Hi, HPC guy for a big US semiconductor firm here. There's a point at which John and Dylan talk about how the tech stack for tools in the industry is old, and people are terrified of touching it. I just wanted to come here to say this is absolutely true. We have entire teams that use things as old as RHEL 5 because they're scared if they update, their ancient versions of Cadence and Synopsys software won't work. There's a little bit of truth to this, but the two also speak about how the entire industry essentially runs like a series of apprenticeships, and I'd like to explain what that looks like in practice. The process usually goes something like this: 1. Someone comes in has to come up with some process for doing something, like setting up a toolchain in a specific way 2. That person teaches the people under them to do things their specific way 3. Said person leaves the company after 20 years 4. The toolchain LITERALLY never updates because the people who work at the company now don't actually know how it works so it's essentially magic Other industries have more public documentation on how their processes work (hell, Big Tech often publishes their software as open-source), so people can come in who understand the processes. This is not the case for the semiconductor industry, so things stay black magic. I can't push people off of RHEL 5 because their black magic probably only works on RHEL 5, so we keep RHEL 5 servers around. It's not that we don't update, we do have RHEL 8, but there are all these small pockets of ancient technology that everything else relies on. The Linux greybeards like me can keep ancient servers running, but I don't understand ANYTHING about modern EDA software, as I'm not an EE guy, and the EE people don't know anything about Linux because they're not software people, so we're a bunch of wizards who don't know how each other's magic works but somehow the magic has to work anyway. Guys, this industry is insane
Sounds like in part Polanyi's paradox, someone has years of experience tweaking a toolchain, gets an intuition for it, but can't transfer that intuition.
It's wild that we reached 40k mechanicum culture in the very highest echelons of our computing industries. That doesn't bode well for smooth continued advancement and basically means any disaster capable of destroying infrastructure or god for bid a large death of the people in the industry due to then being i or effected by war/bio wep breakout(again), or glowie false flag shinanigans could basically send us back to the end of the industrial revolution, utterly without information aged tech
@@musicdev sounds like the Priests in the Foundation Series. How does this reactor work? We say the prayers to it and it just works. So did the ancients build it so that it just never needs maintenance? No, they built it so that it only needs prayers to run.
Please do an episode with Asianometry alone. Jon barely got to speak! I enjoyed the firehose of knowledge from Dylan "delusionally ripping it" Patel but would really appreciate hearing more of Jon's perspective. clarity edit: I'm not calling Dylan delusional; but as he says, "life is just more fun when you're delusionally ripping the bong and feeling the AGI" 😂
@@pajeetsingh Of course, they're both very informative, but with different areas of expertise. We learn more when we hear a balanced amount from both guests. I'm not too bothered by it, the episode was still good; I just hope Jon gets another chance to bring his unique economics & tech history knowledge to the table.
I rarely comment on youtube videos these days, but this video was EPIC. Please dwarkesh! Have these guys on regularly for major industry updates, especially as things get crazier in these upcoming AI cycles.
Finally, a podcast dense with information and energetic enough that I can watch fully immersed and don't even try to do anything in parallel out of boredom
Damn, Dylan’s energy and breadth of knowledge is astounding, he is a machine. Amazing to hear someone so knowledgeable about this topic just blaaarghh it out all without a pause to even breathe. I’d like to hear more of Asionometry perspective as well, maybe invite him for another podcast sometime soon. Incredible podcast Dwarkesh, another nail hit square in the head. Thank you!
The whole discussion around # of GPUs and cluster size is naive. Pre-training and parameter size are already showing diminishing returns, and what you see with the o1/o3 style models is more compute spent during inference, not training. Actually, China centralizing all of their compute resources into one giant cluster would be a mistake. It's much more likely that having a diversity of smaller models will work better during inference time, where you want to explore many different kinds of reasoning and combine them with tree-search or blending in a 'society of mind' like approach. Musk has also fallen into this trap, buying into the fantasy that he who builds the biggest model wins.
@@TheHilariousGoldenChariotI think we hear less from Dylin because Asianometry always talks in his channel so i m fine with that but John is also not an industry guy he is more of an expert of things that have happened like a history guy where as Dylin has some chicky info about what the industry is cooking behind the scenes
@@okman9684 Jon is definitely an ‘industry guy’ if you’ve seen his videos he could practically write the history book. ‘He talks in his videos’ that reply doesn’t make sense, this is a podcast. D.Patel has multiple interviews where he gets to talk. I’m sorry but Jon is a much better speaker and has a much broader range of knowledge.
