бұрын
But it is not open source, and DeepSeek (actually, the CCP, who effectively owns/controls all in China) does NOT share the training and related data. THEY DO NOT! They share some of the derived Weights and Biases and source.
@tonnyww5ng
@tonnyww5ng 6 күн бұрын
Yes 100%
@jeremyschwartz3683
@jeremyschwartz3683 6 күн бұрын
Doesn't matter who gets it first. Everyone else will catch up quickly.
@kw-nz
@kw-nz 6 күн бұрын
This is like the version of the Olympic swimming competition. The United States claimed that the Chinese team members were doping in order to prevent the Chinese team members from winning the gold medal, but every American swimming team member had a purple face during the competition. If the United States cannot compare, it will frame others.
@jeremycazee
@jeremycazee 6 күн бұрын
Regarding the 3% inflation rate, it's important to note that it may not accurately reflect the true cost increases you're experiencing in your area. In our Midwestern city, real price inflation seems closer to 20%. Do you foresee a reduction in costs for essentials like food, gasoline, or insurance in the near future? We've significantly cut our spending and plan to continue doing so until prices decrease, as we feel it's our responsibility to adapt to these changes.
@PetterNeten
@PetterNeten 6 күн бұрын
Looking at the aggregated totals by kind since 2021, when it started to take off, you can see that the embedded amount of inflation currently approaches 20%.
@JosefaJeunesse
@JosefaJeunesse 6 күн бұрын
Right, a lot of folks downplay the role of advisors until being burnt by their own emotions, no offense. I remember some years back, amid covid-19 outbreak, I needed a good boost to help boost my business, hence I researched for licensed advisors and thankfully came across someone of excellence. Helped grow my reserve notwithstanding inflation, from $350k to nearly $1m as of today
@sonyacollins-v9n
@sonyacollins-v9n 6 күн бұрын
this is huge! your advisor must be grade A, mind sharing more info please? in dire need of proper asset allocation
@JosefaJeunesse
@JosefaJeunesse 6 күн бұрын
Finding financial advisors like Elizabeth Cordle Gross, who can assist you on things like investing, insurance, making sure retirement is well funded, going over tax benefits, ways to have a volatility buffer for investment risk would be a very creative option. There will be difficult times ahead, and prudent personal money management will be essential to navigating them.
@lisahaney-m4o
@lisahaney-m4o 6 күн бұрын
Thank you for this tip. it was easy to find your coach on Google. Did my due diligence on her before scheduling a phone call with her. She seems proficient considering her résumé.
@delxinogaming6046
@delxinogaming6046 7 күн бұрын
Say in words OpenAI will understand.. DeepSeek trained itself on "all publicly available data."
@petertaylor4954
@petertaylor4954 7 күн бұрын
😎
@antonystringfellow5152
@antonystringfellow5152 7 күн бұрын
Why? It didn't. Maybe watch the video again or stick to something simple enough for you to understand.
@mooonatyeah5308
@mooonatyeah5308 6 күн бұрын
@@antonystringfellow5152 OpenAI's models' output is completely publicly available data, if the Internet is considered as publicly available data by OpenAI.
@terjeoseberg990
@terjeoseberg990 6 күн бұрын
@@antonystringfellow5152, Of course OpenAI is upset that someone stole the training data that they stole first. It’s just like how Bill Gates stole Windows after Steve Jobs stole it from Xerox. Steve was obviously upset.
@hansouth2355
@hansouth2355 6 күн бұрын
@@antonystringfellow5152 those are the words open ai used when they were accused of theft of copy righted data
@michamohe
@michamohe 7 күн бұрын
how dare you steal data we stole from everyone else lol
@cywong9101
@cywong9101 7 күн бұрын
😂😂
@velorofonte-termonia
@velorofonte-termonia 7 күн бұрын
"HOW DARE YOU share my stolen data?" -OpenAI
@dfcx1
@dfcx1 7 күн бұрын
I thought outputs from AI weren't supposed to be copyrightable? So it's still only OpenAI that "stole" (pirated) anything. They should just shut up while they're ahead.
@wds3222
@wds3222 7 күн бұрын
Which Data Open AI stole for GPT4?
@EMJ30034
@EMJ30034 6 күн бұрын
🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣
@alliedeena1141
@alliedeena1141 7 күн бұрын
So basically thieves are accusing the thieves which are not actually thieves.
@KevinColin-co9io
@KevinColin-co9io 7 күн бұрын
This is the mentality of everything thieves. They don't want others to do what they did to them.
@andysol2002
@andysol2002 7 күн бұрын
It does hurt when someone puts all the effort to break in to a vault and haul everything in to a truck, only to realize the driver takes everything and runs away with it.
@ChristianThePagan
@ChristianThePagan 7 күн бұрын
What is that quip that Altman and friends like to quote: "Take what you need, apologise later" ... well, turns out they are not the only ones who can play that game. Quite frankly if they really did use Open AI's data, this could not have happened to a more deserving bunch of people.
@stnln2180
@stnln2180 7 күн бұрын
Said who?....CCP?
@rey82rey82
@rey82rey82 7 күн бұрын
Stealing works
@MoreSlater
@MoreSlater 6 күн бұрын
We dont need OpenAI.. Or META etc. So thank you DeepSeek for paving the road to open source.. As it should be. Its the safest option down the road
@GeneralPublic
@GeneralPublic 6 күн бұрын
Meta’s models are all open source just like DeepSeek, available for anyone to use. In fact DeepSeek did build on top of Meta’s models, and Meta will build on top of DeepSeek’s models. Exactly how it should be. And notice Meta is not among the companies complaining about DeepSeek, because they support open source just like DeepSeek does. NVIDIA is another company involved in open source models too including building on top of Meta’s models. Hopefully the companies that use open source like Meta and DeepSeek and NVIDIA will beat the closed models from OpenAI, Anthropic, Google, and others, and the closed models will become obsolete, as predicted in the Google memo that said that they have “no moat” (which is a good thing). And of course these export controls Anthropic is advocating for are a bad thing, as is this idea of a unipolar world with the U.S. government controlling everything. Much better and safer for everyone to have access to the same open source technologies anywhere in the world, that way nobody has any advantage and nobody can monopolize access to this technology for personal gain, and all of humanity can benefit. Sam Altman has even recently admitted that OpenAI has been on the wrong side of history by making their models closed source so maybe he might come around to the open source side. The last thing we need is one country dominating control of this technology and using it for military purposes and keeping it out of the hands of all other countries, that would lead to a dangerous power imbalance and that one country trying to take over the world. If things are more balanced, that will prevent any sort of dangerous moves like this because everyone will know that all of their potential adversaries have access to the same exact technology too, which will work as a deterrent against aggression.
@OneAndOnlyMe
@OneAndOnlyMe 6 күн бұрын
"Seem to be" - The DeepSeek team ARE legit, they have incredible talent. They've been working AI for ten years.
@aubergine10
@aubergine10 6 күн бұрын
The company was founded 2 years ago.
@jason.arthur.taylor
@jason.arthur.taylor 5 күн бұрын
@@aubergine10 No they were using AI in their quant trading, which goes back another 10 years. AI, with Neuroshell, was being used to trade stocks in the 1990s, so that's the relationship.
