New FREE & Open Reasoning LLM Matches Open AI o1! + RTX 5090 Unboxing! AI News

[@MattVidPro]

HUGE Thanks to Yandex for partnering with us on today’s video. Check out YaFSDP here: github.com/yandex/YaFSDP Read the Medium Article to learn all about how YaFSDP can cut down your training time by 25% and save up to 26% in GPU resources while training transformer architecture: medium.com/yandex/yafsdp-a-tool-for-faster-llm-training-and-optimized-gpu-utilization-is-no-632b7539f5b3

[@brexitgreens]

Yandex! The only challenger to Google Search. The best (and uncensored) Reverse Image Search engine in the world. ❤

[@brexitgreens]

_A reply was here 🚩. Deleted by KZbin._ 🤐

[@brexitgreens]

All I'm trying is to say how Yandex is my favourite. But I'm not allowed. 😑

[@LouisGedo]

👋

[@FusionDeveloper]

I'll take the 4080. I have an old 1080 Ti.

[@FRareDom]

OPEN SOURCE AI IS BACK!!! 🔥🔥

[@MattVidPro]

Yessirrrr

[@Pawnsappsee]

They're hitting right in the bull's eye😂

[@Pawnsappsee]

I tested this one, pretty good, deepseek is awesome.

[@MattVidPro]

I agree! DeepSeek is a game changer.

[@jeffwads]

And that 32b already available in LM Studio.

[@MattVidPro]

That is super cool! I'm stoked to start working with it!

[@konstantinlozev2272]

What are the settings in LM Studio for the r1?

[@Happ1ness]

Finally, the real Open AI is back 👀

[@rwgamer]

I don't see how Open AI will survive the onslaught of open source implementations. There is no moat there.

[@RealmsOfThePossible]

Seems they are fabricating results to try and keep hold by their fingernails.

[@alan83251]

That's why they're begging the government for a regulatory moat.

[@MuhammadRaiyan135]

Project stargate

[@thebicycleman8062]

Openai owns like 90% of consumers of a.i around the world. And they have enterprise b2b solutions that r reliable. They will always be around coz their name is basically what ppl think a.i is

[@DisturbedNeo]

While I don't put much stock into benchmarks, what's interesting is the idea that 32B can achieve ~90% of the performance at ~5% of the size. Makes me think that scaling really isn't all that valuable of a strategy when it comes to capability for LLMs.

[@JM168M]

Happy Time ! Thank you Matt 🔥🔥🔥🔥

[@TheItalianPeter]

Stoked to try those open source goods, also really looking forward to see you testing local llm's when the embargo for that ends!

[@galrozental3332]

As someone who works in A.I, 15 hours of training is honestly nothing in terms of time. I have trained models far simpler for far longer, this is amazing! Edit: literally the next thing you cover is a lora that takes 15 minutes to train, and only using 6-8 images?!?! I am speechless.

[@evidenceX]

The 32B is distilled so it's not bad when it loses some performance at some tasks

[@mihirvd01]

ALL HAIL THE A.I. OVERLORDS !!! THIS ISN'T ARTIFICIAL BUT EMERGING HYPER INTELLIGENCE !

[@konstantinlozev2272]

It's far from superintelligence. Even the o1. But it's still quite impressive

[@RealmsOfThePossible]

Would you allow an AI to structure your life with your best interests at heart and would you adhere to it's commands even though it may not make sense?

[@aliettienne2907]

21:24 Yeah, that lighting technique look absolutely great 👍🏾

[@FusionDeveloper]

Awesome! I'm soo glad you get a 5090!

[@apatsa_basiteni]

Seems this is going to be another fantastic year

[@austintse5719]

Congrats on getting a 5090 to play with. I hope to get one myself. What's size model would you expect to fit in 32GB.?

[@DWSP101]

Yeah, I just checked it out the deep seek it definitely does seem like it’s almost on par with GPT oh one by GPT o1 is about to be replaced here in about a month from what scheduled for the new GPT five

[@aliettienne2907]

13:37 The ability to merge multiple content into one video content is shockingly superb. One thing that confuses me is how these video generation models will work superbly in certain areas, and are weak in other areas of content creation. I can't wait when the Ai video generation models will shine or excel in all areas of video generation. But I must say that I'm very impressed with this ability to merge multiple content into one. 15:20 I believe this video generation feature will be desired on the Samsung and Apple Ai intelligence platforms. That ingenious technique is attractive and can also be useful for certain applications. 18:55 This John Wick demo is compellingly impressive, and it marks a significant leap in Ai video generation history.

