transcribe

Zavain Dar on Backing Hugging Face Early & Why NVIDIA Wanted It

Newcomer · 49m · transcribed Sep 2026
More from Newcomer Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Section Insights

# 0:00

Nvidia's Open Source Repository and AI's Future

What significance does Nvidia's acquisition of Hugging Face hold for AI and biotech?

Nvidia's acquisition of Hugging Face is seen as a pivotal moment in the AI landscape, marking a shift towards open source development in machine learning and biotechnology. The potential for AI to revolutionize drug design and human health is emphasized, with a focus on the transformative impact of technologies like transformer models.

  • Nvidia's acquisition positions it as a leader in open source AI.
  • The integration of AI in biotech could lead to significant advancements in treating diseases.
  • The speaker draws parallels between transformative tech milestones and the potential future of AI in healthcare.
# 9:52

The Shift in Google's AI Strategy

What changes are occurring within Google's AI divisions?

There is a notable shift in resource allocation within Google, with more TPUs being directed to Google Cloud rather than DeepMind. This has led to a restructuring of leadership within these divisions, indicating a strategic pivot towards cloud-based AI solutions.

  • Google is prioritizing cloud services over its DeepMind division.
  • Leadership changes at Google suggest a strategic realignment in AI development.
  • The speaker sees opportunities for startups emerging from this transition.
# 19:44

Insights from the Biotech Ecosystem in China

What did the speaker learn from their visit to China's biotech ecosystem?

The speaker's visit to China revealed a highly educated and ambitious population driving innovation in biotech. They noted the importance of understanding the interconnectedness of US and Chinese biotech and tech industries, emphasizing that ignoring developments in China could disadvantage US investors.

  • China's biotech ecosystem is rapidly evolving with significant potential.
  • Understanding global biotech trends is crucial for US investors.
  • The speaker compares their insights from China to previous transformative tech experiences.
# 29:36

The Interconnection of US and Chinese Biotech

How are US and Chinese biotech industries interconnected?

The speaker highlights that biotech and tech industries are more intertwined than ever, with significant collaborations and drug developments emerging from China that could benefit the US market. They stress the importance of being aware of these developments to remain competitive.

  • US investors must recognize the growing influence of Chinese biotech.
  • Collaborations between US and Chinese firms are becoming more common.
  • Ignoring China's advancements could lead to missed opportunities for US companies.
# 39:28

AI in Drug Discovery: Key Approaches

What are the primary methods used by AI companies in drug discovery?

AI companies in drug discovery utilize generative models to design new molecules and therapies. These models analyze biological data and literature to create targeted drug designs, enhancing the efficiency of the drug development process.

  • Generative models are crucial for designing new drugs and therapies.
  • AI is not just about drug design but also understanding biological functions.
  • The integration of AI in drug discovery can significantly streamline the development process.

Transcript

0:00 What do you think Nvidia sees with this deal? >> I think they now have the the most important open source repository in the history of software and and and definitely for kind of ML and AI. Like a face shift. There was my life before, you know, transformer models and and and after and in a lot of ways I kind of think of my life before Shanghai 2025 and after Shanghai 2025. The vast majority of human disease should be solved, should be treatable.

0:23 >> >> and I I I think we're on that path. >> Zavain Dar is one of the smartest investors I know. Back when he was at the venture capital firm Lux, he helped to invest in Hugging Face long before Nvidia scooped it up for some $13 billion. He unpacks Nvidia's open source endgame, China's AI and biotech surge, and whether AI can deliver on its boldest promise yet, designing better drugs and transforming human health. I'm Eric Newcomer, host of this podcast.

0:50 Go read our Substack at newcomer.co. Let's get into it. Zavain Dar, welcome to the Newcomer podcast. >> What's up, Eric? Thanks for having me. >> Yeah, thrilled to have you. There's so much I want to talk about. You and your firm Dimension just went on a journey to China and wrote a note digesting that. You were part of a investment when you were at Lux in Hugging Face, which is this phenomenal, insane exit to Nvidia, something like 12.9 billion plus >> announced today.

1:18 >> plus retention. >> Yeah. >> And then you're sort of at the heart of investing in AI companies that are trying to transform science and biology. And that's at the core of some of the like the biggest promises that the AI industry is making. So I want you to sort of play tour guide a little bit and explain what's going on there and what what companies can do and can't do. You were an investor in Coefficient Bio, which I reported sold for something like $400 million to Anthropic early on, a great early return for Dimension. Let's let's start off on Hugging Face just given it's so fresh in everyone's mind. What did you guys see, you know, Brandon Reeves at Lux invested with you? What did you guys see when you made the investment? And then yeah, what do you think Nvidia has seen with this deal?

2:09 >> Yeah, yeah, I think at Lux we'd been investing in and around kind of ML and AI really since 2014 around when I joined the firm and a lot of it was spurred by advancements we were watching at least at that point just in our backyard in Stanford with kind of the emergence and maturation and scale of neural nets which then became kind of became kind of deep nets and deep learning. Hugging Face was a company we had been aware of and had been tracking. It was formally kind of a chatbot, transitioned into a kind of a an open-source repo for transformer models specifically at that point not to get too technical BART and BERT models.

2:51 it was actually sourced by our former partner >> Zack Shulman. >> I think an email from Betaworks if I'm remembering correctly saying hey, one of our companies used to be a chatbot they're bringing it to the enterprise space. It's kind of deep tech and AI we think Lux might be interested in it. and Zack kind of shared it around internally. Brandon and I raised our hands. Brandon gets truly kind of all of the credit here. He was uniquely the champion, the table pounder, vociferous that kind of we should be involved. we ended up leading the series A after spending time with Clem and Julian and Thomas.