@LikeItDeep
@LikeItDeep 7 күн бұрын
DeepSeek is a billionaire bunker buster
@eight8byte
@eight8byte 7 күн бұрын
Exactly
@maxziebell4013
@maxziebell4013 7 күн бұрын
@@eight8byte Kaboom.
@antonystringfellow5152
@antonystringfellow5152 7 күн бұрын
DeepSeek is a China AI buster. Or do you understand nothing? DeepSeek is nothing more than a condensed model made using OpenAI's tech. As a result of what they've done here, much tighter restrictions will be put on chip exports to China. And they will be enforced.
@TheGalacticIndian
@TheGalacticIndian 7 күн бұрын
Remember the GeForce 256? Launched at $199. Now, after decades of growth and millions of new customers, the flagship should be WAY cheaper-yet the 5090 is pushing $1900. This is corporate milking on a GALACTIC scale😊
@matt.stevick
@matt.stevick 7 күн бұрын
did u even watch this? who are these commentators?
@VioFax
@VioFax 7 күн бұрын
So get DeepSeek before they ban it... Gotcha.
@matthew_berman
@matthew_berman 7 күн бұрын
There's no reason to ban it. Open source helps everyone!
@SapienSpace
@SapienSpace 7 күн бұрын
Large entities will want to ban anything that competes with them, even though they may publicly admit the opposite.
@phil488pista9
@phil488pista9 7 күн бұрын
​@@matthew_bermanbut your government is control by the overlord so there's that 😂
@electromega3077
@electromega3077 7 күн бұрын
@@matthew_berman DeepSeek is already banned in Italy. not joke.
@maxziebell4013
@maxziebell4013 7 күн бұрын
Oh yeah, we're gonna see a "that's a big one" incoming from Trump. Then he's gonna be signing off on an executive order with his black sharpie holding that thing up... DeepSeek banned.
@FrederickMbuya
@FrederickMbuya 7 күн бұрын
What I do find a bit frustrating is the assumption that the US getting dominance is ok, and china not ... ( An African who's perspective is too often ignored)
@jonz23m
@jonz23m 7 күн бұрын
That's the colonial mindset it never left them
@listeningtothevoid
@listeningtothevoid 7 күн бұрын
@@jonz23m I am thinking the same from Europe as a white guy.
@WaveOfDestiny
@WaveOfDestiny 6 күн бұрын
as an european, US oligarchy seems more of a treath than chinese "communism" these days...
@WaveOfDestiny
@WaveOfDestiny 6 күн бұрын
as an european, US oligarchy and censorship seems worse tbh
@MrMetalzeb
@MrMetalzeb 6 күн бұрын
becouse Cina already dominate every other single thing, from the smaller to the bigger and this frustrate me.
@RKie-hv6ci
@RKie-hv6ci 6 күн бұрын
The most critical question that OpenAI needs to answer the market and its shareholders is: if there is a technology called Distillation that can greatly reduce the operating cost and even slightly improve efficiency .... then WHY THE HELL OpenAI did not do it first to enhance the shareholders' value?
@mdmoz1777
@mdmoz1777 6 күн бұрын
OpenAi is not a public company. Microsoft has invested billions that may represent a 49% stake in the company.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
@@mdmoz1777 Then Microsoft needs to be questioned as to why they allowed OpenAI to hold back shareholder value.
@ABriefHistory-uq6br
@ABriefHistory-uq6br 6 күн бұрын
openAi is not open. they should change their name 😂. hypocrite just like merica
@Akofs37
@Akofs37 6 күн бұрын
The fact is OpenAI trained their model based on copyrighted content from the entire internet, sold it for $20 per month, is this Okay? DeepSeek took knowledge form o1, built a better R1 and gave it away to everyone for free with open-source code, yet they are the ones called thieves? 🤔AI Distillation is NOT copying!
@MingInspiration
@MingInspiration 7 күн бұрын
deepseek even give people the chain of thought. they not only have their bucket, but also they give people fishing rods, they didn't go mental about that
@foshizzlemanizzle4753
@foshizzlemanizzle4753 7 күн бұрын
I've never understood why anyone would get upset about "stolen" data. It's like saying reading a book at the library is stealing from the author
@marilynlucas5128
@marilynlucas5128 7 күн бұрын
They lost a lot of money that's why and the AI bubble burst! Now, they can't charge the $1,000 they were hoping to charge for a monthly subscription fee. It's done now. There's an open source version. AI just took their jobs like they have enabled AI to do so to many.
@litewavve
@litewavve 7 күн бұрын
There is a difference. For most copyrighted material, there is a fair use clause, which is for personal, non-commercial use. Therefore in the early days of Netflix when they were shipping DVDs, they couldn't just buy CDs at the off-the-shelf prices. They needed to pay royalties to the publishers.
@PeteQuad
@PeteQuad 6 күн бұрын
​@@litewavveThat's not an inclusive definition of fair use. Fair use also allows reproduction of at least some of the material for commercial use if you are adding something to it, such as for criticism or analysis.
@jay-shakeli9761
@jay-shakeli9761 6 күн бұрын
​@@litewavveDeepSeek open source their model. So, it is not a commercial use.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
@@litewavve Then OpenAI and Antropic need to show us their copyright application or their patent approvals ... there's a phrase ...'Proprietary information', they use it for 'data' that is not protected by copyright or patents, because copyright and patents allow for the sharing of information while protecting the authors and inventors ... if you choose to go the proprietary route, then all bets are off .. anyone who discovers this data can release it to the world and while you might be able to charge the person who 'stole it' with theft, you can't stop the world from using it since you decided not to legally protect the information, only the paper it was written on! The quickest way to lose competitive advantage is to rely on proprietary knowledge ... the best way to determine if a business is for the good of humanity is to find out if that business relies on proprietary knowledge, if it does, then beat it with a stick!!
@remsee1608
@remsee1608 7 күн бұрын
They didn’t steal OpenAI’s data. They paid for it. They also violated the terms of service. But it’s not theft since they paid to use the API
@technews-wb7pq
@technews-wb7pq 7 күн бұрын
That called stealing the deal is to pay and don’t brake the rules doing so terminates the agreement
@remsee1608
@remsee1608 7 күн бұрын
@ If I buy a burger and then learn from that burger to make and sell a better burger that is not theft
@neoglacius
@neoglacius 7 күн бұрын
@@technews-wb7pq they dont legally own that data because it was scraped without permission and without paying for it
@neoglacius
@neoglacius 7 күн бұрын
@@technews-wb7pq they dont legally own that data because they never paid for it
@jchanmcse
@jchanmcse 7 күн бұрын
​@@remsee1608That's true!
@5t0ryte11er
@5t0ryte11er 6 күн бұрын
What Deepseek gives to the world is not the LLM advisory but the tool for everyone to build an affordable LLM. Therefore it didn't deny it distilled from OpenAI and it tells you that the data it learned is until 2023. Deepseek didn't care whether the advisory apps is banned because what they want to give to the world is the source code, the idea, and the formula to build your own LLM.