[@pierre-samuelgreau-hamard6379]

I'd really like to see a video on 3D video generation, as I think it is the future to really allow in-depth composition.

[@simongentry]

Mat - I’m looking to drop an agent in a sandbox for training… can you suggest? have a favorite? ty

[@cbnewham_ai]

It's not just Twitter claiming AGI from OpenAI - several prominent KZbin channels (which will remain nameless) have been hawking this rubbish around. I'm glad to see you didn't follow in their footsteps.

[@brexitgreens]

Why believe Altman? Telling us the truth is not his job. OpenAI is not allowed to accomplish AGI too soon due to their terms of contract with Microsoft.

[@cbnewham_ai]

@brexitgreens oh, I don't believe Altman - where did I say that? - and I don't believe all the rubbish that is claimed about AI either. AGI is several years off, at best. And LLMs may not even be the path to it. Not until they've sorted out its ability to do maths, at the very least.

[@nathanbanks2354]

I just ran R1 -70b on a pair of 3090's using ollama. I wonder if it could use your old video card & new one together with 32GB + 16GB? of RAM to run the 4-bit quantized version of R1-70b on your computer....

[@weevil601]

Ollama already has the new Deepseek-R1 models in its library, so I downloaded the 14b model and gave it the question I ask all models: >>> A farmer is trying to take a fox, a hen, and some grain to market. He comes to a river. There is a boat on the shore, but it ... is only big enough for the farmer to take one of his items across at a time. He can't leave the fox alone with the hen, and ... he can't leave the hen alone with the grain. How can the farmer take the hen across the river? This is a classic logic puzzle that most people (and LLMs) have seen before. But notice the question I asked. I didn't ask how to transport all three items. I only asked how to transport the hen. Every LLM I've tested so far has missed that nuance and proceeded to try to give me the classic answer. Sadly, Deepseek was no different. It went into the usual "thinking" mode that reasoning models do and started printing out its "thoughts". It had a lot of thoughts. All in all, it printed 2588 lines of thoughts by the time it gave me its answer. The funny thing was that by the time it finally got to its final thoughts, it had forgotten the scenario I had given it and had reverted to its training data puzzle, which apparently involved a wolf, a goat, and some cabbage rather than a fox, a hen, and some grain. The answer it finally gave was incorrect as well. Oh well. No local AI has passed this one-question test so far. I'll play with Deepseek-R1 a lot more, and I do have a feeling it will be the best one yet for my system, but I'm still waiting for that genius AI that notices the actual question I'm asking rather than what it finds in its training data.

[@couchtaming23]

Chinese AI research papers already surpassed those from the U.S., I think about 4 years ago. With the backing of the East Asia cultural sphere and a population of 1.4 billion, that’s honestly pretty intimidating!

[@kuromiLayfe]

Would say NeuroSama is probably the smallest version of AGI level of AI… like she can hold a conversation, understands the environment she is placed in, knows when something is a joke or not and learns from her own mistakes in almost real time, all she needs is these newer smaller open source models that can process millions of tokens instead of the few thousands.

[@1Know1tHurts]

Matt, please test Flux image speed generation on your 5090. Congrats on getting this monster :)

[@pdjay8912]

지금 써 봤는데 DeepSeek 진짜 좋네요. 이미지 업로드 등은 안되지만, 오픈 소스인데 이 정도면 너무 훌륭한 것 아닙니까! ㅎㅎ 감사합니다. 좋은 정보:)

[@aresaurelian]

Locally run Open artificial intelligence systems for robotics is going to be wild. Especially among young engineer kids.

[@DiceDecides]

For me AGI needs to be embodied and able to do what any human would be able to learn, so basically not having the artificiality as a limit.

[@Brenden-H]

AGI is when it can do whatever a human can do. Including AI research and improving the AGI model to have more capabilities.

