Which GPUs are best for running AI models | Lex Fridman Podcast
HTML-код
- Опубликовано: 10 фев 2025
- Lex Fridman Podcast full episode: • DeepSeek, China, OpenA...
Thank you for listening ❤ Check out our sponsors: lexfridman.com...
See below for guest bio, links, and to give feedback, submit questions, contact Lex, etc.
GUEST BIO:
Dylan Patel is the founder of SemiAnalysis, a research & analysis company specializing in semiconductors, GPUs, CPUs, and AI hardware. Nathan Lambert is a research scientist at the Allen Institute for AI (Ai2) and the author of a blog on AI called Interconnects.
CONTACT LEX:
Feedback - give feedback to Lex: lexfridman.com...
AMA - submit questions, videos or call-in: lexfridman.com...
Hiring - join our team: lexfridman.com...
Other - other ways to get in touch: lexfridman.com...
EPISODE LINKS:
Dylan's X: x.com/dylan522p
SemiAnalysis: semianalysis.com/
Nathan's X: x.com/natolambert
Nathan's Blog: www.interconne...
Nathan's Podcast: www.interconne...
Nathan's Website: www.natolamber...
Nathan's RUclips: / @natolambert
Nathan's Book: rlhfbook.com/
SPONSORS:
To support this podcast, check out our sponsors & get discounts:
Invideo AI: AI video generator.
Go to lexfridman.com...
GitHub: Developer platform and AI code editor.
Go to lexfridman.com...
Shopify: Sell stuff online.
Go to lexfridman.com...
NetSuite: Business management software.
Go to lexfridman.com...
AG1: All-in-one daily nutrition drinks.
Go to lexfridman.com...
PODCAST LINKS:
Podcast Website: lexfridman.com...
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com...
Podcast Playlist: • Lex Fridman Podcast
Clips Channel: / lexclips
SOCIAL LINKS:
X: x.com/lexfridman
Instagram: / lexfridman
TikTok: / lexfridman
LinkedIn: / lexfridman
Facebook: / lexfridman
Patreon: / lexfridman
Telegram: t.me/lexfridman
Reddit: / lexfridman
Lex Fridman Podcast full episode: ruclips.net/video/_1f-o0nqpEI/видео.html
Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/cv8484-sa
See below for guest bio, links, and to give feedback, submit questions, contact Lex, etc.
*GUEST BIO:*
Dylan Patel is the founder of SemiAnalysis, a research & analysis company specializing in semiconductors, GPUs, CPUs, and AI hardware. Nathan Lambert is a research scientist at the Allen Institute for AI (Ai2) and the author of a blog on AI called Interconnects.
*CONTACT LEX:*
*Feedback* - give feedback to Lex: lexfridman.com/survey
*AMA* - submit questions, videos or call-in: lexfridman.com/ama
*Hiring* - join our team: lexfridman.com/hiring
*Other* - other ways to get in touch: lexfridman.com/contact
*EPISODE LINKS:*
Dylan's X: x.com/dylan522p
SemiAnalysis: semianalysis.com/
Nathan's X: x.com/natolambert
Nathan's Blog: www.interconnects.ai/
Nathan's Podcast: www.interconnects.ai/podcast
Nathan's Website: www.natolambert.com/
Nathan's RUclips: youtube.com/@natolambert
Nathan's Book: rlhfbook.com/
*SPONSORS:*
To support this podcast, check out our sponsors & get discounts:
*Invideo AI:* AI video generator.
Go to lexfridman.com/s/invideoai-cv8484-sa
*GitHub:* Developer platform and AI code editor.
Go to lexfridman.com/s/github-cv8484-sa
*Shopify:* Sell stuff online.
Go to lexfridman.com/s/shopify-cv8484-sa
*NetSuite:* Business management software.
Go to lexfridman.com/s/netsuite-cv8484-sa
*AG1:* All-in-one daily nutrition drinks.
Go to lexfridman.com/s/ag1-cv8484-sa
*PODCAST LINKS:*
- Podcast Website: lexfridman.com/podcast
- Apple Podcasts: apple.co/2lwqZIr
- Spotify: spoti.fi/2nEwCF8
- RSS: lexfridman.com/feed/podcast/
- Podcast Playlist: ruclips.net/p/PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
- Clips Channel: ruclips.net/user/lexclips
*SOCIAL LINKS:*
- X: x.com/lexfridman
- Instagram: instagram.com/lexfridman
- TikTok: tiktok.com/@lexfridman
- LinkedIn: linkedin.com/in/lexfridman
- Facebook: facebook.com/lexfridman
- Patreon: patreon.com/lexfridman
- Telegram: t.me/lexfridman
- Reddit: reddit.com/r/lexfridman
“We should continue to catch ‘tigers’ as well as ‘flies’ when dealing with cases of leading officials in violation of Party discipline and state laws as well as misconduct and corruption problems that directly affect the people’s interests. All are equal before the law and Party discipline; whoever is involved in a corruption case must be thoroughly and impartially investigated.”