3:25 My memory is that it was a little bit of a competitive series A so we kind of took out you know, Clem and Julian and Thomas a French spot here in the West Village if you've been there had cocktails at Garfunkel's and how deep the conversation was. >> of the courting. I was good at kind of yeah. >> >> It was a good dinner companion. >> parlez-vous français and deep oh we you know >> >> you could kind of conversate with the founders in their language quite literally.

3:52 >> They'd been in like sort of chatbot territory before, right? But then they had But they were pretty technical for what they were doing. >> Yeah, I mean, Julian did his grad school in computer science at Stanford. Toma, was at Cal Polytech, computer science PhD. they were amongst some of the earliest true kind of innovators in application engineers atop of, transformer models. And so, their initial, application was more of a chatbot Tamagotchi like thing for kids. frankly, we met the company after they had already transitioned, but right after they had transitioned.

4:27 And what was immediately evident then was, just an insane amount of, developer kind of center of gravity, developer pull, for our what they were, kind of tracking. I remember Brandon coming into a Monday IC and showing kind of the GitHub stars over time for the Hugging Face repo, and it was kind of outpacing any other prior open source repo we'd ever seen. so we ended up leading, the A, Brandon, again, truly getting all the credit here, took the board seat. I was, kind of shadowing along as a board observer with him. and then obviously, you know, I left in 2022, but like what he and Clem and Julian and Toma, and then others around the the firm, too. Coatue came in, Sequoia came in, Addition, Leaf Excel came in, too. get just a bunch of credit. I I posted on Twitter a little bit of a, a funny kind of tongue-in-cheek comment, which is we we actually approached Beta Works multiple times after we joined the board, to try to buy their equity. at a premium to where we had just invested, but kind of seeing just how special this team was. And to their credit, they, they definitely, shirked our our offers. So, I think a really good day and outcome for everybody.

5:40 >> And what what do you think Nvidia sees with this deal? >> Yeah, I mean, I think they now have the the most kind of important open source repository in the history of software and and and definitely for kind of ML and AI. If you think of where an Nvidia goes, Nvidia is successful almost one-to-one with kind of the the rise of open source. And so, the ability to standardize essentially the the medium between software and silicon. if that can be kind of intermediated with open source, Nvidia's in a really good spot to make sure that their chipsets kind of reign supreme. it also I think helps counterbalance a lot of the pull that Anthropic and OpenAI have right now. And all you have to do is see OpenAI launch something like was it jalapeno? Their own chipset. to see >> Anthropic's thinking about it and obviously >> I imagine Anthropic's thinking yeah, Google's got TPUs, Amazon's got Trainium. and so >> There's a take I've seen online that's sort of the traditional, you know, business school wisdom is that you commoditize your complement. You make something that you need to buy my service really cheap and easy to get so that you're happy to buy my service. But Nvidia seems to be commoditizing their customer, which you think they would want their customers to thrive to continue to spend money on Nvidia. I know you get to live in the world of early stage startups and Nvidia's like corporate strategy is a little far-field, but do you have any thoughts on you know, Anthropic and OpenAI that Nvidia needs those companies to succeed.

7:19 Like it's going to be pretty brutal >> Yeah. >> if they don't. I imagine it's a huge percentage of the spend has to do either from OpenAI or because people are using OpenAI or whatever. >> Yeah, yeah, like like I think 2 years ago that probably was true where the lion's share of Nvidia's revenue came from the hyperscalers and the frontier labs. Today it's the hyperscalers vis-a-vis inference not necessarily just on their models, but on open source models, and then also the Neo Clouds.

7:46 so, you know, the Models, the Base Stands, the Fireworks, the Nebiases, the the Core Weaves of the world. >> but in some ways some of those are Nvidia too in certain ways. Nvidia's funded them, they're using Nvidia >> them in business. I mean, at this point they're, you know, runaway freight trains in terms of business momentum, revenue, customer traction, so on and so forth. Nvidia again, probably in a strategically shrewd shrewd way, Jensen gets a lot of credit for this, made the ecosystem more diverse. Like, deliberately started investing in a diverse ecosystem 2 or 3 years ago. So, no one customer or provider could get too much leverage over them. If you think about what it means to make the soft to commoditize the software layer, and and we can debate whether or not that's actually happening or to what extent or what the various shades of gray there are, but if you think about what it means to commoditize the soft software layer, it means it's much easier now for someone to kind of set up an inference company.

8:42 And how do you set up an inference company? You're going to need necessary >> Buy a bunch of chips from Nvidia. >> a bunch of chips. And so, I think it's helping their customer. Actually, it's helping their new cohort of customers more. >> are the Neo Cloud, you're saying. >> Exactly. Yeah, or the hyperscalers, Google obviously with their kind of potentially shuttering DeepMind instead to focus more so on kind of the the Google Cloud. >> You think they're going to shutter DeepMind?

9:03 >> Well, they've divested some of their resources into into DeepMind, and I think that's tantamount to some sort of reflection that they're probably saying at least today the revenue's coming from inference. And so, >> So, it's more like Google Cloud than it is be the next Open AI. >> That's right. >> You think that's how they're committed? >> I think that's that's certainly the body language that that you're reading if you're observing them. Yeah. >> Interesting. That's I mean, I saw obviously with Demis Hassabis chair, you know, Jeff Dean leaves, tons of negative signals.

9:34 >> Chavez on on on your podcast? >> Yeah, yeah, he spoke of our events and yeah, podcast feed, yeah. >> I imagine the debate internally at at the Google board level and then between Google Cloud and DeepMind and certainly you were starting to see these tensions arise, which is hey, we're spending a ton of money on DeepMind. we're in like third or fourth place. It's not generating a ton of revenue. And oh, by the way, if you look at Google Cloud right now, the the money from inference on Gemini and open source is rocketing.