Semi Analysis & Stratechery is like getting an MBA for $30bucks a month. You hear enough of them consistently, you will begin to pick up things in the tech domain. Then you can research topics individually on Google to piece things together. Or you can just buy QQQ when it dips < 10% 😂
I'm 30 minutes into the video but have 2 hours already spent watching. I've had to google almost every sentence and learn the backstory as they tell it. What a great conversation
The guy is wrong about a lot of things. Like you can't just produce more power whenever you want to from a hydroelectric damn. The water and the turbine has to be there. If you drain the reservoir all the other turbines lose capacity. The world would notice if 5gigawatts of aluminum production was shutdown or shifted. It would affect the price world wide within a week, if leaked within minutes. The heat signature of the plant would change the second something changed. Way to easily notice if someone was looking. Google was building 200megawatt buildings nearly 20years ago with plans for multiple buildings per site. Twerking the building design to 250megawatts is easy. And many sites had plans for four or more buildings. Facebook was doing very similar scaled stuff more than 10 years ago. Plant scherer and quite a few other coal fired plants on the eastern sea board are running under capacity or have shutdown turbines. Plant vogtle is coming online that's 4gigawatts. It's estimated that plant scherer is running nearly a gigawatt under capacity. That's 5 gigawatts that could be turned up pretty quickly. Also google and Facebook can turn around hardware very quickly. Shoving video cards into servers is too easy. I'm talking about designing new motherboards and swapping them in. If they were very motivated it could happen within 6months. That's from spending approval to 1st 10k units on site powered up. Both of those companies are no stranger to marking things happen.
Also he mentioned early huawei network equipment having Cisco code on them. Huawei was the contract manufacturer for Cisco's switch products at the time. They had/ran Cisco code because they were the same switch hardware in a box painted a different color. It's not espionage when they give you the design and software then pay you to manufacture it.
@@dimedriver Cisco disagrees. Google this - To that end, the following are verbatim excerpts from the Neutral Expert’s Final Source Code Report, dated June 15, 2004: From a section entitled Comparison of Cisco STRCMP and Huawei’s [CODE NAME REDACTED]: “It must be concluded that Huawei misappropriated this code.
Amazing episode. It was jammed with so much information. I needed to take a breath outside to recoup after this epsiode and move some pieces in my brain.
First time viewer, watched every minute at 2x obviously. Holly shit. my people. I could watch another 5 hours of you guys melting my brain. Thnx for a great vid
Dylan said so many incorrect things in this video... It might be good to add some corrections. I'll just point out a couple that were pretty bad: at 47:30, he said the Magnificent 7 make up 60-70% of the SP500. No it's not even close to 60%, they make up about ~31% of it by capitalization as of now in 2024. Earlier on, he said China adds as much electric power per year as half of Europe. Not correct, they add about 20-25% (still a lot but it's grossly inaccurate). I might add more later
To be fair to Dylan, he said many many things. Not sure but probably over 1000 claims so it's not surprising there would be a some errors. I did learn from the podcast, so thank you for sharing it. Just be sure to do your own verification, especially on claims that sound surprising
We all knew a "Dylan" in college, and usually called Dylan, too. He's a hype bro, he's just not that knowledgeable, but he talks loudly and quickly enough that others assume he knows what he's talking about.
With 2 people it probably not easy to make sure they both talk same amount of time. I liked watching this alot. I'd really like to see both of those guys back
Was there any point where what Dylan said was useless or boring - that's when I feel like you can say Dylan took too long and overextended. Else it's equality for equality sake. Let's not take away agency from Asianometry - he can jump in whenever he wants. When the host was switching subjects and Dylan had more to add, he said hold up and went to complete his thought. Asianometry has a UA-cam channel - he speaks through that as well.
This is some sage career advice! "it's really about: what are you good at, where can you vibe, enjoy your work,, and where can you make an impact in society...what engages you? Because if you're interested in it, you"ll work harder" Dylan
Would have WAAAAAYYYY preferred this if dylan shut up a bit more and Asionometry got to speak more than 4 syllables at a time. Would have made this a lot more listenable. Feel like Asionometry’s analysis would have been a lot more grounded too rather than this super loose stuff coming out of Dylan’s mouth. He talks about OpenAI raising 50-100B when they only raised 6.5B. Interviewer should have moderated the convo better imo, this had so much potential and felt like it only came half way to reaching it. 😊
I knew that Dylan and Jon are amazing but Dwarkesh keeping up (or sometimes slowing them down a bit ;) and asking useful questions is great! Anyone who likes this episode should check out Dylan, Jon and Doug's podcast "Transistor Radio" (unfortunately it's a bit more rambly at times, because no Dwarkesh, but you also get random deep dives on stuff).
Something everyone should know about the semiconductor industry is that terms like 2nm, 5nm, etc mean nothing and are nothing more than marketing terms. The smallest features on the 2nm chips have an actual feature dimension of 20nm.
Bro, I didn't want to like your podcast because my immediate impression was "hustle and grind" culture, but you keep having such phenomenal guests that I can't stop listening.