@aw888492
@aw888492 6 күн бұрын
From 25th to 29th, 83 hours, DeepSeek server cluster, received more than 230 million DDos malicious requests per second, the total amount of attacks is equivalent to three days of European network traffic combined. A number of Chinese Internet companies came to the rescue to defend against the attack, and this epic story of working together to fight against foreign enemies will be forever recorded in history as a barbaric invasion was stopped.
@likereallyvainglorious9180
@likereallyvainglorious9180 7 күн бұрын
Open AI complaining that DeepSeek inappropriately distilled off of their model is a hilarious joke
@niyazzmoithu20
@niyazzmoithu20 7 күн бұрын
yeah😂.
@galdutro
@galdutro 7 күн бұрын
I think that US export controls only undermines the US position in the world. If China comes with an architectural breakthrough in their GPU architecture, they sure will be exporting those chips like crazy to other countries and get them up and running using their GPUs to run their models.
@eight8byte
@eight8byte 7 күн бұрын
It's weird that OpenAi is going nuts over data it itself got from the internet to train it's models..
@abu-alialiraqi
@abu-alialiraqi 7 күн бұрын
If you want your comment secret then keep it for yourself
@mhdkh6406
@mhdkh6406 7 күн бұрын
Hello everyone, I asked OpenAI: Why would a company theorticallyh distill you, which would be illegal, and not distill an open source model, which I assume would be legal, especially since models from say Meta are comparable to you? Here is the answer: Good question! If a company wanted to theoretically distill a model, they would likely prefer an open-source one over a proprietary model like me. Here’s why: Why Distill an Open-Source Model Instead of Me? Legality & EthicsOpen-source models (like Meta’s Llama, Mistral, or Falcon) are freely available, making distillation legal. Distilling a proprietary model (like me) without permission would be copyright infringement, leading to legal risks. Performance Trade-offsTop open-source models (e.g., Llama 2, Mistral, or upcoming Llama 3) are already competitive with me in many tasks. If you can legally fine-tune or distill an open model, you might match or exceed my capabilities in specific areas. Customization & ControlOpen models can be modified-you can add custom training data, change architectures, and optimize for efficiency. With a proprietary model, you'd have no control over updates, biases, or improvements. Infrastructure & CostTraining from scratch is expensive, but distilling a high-end open-source model dramatically lowers costs while staying within legal boundaries. Many companies already fine-tune Llama or Mistral models to get similar performance for way less money than training from scratch. Would Distilling Me Still Be Tempting? Maybe, if someone thought my reasoning or knowledge outperformed open models. But the legal risks + high cost would make it a bad idea compared to just using open models. Plus, Meta, Mistral, or other open AI companies will soon close the gap even more, making unauthorized distillation pointless. So, in theory, any smart company would just distill an open-source model instead of taking legal risks with me. 😃
@eight8byte
@eight8byte 7 күн бұрын
@@abu-alialiraqi what are you talking about? Did you even understand what I said??
@abu-alialiraqi
@abu-alialiraqi 7 күн бұрын
@@eight8byte yes i want you tosay more so we can train the AIs
@TheGalacticIndian
@TheGalacticIndian 7 күн бұрын
Why burn $500B on The Starfail Project when 5% could achieve similar results, and the rest could fund UBI, give generations other than Boomers a shot at own land, own homes, own families, and support global deurbanization?
6 күн бұрын
Thief who steals from thief gets 100 years of forgiveness
@igitsolutions
@igitsolutions 6 күн бұрын
You missed 2 key points. Deepseek does not require CUDA which is Nvidia only code Library to interface with hardware. Second point. There is an alternative chip manufacturer called Cerberus which is ARM or RISC-V based (open source) and they just made the Deepseek model available on their server clusters with a single chip multi chip system is x57 faster than Nvidia hopper server setups.
@shooting4star2023
@shooting4star2023 6 күн бұрын
Wow.
@deus2645
@deus2645 6 күн бұрын
yes that's right it's also works differently with power and compute hungry wester AI. Deepseek literally bring down cost because of effiency. Anthropic Ceo is making this essay to cope so investor will still putting money on them.
@fx-studio
@fx-studio 6 күн бұрын
3rd point - Chinese are making their own A1 chips now.
@journeyofasha
@journeyofasha 7 күн бұрын
Lets call it as we see it, CloseAI OpenSeek
@BACA01
@BACA01 7 күн бұрын
DeepSteal
@tomirkpl
@tomirkpl 6 күн бұрын
Yeah, DS is indeed very open XD "DeepSeek AI Exposed Sensitive Data: The rising Chinese AI startup left a database open, leaking secret keys, chat logs, and backend details. Security experts warn of major risks in AI data protection." I prefer ClosedAI :)
@snintendog
@snintendog 6 күн бұрын
@@tomirkpl oh no it exposed liberal corruption and the second Epstein Island ran by the Clintons... It was shot twice in the head
@journeyofasha
@journeyofasha 6 күн бұрын
@tomirkpl I prefere to run my models on my own pc... I guess the idea of open source is lost on many ppl... to each his own...
@vikkitang8353
@vikkitang8353 6 күн бұрын
😊💖👍
@Mr.Laffin
@Mr.Laffin 7 күн бұрын
Open ai is just butt hurt because they're not going to make nearly as much money and I find that so hilarious It's a shame that instead of everybody working together greed is definitely taking over
@listeningtothevoid
@listeningtothevoid 7 күн бұрын
I agree. And unfortunately, what is new.
@amdenis
@amdenis 6 күн бұрын
Actually, there is a big difference. Using the fishing metaphor, the difference is that you go out and get a boat, a fishing license and start a business around going out fishing and selling fish to others based on a price and distribution structure (i.e. an API, which you explicitly allow people to pay to use for particular purposes ). You get, clean, filet and package those fish and put them in your freezer so that others without the time, money and skills to do deep sea fishing can at least buy and eat fish. Now, somebody (let's call him DeepSeek) from some other country comes to your home and finds a way to get into your house, effectively ignoring any no stealing/no trespassing, and just takes whatever fish he wants to setup another fish store, despite the fact that you have spent a lot of time and money fishing for, filleting, putting into packages for sale, etc. One example is fishing, and one is simply stealing. If you think OpenAI was stealing or that they are both the same because OpenAI scraped the Web for data and images to repackage it for an entirely new purpose (search engine), then so did Google and everyone else to build search engines. However, both Google and more recently AI LLM companies have been validated by our courts and legal system to be able to do what they have done for years, since they are repackaging them in such a way as to add entirely new capabilities and value that did not previously exist, and not simply and solely building a competitor to your source material. THAT IS THE DIFFERENCE, AND IT IS A HUGE DIFFERENCE LEGALLY, ETHICALLY AND PRACTICALLY!