10:02 >> Hmm. >> we are giving more of our TPUs now to Google Cloud than we are DeepMind. And there was tension between those two factions. And what you ended up seeing was essentially Demis and John Jump and Jeff Dean, Quoc Lee, many of kind of the leaders of the DeepMind Google Brain org slowly leaving and or being shown the door. I think it's probably a little bit of both in favor of Google Cloud. >> You must be startup hunting out of all these DeepMind people then.

10:27 >> For sure. Yeah. I I mean, I think we're startup hunting just out of the you know, the most talented people on the planet right now. Of course, cuz something like Google DeepMind is is a really interesting place to be to be hunting. >> What I mean, the other piece of this I I broke the news and Newcomer was a huge scoop for us that Nvidia was doing this weird partnership with Poolside. 6 billion to acquire a lot of Poolside and then invest some in the remaining company which is sort of continuing hard at work and they have a separate infrastructure company. So a very complicated deal.

11:01 But so yeah, between Hugging Face and Poolside, Poolside had really leaned into American open source. >> Yeah. >> Nvidia is going to be sort of like a key player in this American open source thing. That sort of fits into the next piece of the conversation which is you go to China where a lot of the open source work is being done. I guess before we get to the journey to China, what do you make of American open source today? Like how credible are American open source models?

11:28 >> It's definitely I mean this with all respect. It's it's you know, it's quantitatively a half step. >> Yeah. Yeah, and for a whole milieu of reasons, talent, capital, resources, whether or not one can legally here distill against the frontier models or not. >> get to compete on a different >> slightly different playing field. How much the government is investing in and in what chip sets and what data you know, the the two continents have and so >> I think there's a whole other question in in in terms of like what it means to be American open source versus European open source or Japanese, you know, Chinese, so on and so forth and like and why that necessarily matters. I think we there's somewhat of like a a false focus on it, which is >> What do you mean?

12:14 >> >> If you go to Europe they're all about sovereign AI. They're like, you know, you know, I mean Cohere is literally, you know, a Canadian company building ties to Germany, set it spending time in London with the idea that like, we need every democracy needs that's not America needs to be tied together. So, I I do think Europe is very mindful obviously of where this stuff is coming from. >> Closed source sovereign AI makes a bunch of sense and you could argue AI and Anthropic are closed source American AI.

12:45 And even just in terms of the Trump admin's ability to say, "Hey, don't open up that, shut down that, work here. We want this before you open it to others." etc. etc. You're already starting to see potentially even taking equity positions in Open AI and or Anthropic. >> Though Anthropic did successfully sue the Defense Department at least. Yes, but but definitely there's a lot of influence. >> Yeah. Open source is a little bit different. You know, if something is definitionally you know, open, assuming the license, whether it's an MIT license or something else, doesn't doesn't back prop in some way, shape, or form to some sort of kind of country preference, which they don't right now, it's unclear to me why it matters. And I can imagine and just to kind of say what I think an intelligent computer scientist who might be listening to this would say is, what about backdoors built kind of in the open source? What if what if there's kind of latent representations that are you know, quasi propaganda in terms of what the model would respond back with or suggest based on how it's trained or kind of how it's productized for you.

13:51 And so we should be wary of that, but I think there tends to be sometimes a little bit of a false precision when there's too much focus on the the sovereignty of where open source comes from. >> Right. And you know, open source in AI can be different than open source elsewhere in that there are like layers to it and people you know, we've seen with meta and others where they all say open source, but that doesn't really mean you can just sort of like rip it and go. But now the distillation where you can basically have somebody else's model and easily sort of like oh figure out how it does things and do it yourself. There's there's almost been less of a concern where you know, this has come up in the US government's interest in like should we ban Chinese models?

14:32 I think there's been this response of what does that even mean? We just like rip off their model and call it our US one and it'll be slightly different and I guess hopefully we'd weed out whatever propaganda was built in, but maybe we don't understand it well enough, but you know, what Sue us and say that that's actually the old model? It's like well we say it's a new model, you know, I don't know. What do you make of that that debate? Yeah, like I think any talk of >> in some sense, you know, severing this figurative cord that crosses the Pacific in terms of silicon and data and compute and models and weights seems you know, far-fetched at best.

15:10 And just to give kind of a brief vignette of I think how most you know, vertical AI companies today are using something like Kimmy K3 or you know, GLM from ZAI. >> They talk about using, you know, Fireworks or Base 10 >> Yeah. >> Model, whatever, but a lot of times they're sort of the route to use these Chinese models. >> like often times what happens and and at least up until now this is what has been happening, which is Anthropic or Open AI will use American chips. So, Nvidia, AMD, so on and so forth to train frontier models, largely with American data using Merkle and Surge and Scale, so on and so forth, the web, of course, maybe the largest biggest free data set out there.

15:53 They will train these models. A Chinese lab, maybe Moonshot, which produces Kimi, will distill against it. So, they'll ask it a bunch of questions in quick succession and they're kind of build a representation. Rip it off. Yeah, build a representation of the model. And so that representation already has embedded in it the the data that was American vis-a-vis kind of the silicon that was American vis-a-vis the talent that was American, so on and so forth. They'll then open source that model.

16:23 And then this has happened, you know, with with Harvey or with Cursor or with Cognition, you know, any of the vertical AI companies now that Silicon Valley's pouring hundreds of millions if not billions of dollars towards will fine-tune or post-train against these Chinese models. And so you what you end up with is this, you know, the this technology crossing the Pacific twice before it comes to the end consumer. So, if I'm a lawyer and I'm using Harvey, it's already crossed the Pacific twice and it's already touched both Asian and American technology and silicon and data and talent. And so, to cut it, what I think really chop at, you know, at the knees, vertical AI and American kind of, you know, vertical AI software's ability to innovate right now. Maybe unless like truly US open source catches up. If US open source catches up, then I think you can start to think about ways to kind of maybe cut the the the technological tether.