Big cheers to dwarkesh, throughout the interview with these two giants he´s clearly done its research, yet he´s constantly pushed to the edge of his knowledge, while still keeping them to getting into semiconductor cloud nine.
this should probably have been 2 podcasts tbh. They are both extremely knowledgeable but Dylan is a force of nature when he's on roll so Jon was silent a bit too much. Would love to hear more from him!
There should be a separate podcast episode with just Jon. Maybe he can talk about his process about how he makes such in-depth detailed videos on a topic.
I only watched that video because of Asianometry. I never seen other two guys and may be would not watch them again. But he worried when that "unknown guy" referred him as someone who made video about zips. It was most emotional moment in a video.
I feel like this is the most objective conversation I've heard about Chinese manufacturing, it's culture, and the challenges they face in my life. Thank you very much.
Nah they have very biased view. China did well because of espionage ? What about the world’s largest fleet of stem talents and largest public funding for semiconductor industry ? $$$+ Chinese talents = success. There is 0 chance for failure.
Great talk, plenty of grounded perspective on Geopolitics and the companies involved in this right now. Refreshing to hear as someone in the Electronics contracting space who sees folks simply not adapting fast enough.
One of the unique things that came out for me in this excellent conversation was how your two guests described, and in their own ways, represented, two different kinds of fanaticism as motivation: with Dylan, a fanaticism of opportunity, and with Jon (though he was obviously having fun too) a fanaticism of fear.
This is one inconvenient truth that I think DC policy makers are missing. I studied at a Taiwanese University for an intensive study abroad, one of the top schools in Taiwan, and the school is a pipeline feeding China, not Taiwans, tech sector. Those hired by Taiwanese companies plan to work there, and then cash in on a job offer in China. For all the Pro Taiwan nationalism, very few Taiwanese won’t accept a higher paying job in Shanghai or Shenzhen. Once they get rich in China they bring their wealth back to Taiwan. This is a huge economic Achilles heal for Taiwan, a significant chunk of their economy depends on Taiwanese people getting massive salaries in Mainland China and coming home when they retire. Taiwans economic dependence on China is massive and the country would be on life support without trading with the greatest adversary. This also means the best and brightest semi conductor researchers have either been poached by China or will be. It’s only a matter of time before SMIC matches TSCM at this point.
They all need to limit their use of “like” I think they could have shaved 40-50 minutes off the conversation. Plus, it’s slightly, not overwhelming, annoying.
This was a fantastic episode. None of the AI stuff matters if you can't build the physical infrastructure to develop it. I'm left confused about the timelines though. GPT 5 is being trained right now, so doesn't that mean it won't be trained on Blackwell? Is Blackwell for GPT 5 turbo/mini/o/5.1? Or is it enough for the next order of magnitude model (GPT 6)? And where do the $100 billion data centers come into the picture? GPT 6? GPT 7? And will they be stocked with GPUs more advanced than Blackwell? I feel like this is relevant because we can only be reasonably certain that scaling holds up to GPT 5. In order for these mega investments to be worthwhile, it seems like scaling might need to hold up for several orders of magnitude more than GPT-5.
Hearing Dylan talking reminds me of the crypto bros, the fast talking excitement, techno optimism, what-if scenarios. I don't mean that in a bad way, just saying the way of speaking is the same. If we did have to make the parallels, a hype definitely exists and you have to believe to invest, as these large investors are doing at the moment and question is if they will continue to do so, it has to show (enough) usefulness aka lead to that ROI. Definitely want to see a podcast with just Jon to let him speak more.
Question: Currently, Chinese scholars appear to be publishing openly. What happens if they stop doing so in a way similar to how those in the USSR did during the Cold War period from 1945-1991? I wish I had seen this interview sooner. it is unique and golden. Thank you for setting up it, Dwarkesh.
01:09:30 reminds me of how Jian Yang from Silicon Valley figured out an architecture on how to make the PiedPiper product without using its IP and Gavin Belson was just blown away! 😂
I kinda feel lucky to have the background that I have. Power Systems/Electronics, Digital IC design, Analog circuits, Control/Signals Systems, and lastly AI design. This episode is such an overload that it is giving me anxiety.
The main money source of big US tech companies is advertising. If they have extremely good generative AI they could use it for proactive advertising, where their secret AI project user actions and serve ads _before_ people click on anything related. that's enormous money opportunity and completely separated from their public AI systems.
For those people who thinks Huawei is stealing tech/manufacturing secrets from somwhere, they don't know anything about how hard they work, how many super competent technical talents Chinese university supplies every year, how good the supply chain of everything in China is. Here, in the united state, it takes 5 months to redesign a pilot scale sanitation pipe, the China speed is two weeks and way cheaper.