The results, after 12 years of Xi’s anti-corruption campaign?
“Corruption is RAMPANT in China!”
-Victor Gao, Al Jazeera, August 2024 in front of a live international audience.
The land of arbitrary law enforcement™️.
DeepSham™️
I think even Lex was freaking lost lol these guys are next level and the piggyback off each other flawlessly
too niche, too hyperspecialized... not always a good thing
I swear, nifalixo Money's Untold Mysteries is one of the best books I’ve read. It’s life-changing.
Necessity is the mother of invention the fact that they were keeping the best tech from the Chinese made them have too find work around and come up with better software.
What deepseek is showing is how us foreign policies have gotten lazy
they used Ai to build AI. Nothing is "better". Now they need more hardware or they get flooded by 10 k users.
From what I've read from Altman and Dario. They don't think code efficiency to skirt around sanctions gets you to AGI/ASI. They think HORSEPOWER(scaling) gets to the finish line.
So that's why the US policy is there. Horsepower limiter.
Which is also eye opening due to the fact everyone goes off the notion ‘ China makes everything’. They worked with what they had
@@lotyogipityu7992 Yeah, I wonder if they have smart people like you in their company!!!
I understand 1% of what they are saying but still love 100% of the conversation 😂👍🏽
This guy intelligence is on another level.
Beautifully said 😂 truly next level
Right? Right.
Left
Yeah definitely a reason I’m in the Navy.
There's a small mistake in the explanation on KV-caching. The size of the KV-cache is linear with respect to sequence length, not quadratic. The total attention computation is indeed quadratic wrt seq_len, but the KV-cache itself is just a cache of the all Key and Value projections which there is one projection per layer per head per token.
These two are one of few that actually understands ai and models work.. I would say about 10% ai experts actually understand how it all works..
You did not mention the exchange rates and electricity cost differential.
Right?
The sad part is all of the law makers have no idea what they are talking about.
Don't worry musk is there
They do. They do what guys who buy a lot of GPUs tell them.
One three just overflowed the conversation.. 😅
Super interesting
What episode of Silicon Valley is this??
Deepseek can be trained on Huawei GPU too
Huawei can only make 7nm. While Nvidia 3nm
@@carkawalakhatulistiwa The performance gain from 7nm to 3nm is considered "small" because as transistors shrink to extremely small sizes, the potential for significant performance improvements diminishes due to physical limitations, increased heat generation, and the complexity of manufacturing such tiny components, leading to a phenomenon called "diminishing returns" in miniaturization; meaning each further reduction in size provides proportionally less performance benefit.
@@carkawalakhatulistiwa Training AI models doesn't solely depend upon sizes of chips. Depends upon FLOPS, CUDA memory, cores, and bandwidth. What chinese devs did is wrote highly efficient low level code to tackle shortage of Nvidia GPUs. Hence, cost and memory efficient.
@@carkawalakhatulistiwaTSMC are the foundry, not Nvidia.
Whats up Lex..thanks for feeding my mind!
CEREBRAS AI Chip is best 1600 tokens /Sec
The question that everyone wants to know. Should i buy nvdia 😅
Where is the electricity for these Chinese AI data centers coming from?
power plants
electricity in china is much cheaper than in the usa.
Coal
Coal, solar, hydro, wind, nuke, gas...
They do it all
humans running on treadmills? 🤷🏻♂️
Unfortunately Lex is kind of slurring words, guests speak much clearer
GPU RTX 3090 4090 5090 architecture ruclips.net/video/U4TcPPf5vaE/видео.htmlsi=jelKVOd4h1HoJAjD
"Right? Right? Right?"
Mofo if you're so unsure about what you're saying maybe don't say it.
I think I should sell my deepseek account. I have two accounts
Nvidia
I think this guy is Italian.
Bro needs to stop saying "right" like 5 times for every sentence. Yeesh.
Right.
ruclips.net/video/_Alq_xh7rAk/видео.html
Safety ? AI doesn't generate sticks and stones and doesn't do migrant farm work so IDGAF about safety. Safety from some words it generates that I'm probably not gonna read anyway is kinda funny. 🙄🤪😜
ban nvidia.
banger
total chaos. Can you have someone that knows how to explain. This is total ego trip or they don't know what they talking about.
Seems more likely YOU don’t know what they’re talking about
@@Vaquero_69 exactly and I'm working in this field. That's a problem
@@mikechannel5026 That's a big problem for you. Why the negative comment about them? Focus on you lol.
@ more then 50% of the talk is world salat 5 h could be shortened to 2 very easy. But enjoy focusing on yourself.😂
@@mikechannel5026👍
The prices of r1 api changed drastically today, the information in this podcast is already misinformation, @lexfridman, maybe add an overlay with a warning ?