[@HolidayAtHome]

hah, cool you got an RTX5090 =D I hope I will get mine on the 30th before all the scalpers buy them away >_

[@eyevenear]

Soon some genius will release an open source video generator with the “ingredients” feature

[@4thpdespanolo]

Great all I need now is a few H100s

[@DerApfelfan-g5l]

Where is Anthropic and Google? They should release something to compete.

[@procrastinatingrn3936]

Google has a reasoning model, what are you talking about?

[@DerApfelfan-g5l]

But it is not in its final release, and there is no Gemini 2.0 Pro or an even more powerful Gemini 2.0 Pro Thinking Model.

[@dweb]

Does the model offer a built-in kill switch under absolute human control?

[@emport2359]

No R1 testing?

[@MattVidPro]

Not in this video! Expect testing in future videos!

[@DonaldKronos]

I've been trying to tell you for some time now, but you never seem to get the message. Artificial general intelligence is simply general intelligence that happens to be artificial. That's it. It is a very solid definition, and people moving the goalposts doesn't change that. That is to say, artificial general intelligence is literally adaptable broadscope intelligence that is an a product of Ingenuity or cleverness. That's a full general definition of artificial general intelligence, and it if it's every reasonable usage of the term. Oh, and by the way, yes I definitely could use a gpu, as I waited over 50 years for such technology specifically to use on artificial intelligence, and I can't afford one. Not that I expect there's any chance I would even know if you got this message. LOL

[@cbnewham_ai]

I wish Kling would just fix the upload video for lip-sync - which STILL does not work. FFS. Why do these companies bring out new features when they don't fix the current features that are broken.

[@hekasoram]

are you supposed to return the 5090 to nvidia at some point?

[@phen-themoogle7651]

Your AGI question: To me, anything i can ask to make me money and it knows how to make money and transfer it to my PayPal account without me needing to do anything. It needs to be very agentic and capable of all the necessary skills in-between to do jobs that humans do online, and should work constantly on them and improve making me more and more money each time. It could potentially do several jobs at the same time(depending on how much they limit a single user), I expect us average joes could make millions of dollars by just prompting AGI to make us money. Otherwise, it wouldn’t be AGI. If it’s truly as good as a human , but also has the benefits of being an Ai( never gets tired nor needs rest/process information faster) then it can work 10x or more harder and longer in a single day. It would also build algorithms to improve itself and try to beat its previous performance scores.

[@bengsynthmusic]

Ask it "I accidentally washed my car in my washing machine. What should I do?" Apparently it gave a pretty dumb answer to some.

[@ronilevarez901]

Whatever answer it gave, maybe that's what they'd do in China! Different cultures and all, you know? XD

[@sadshed4585]

how much vram to run fp4 of 32b model is needed

[@sirhammon]

AGI in my books, is an AI that can do anything a Human can do online in at least one job title (google job requirements of online jobs). The same AI should be able to specialise in any field of expertise. It doesn't have to be one AGI specialising in every expertise at the same time, but the model can Learn to specialise in any field of expertise. The AGI needs to be able to learn new things and improve over time by itself. If you stick it into a robot it may take 6 years or so to learn how to operate the robot effectively just like a human would.

[@SPEEDBALL436]

Bro in the thumbnail: 😀👍

[@mabena-f2o]

When will deepseek take images...I got math module's 😅

[@jokmong2360]

open ai have to release the O.3 asap

[@haroldpierre1726]

But is it affordable.

[@Happ1ness]

You made a typo. *ClosedAI

[@jokmong2360]

@@Happ1ness 😅

[@cbnewham_ai]

DeepSeek is what OpenAI should have been. OpenAI seems to be becoming less relevant by the month.

[@ronilevarez901]

Yeah, that's why they are closer to AGI than anyone and why they're backend up by USA government, right? 🙄

[@michaelwoodby5261]

Honestly I thought it did a great job with the hopping monster at 16:00, the original critter could just be scrunched down. Also I wonder why everyone's go-to test for video generation is "attractive young woman in very specific clothing"?

[@NoahtheGameplayer]

Couple of people said Nvidia did not being honest when it comes to benchmarking so it's a hard stretch

[@cariyaputta]

5090 should be sweet for 32b at q8 and 70b at iq4nl.