@@directxxxx71 no. Companies I know of for advanced packaging are: TSMC, Intel, Samsung, ASE, Amkor, and SPIL which are located in Taiwan, USA/Israel, South Korea, Taiwan, USA, and Taiwan respectively. Depending on how advanced you mean by "advanced packaging" you could argue that only TSMC and Intel (maybe Samsung also) have truly advanced packaging. Amkor claimed to be doing volume CoWoS on an earnings call, but that claim was embedding in insulting ASE, so I'm not sure how seriously to take that claim; and if they're literally calling it CoWoS then it's probably licensed from TSMC. ASE's newest offerings are also opaque to me. SPIL is small so I haven't paid much attention to them, maybe that was a mistake. Samsung 🤷 should be doing better than they are. I don't know anything about STATS ChipPAC Ltd
Feels like you need a security clearance for this episode
It's all public knowledge
Not much here. These guys are zoomers of inside information; at least of what they are willing to share.
@@pajeetsingh Asianometry is highly knowledgeable, I don't know much about the other guys but they could keep up pretty well in the convo.
@@amazin7006 He is indeed, but both of them are avoiding some sensitive topics related to espionage.
These speakers have no knowledge of Financials in China. The economy is collapsing and yuan printing has been extreme and has made it worthless.
Hi, HPC guy for a big US semiconductor firm here. There's a point at which John and Dylan talk about how the tech stack for tools in the industry is old, and people are terrified of touching it. I just wanted to come here to say this is absolutely true. We have entire teams that use things as old as RHEL 5 because they're scared if they update, their ancient versions of Cadence and Synopsys software won't work. There's a little bit of truth to this, but the two also speak about how the entire industry essentially runs like a series of apprenticeships, and I'd like to explain what that looks like in practice.
The process usually goes something like this:
1. Someone comes in has to come up with some process for doing something, like setting up a toolchain in a specific way
2. That person teaches the people under them to do things their specific way
3. Said person leaves the company after 20 years
4. The toolchain LITERALLY never updates because the people who work at the company now don't actually know how it works so it's essentially magic
Other industries have more public documentation on how their processes work (hell, Big Tech often publishes their software as open-source), so people can come in who understand the processes. This is not the case for the semiconductor industry, so things stay black magic. I can't push people off of RHEL 5 because their black magic probably only works on RHEL 5, so we keep RHEL 5 servers around. It's not that we don't update, we do have RHEL 8, but there are all these small pockets of ancient technology that everything else relies on.
The Linux greybeards like me can keep ancient servers running, but I don't understand ANYTHING about modern EDA software, as I'm not an EE guy, and the EE people don't know anything about Linux because they're not software people, so we're a bunch of wizards who don't know how each other's magic works but somehow the magic has to work anyway.
Guys, this industry is insane
Sounds like in part Polanyi's paradox, someone has years of experience tweaking a toolchain, gets an intuition for it, but can't transfer that intuition.
It's wild that we reached 40k mechanicum culture in the very highest echelons of our computing industries. That doesn't bode well for smooth continued advancement and basically means any disaster capable of destroying infrastructure or god for bid a large death of the people in the industry due to then being i or effected by war/bio wep breakout(again), or glowie false flag shinanigans could basically send us back to the end of the industrial revolution, utterly without information aged tech
@@musicdev sounds like the Priests in the Foundation Series.
How does this reactor work?
We say the prayers to it and it just works.
So did the ancients build it so that it just never needs maintenance?
No, they built it so that it only needs prayers to run.
Why don't you guys just call IBM's guys and let help you with RHEL upgrade?
Ya, that sounds like every single industry.
You know a podcast audience is legit when the sponsor is freaking Jane Street looking for employees
Bro I got an ad for legos wtf.
@@poetac15 what is your watch history
@@Brodragon2225 it was a joke chill :)
Even I was thinking that- Jane Street is a big myth in India.
Same Jane street SBF came from?
Dwarkesh achieved rare lore for getting these 2 guys together. +10k aura.
Dylan was one of the speakers at the Asianometry meetup last year, so they have some history ))
@@5anjuro transistor radio is a regular podcast they are on with Doug from fabricated knowledge. Thats the dream team right there
@@gireboyOh good to know, thanks
I didn’t expect Jon to look like a Taiwanese uncle who is a high roller at several major casinos
Hahaha
I love Asianometry
you geh?
It's a great channel. Quality work.
Sure
@@maols43 still i feel this guy is puppet of usa gov just like twitter youtube fox news etc
Now we know why his audio is so bad, freaking face mask.
Asianometry's dressed like Incognito mode
I look great, don't I?
@@AsianometryOoooh shit dawg it's the G himself!
@@AsianometryLove your work! Keep it up!
Dunno why I found it so funny. But respect for him wanting to maintain his privacy. There's probably logistic reasons too as to his jurisdiction.
@@lephtovermeet Jon is actually hiding from the California IRS...
So odd to see Asianometry speaking so casually. Great interview!
yes
I mean, he generally has a pretty casual cadence on his videos.
In his channel he looks an asian speaking his second language. He sounds very american here.
@@westrimbut he talks in a specific tone, that has not changed that much other the years.