17:19 >> But we can't stop distillation, right? Or >> You can't stop distillation per se, I think I think the the labs can probably be far better at tracking it. >> There who's aggressively using our model to >> Yeah, it's like a DDoS attack. You know, you should be able to track when it's happening and you should be able to shut down that API. >> we're seeing reporting now about trying to hide some of the thinking potentially to sort of >> Yeah.

17:40 >> how stuff works. >> And there's like there's numerous theories. No one's come out and said it here's exactly why this is happening, but like a very reasonable theory is that Anthropic and OpenAI have been chasing revenue. And of course, even if you're distilling a model, you're calling kind of inference to Anthropic or to OpenAI quite a bit. And that helps the revenue numbers. >> >> In the short run, let them distill it and then ban it after you get paid for it.

18:03 >> goes at the limit though is and you're again already starting to see this, Anthropic and OpenAI not opening up vis-à-vis API their frontier models. Which if if a frontier model isn't programmatically available, you can't kind of query it at the volume and scale that one would need to distill against it. and so you know, I you might end up seeing a little bit of a kind of a growing gap between Chinese frontier models and American frontier models simply because American frontier models are no longer longer distillable simply because they've shut down API access.

18:36 >> So we sort of explained one of the many reasons why it'd be interesting to go to China right now. Tell, you know, explain a little bit about the trip. Where'd you go? Who showed you around? Yeah, what what what what what were you trying to learn? >> you know, so so that maybe as a little bit of context, we we went out to Shanghai last year and at least from our second fund ended up making one of the bigger capital investments that we made from the fund into a Kymeta Corp.

19:04 But essentially, you know, Chinese kind of footprinted company based in Shanghai. And that was a company that sat at the intersection of, you know, tech and biotech. So >> What's it called? >> Yurodel or Helixson. It's got like kind of two names. >> And how's it doing? >> Extremely well. Yeah, I know. I So it was a founded by a Chinese-born, US-educated, formerly kind of tenured UIUC computer science professor who during COVID moved back first to Beijing and then Shanghai, decided to start this company. Ended up raising 787 million from us and DST and Hillhouse and a number of kind of global tech and and healthcare-oriented investors.

19:45 Have a number of large pharma partnerships and collaborations with Sanofi, for example, a few that they haven't yet announced. Have multiple drugs in the clinic and frankly have a technology that's working. >> Hm. >> and so we invested in them last year based on based on a visit there. Again this year we decided to go back out, spend time with them and also kind of get more ingratiated not only in the biotech ecosystem in China but then also in the AI ecosystem. We ended up kind of landing in Shanghai, spending time in Shanghai and Beijing and Hong Kong. and mostly spent time with with investors and entrepreneurs and labs and and kind of everyone in and around the ecosystem so like bankers and advisors and kind of the whole the whole the whole spiel. Anything one would do if, you know, for example, if we were going to go visit London and wanted to get smart on the tech ecosystem. Probably the same kind of cadre or, you know, collage of different kind of sorts of people that we met.

20:43 >> And what were some of your key takeaways? >> >> You wrote a letter. >> Yeah, yeah. >> You structured your thoughts here. >> Yeah, going to China last year was as big of a oh wow, this is happening and it's going to be a tsunami wave when it happens for me as as as seeing the first results from transformer models in the transformer papers were in like 2017 for me. And so like a face shift, there was my life before, you know, transformer models and and and after and in a lot of ways I kind of think of my life before Shanghai 2025 and after Shanghai 2025 a young, a highly educated, highly disciplined, highly hard work, highly intelligent population with deep and aligned ambitions to to do amazing science and technology to create and capture value and and an ecosystem and a culture that does a thing and instead of then tweeting about the thing, which is I think often times like what American entrepreneurship and technology has become, they do the next thing. And and and in that way it was a little bit refreshing. I grew up in Berkeley in the '80s, '90s, and early 2000s and my mom was actually a com- a computer science student at Cal and so a lot of her friends were early Google employees and and and and early tech Silicon Valley folks. And this was before tech in Silicon Valley had really jumped the shark and it reminded me of that feeling like of just kind of being in and around a milieu of deeply ambitious, very intelligent people.

22:24 >> Valley has become look at me sort of so so much promotion. >> Quasi quasi navel-gazing. I I know you had the TVBN guys and I I I I really like what they're doing, but I got a text from a friend yesterday showing me a screenshot of like a paparazzi style picture of like some entrepreneur or or or or investor and and it's just become a little bit navel-gazing. And and don't get me wrong, we should celebrate entrepreneurs and investors and people taking risk and people creating and inventing, but you also need to give space and time for like long-form thinking and focus and space to actually do work.

22:57 China can often make America feel like Europe >> Hmm. >> in the same way that when we sit here and kind of think about like the the work ethic or the drive or the pace of business kind of velocity in Europe it's a step slower. >> I don't know. Having an 11-month-old, I'm now decidedly anti-996. Or it's like I >> >> I understand that there are different phases of life, but it just feels that a company, you know, would have that expectation. It's just like I mean this is how Europeans feel about Americans.

23:24 It's like we achieved all this >> Yeah. Yeah. Yeah. >> to not have to do that, you know? There are some things that that matter more. and so, you know, I guess being America somewhere between Europe and China, you know, you get to feel a little good about like, "Oh, maybe it's the sort of happy medium." >> yeah, and some like look, I I wonder if a normative assessment of 996 is like is the question that we should be asking or if it's just like observational. Oh, okay, this is 996 and and one can and should, of course, ask ask the question, "Do we as a society have enough to take care of everybody?"