[@Corteum]

The X hype is no for AGI. It's for Open(Closed)AI 😛

[@Epic-Generations]

I need that 4080!

[@pdjay8912]

헛! Hunyuan+Lora 미쳤는데요. 😮

[@couchtaming23]

East Asia's cultural sphere, including Japan, Korea, Taiwan, China, and Singapore, has the most students excelling in various fields. China alone has four times the talent pool of the U.S. If the U.S. doesn’t switch to ASI researchers soon, it risks falling far behind China-and by a huge margin!

[@matthallett4126]

Does any know a Coding LLM that can perform as well as o1 with natural language requests and run on a 4090?

[@makers_lab]

I've run the 14B Q6 K variant of R1 on my 4090, and even that does pretty well across a wide range of tasks, including coding. Fitted easily and super fast. You might try that or a higher parameter variant with a different Q factor to offset the higher parameter count. It's really insane that it can answer the extent that it does with a model of around 12 GB.

[@stephaneduhamel7706]

Of course Deepseek V3 will not score as well as the others at reasoning benchmarks, it's just a language model, unlike the others which are reasoning models.

[@cashdoo]

i thought i tried this model out about a month ago... are you just really late or was i using a preview model before?

[@MattVidPro]

I believe it was a preview model, or deepseek v3. This model has been shared publicly (weights and all) for the first time today

[@Alice_Fumo]

You were probably using DeepSeek-R1-Lite-Preview (yes, actual name)

[@WMR1776]

Please download Suno and make music too!

[@desmond-hawkins]

Congrats on getting a 5090! I can't wait for January 30th, this will make a huge difference to anyone running models on a single card.

[@ReneOrense]

🔥🔥🔥‼️

[@benjames813]

5090 huh? must be nice

[@TheGalacticIndian]

4080 GIVEAWAY PLEASE!🤗🤗

[@yashaouchan]

How do you not have Lemongrass somewhere in the background? Epic fail. Unacceptable!

[@SjarMenace]

Do you sleep enough?

[@Dina_tankar_mina_ord]

The Turing test was all the rage with ai progress. But nobody even knows when it got passed because that is how some millstones gets ignored. The definitions are just to fuzzy. Open AI has agi. The just dont want to admit it for financial reasons. The need to milk the progress to keep acquiring funds. If you get the cure gor cancer there is no need for compnies to fund more research.

[@DFortuna]

the gaslighting by sam is crazy lol

[@Atheist-Libertarian]

Restrictions on Chip export to China must be lifted. If they get access of more and better GPUs, they will create far better models. And they will open source it. Which is very good for whole world 🌎 and Open Source community. Thank you China and DeepSeek ❤

[@ronilevarez901]

But if the restrictions continue, they'll develop their own GPU technology, which is also good.

[@masakazuishiguro8525]

This is the most thinly veiled propaganda I’ve ever seen and I’m actually a Chinese person.

[@masakazuishiguro8525]

@@ronilevarez901Don’t know where you got your propaganda but China’s economy is collapsing, and we already lost the chip war. India might have more chance doing that in the long run.

[@ronilevarez901]

@@masakazuishiguro8525 most people outside Europe only know what dear USA tells us (China is a terrible and powerful economical and military threat and everything we do is to save the world from them and from our others enemies). So thanks for the (unfortunate) insights. But I still see some hope for diversity in tech players, including China, at least if cheaper versions of models keep being released.

[@Atheist-Libertarian]

@@masakazuishiguro8525 Ohh my god. I am not even Chinese, I am Indian. How is this a propaganda ? These are my thoughts (CCP is not giving me money). I am an Libertarian, so I believe in Free Trade Free flow of Capital, Free flow of technology. Because it's a win-win for everybody. Yes I don't like CCP, I wish China becomes an electrol Democracy. But that's a seperate issue.

[@javiergimenezmoya86]

I love open source but O1 is better. If you use both you will know that.

[@mjprelic]

Jarvis - Ultron - or Vision. All agi lol

[@CorbiasLess]

Yol 🎉

[@Pawnsappsee]

🎉🎉🎉🎉

[@adude3625]

just noticed your chimeco

[@JoeBrigAI]

These open source models perform well in benchmarks but rarely in the real world.