@@AM-fs7yhhe is Taiwanese-American if I am not mistaken.
Please do an episode with Asianometry alone. Jon barely got to speak! I enjoyed the firehose of knowledge from Dylan "delusionally ripping it" Patel but would really appreciate hearing more of Jon's perspective.
clarity edit: I'm not calling Dylan delusional; but as he says, "life is just more fun when you're delusionally ripping the bong and feeling the AGI" 😂
Right? Really felt like I was the only one
for sure
He is there to assess.
The other guy's takes are also very informative. We are here to learn and know ; why does it matter who speaks?
@@pajeetsingh Of course, they're both very informative, but with different areas of expertise. We learn more when we hear a balanced amount from both guests. I'm not too bothered by it, the episode was still good; I just hope Jon gets another chance to bring his unique economics & tech history knowledge to the table.
I rarely comment on youtube videos these days, but this video was EPIC. Please dwarkesh! Have these guys on regularly for major industry updates, especially as things get crazier in these upcoming AI cycles.
Imagine the trade secrets you can get from Dylan when he is drunk.
He wouldn't be here if he had trade secrets.
Serious question, why is it assumed that he has a clue? Has he worked in the industry? Made prescient predictions that turned out true?
@@stri8ted Internet invites all kinds of people. Maybe the poster and like-actors are just naive souls.
you do need people who can observe the industry as a whole
He will be one of those people who will revea somel military secret just to win a random argument on the internet 😁
unironically the most dangerous podcast put out in a while -- love it
@@TheAudioSaur yes and the one with Yergin also flows into this. I hope Dwarkesh keeps it up and not be like Lex Fridman
Jon Y brings a much needed reality check.
Finally, a podcast dense with information and energetic enough that I can watch fully immersed and don't even try to do anything in parallel out of boredom
Damn, Dylan’s energy and breadth of knowledge is astounding, he is a machine. Amazing to hear someone so knowledgeable about this topic just blaaarghh it out all without a pause to even breathe. I’d like to hear more of Asionometry perspective as well, maybe invite him for another podcast sometime soon. Incredible podcast Dwarkesh, another nail hit square in the head. Thank you!
And this was him being conservative. If he werent to talk in laymans terms we'd all be on the floor having seizures.
The whole discussion around # of GPUs and cluster size is naive. Pre-training and parameter size are already showing diminishing returns, and what you see with the o1/o3 style models is more compute spent during inference, not training. Actually, China centralizing all of their compute resources into one giant cluster would be a mistake. It's much more likely that having a diversity of smaller models will work better during inference time, where you want to explore many different kinds of reasoning and combine them with tree-search or blending in a 'society of mind' like approach. Musk has also fallen into this trap, buying into the fantasy that he who builds the biggest model wins.
This feels so weird. A: never seen asianometry and b: hes not talking in his expert tone.
Yeah threw me off heavily. Also he barely was given the chance to speak 😂.
@@TheHilariousGoldenChariotI think we hear less from Dylin because Asianometry always talks in his channel so i m fine with that but John is also not an industry guy he is more of an expert of things that have happened like a history guy where as Dylin has some chicky info about what the industry is cooking behind the scenes
@@okman9684 Jon is definitely an ‘industry guy’ if you’ve seen his videos he could practically write the history book. ‘He talks in his videos’ that reply doesn’t make sense, this is a podcast. D.Patel has multiple interviews where he gets to talk. I’m sorry but Jon is a much better speaker and has a much broader range of knowledge.
This episode alone is more value than all All-in, Gurley, A16Z, Barrons pods combined
Or Andrew Tate's course on how to get hot girlfriend.
Glad I don’t know who any of those people are
Semi Analysis & Stratechery is like getting an MBA for $30bucks a month. You hear enough of them consistently, you will begin to pick up things in the tech domain. Then you can research topics individually on Google to piece things together.
Or you can just buy QQQ when it dips < 10% 😂
@@pajeetsinghunderrated course tbh
bg2 is pretty good
This is gold! Dude like what the hell. I didn’t know Asianometry is this cool
These conversations are so heady and intense it reminds me of spending hours arguing with my friends in college. Keep it up!
This was excellent content. Asianometry and specially Dylan aer so knowledgeable, it was great!
I'm 30 minutes into the video but have 2 hours already spent watching. I've had to google almost every sentence and learn the backstory as they tell it. What a great conversation
I have not seen Dwarkesh have as much fun as in this podcast. Very engaging!
51:20 He is wrong, TSMC publishes that data, out of their 77000 employees, only 3000 are PhDs. (2024).
The guy is wrong about a lot of things. Like you can't just produce more power whenever you want to from a hydroelectric damn. The water and the turbine has to be there. If you drain the reservoir all the other turbines lose capacity. The world would notice if 5gigawatts of aluminum production was shutdown or shifted. It would affect the price world wide within a week, if leaked within minutes. The heat signature of the plant would change the second something changed. Way to easily notice if someone was looking.