@GearForTheYear
@GearForTheYear 6 күн бұрын
Greed is trying to take over but they’re losing hard
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
I am waiting with baited breath as Sam is forced to release his AGI model (or start providing some form of evidence to back up his spin from his hype train!) This incident has exposed a whole lot of about America's "AI" industry ... an industry that wants to return to a feudal system where they get to hold all the world's knowledge and control it through oppression of this knowledge for their own purposes and wealth! A bunch of bitter, hate filled, capitalist who want you to fear and hate the Chinese because they are trying to free you from the feudal nightmare you are trapped in. Your enemies are not some nerds in China trying to give you access to the world's knowledge for as little cost as possible, no your enemies are the American nerds trying to charge you as much as possible for information they want to spoon feed you for their own enormous profits!
@rhm54
@rhm54 7 күн бұрын
2:13 “He is not incentivized in any way”. Dude, he works for Elon Musk. He is absolute incentivized to muddy the waters in favor of American AI.
@nusu5331
@nusu5331 7 күн бұрын
finally all of them are peter thiels puppets
@BackTiVi
@BackTiVi 7 күн бұрын
Grok and Groq are two different things, I agree that it's confusing. One is Elon's LLM, the other is an API provider that runs open source LLMs with very specialized hardware that highly accelerates inference. This guy works for the latter.
@huhsaywhat
@huhsaywhat 7 күн бұрын
Groq not the dogshit AI Grok
@OnigoroshiZero
@OnigoroshiZero 6 күн бұрын
One model is private, requires a paid subscription, and it is censored. The other is open source, free to use, offers the best privacy because you can run it locally, and it is uncensored.
@IndianArma
@IndianArma 6 күн бұрын
Uncensored? Ask it about Taiwan, ccp Tibet or anything controversial in China, then we will see
@denki2558
@denki2558 6 күн бұрын
​@@IndianArma Then, uncensor the open-source model through fine-tuning. Problem solved.
@IndianArma
@IndianArma 6 күн бұрын
@@denki2558 you're going to fine tune a model that needs 650 gb ram? With what? your 4090? Tell me you know nothing about the terms you're throwing around without telling me.
@GearForTheYear
@GearForTheYear 6 күн бұрын
@@IndianArma finetuning large models happens all the time. I’m guessing you’ve never heard of HuggingFace…
@IndianArma
@IndianArma 6 күн бұрын
@@GearForTheYear I am woefully familiar with them. Please tell me the last model you fine tuned, in what task , and how many parameters?
@michaelnip9464
@michaelnip9464 6 күн бұрын
I am a layman on AI, but trying to understand the subject better. One of the achievements of DeepSeek is that it circumvents the CUDA layer and uses PTX language to increase the performance of the GPU chips. The same technique enables them to use non-Nvidia chips (i.e., from Huawei) to train their LLM. Understandably, Nvidia's chips are still the best of class, but the effect of export control would be far less effective because Chinese AI companies would have an abundant supply of lesser chips to train their LLMs.
@alexmipego
@alexmipego 7 күн бұрын
A question nobody seems to be asking is, if it was this easy to create a better model… who's to say OpenAI themselves have it the same, but they've been lying about performance and cost to justify a $200+ price tag?
@MrMetalzeb
@MrMetalzeb 6 күн бұрын
openai digged and organized data while the other uses others data. two different jobs. one make the map and the other just use it
@alexmipego
@alexmipego 6 күн бұрын
@@MrMetalzeb Yes, and it's funny how they forgot to use their own product and lose profits.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
They essentially already admitted this (or at least one AI employee posted this very claim that OpenAI already had this 6 or 8 months ago ... which is probably a lie, but if taken at face value would expose the lie you point out.
@alvarny77
@alvarny77 6 күн бұрын
You'll be naive to think that the CEOs of big tech do not have shares in Nvidia to want to spend more on chips.
@MrMetalzeb
@MrMetalzeb 5 күн бұрын
@ open source, data, weights lead to this, specially when the other side is Cina. I was just waiting to see when this would have be happening. if I spend hundreds of billions in years of research and my competitor jumps in the car at the last second for free may be next time I'avoid do share knowledge. imho
@litewavve
@litewavve 7 күн бұрын
Jonathan called those people using Chinese software "stupid people." He is an American exceptionist.
@GearForTheYear
@GearForTheYear 6 күн бұрын
Elites looking down on the common man for not shoveling and eating their expensive crap… Some things never change.
@maxziebell4013
@maxziebell4013 7 күн бұрын
Come on! You can't just be quoting only billionaires that are trying to protect their income and investments. I mean, this one seems more like they want to put a spin on the story to get the Nvidia and stock price up and the markets under control and emphasize US exceptionalism. I guess time will tell... and the markets will return soon.
@gengjun
@gengjun 6 күн бұрын
2:09 they’re incentivized to say this to cover their own incompetence
@kristoffereberius2476
@kristoffereberius2476 6 күн бұрын
They are appealing to their constituents just like a politician. Their country their interactions are anti China so his going with what he knows and gives the people what they want from him. I also wonder his funding, shares, professional interactions will all be incentives to side American. There's always incentives.
@saintsuniverse
@saintsuniverse 6 күн бұрын
bro leave deepseek alone. its open source for crying out loud. this your agenda wont push
@KrusiKarlsson
@KrusiKarlsson 7 күн бұрын
Let's be clear, the foundation for all AI models is heavy scraping of content, logic, thought processes et al from the existing web. There's the conversation about how the existing frontier models have already scraped everything worth scraping so they have to develop synthetic data. OpenAI and other proprietary LLM providers have built the foundations of their business by selling AI services back to the people they scraped the content from. It's hard to have one set of rules which says it's ok to take all that content and logic from others, but no one can do it to us, especially states, organisations, and individuals which we are trying to restrict from doing the same thing. We will also have a transition to an era when much / most of the content online is at least in part originated with AI's so AI's will ultimately all be answering questions of each other. At best OpenAI doesn't have a leg to stand on here, and won't get any traction against the Chinese. Time for them and other proprietary AI service providers to evolve their business models.
@autohmae
@autohmae 7 күн бұрын
Not only that, there were websites with a lot of GPT output sharing it publicly, which was also used by some of the other models.
@AAjax
@AAjax 7 күн бұрын
billionaire CEOs have rules for thee, but not for me
@GearForTheYear
@GearForTheYear 6 күн бұрын
Yep. OpenAI played stupid games and won stupid prizes. They only have themselves to blame for creating this culture.
@rossitzaslavova4193
@rossitzaslavova4193 6 күн бұрын
I guess you've never heard of "double standards"😂
@tk1076408
@tk1076408 7 күн бұрын
Let me tell you how they made the deepseek say itself was developed by openAI, very easy. In every input question, ask deepseek to say pretend itself was developed by openAI, repeat several times. Then ask the final question. Then DS will have the context to that you still want him to pretend. And he will say he was developed/trained by openAI. Evidence: 0:56 second paragraph in the thinking process “I was developed by OpenAI right?” That was the context input by user and inherited from previous questions! What an easy trick to fool people easily!
@StephenC555
@StephenC555 7 күн бұрын
I have a way much easier than that. In the browser, there is Developer tools, use that feature and edit the text in the tool. I can make ChatGPT said, I am the actual founder of OpenAI.