23:59 and or have we become like a perpetual, you know, you know, hamster wheel. and is there any way can we imagine a world where you find a little bit more balance? >> For sure. But if you're trying to find the best investments, the ones doing 996 might >> Yeah, a little bit of both. I think you have like the best entrepreneurs also tend to be creative and spunky and I think oftentimes are spiky to use the maybe the the word of the year in the kind of venture entrepreneurial community in all sorts of ways that aren't just work related.

24:27 >> I mean, one of the observations that really resonated with me >> Mhm. >> is that Ameri you know, I feel like every every work of fiction you sort of like the villain is created by the good intentions of >> Yeah. >> the heroes or whatever or this is you know, it's opposition and it's like the US was like, "Okay, we want to stop China. We're going to crack down on these Nvidia chips." And then through the deprivation of not having access to the best chips trying to sort of been able to do things in sort of more humble ways, which has allowed those models to be cheaper, which means that they're more sustainable, more usable, better for these vertical application companies we're talking about. Yeah, I mean that's sort of it's a nice it's a nice story, but I mean you also want to have like the best chips in the world. Do you think Do you US made a mistake by by restricting the access or this would have happened anyway or what Yeah, what?

25:19 >> I think the jury's out. like that certainly like with no doubt what has happened is we cut off supply of cutting-edge chipsets namely through Nvidia and and that forced the frontier labs and engineers in in Beijing and Shanghai and a lot of the AI work specifically happens in Beijing or in and around Beijing because that's where Tsinghua and Peking University that kind of academic talent is. it forced the you know, the compilers engineers and the systems engine engineers in the low-level assembly line and assembly code engineers to really innovate all the way down to the bare metal metal in a way where the American labs aren't and in some ways that circumvented the export controls that we that we'd put on. in other ways the Chinese labs are still struggling. And so like what are the two worlds where where we'll look back and say was this the right decision or or or the wrong decision? If if right or wrong this is depending on like whether or not we win the AI race win the AI race against China.

26:19 if Chinese kind of engineers and computer scientists are able to innovate such that they make irrelevant kind of the like the US's lead on on on silicon, then yeah, for sure. We'll look back and say man, that was the that was the wrong decision. That's not the case right now. Our frontier labs are meaningfully ahead of the Chinese frontier labs in no small part because we have better silicon. We don't have as much power and at some point we'll kind of run up against that wall, but right now we have a meaningful lead.

26:50 China, even if it has enough power, it doesn't have enough chips, it doesn't have enough compute. And so, right now they've done an amazing job optimizing at a very low level. It still isn't enough to catch up to our frontier labs. But it's certainly something that we should be paying attention to. And and the question is are we creating an enemy in the future or somebody who wants to catch us up on the silicon level, then they'll both have power and better low-level software engineering than us.

27:14 That's kind of a scary world, and so we should be tracking those two. I think probably what's the most interesting is to watch the the velocity and the cadence of of innovation from the Chinese semiconductor industry. And right now that's I think moving faster than what most industry kind of, you know, kind of viewers would have predicted 24 months ago. And so, and I I think that's probably where we want to be paying the most attention to.

27:36 >> Just do you think they will catch up or >> In the fullness of time, when you have to think that they will catch up. Nothing is fully defensible forever, and if you have, again, a a large, extremely intelligent, extremely motivated, and capitalized population to go and pursue these things, they will. China graduates more STEM more STEM graduates than the world globally every year. And so there's just like hundreds of thousands of extremely intelligent people who will go and pursue these these questions.

28:05 >> How To what extent do you feel free to be a global investor? You know, I'm I'm sort of a globalist. Like I want a world where the US and China can >> Yeah. >> share technology. Obviously, I want Chinese to follow agreed upon rules and compete on a level playing field. There's some protectionism issues I have, but generally I'd like to see a world of cooperation and fair competition. but you know, just a few years ago Sequoia Capital, you know, had to spin out of China. Literally GGV, which was known for being this cross-border investor, split split the firm. But now it does seem like there's some, you know, some comfort. I mean, the information, which obviously journalism is different, but Jessica Lessin just sort of talked about hiring a new reporter to really focus in on China.

28:50 And there's just like a mood that people feel a little bit more comfortable saying, you know, that they spent time there, and you've invested around those companies. Like, what's your read on how free you are to invest in Chinese companies and what the what the mood is right now? >> maybe I'll start with saying like, you know, we we invest in companies largely globally for the vast majority in the West. and largely avoid, you know, regimes and areas that we kind of view as unethical or kind of adversarial to Western values. And so, like, we don't have investments in Russia, for example, or Iran, for example, right now.

29:31 I I think it's a shame that people are a fearful such that they shunt their curiosity in these situations. I think that being open and honest with your curiosity and learning from how others do things is is really, really valuable. We don't have a China strategy per se. And again, we have one investment that we made in a company that's a Came in Top Co, happens to be, you know, footprinted in Shanghai. but I do think that there's a lot that we can learn.

29:59 And especially right now, if if you read the letter, I think one of the the points I really tried to convey was biotech and tech today are more interwoven between the US and China than arguably they've ever been before. And so, to have your eyes shut off to what's happening there, I think it's positioning you, even if you only focus on the US, you are disadvantaged to be ignorant or naive about what's happening in China. >> Have Have This is going to This is not my area of expertise, but have Americans really benefited from like novel Chinese drugs to date, and is that a world we're about to enter?