Google was building 200megawatt buildings nearly 20years ago with plans for multiple buildings per site. Twerking the building design to 250megawatts is easy. And many sites had plans for four or more buildings. Facebook was doing very similar scaled stuff more than 10 years ago. Plant scherer and quite a few other coal fired plants on the eastern sea board are running under capacity or have shutdown turbines. Plant vogtle is coming online that's 4gigawatts. It's estimated that plant scherer is running nearly a gigawatt under capacity. That's 5 gigawatts that could be turned up pretty quickly. Also google and Facebook can turn around hardware very quickly. Shoving video cards into servers is too easy. I'm talking about designing new motherboards and swapping them in. If they were very motivated it could happen within 6months. That's from spending approval to 1st 10k units on site powered up. Both of those companies are no stranger to marking things happen.
Also he mentioned early huawei network equipment having Cisco code on them. Huawei was the contract manufacturer for Cisco's switch products at the time. They had/ran Cisco code because they were the same switch hardware in a box painted a different color. It's not espionage when they give you the design and software then pay you to manufacture it.
@@dimedriver Cisco disagrees. Google this - To that end, the following are verbatim excerpts from the Neutral Expert’s Final Source Code Report, dated June 15, 2004: From a section entitled Comparison of Cisco STRCMP and Huawei’s [CODE NAME REDACTED]: “It must be concluded that Huawei misappropriated this code.
This felt like i was sitting with you guys at a bar for how informal everyone's language was, i absolutely loved it.
40 minutes in lmao. These guys are absolutely cracked
Good to see Asianometry. Love his YT channel!
Incredible talk!
Amazing episode. It was jammed with so much information. I needed to take a breath outside to recoup after this epsiode and move some pieces in my brain.
First time viewer, watched every minute at 2x obviously. Holly shit. my people. I could watch another 5 hours of you guys melting my brain. Thnx for a great vid
Dylan said so many incorrect things in this video... It might be good to add some corrections.
I'll just point out a couple that were pretty bad: at 47:30, he said the Magnificent 7 make up 60-70% of the SP500. No it's not even close to 60%, they make up about ~31% of it by capitalization as of now in 2024.
Earlier on, he said China adds as much electric power per year as half of Europe. Not correct, they add about 20-25% (still a lot but it's grossly inaccurate).
I might add more later
To be fair to Dylan, he said many many things. Not sure but probably over 1000 claims so it's not surprising there would be a some errors. I did learn from the podcast, so thank you for sharing it.
Just be sure to do your own verification, especially on claims that sound surprising
We all knew a "Dylan" in college, and usually called Dylan, too.
He's a hype bro, he's just not that knowledgeable, but he talks loudly and quickly enough that others assume he knows what he's talking about.
This crew should be its own podcast this was epic 😂
And it would be called the Dylan Patel Show lol
There's no doubt Dylan is super knowledgeable, but he could've let Asianometry speak a bit more too 😢, please do another episode with Asianometry
That’s a job for the host.
@@TheDrokon there's not much the host and participant can do, when the dominant person is speaking over you
With 2 people it probably not easy to make sure they both talk same amount of time. I liked watching this alot. I'd really like to see both of those guys back
Was there any point where what Dylan said was useless or boring - that's when I feel like you can say Dylan took too long and overextended. Else it's equality for equality sake. Let's not take away agency from Asianometry - he can jump in whenever he wants. When the host was switching subjects and Dylan had more to add, he said hold up and went to complete his thought. Asianometry has a UA-cam channel - he speaks through that as well.
Well, this video turned out a lot better than I thought initially it would. ❤❤ Mr. Patel
This is best combo of guests and host of 2024. 🎉❤
This is some sage career advice! "it's really about: what are you good at, where can you vibe, enjoy your work,, and where can you make an impact in society...what engages you? Because if you're interested in it, you"ll work harder" Dylan
That was so good! A lot of insights and information that humbled me, really.
Jon deserves his own podcast, Dylan didn't let him speak at all!
Would have WAAAAAYYYY preferred this if dylan shut up a bit more and Asionometry got to speak more than 4 syllables at a time. Would have made this a lot more listenable. Feel like Asionometry’s analysis would have been a lot more grounded too rather than this super loose stuff coming out of Dylan’s mouth. He talks about OpenAI raising 50-100B when they only raised 6.5B. Interviewer should have moderated the convo better imo, this had so much potential and felt like it only came half way to reaching it. 😊
Thank you for recording this!
I knew that Dylan and Jon are amazing but Dwarkesh keeping up (or sometimes slowing them down a bit ;) and asking useful questions is great! Anyone who likes this episode should check out Dylan, Jon and Doug's podcast "Transistor Radio" (unfortunately it's a bit more rambly at times, because no Dwarkesh, but you also get random deep dives on stuff).