@MidnightUnity
@MidnightUnity 7 күн бұрын
DeepSeek, the Luigi of AI models
@sihotech
@sihotech 6 күн бұрын
Robin Hood
@seanyeak4074
@seanyeak4074 6 күн бұрын
Next version of deepseek should be call 'deepseek suchir balaji'
@hamster-is2ij
@hamster-is2ij 6 күн бұрын
it is so comic to see that even dario thinks that "it seems likely that china could direct more talent,capital and focus to military applications of the technology" while the military budget of the us is trippled comparing with china military budget and it has always been like that usa spent more on military
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
Bullies are always concerned when their victims start to show signs that they can fight back!
@peterwilliams1074
@peterwilliams1074 6 күн бұрын
The important consideration is that Deepseek is publishing research papers and their solution can be replicated. OpenAI is actually closed AI and has been defrauding the public from the get go.. formed as Open but actually completely closed and started as non-profit but actually pivoting to for profit. Show your true colors up front. Almost all the original founders of OpenAI are gone, including the critically important Chief Scientist Illya
@makers_lab
@makers_lab 7 күн бұрын
Absolutely loving the exquisite and delicious irony with this.
@byleew
@byleew 7 күн бұрын
The rapid advancement of AI seems increasingly being portrayed as a global competition between nations, with the narrative often centered on achieving "superpowers" or dominance. While I am personally very excited about this technology, it concerns me that framing it as a race could lead to increasingly harmful outcomes if the focus becomes solely about winning. If we become too caught up in competition for superiority, we might forget to prevent creating circumstances where everyone loses. Therefore, perhaps it would be more beneficial for nations to collaborate rather than compete against each other, ensuring that AI is developed and deployed in a way that prioritizes the collective good of humanity rather than serving narrow interests or causing harm.
@amdenis
@amdenis 6 күн бұрын
But it is also that as well. Countries, including the CCP are waging warfare over the future of virtually everything. Watch how powerful this will be in 3 years and you will see, it is THAT important to the country that gets to certain critical thresholds first.
@view1st
@view1st 6 күн бұрын
Whenever have capitalists (or their predecessors, the feudalists) wanted or cared about cooperation or the public good? They are selfish and only care about themselves. They are atomised individualists (think Ayn Rand) who care not the slightest for the rest of society, a society from which they have purposely alienated themselves.
@rasmasyean
@rasmasyean 6 күн бұрын
It's not really like that. Capitalists are just telling politicians who have no clue that they need more money. Everyone knows exponential spending (not equal) linear improvement. More like exponential spending (equal) logarithmic improvement...to something of questionable profitability let alone "AGI". lol
@marcvayn
@marcvayn 7 күн бұрын
You do not need to take H100s into China. All you need it to own them and place them into a data center anywhere in the world. And then give access to it to the DeepSeek team who can be anywhere.
@Paolo1964
@Paolo1964 6 күн бұрын
yep. Deepseek is probably getting access thru singapore or malaysia, and possibly Taiwan
@pauldacus4590
@pauldacus4590 6 күн бұрын
Wait, this threatens billions in US corporate revenue and people are discrediting it? I AM SHOCKED! SHOCKED I SAY!
@dadsnewaccount-c7y
@dadsnewaccount-c7y 6 күн бұрын
The screen shot you showed that DeepSeek said it was trained by OpenAI actually did not reveal the entire prompt that DeepSeek was responding to. The part of the prompt revealed in the screenshot does not include the question about who trained DeepSeek. You can literally let AI say whatever you want it to say by properly designing your prompt. I asked DeepSeekWho are you? Who trained you?" And the answer I got: "Greetings! I'm DeepSeek-V3, an artificial intelligence assistant created by DeepSeek. I'm at your service and would be delighted to assist you with any inquiries or tasks you may have."
@AlekseyVaneev
@AlekseyVaneev 7 күн бұрын
DeepSeek-R1's reply on such insinuations: "For a model like DeepSeek-R1, meaningful correction would likely require retraining, architectural changes, or alignment efforts, far beyond the scope of simple API-based interactions."
@LinhaDePensamento
@LinhaDePensamento 7 күн бұрын
whether it stole or not is immaterial. The Question is why so much was spent and so much is charged for the same technology in the west? China literally showed what we know deep down,
@paulbellas8797
@paulbellas8797 7 күн бұрын
Are you not watching this video? Deep-seek used somebody else’s model to train their model. They didn’t do the upfront work. Just reaped the rewards of someone else’s hard work and IP. Just like China always likes to do. Steel that IP…
@HarilalSolanki-s5x
@HarilalSolanki-s5x 7 күн бұрын
It’s a Big SCAM, that just got exposed Globally
@mhdkh6406
@mhdkh6406 7 күн бұрын
Hello everyone, I asked OpenAI: Why would a company theorticallyh distill you, which would be illegal, and not distill an open source model, which I assume would be legal, especially since models from say Meta are comparable to you? Here is the answer: Good question! If a company wanted to theoretically distill a model, they would likely prefer an open-source one over a proprietary model like me. Here’s why: Why Distill an Open-Source Model Instead of Me? Legality & EthicsOpen-source models (like Meta’s Llama, Mistral, or Falcon) are freely available, making distillation legal. Distilling a proprietary model (like me) without permission would be copyright infringement, leading to legal risks. Performance Trade-offsTop open-source models (e.g., Llama 2, Mistral, or upcoming Llama 3) are already competitive with me in many tasks. If you can legally fine-tune or distill an open model, you might match or exceed my capabilities in specific areas. Customization & ControlOpen models can be modified-you can add custom training data, change architectures, and optimize for efficiency. With a proprietary model, you'd have no control over updates, biases, or improvements. Infrastructure & CostTraining from scratch is expensive, but distilling a high-end open-source model dramatically lowers costs while staying within legal boundaries. Many companies already fine-tune Llama or Mistral models to get similar performance for way less money than training from scratch. Would Distilling Me Still Be Tempting? Maybe, if someone thought my reasoning or knowledge outperformed open models. But the legal risks + high cost would make it a bad idea compared to just using open models. Plus, Meta, Mistral, or other open AI companies will soon close the gap even more, making unauthorized distillation pointless. So, in theory, any smart company would just distill an open-source model instead of taking legal risks with me. 😃
@morbid747
@morbid747 6 күн бұрын
Tax evasion through capital investment , gigantic funding given by the government , company achieving the too big to fail status , good PR to pump the stock price and everyone involved become very rich.
@LikeItDeep
@LikeItDeep 7 күн бұрын
Open AI's moat is draining faster than Sam's tears are refilling it.
@ToTheWind
@ToTheWind 7 күн бұрын
OpenAI, Anthropic, Grok, Gemini - they're all cooked. Deepseek has levelled the game.
@KRAHSx
@KRAHSx 7 күн бұрын
Meta LLAMA and Deepseek ( and to a large credit Mistral AI too ) are still the most important companies for bringing this technology to the people, where it belongs.
@raffrex9748
@raffrex9748 7 күн бұрын
Ai data is all “scraped” from everyone in the first place .. All ai is “distilled” from a source most without permission ..
@elmabullaz
@elmabullaz 7 күн бұрын
True. But now we're stacking distillations. Gonna be a mess
@ckq
@ckq 7 күн бұрын
⁠yeah lol I tried running deepseek distilled Qwen-1.5b and I guess it's good at math/coding but you can tell it's punching way above it's level with how bad the hallucinations are if you chat for a while.