30:34 >> Yeah, we are certainly entering that world. if anyone has been following Summit, which is a a US biotech company that's got a drug that's going after PD-1 VEGF, that molecule was out-licensed from Acaso, which is a a Chinese biotech. >> Hm. >> >> >> a BeiGene, which was formerly known as BeiGene, now European domiciled, founded and and really kind of brought to maturity in in in China, now has kind of commercial global ambitions with their kind of commercial drug pipeline. as well. if you look at like where the at this point the majority of pharma out-licensing and BD dollars are going, it's going into China. And so, you know, all you need to do is look at the large billions of, you know, upfront and biobucks dollars that maybe, for example, GSK is spending on Chinese assets to see, okay, if those assets are now entering GSK's pipeline in the next 3, 4, 5 years at some point, assuming that they're approved, they will be commercially available, and we'll all be at that point taking an antibody that was maybe invented in a lab in Suzhou, China.

31:45 >> And so. >> So, it's coming. >> It's coming. Yeah, yeah, yeah, it's coming. >> And and so, this this sort of transitions us to this you know, field that's you know, just hard to understand for many of us or like you know, AI and biology. It's like you know, I think some of the largest pronouncements from people like Dario at Anthropic are like, oh, you know, I mean, literally said something the other day that's like oh, we want to win people over by curing cancer or something, you know, then then we'll really earn goodwill, and that's the kind of stuff that'll move the needle.

32:15 And I think some people reacted to that and like even if that's true, it's going to take time. So long, and obviously, you have to run experiments. You you might think you have the answer, but it could take years and years till people really believe it or, you know, not believe it's even the wrong word till it's proven and goes through sort of the human processes that it has to, but where, you know, what makes you most optimistic about sort of drug discovery and disease attacking diseases with AI and and where where are you sort of like this going to take a long time?

32:47 >> I'll maybe I'll preface it in that I've lost in my investment career which like really started in 2013 lost far more money underestimating the velocity of technology and science progress than overestimating the velocity. >> Like you missed deals you should have done. >> I missed deals. I thought this would take too long. I didn't think it was tractable and so on and so forth. And so I think I have learned through a negative Pavlovian response to to really attempt to like understand why someone who's really intelligent even if they're saying something that I think is far-fetched to try to like understand where they're coming from.

33:27 I think what I already said which is really inspiring is you know, it's it's not it's one thing to talk about it. It's another thing to actually go do it and we can talk about the coefficient and the team and you know, kind of our views on on on what they're doing. but I think in our lifetime I don't think it's it's hard for me to have you know, the fidelity to say seven or 10 years or 15 or 20 years but certainly in our lifetime assuming you know, we live to ripe healthy ages the vast majority of human disease should be solved should be treatable and I I think we're on that path and you know, we can talk talk about a company like Chai that essentially has developed generative technologies to perturb biology in exactly kind of the way one would want to so a biologist can say hey I want to perturb this protein at this site by up regulating it or down regulating it.

34:17 >> Sorry, perturb is this a model or is this in a lab or >> in a in a human or an animal setting. So I want to get a protein I want to get this protein into this cell type or into this organ type and then at that organ type, I want it to hit another protein. And the protein that I want to hit it on, the target in in kind of, you know, biotech pharma speak, I want to attach my protein that I generated onto this part of that. It's called an epitope.

34:45 and so that's kind of amazing. That's like a that's a toolkit that we just never had before and you're seeing it in terms of the pharma dollars and the revenue and the traction that they're getting. but that's kind of like one step of many that's all simultaneously kind of happening all at once. >> So, in our lifetime we have the ability to sort of treat all these diseases. Yeah. Does that mean we're living forever? It means we're living to 200. It means >> Yeah. and and and like and maybe before we get to that question like I think the long pole is and will continue to be testing in humans. Like we we don't have we don't have in silico systems to replace. nor do we have animal models to replace in totality. Like the necessary requirement to test in a stepwise function. First safety, then efficacy, then really testing safety and efficacy in a phase one, two, and three kind of trial for most for most drugs. I I actually saw Jacob Kimmel who's the the founder president CEO of New Limit kind of comment I think on Twitter recently that that this kind of long pole will continue to exist. We don't really have line of sight to something that's going to meaningfully step function, change it, or decrease kind of the the timing of it. We might be able to multiplex it. So, when you're kind of enrolling in a clinical trial, can we be creative or intelligent about testing multiple drugs in one person at the same time to get kind of a >> And to say in sort of the dumbest way possible, if a drug professes to keep you alive forever, you have to run that experiment. You know, it's like if it's like, "Oh, this is going to help you, you know, 30 years from now."

36:20 How do you really you know it's not hurting you in the near term? Maybe it shows some immediate benefit, but you don't necessarily know, know. You might have some theory of the case why AI believes it'll do X, Y, and Z. >> Novo Novo just had a failed phase three antibody that was going after an interleukin. So, kind of extracellular protein called IL-6. And we had every biological and theoretical and kind of in vivo and even in human reason to believe that if we kind of shut down or down regulated IL-6, it would reduce the rate of like heart disease, essentially ASCVD, various forms of like heart attack and heart disease.

37:02 We we took We we took this kind of drug through the preclinical setting. We took it into phase one. It proved safe. We took it into a phase two setting which setting which is a little bit shorter. So, you know, you didn't have the full three or four years required to see, "Hey, on average is are people having less heart attacks?" But we had all of the biomarkers that we thought were highly correlated to heart heart health. And and it and we were taking down IL-6 and it was taking down all of the biomarkers that we thought were highly correlated to heart disease.