Something everyone should know about the semiconductor industry is that terms like 2nm, 5nm, etc mean nothing and are nothing more than marketing terms.
The smallest features on the 2nm chips have an actual feature dimension of 20nm.
Never seen Asianometry before, ready for 90% of this to go over my head!
This episode was a riot! Asianometry and Dylan are hilarious together plus they know a bunch about lithography and geopolitics and whatever
Incredible episode. The Pascal’s wager discussion alone is incredible!
Bro, I didn't want to like your podcast because my immediate impression was "hustle and grind" culture, but you keep having such phenomenal guests that I can't stop listening.
Legit, dwarkesh always has a bit of an aura of hype but the guests are always on point
the slight hype attracts people because they like feeling lowkey hyped
Why would you hate hustling and grinding? You Democrats are ruining the country in every aspect
I'm curious how you got that immediate impression.
@@biesman5 Hes a Democrat, they literally believe being a productive member of society is evil
This one feels more “hanging with the bois” than the other interviews
I’d love to be geeked off the blue bottle with these boys
In a good way
great episode, really fun yet super insightful 👍
Came here for Asianometry but came out more curious about Dylan and his newsletter.
Big cheers to dwarkesh, throughout the interview with these two giants he´s clearly done its research, yet he´s constantly pushed to the edge of his knowledge, while still keeping them to getting into semiconductor cloud nine.
this should probably have been 2 podcasts tbh. They are both extremely knowledgeable but Dylan is a force of nature when he's on roll so Jon was silent a bit too much. Would love to hear more from him!
Honestly
Loved the energy of this interview. Love asionometery so this was just fantastic
The new paper about scaling law really applies to this video!
Having 1:1 information to entertainment ratio on this topic is insane
There should be a separate podcast episode with just Jon. Maybe he can talk about his process about how he makes such in-depth detailed videos on a topic.
I never expect Jon to go on a podcast. Happy to see him anyway.
Really love dylan patel. He seems super knowledgeable about this you should have him on again
I only watched that video because of Asianometry. I never seen other two guys and may be would not watch them again. But he worried when that "unknown guy" referred him as someone who made video about zips. It was most emotional moment in a video.
Amazing to have Jon on
Dwarkesh is easily goated now.
I feel like this is the most objective conversation I've heard about Chinese manufacturing, it's culture, and the challenges they face in my life. Thank you very much.
Nah they have very biased view. China did well because of espionage ? What about the world’s largest fleet of stem talents and largest public funding for semiconductor industry ? $$$+ Chinese talents = success. There is 0 chance for failure.
On youtube maybe, but Dwarkesh is so biased, glad the other 2 reined him in.
@@GWT1m0bitch just jealous
@@Unclesam404 Lol China bot.
@@Unclesam404 at least pretend a little bit not to be a chinese agent, introduce some subtlety next time you comment
cocaine’s a helluva drug
Great talk, plenty of grounded perspective on Geopolitics and the companies involved in this right now. Refreshing to hear as someone in the Electronics contracting space who sees folks simply not adapting fast enough.
This was a really good episode. ^^ *Thumbs up*
One of the unique things that came out for me in this excellent conversation was how your two guests described, and in their own ways, represented, two different kinds of fanaticism as motivation: with Dylan, a fanaticism of opportunity, and with Jon (though he was obviously having fun too) a fanaticism of fear.
This is one inconvenient truth that I think DC policy makers are missing. I studied at a Taiwanese University for an intensive study abroad, one of the top schools in Taiwan, and the school is a pipeline feeding China, not Taiwans, tech sector. Those hired by Taiwanese companies plan to work there, and then cash in on a job offer in China. For all the Pro Taiwan nationalism, very few Taiwanese won’t accept a higher paying job in Shanghai or Shenzhen. Once they get rich in China they bring their wealth back to Taiwan. This is a huge economic Achilles heal for Taiwan, a significant chunk of their economy depends on Taiwanese people getting massive salaries in Mainland China and coming home when they retire. Taiwans economic dependence on China is massive and the country would be on life support without trading with the greatest adversary. This also means the best and brightest semi conductor researchers have either been poached by China or will be. It’s only a matter of time before SMIC matches TSCM at this point.
This is awesome. Been watching Asianometry for years and it's great to put a face to Jon. Subscribed.
Amazing episode, I had no idea how opaque the semiconductor world is. Great original content 😁
haven't watched this yet, but i already know this is going to be good
thanks for all your hard work!
hey Dylan, chill with the word "Like". Nice convo guys, many thanks!
They all need to limit their use of “like”
I think they could have shaved 40-50 minutes off the conversation. Plus, it’s slightly, not overwhelming, annoying.
Super interesting guests and you did a great job steering the conversation
Finally all my best friends in the same room
can't you just invite them all out to dinner together if they're your best friends
Love Asianometry ! Very pleasantly surprised
surprised to see Asianometry and Patels together !!