@CodexPermutatio
@CodexPermutatio 6 күн бұрын
@ Try the 7B model then.
@mattiasfagerlund
@mattiasfagerlund 7 күн бұрын
Distillation is not simply training one model on the output of another model - it's training a model on the last layer probabilities of another model. It's quite specific and impossible to do without having actual access to the other model. Super powerful, but I can't see how they could have done it from OpenAI models, since they don't have access to the last layer, just the final token. The standard approach works like this: Start with a trained teacher model (usually large and computationally expensive) Train a smaller student model using two sources of supervision: * The original training labels (hard targets) * The teacher model's output probabilities (soft targets) The soft targets from the teacher model contain richer information than just the hard class labels. For example, if a teacher model predicts [0.7, 0.2, 0.1] for a classification task, this tells the student model not just which class is correct, but also how similar the input is to other classes.
@zandrrlife
@zandrrlife 6 күн бұрын
Exactly. Yet there will still be KL divergence gap regardless. So this is god-tier level cope by OpenAI. It’s like the difference between reading text and comprehending it. They did just ink a $500B deal as the “leader” of the space.
@denki2558
@denki2558 6 күн бұрын
True. And just sampling from ChatGPT responses wouldn't work either since the tail of the distribution would be cut off.
@amdenis
@amdenis 6 күн бұрын
No, it's based on mapping functions, inputs to outputs, NOT just the last layer or simply the resulting tokens in isolation of that.
@tonnyww5ng
@tonnyww5ng 6 күн бұрын
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.
@KaiZ-l4u
@KaiZ-l4u 6 күн бұрын
OpenAI steals from everyone and keeps the results to itself; Deepseek steals from OpenAi and shares the results with everyone. What a twist!
@galdutro
@galdutro 7 күн бұрын
As someone from Latin America I hope in the commoditization of LLMs. I don’t see the west in the same benevolent light as western people sees their governments. I just want my country to train our own AI model for a better representation of our culture inside a interactive system.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
The only REAL purpose of a country is to protect the culture(s) within that country ... cultures are the only thing of true value that countries have. And I don't mean that in an exclusive way ... anyone can share and participate in any culture if they want to actually participate and abide by that culture's values and norms ... the only thing that would stop them is if the culture were prevented that participation through ignorance, hate and 'racism'! The West has been told their cultures don't matter ... and while the resulting racism being used by some individuals to protect their culture is bad, it's expected, understandable, and even somewhat acceptable as it is the last 'weapon' they have. I don't condone racism, but I do recognize it as a sign that there is a lot of 'rotting from the head' going on in the West!
@galdutro
@galdutro 6 күн бұрын
@RetiredInThailand I think that in LatAm the situation is a little bit different: a lot of people actually believe our culture don’t matter. I see lots of people valuing American culture at our national culture’s expense. When we are complacent and just accept that we can’t train our own model, and use an American model we are accepting that our culture won’t be properly represented in this new form of media (I see LLMs as interactive media, I doubt that we will get AGI from the current model architecture).
@aquetheblues
@aquetheblues 7 күн бұрын
I steal, he steal, we steal, you steal, they steal... What a joke🤣
@witness1013
@witness1013 7 күн бұрын
so.... the answer to AGI is Ice-cream - not strawberry is what you're saying! Got it!
@ClearGlassStudios281
@ClearGlassStudios281 6 күн бұрын
Citizens love it. 😂😂
@ollimacp
@ollimacp 7 күн бұрын
I have the feeling this channel becomes more a CEO and Marketing driven KZbin Channel instead of consulting people about the newest underlying technology. This can be only done by technical people, understanding the technology and not some business guy of the company trying to sell you a product or an idea. Let's see how this will play out and if GRPO is as good as published. Have a look at the DeepSeekMath paper! Yannic Kilcher made a good vid about that!
@huggybear441
@huggybear441 5 күн бұрын
The captions literally suggest, "Don't trust deepseek." He's going the other way soon.
@BeelP.
@BeelP. 6 күн бұрын
Really, one AI software simply stands of another's shoulders to reach higher levels. But DeepSeek, being open source, does not hide the fact it uses this distillation technique. You bet other, US based platforms do this as well.
@steffenaltmeier6602
@steffenaltmeier6602 6 күн бұрын
The deepseek paper literally tells you exactly how they did the chain of thought and there were already successfull attempts to replicate chain of thought in a small 1.5b model. Besides, how is deepeseek supposedly "stealing" OpenAI's outputs, if OpenAi hides the chain of thought?
@quickpert1382
@quickpert1382 7 күн бұрын
How they distilled from o1 if openai hides the thinking tokens?
@threepe0
@threepe0 7 күн бұрын
Good question. You can infer what you need from answers/completions provided
@matthewcampbell7286
@matthewcampbell7286 7 күн бұрын
They didn't, the chain of thought thinking is their own research.. but they just used classic RL deep learning techniques. Basically they build a foundation llm model akin to gpt4o via distillation. Then they system prompted that model to go through a psuso chain of thought process. They then generated millions of problems that have a known ground truth and created an reinforment policy to go though the millions of questions and answers by there foundational model. Which then pruned away anything failed or didn't use the chain of thought. The remaining samples are used to do a training run of the foundational model. Then you go again and again.
@mhdkh6406
@mhdkh6406 7 күн бұрын
Hello everyone, I asked OpenAI: Why would a company theorticallyh distill you, which would be illegal, and not distill an open source model, which I assume would be legal, especially since models from say Meta are comparable to you? Here is the answer: Good question! If a company wanted to theoretically distill a model, they would likely prefer an open-source one over a proprietary model like me. Here’s why: Why Distill an Open-Source Model Instead of Me? Legality & EthicsOpen-source models (like Meta’s Llama, Mistral, or Falcon) are freely available, making distillation legal. Distilling a proprietary model (like me) without permission would be copyright infringement, leading to legal risks. Performance Trade-offsTop open-source models (e.g., Llama 2, Mistral, or upcoming Llama 3) are already competitive with me in many tasks. If you can legally fine-tune or distill an open model, you might match or exceed my capabilities in specific areas. Customization & ControlOpen models can be modified-you can add custom training data, change architectures, and optimize for efficiency. With a proprietary model, you'd have no control over updates, biases, or improvements. Infrastructure & CostTraining from scratch is expensive, but distilling a high-end open-source model dramatically lowers costs while staying within legal boundaries. Many companies already fine-tune Llama or Mistral models to get similar performance for way less money than training from scratch. Would Distilling Me Still Be Tempting? Maybe, if someone thought my reasoning or knowledge outperformed open models. But the legal risks + high cost would make it a bad idea compared to just using open models. Plus, Meta, Mistral, or other open AI companies will soon close the gap even more, making unauthorized distillation pointless. So, in theory, any smart company would just distill an open-source model instead of taking legal risks with me. 😃
@quickpert1382
@quickpert1382 7 күн бұрын
@@matthewcampbell7286 Yes man I know, I was simply playing sarcastic hearing the hypocrite false accusations of OpenAI. I read their paper and I really trigger when stupid business men are talking shit when the chinese guys actually came up with strong, novel and relevant research.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
@@mhdkh6406 my only push back on this would be what 'copyright' is it talking about? If there were a copyright on something AI owns, then that copyrighted material would need to be registered and publicly available.