37:30 We then ran it in a in a phase three setting. It cost hundreds of I think over 500 million dollars was the aggregate cost of this phase three. It took about four years to run. And there was zero correlation. There's there was there was 1% correlation. It was a 0.99 hazard ratio which is the equivalent of 1% correlation between that drug and decreasing heart disease. And so, what does it mean? It means that biology is really hard. We thought we had all of the biomarkers. Those ended up being correlative not causal. The drug did what we expected it to do. It was taking down IL-6. It was taking down kind of that pathogenic protein, but it wasn't actually reducing the disease. And so, until we figure out better biomarkers, until we better understand human disease biology, it still going to take time and there still will be failure.

38:15 >> But if we cure all disease or if we or just cancer Does that just There's still a clock or this is an attempt to sort of we will will not die of natural causes. >> It That's somewhat of a philosophical question. Do you think aging is a disease or is it you know, part of the human condition? I think there there are many smart people who would answer on both sides. >> Oh, because your claim might be accept aging or like are you sort of leaving aging out of >> There there are there are reasons to believe that aging might be helpful to some extent you know, in the first 20 or 30 years of our lives, but then we haven't really evolved. There's very little in human evolution that's based on the years 30 plus simply because the vast majority of our ancestors never made it past those years. And so yeah, could we could we imagine a world where we maybe freeze the aging function at around 27 and 33 in that range perhaps if that's possible. Companies like New Limit by the way like not in this fantastical let's go do it all at once freeze at 27, but like let's be organ or cell specific and figure out a way to perturb the age of particular cell types or organ types with exquisite precision and control. Like that's what they're going after.

39:32 >> And how much is it sort of generative models like the LLMs we all know like is it I go into open AI you know, some sort of private version of open AI and I ask like what should we do and like it spits out some answer and then scientists run around and run experiments or what? How should we imagine what these like sort of AI drug discovery companies are doing or like what's what are their key approaches?

39:58 >> so there are I think you know, a handful like maybe to name a few of them. There's companies like Ty again for example who are developing technology to help you design it what in industry is called composition of matter. So I want to design a a molecule or an antibody or a macrocycle, or you know, radioligand therapy, help me generative algorithm design the actual thing. there are So, that's I build a specific model that is sort of thinking through it around these problems, trained on it, and then it's coming up with answers itself. Exactly.

40:33 And instead of saying, "Hey, write me an essay, or write me a poem," as one might do to Claude or Anthropic, it's saying, "Design for me the antibody sequence of a drug that binds to that epitope of that protein, or design for me a small molecule that hits into that allosteric pocket of, you know, that intracellular protein," so on and so forth. there are companies that are trying to map biology. So, it's not just about designing the medicine, it's about knowing what the medicine medicine should do. And so, that's often reading literature, it's understanding, you know, large multimodal biological and human health data sets, so multiomic, human health, EHR, EMR, kind of coalescing all of that and making sense of it. we we started actually a company internally at Dimension called Coefficient.

41:21 And one of the ways we used it internally before kind of spinning it out into a newco, was to make sense of new papers, new ideas, new pathways, new targets, new potential remedies in in in ways that we could very quickly make make sense and get smart in these areas, even if we never kind of seen them >> And could have helped you just invest in great companies. >> Exactly. No, and and and and that's frankly how we used it. And and and every day across, you know, the 10 >> This Coefficient Bio, which sold to Anthropic for like 400 million dollars.

41:53 >> Exactly. And and and every day across the >> Am I making this up, like a 30,000 IRR something? >> I I I probably shouldn't say the the the specific financials of it, but al- although again, here too, my my Dimension co-founder, Adam Goldburn, a bunch of credit. He was kind of the entrepreneur investor co-founder of the business. It was his idea to spin out what we were developing internally into a company. He was single-handedly responsible for recruiting the founding team, galvanizing them, financing them, and then working with them through kind of the early innings of the of the coefficient build. We were extremely lucky to partner with Nathan and Aris and and and Sam and Jesse and the team there.

42:36 And And and to give you just maybe like a quick sense of like where at least we were kind of thinking about building it. Every day there's, you know, 10,000-plus, tens of thousands of universities that are all pursuing different parts of science. Every day they're putting online new results, new data. If I am, let's say, Lilly and I'm running a trial against IL-6, maybe, and there's a a small research lab in Germany that publishes a result. what are the odds that I as the, you know, that the trial investigator will happen to see that result? Zero to none, right?

43:13 But if I have an agent, quote unquote, online scouring the world's kind of corpus of knowledge in real time, so being able to incorporate in real time against your world model, what you know, what do you think the probability distribution of different biological phenomena to be true is? If you see a result from a small, you know, research university, you know, in the suburbs of Berlin come out come online, you should be able to the next day update what your thinking is, update the Gantt chart. No, actually, you know what, we should look at that patient criteria, not that not what we had before, so on and so forth.

43:44 And so I think that's where we're trending. It makes me really hopeful that we will get more accurate with our kind of scientific predictions. >> and it's just like the average American is so underestimating how revolutionary what's happening with generative AI is. Like, I mean, Americans are pretty hostile to it. And partially the, you know, AI industry gets some of the blame for how they talk about it, but just just on on science alone, the amount that people see opportunity, it's it's sad that yeah, the American public sort of hasn't gotten that message.

44:20 >> cool. Like and like what's happening in math right now? Like we're we're we're we're potentially nearing like the asymptote of the full closure closure of, you know, a priori human knowledge. And man, like as as as a nerd and as a geek growing up, like that's like the coolest thing. >> Translate that. You just You just think AI is about to sort of solve all these unsolved truths. >> that you can deduce with a pen and paper. So like mathematics, for example, you should we should have the full closure of it. We can in theory, I think, sooner rather than later be on a path towards essentially like having a pamphlet of here here's all of mathematics and here's kind of like the proofs. To the extent that I mean, there's no complete representation of it, but to the extent that for any any statement, we should be able to kind of give a true or false proof for the most part. We're on that path. and that's that's done.