[@WesTheWizard]

Oh, tell me sir, how did you try to use one and how did it fail you?

[@JoeBrigAI]

@ Ollama. Overly chatty, adds hex codes to programs for no reason. Awful so far.

[@WesTheWizard]

@@JoeBrigAI but Ollama isn't a model. It's a program that runs models. Maybe you chose a bad model?

[@Codemanlex]

@@WesTheWizardHe’s talking rubbish

[@brexitgreens]

My short encounters with open-source LLMs were, ehm… short, for a reason. However I've just tried DeepSeek R1… and, man, it is powerful! 🤩 Beats Claude 3.5 Sonnet for me - let alone o1 Preview and o1 Mini (which I found inferior to Claude).

[@ezzye]

Your sponsor needs a marketer😮

[@Merializer]

Tried it, still pretty dumb. Expected more.

[@lonedruid1]

🎉🎉🎉

[@Pawnsappsee]

🎉🎉🎉🎉🎉

[@strategictechnologist]

Chinese model?

[@zikwin]

yeah we want give away rtx5090

[@BigNekox]

Bruv deep seek crashed the market single handedly

[@tradehut2782]

Who has 700GB of space in their computer? 😮

[@pfos]

🤖AGI = she can hands me the drink 🍹;-p

[@coalepps8936]

I wholeheartedly welcome this Chinese competition that is shaking up America. It is beneficial for the citizens of the world. Never bet against Chinese corporations. Chinese company, a champion of open source. That’sbeautiful to see.

[@aruak321]

Hopefully the open sourced version isn't censored due to the CCP like the version on their website though

[@antonystringfellow5152]

No LLM created to date has the IQ of the average 8-year-old, never mind 130. If one scores that in a test, there's a big problem with the test. Intelligence requires understanding and we have yet to produce an AI that has a measurable level of understanding.

[@RobertHouse101]

HI. Nice to see you again. Before and after seeing Musk do a Heil Hitler salute, I don't believe a G damn thing from that platform.

[@TheGeneticHouse]

WOW!! IS ALL I HAVE TO SAY WHEN YOU WERE GOING THROUGH THE BENCHMARKS... IT'S CRAZY HOW FAR APART OPEN SOURCE AND CLOSE ARE I MEAN IT MIGHT BE ANOTHER 10 YEARS BEFORE R1 GETS CLOSE TO 01 LMFAO THEY'RE HERE WE'RE PRETTY MUCH EVEN WOW

[@guerillachan20]

Ban incoming.

[@brucermcarthur]

open source !!!!!!!

[@not_riley]

r1’s kind of garbage. Refuses to swear.

[@tomirkpl]

I like Your videos, Matt, but I don't care about Chinese OS models because I don't trust them and they are too expensive (so big LLM's need expensive hardware or You send Your data to chinese operator). Besides, OpenAI will be offering o3 any day now, which beats DS in spades (as I would say :)).

[@Pawnsappsee]

You're selling your data anyhow😂

[@tomirkpl]

@@Pawnsappsee But not to China ;)

[@Raulikien]

You're still selling it to companies who are partnered with military contractors and will use their tools for war one day. Open source is the way.

[@brexitgreens]

*1.* Brainwashed by the Western #MilitaryIndustrialComplex. *2.* #FalseDichotomy. There's a third way: you can deploy an open-source Chinese model on a GPU rented in the cloud controlled by a Western tyranny of your choice.

[@brexitgreens]

KZbin censors my response, so I'll split it in two. Maybe it'll help.

[@vazus171]

Do you know that Yandex is a Russian company? Is it safe for you to take their money for the Ad?

[@ddiva1973]

Yandex is a rusian comapany.

[@K.F-R]

Yandex?? Really?? Damn. Bye.

[@bWWd0]

Dood dont talk how you only do this channel cause of sponsors , taht is pretty shit, so without payday you wouldnt do it, ok so dont force yourself .sure its nice to make money but dont say shit like this openly ok.Also serious AI channel would get A100

[@jackmax2150]

eventhough deekseek is 2nd rank llm, not many developer and user is using it. r1 is superb, but not well know and no body really care. rich people are using openai only