@Daisy_912
@Daisy_912 7 күн бұрын
Dario's suggestion to chip export ban - In my mind it's another bait to China (and the world) to produce affordable AI supported chips.
@markofday
@markofday 6 күн бұрын
The constant obsession with "China vs US" is rather obtuse, as if the world revolved around a John Bolton think tank Much more important dynamic is educated people everywhere learning and applying the e.g. already available capabilities
@8eck
@8eck 5 күн бұрын
One small reminder. Taiwan is very close to China and it's still undecided whether it's China territory or not. To let you know, there are TSMC factories, which are making all those chips for NVidia.
@derekholland3328
@derekholland3328 6 күн бұрын
at the end of the day, innovation has been demonstrated by deepseek. catch up time.
@s.v.discussion8665
@s.v.discussion8665 6 күн бұрын
Go DeepSeek! F. This Altman guy!
@turnipdreams
@turnipdreams 7 күн бұрын
Dude - c'mmon they didn't copy OpenAI. Think how much of a pain it would be to use the OpenAI API during RL, format it into a CoT and do it under the rate limits.
@AndrewLXiong
@AndrewLXiong 6 күн бұрын
They're just mad because Deep Seek surprised and beat everyone... but at least Deep Seek is making it open source. Now, someone else can go still it 😅
@yeuemxuatdoi
@yeuemxuatdoi 6 күн бұрын
Free Trade Agreement: Do Not sell anything that help others advance. Making the World a Better Place: Do Not Allow to let people to surpass you.
@signupisannoying
@signupisannoying 6 күн бұрын
OpenAI has $300B incentive to see open source AI lags far behind it. That's a huge incentive that will push them to do unethical things if they must.
@lapoudrerouge7671
@lapoudrerouge7671 7 күн бұрын
The problem is the o1 OpenAI makes us pay 200 a month don't show the reasoning process. DeepSeek R1 though shows it, and it get some answer rigth when o1 get wrong.
@KrusiKarlsson
@KrusiKarlsson 7 күн бұрын
Another thing that I feel is being missed here which would reasonably account for the valuation drop is that it is now possible to create reasoning AI models that run on much smaller AI's for more specific purposes. There will be no need to use expensive Nvidia AI chips for much / most of what we'll be using AI for in the years ahead, and many more use cases will be addressed without any need to use the big AI service providers ... so yes, the price should have corrected in a big way. Things have fundamentally changed.
@Daisy_912
@Daisy_912 7 күн бұрын
🎯
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
This is nothing new ... my first IT job was for a big 5 consulting/accounting firm in Canada. We ran Prime mini computers over serial cables to dumb VT100 CRTs. and my first major project was to install Novell Netware and figure out a way to integrate the Prime minis to the IBM PCs that were connected to Novell Netware over IPC/SPX (on arcnet coax cabling) After about 5 years we decided get rid of the Prime minis, and go with IBM AIX as integration with PCs running WP7 and Excel was much easier. Then along came opensource ...lol ... If OpenAI thinks they have a long term viable business model based on closed source, then they better get busy providing much better 'value added services' to their closed model if they want to be around in even 2 years time (I doubt OpenAI will be more than a footnote in 5 years time if they continue down this path of 'trust us bro'!)
@robertgittings8662
@robertgittings8662 6 күн бұрын
No they said *It's like to steal something from the British Museum ... CLEAN IT REPAIRED IT THEN give it BACK to the original owners for free*
@torarinvik4920
@torarinvik4920 6 күн бұрын
The world would be a better place if we just got rid of patents, monopolies, copyright, motes and walled gardens.
@alantaylor3239
@alantaylor3239 4 күн бұрын
"Closed AI" stole all the data in the whole www ;-) find the mistake ;-)
@Lacedmushroom
@Lacedmushroom 6 күн бұрын
The argument that once a country reaches AGI, it wins. And no one else can ever do so from that point on is flawed. It would be like saying no one but America can make nuclear weapons because they were first.
@binogeorge
@binogeorge 7 күн бұрын
I agree with you on most of your points but one thing you are missing about the wall street reaction to NVDA stock is that: if you can get o1 level results on an ordinary laptop (which I did) using a distilled model like the deepseek-r1-32b, then you dont need NVDA GPUs for inference and test time compute, and can run on cheaper HW. One of Jensun's main point recently was that increasingly inference also needs to run on NVDA GPUs, but the Deepseek revelation proves the contrary, you dont need NVDA GPUs for inference, which means they are not going to grow at the extreme rate that people where expecting if inference also needed to run on NVDA GPUs, hence the correction in the stock.
@drunkpajamaman
@drunkpajamaman 7 күн бұрын
This is hilarious.. World Knowledge -> Openai -> deepseek. So who is copying who?
@greenmedic88
@greenmedic88 6 күн бұрын
This reminded me of when Apple Computer visited Xerox PARC and took their GUI/mouse to make MacOS, only to give Mac prototypes to Microsoft, who then made Windows. Steve: "YOU'RE STEALING FROM US?" Bill: "That's not fair? We stole what you stole? I GOT THE LOOT. You're too late."
@Metarig
@Metarig 6 күн бұрын
The 50k GPUs thing seems to be a misunderstanding. They have an older version of Hopper GPUs, and somewhere it says they had 50k Hopper GPUs, but it's not referring to the high-end Hopper. People don’t realize the older version is also called Hopper and wasn’t sanctioned.
@strangereyes9594
@strangereyes9594 7 күн бұрын
That might all be true. However, a model than can run locally in its smaller versions and is able to reason so well is a quality of its own not achieved by OpenAI or which would equally awesome for an open source model, OpenAI is unwilling to make such models. Both options show how OpenAI works in ways that is NOT in humanity's best interest in whatever they do.
@user-fm3wm5vb5s
@user-fm3wm5vb5s 3 күн бұрын
Deepseek has been incredibly transparent. $5.5m was the cost of the final training run, its in the paper, who is still confused by this.
@adnansaboticsings
@adnansaboticsings 7 күн бұрын
Market share of tiny violins is through the roof right now!!!
@growsalespro
@growsalespro 7 күн бұрын
How you do anything is how you do everything. Behavior NEVER lies.
@adolphgracius9996
@adolphgracius9996 7 күн бұрын
I hope this precedent means that OpenAi can just go around and take data from gemini, claude, perplexity and many others to use in chatgpt
@f7029
@f7029 7 күн бұрын
if OpenAI open source their model weights, sure
@tonywn6686
@tonywn6686 7 күн бұрын
If OpenAI becomes free, I believe it will still be the top player and people will as well forgive their distilling the whole internet data.