45:13 >> Where it's sort of you could solve it with binary is sort of the claim or >> can solve Yeah, in in philosophy, it's called a priori. So it's it's things that you can deduce without kind of empirical testing. >> Hmm. >> physics, chemistry, astrology, space, so on and so forth. Astronomy, space. those are more empirical sciences. So it's a posteri a posteriori. So you need to be able to kind of then query the universe, get more data, and then help guide your knowledge back. and that's where there are two major kind of branches of AI research right now. One is in and around kind of mathematics, which is like a priori knowledge. And the other is around kind of empirical knowledge. So you're seeing so many kind of labs being spun up to essentially have closed-loop interfaces between kind of the the AI model and some sort of set of experiments that they can run and then have it come back and update the priors of the AI model.

46:04 >> I feel like, you know, what do you what do you make of Silicon Valley's sort of I guess silence on some of the US turn against immigrants or >> >> You must So weird. It's so weird. >> It's sort of dark, right? Like >> the fact that we can I like I'm all for opening the over 10 window. and but it but I find it deeply troubling that there's not more pushback against you know, essentially a a view that we should have some sort of reverse DEI for Americans. Like we are I believe we are as a country and as an economy certainly who we are because we've been the beneficiaries of global brain drain. And so we've you know, for every role in every company, in theory all things equal, we should be able to you know, find the best person, source the best person globally. and that's why our companies by and large have been successful and >> It's good for America. I mean, Palmer Luckey you know, has been fighting with Paul Graham probably fighting as we speak for days. And it's like successful American companies will either hire everybody or they'll open up offices somewhere else. So like a global multinational is still going to hire people outside of the US.

47:21 And so I don't even understand it's better for Americans that the talented people are here paying our taxes like >> starting our companies here and being aligned with our government. hiring with and building with our values. >> But the defense I mean, Garry Tan I understand you have like the you know, the White House basically or like the FTC or whatever it was like zeroed in on you. But just like not even immigrants have given the Silicon Valley so much and to not even like defend them at this moment just feels >> Yeah, like I like I was deeply disenchanted by the large, you know, sound of a pin dropping from many of my Silicon Valley venture peers in the face of this debate. And I don't think that would have happened like 10 years ago or 15 years ago.

48:13 Frankly, it's for the most part it's illegal to to hire on the basis of citizenship. So it's like >> >> Like there and and there are times when like you legally need to find an American citizen because like it's CIA or government related, it's classified, so on and so forth. But for the most part, like again, all things equal, it is actually illegal to like to make a hiring decision based on citizenship. And so yeah, we're just in a weird moment in time. I I hope we're I hope we swing out of it.

48:47 I and like I would like I yeah, I hope we swing out of it. >> Well, great to have you on the show. A lot lot to be hopeful for and so much change in the world. So really appreciate you coming on. >> Awesome. Thanks, Eric. >> That's our episode. This is the Newcomer podcast. Thanks for listening. Please like, comment, subscribe. Go find us on our Substack at newcomer.co. If you're hungry for more podcasting, I also have a talk show called the Silicon Valley Show.

49:14 We love your comments. Thanks for all your support and see you next week.

Summary

Nvidia's acquisition of Hugging Face for approximately $13 billion is seen as a strategic move to secure a dominant position in the open-source AI and machine learning landscape. Zavain Dar discusses the implications of this deal, the evolving AI ecosystem in China, and the potential for AI to revolutionize drug discovery and human health.

- Nvidia aims to standardize the interface between software and silicon through open-source models, enhancing its chipset dominance.
- The acquisition of Hugging Face positions Nvidia against competitors like Anthropic and OpenAI, which are developing their own chipsets.
- The American open-source AI landscape is perceived as lagging behind China's, which is rapidly advancing in AI and biotech innovation.
- Dar emphasizes the importance of understanding the interconnectedness of global AI and biotech ecosystems, particularly between the US and China.
- AI's potential to transform drug discovery is significant, with advancements in generative technologies enabling precise biological perturbations.
- The long-term challenge remains in human testing and regulatory processes, which can slow down the pace of innovation.
- Dar expresses optimism that many human diseases could be treatable within our lifetime, driven by AI advancements.
- The conversation touches on the need for a balanced approach to immigration and talent acquisition in the US tech industry to maintain its competitive edge.

Questions Answered

What significance does Nvidia's acquisition of Hugging Face hold for AI and biotech?

Nvidia's acquisition of Hugging Face is seen as a pivotal moment in the AI landscape, marking a shift towards open source development in machine learning and biotechnology. The potential for AI to revolutionize drug design and human health is emphasized, with a focus on the transformative impact of technologies like transformer models.

What changes are occurring within Google's AI divisions?

There is a notable shift in resource allocation within Google, with more TPUs being directed to Google Cloud rather than DeepMind. This has led to a restructuring of leadership within these divisions, indicating a strategic pivot towards cloud-based AI solutions.

What did the speaker learn from their visit to China's biotech ecosystem?

The speaker's visit to China revealed a highly educated and ambitious population driving innovation in biotech. They noted the importance of understanding the interconnectedness of US and Chinese biotech and tech industries, emphasizing that ignoring developments in China could disadvantage US investors.

How are US and Chinese biotech industries interconnected?

The speaker highlights that biotech and tech industries are more intertwined than ever, with significant collaborations and drug developments emerging from China that could benefit the US market. They stress the importance of being aware of these developments to remain competitive.

What are the primary methods used by AI companies in drug discovery?

AI companies in drug discovery utilize generative models to design new molecules and therapies. These models analyze biological data and literature to create targeted drug designs, enhancing the efficiency of the drug development process.

© transcribe · For agents Built with care and craft by Gokul Rajaram