The nerdiest, best conversation about literal magic
Suddenly this video suggestion come up, I loved every second of it. Awesome content.
This was a fantastic episode. None of the AI stuff matters if you can't build the physical infrastructure to develop it. I'm left confused about the timelines though. GPT 5 is being trained right now, so doesn't that mean it won't be trained on Blackwell? Is Blackwell for GPT 5 turbo/mini/o/5.1? Or is it enough for the next order of magnitude model (GPT 6)? And where do the $100 billion data centers come into the picture? GPT 6? GPT 7? And will they be stocked with GPUs more advanced than Blackwell?
I feel like this is relevant because we can only be reasonably certain that scaling holds up to GPT 5. In order for these mega investments to be worthwhile, it seems like scaling might need to hold up for several orders of magnitude more than GPT-5.
why did I enjoy listening to this so much, yet understand so little
Hearing Dylan talking reminds me of the crypto bros, the fast talking excitement, techno optimism, what-if scenarios.
I don't mean that in a bad way, just saying the way of speaking is the same.
If we did have to make the parallels, a hype definitely exists and you have to believe to invest, as these large investors are doing at the moment and question is if they will continue to do so, it has to show (enough) usefulness aka lead to that ROI.
Definitely want to see a podcast with just Jon to let him speak more.
After this episode, Dylan was on oxygen for the subsequent 3 hrs.
We need an AI to filter out how many times Dylan says "like". Holy shit
Question: Currently, Chinese scholars appear to be publishing openly. What happens if they stop doing so in a way similar to how those in the USSR did during the Cold War period from 1945-1991?
I wish I had seen this interview sooner. it is unique and golden. Thank you for setting up it, Dwarkesh.
"If you use AI to optimize your business, you will be successful".
AI: please maximize our production of paperclips.
It's so cool that you've invited my Uber Eats driver
Excellent guest! I eagerly look forward to listening this
5 minutes in and I’m loving this episode. Also feeling kinda paranoid.
Heard a lot about SemiAnslysis. Good to finally hear from Dylan
01:09:30 reminds me of how Jian Yang from Silicon Valley figured out an architecture on how to make the PiedPiper product without using its IP and Gavin Belson was just blown away! 😂
I kinda feel lucky to have the background that I have. Power Systems/Electronics, Digital IC design, Analog circuits, Control/Signals Systems, and lastly AI design. This episode is such an overload that it is giving me anxiety.
😂
The collaboration I just realized we needed
The main money source of big US tech companies is advertising.
If they have extremely good generative AI they could use it for proactive advertising, where their secret AI project user actions and serve ads _before_ people click on anything related.
that's enormous money opportunity and completely separated from their public AI systems.
This whole interview was an unholy concoction!
For those people who thinks Huawei is stealing tech/manufacturing secrets from somwhere, they don't know anything about how hard they work, how many super competent technical talents Chinese university supplies every year, how good the supply chain of everything in China is. Here, in the united state, it takes 5 months to redesign a pilot scale sanitation pipe, the China speed is two weeks and way cheaper.
WOWOWOW NOW this is a crossover I did not expect!
20:50 what advanced packaging company are they talking about as moving from Hong Kong to Singapore?
been trying to figure that out too. a lot of money is moving into singapore semi but nothing stands out to that description
Advanced packagings are mostly in China, aren't they?
@@directxxxx71 no. Companies I know of for advanced packaging are: TSMC, Intel, Samsung, ASE, Amkor, and SPIL which are located in Taiwan, USA/Israel, South Korea, Taiwan, USA, and Taiwan respectively. Depending on how advanced you mean by "advanced packaging" you could argue that only TSMC and Intel (maybe Samsung also) have truly advanced packaging. Amkor claimed to be doing volume CoWoS on an earnings call, but that claim was embedding in insulting ASE, so I'm not sure how seriously to take that claim; and if they're literally calling it CoWoS then it's probably licensed from TSMC. ASE's newest offerings are also opaque to me. SPIL is small so I haven't paid much attention to them, maybe that was a mistake. Samsung 🤷 should be doing better than they are. I don't know anything about STATS ChipPAC Ltd
dwarkesh moves all discussion to llms. These guys are way bigger. Dont limit your brain to LLMs. AI will grow that out mindset.
a most fascinating story of human collaboration mechanics!
Saw both Dylan and Jon speaking at the Asianometry meetup in Taipei last year.
Would love to know what resources were used by Dwarkesh to prepare for this conversation.
Wouldn't the 24x7 infrared heat signature of a Chinese data centre with 1M GPUs be somewhat visible via satellite?
Dylan hypes it so much that I feel like starting a AGI company casue it's my last fucking chance...
Thanks for introducing dylan to me
37:34 Why is it surprising Huawei is so successful? A cornered beast is the most dangerous.