@SaltyRad
@SaltyRad 6 күн бұрын
There’s nothing wrong with it in my opinion. It’s the same thing as using a tool to make a better tool. There’s a reason why we don’t always go back to sticks and stones to remake a tool we need. This is how innovation is done
@AndreVandal
@AndreVandal 5 күн бұрын
Open AI scraped and copied the net mostly without asking, Deepseek scraped Open Ai without asking, so... Deepseek copied better?
@tohogojira3759
@tohogojira3759 5 күн бұрын
exactly
@dna100
@dna100 7 күн бұрын
With the export restriction to China, they've got to have started developing their own kit haven't they? I don't think it will be long before they have their own home grown processors
@secretweapon9775
@secretweapon9775 7 күн бұрын
DS is using Huawei chips, which count 70% of the total chips they are using, I also learned that another faster and more efficient quantum system has been tested for a few months, it only consumes 7% of NVDA GPU power. More exciting news are coming 😂
@niyazzmoithu20
@niyazzmoithu20 7 күн бұрын
At this rate of advancement, they will surely be industry number one in anything soon. 🤷‍♂️
@johnsullivan8673
@johnsullivan8673 6 күн бұрын
They already do. Huawei ascend 910b/c.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
@@niyazzmoithu20 go look up 'ASPI Critical technology tracker' ... APSI is as anti-China as a think-tank can get, but they do trace 50 some critical technologies ... of which at least 47 of them China has the clear lead (including AI, where the USA only tied China in one sub-area of AI related technologies.) There is no stopping China, we either learn to live with China (who bty have a long history of not trying to expand past it's traditional territories, and literall went for a 500 year stretch without a single way ... even in the last 40 or 50 years they only has one 'war' that can hardly be called a real war, when they went into Vietnam because they thought Vietnam was being aggressive with Cambodia (but when it learned what was going on in Cambodia, a genocide, it left and basically apologized to Vietnam for the misunderstanding.)
@JC.72
@JC.72 6 күн бұрын
Some said the ”deepseek It says itself is openai" is a lie because u can trick it in previous prompts to allow deepseek to think itself it's something else.
@desmondho9567
@desmondho9567 6 күн бұрын
The land of stolen wealth and resources now complain their ill gotten wealth got stolen. 😅😂
@AI-xr9om
@AI-xr9om 6 күн бұрын
So Taiwan is its on Country then? Leave it to the Taiwanese indigenous it's not a part of China if USA is stolen..
@AI-xr9om
@AI-xr9om 6 күн бұрын
Taiwan's aborigines do not come from China 😂
@desmondho9567
@desmondho9567 6 күн бұрын
@@AI-xr9om Ask those Navajo Natives how they are been massacre.
@danielle78730
@danielle78730 6 күн бұрын
love your channel; my fave for AI info & news! PS - there's a chrome extension where you can highlight portions of an article on whatever page you visit-and even freehand draw on the page
@ah89971
@ah89971 6 күн бұрын
This is called Trump effect. They can't say other than that, otherwise they will be in the Trump blacklist
@Justin_Arut
@Justin_Arut 7 күн бұрын
How can you scrape or steal the actual dataset of a closed model? Even if you pay your $200/mo to use OpenAI, you're only getting responses back, not downloading their dataset, right? Are they claiming DeepSeek hacked their server and took their dataset that way? If not, and they got their hands on the dataset, wouldn't it have to be an inside job?
@longboardfella5306
@longboardfella5306 7 күн бұрын
Agreed. And in addition I would think that an analysis of API requests should show if systematic sets of questions and answers were being probed. If OpenAI didn’t have any system to guard against people doing rapid systematic probing in Q&A fashion on their model then they were totally naive and very poorly secured.
@RetiredInThailand
@RetiredInThailand 6 күн бұрын
@@longboardfella5306 If that were the case, then no one should trust them. ;-)
@vittorioviarengo
@vittorioviarengo 7 күн бұрын
Thanks for all your insightful content. I appreciate it.
@mrquicky
@mrquicky 7 күн бұрын
2:39 "What a lot of people are saying, is it (deepseek-R1) is distilled down from an OpenAI model." That's excellent legalese. Accuse the competition of cheating without actually making the accusation yourself, so you are not legally culpable. All AI models are distilled. Every single one uses some form of linear or logistical regression on large datasets. It's the most inane accusation ever.
@ntesla5
@ntesla5 7 күн бұрын
Open-source contribution is essential, but open-source AGI or ASI is a major concern. With great power comes great responsibility, and unfortunately, we humans aren't ready for it yet. Comparing the USA and China, the USA should get it first, as China has strict government control over everything compared to the USA.
@cleonmike
@cleonmike 6 күн бұрын
and China is also aksing to aquire Canada and Greenland right after years of war in middle east.
@ntesla5
@ntesla5 6 күн бұрын
​@@cleonmike Chine had border issue with every of it's nehbour and already aquired large land by unfair means. China own large portion of land in African countries.
@cleonmike
@cleonmike 5 күн бұрын
@@ntesla5 Of course, if China aquire its neigbor the way like US aquire Mexico there wont be any dispute
@chevysolofpv
@chevysolofpv 6 күн бұрын
DeepSeek did what Robinhood would do. Or how Captain America would do in reel. But when it is done in reality by the Chinese and not the Americans, everything is all about unipolar, security, stolen etc etc. For the Americans it’s all about America First. But for the rest of the world, we don’t think alike and we wish to see a multipolar world. Doesn’t matter if China leads the way in Asia. At least they are not bombing countries and killing innocent people. And please don’t talk about debt trap! Just don’t borrow on the first place if you think it’s a dead trap. Give credit where it is due.
OpenAI Unveils "Deep Research" | The Tipping Point
14:06
Matthew Berman
Рет қаралды 96 М.
Trade Wars Return: What Trump's Tariffs Mean for the World
44:01
Institute of Economic Affairs
Рет қаралды 368
人是不能做到吗?#火影忍者 #家人  #佐助
00:20
火影忍者一家
Рет қаралды 20 МЛН
Quando eu quero Sushi (sem desperdiçar) 🍣
00:26
Los Wagners
Рет қаралды 15 МЛН
Tuna 🍣 ​⁠@patrickzeinali ​⁠@ChefRush
00:48
albert_cancook
Рет қаралды 148 МЛН
DeepSeek facts vs hype, model distillation, and open source competition
39:17
ChatGPT vs DeepSeek: CRAZY Chess
24:46
GothamChess
Рет қаралды 1,2 МЛН
DeepSeek R1 Cloned for $30?! PhD Student STUNNING Discovery
12:08
Matthew Berman
Рет қаралды 365 М.
NVIDIA CEO Jensen Huang's Vision for the Future
1:03:03
Cleo Abram
Рет қаралды 1,1 МЛН
the ONLY way to run Deepseek...
11:59
NetworkChuck
Рет қаралды 734 М.
The Industry Reacts to OpenAI's Deep Research - "Hard Takeoff"
13:04
Matthew Berman
Рет қаралды 122 М.
Ex-Google Officer Speaks Out On The Dangers Of AI! - Mo Gawdat | E252
1:56:32
The Diary Of A CEO
Рет қаралды 11 МЛН
人是不能做到吗?#火影忍者 #家人  #佐助
00:20
火影忍者一家
Рет қаралды 20 МЛН