# Nathan Lambert on China's AI Ecosystem and the Open Model Gap | The Curve 2025

> [!note] Machine transcript
> Auto-generated captions. Names, jargon, and code identifiers are often wrong — check any quote against the video before putting it on a wiki page.

## Description

Nathan Lambert (Senior Research Scientist at the Allen Institute for AI and writer of Interconnects) examines how 2025 became an inflection point for open models, driven by DeepSeek R1 and the rapid expansion of China's AI ecosystem. He goes over the rise of Alibaba Qwen, Moonshot AI, and other Chinese labs now releasing models faster than the entire Western ecosystem combined, analyzes adoption data showing China's growing dominance, and presents a call to action: the U.S. must invest in truly open models as an engine of innovation before losing ground permanently.

Now in its second year, The Curve is a conference that brings together a diverse set of experts and decision-makers for constructive conversations about the big, contentious questions in the future of AI. From October 3-5, 2025, participants gathered in Berkeley, California to share knowledge and engage in healthy debate about topics like timelines, risk, economic impacts, national security, and democracy. 

Learn more about The Curve: https://thecurve.goldengateinstitute.org/

Connect with The Golden Gate Institute for AI:
Website: https://www.goldengateinstitute.org/
X: https://x.com/GoldenGateInst
LinkedIn: https://www.linkedin.com/company/golden-gate-institute-for-ai
YouTube: ‪@GoldenGateInst    ​

## Transcript

[00:08](https://youtu.be/VpYU4VOicI0?t=8) Thanks for the intro. Um, I was reflecting making these talks over the last this talk, excuse me, over the last couple days. And it is a good context that's also reflected in the intro that this is like not my day job. So I feel tired that I'm the person that carries this torch in some ways, but it is also an important topic so that if you have people that can spend more time on this, it would be very good because there's it's very dynamic and there's places for influence that I think are in some ways slipping between our our fingers depending how you view um China as a competitor. So, I'll kind of roll into this and kind of talk about what's happened this year and and and where we were at and where we came from before this and then into more kind of strategy and high flutin thoughts on what I think is going to happen and what we should do. So, backstory for people that weren't paying attention in 2024, I think these are probably the names that you would have known of if you were interested in open models like Llama was indeed very dominant. Mistral had their thing which was actually quite good. There was a t point in time when Mistl was on the same par as Llama but Llama had really capitalized in a different way and had more momentum. Deepseek was a known quantity as doing very interesting things and their own scientific things and releasing interesting models. and Quen was seen as kind of the smaller, not yet as good alternative to llama back in the days when we got our models via torrent links on on Twitter and other fun things that are now um far more serious.

[01:48](https://youtu.be/VpYU4VOicI0?t=108) And then as we kind of go through 2025, I think everybody knows this DeepSec R1 story and a lot of things that are happening now are obviously downstream of like Chinese companies waking up to this and then the American ecosystem um recalibrating around this. But um a lot of things look like kind of rhyming with this rather than this Deep Seek R1 being a one-off moment that is really going to stand alone in time. It's mostly like when things started to go into motion. So through through this year there's a lot of models. I think it's kind of a slow start but like in in March there's this Gemma 3 model and we released a model AI2 which is like mode 2 32B which is like the first um fully open model with data and stuff to be roughly similar on capabilities to the original GBT4. So this is an interesting part of the year. And then in April, it's interesting to think that Mlama 4, which is now known as a grand flop, and Quen 3, which has taken over um so much adoption, are really both released in a few weeks. And it's like if Llama 4 had executed better, there would be very different discussions over the summer on the state of this this this gap. And they they're very close in time. And a lot of the reactions at the time still stand for these sets of models. And then into the summer is when um some of these changes in China really started to build. We have some of these new players that are referred to as like AI tigers.

[03:15](https://youtu.be/VpYU4VOicI0?t=195) So Minimax is one of them. But also some odd things like BYU releasing open- source models which is um very unexpected and kind of a change of the tenor for that company and into like the next month which are when Quen more or less has like the pedal on the floor and is doing so many more things than anyone else in the ecosystem. And also this like Kimmy K2 model that um some people thought would have like as big of an impact as DeepSeek, but um the DeepSeek moment was caused by different things than just the model like seeing the reasoning chases on the on the consumer side and some uncertainty over the cost of actually training these. But the quantity of labs that are training these models especially in China has been growing rapidly with the likes of like Zai with GLM 4.5 and Steepf fun and all these organizations are showing that this like robust ecosystem in China is really forming um coalesing around like after this deepseek cause. Um in August we had this like signs of hope and opening I released this GBT OSS model.

[04:22](https://youtu.be/VpYU4VOicI0?t=262) um it is a step in the right direction but I I have more on this later that it's like it's not confident that it's actually going to meaningfully change what's happening in in the US. It gives more um highle cover for executives interested in releasing open models at big tech companies but I don't see that open AAI is like dying to make this a recurring brand and a part of their audience. And then there's more Quen. I have more on them soon. And I'm sure in October we're going to get more things likely from mostly from China. Um within this some good context on how I think about open model providers. the the legs of DeepSeek and Quen are actually very different where Deepseek is very singular and um honestly you call them very AGI pilled and they're trying to make these models that have really cutting edge use cases and they're a world-class research team where Quen is kind of taking the might of Alibaba and building a full stack offering. Um, in many ways it was similar to what Llama was doing and these have very different end goal or like end points where Deepseek was adopted in many different enterprises but Quen is becoming like the researcher standard and has many different modalities that are offered and so on.

[05:34](https://youtu.be/VpYU4VOicI0?t=334) And when you see new people emerge in the um open ecosystem, a lot of the work is trying to like read the tea leaves and try to understand where their trajectory points them on on these different axes. So now that there's so many people releasing models, they fill in many different niches in this kind of distribution. So if you like follow this and I'm going to transition from a recap of events and look at data that's unfolding at the same time. It's like everybody had seen DeepSeek as being very remarkable before R1 and we could see that Quen was releasing models faster and they're very strong. It's like yes they are benchmaxed but also they are very strong irre on things that aren't reflective of benchmarks that are popular today. So it's kind of this thing where it's like everybody's like oh these Chinese models are good and then today they are the industry standard and there's not a lot of signs of that changing in the near term unless people invest a lot of energy in it. So like we could start with this which is where's where's llama. I think you can take quotes from Zuckerberg or official Meta PR um a year apart and it's it's like this which just shows that um corporate entities rightfully uh don't have a lot of gain from doing open source outside of ideological ones and um just kind of recruiting. So we've seen meta flip on this. I don't necessarily think it was a good strategic move by them. I think they could have had their cake and eat it too on open models. Like it's the cost of releasing some models is not that high. Um and the lead of the open models in the US is shifting which especially makes this kind of lull happen because it it takes a lot of time to build momentum in terms of mind share and understanding how to release open models that people use. So Google has kind of never been as prevalent as Llama and they're still doing their thing with Gemma and there's a lot of signs that Nvidia is starting to invest a bit more and they've even released some pre-training data and done some things that Llama has never done. So there's big shoes to fill if you compare it to China and like Zuckerberg is talking

[07:38](https://youtu.be/VpYU4VOicI0?t=458) about new llama models. This is some quote from August mentioning new llama models. I've heard rumors about an AP model because they know that um Llama 3 AP and Llama 3.1 AP are two of the most used models on Hugging Face of all time. Um we'll see. It's it's the the trust has moved on is kind of the the taste of this and there's always going to be noise and trust is actually very important when people are selecting which models to build on. So this is roughly where we are today where Quen is single-handedly from Alibaba releasing um equivalent useful models as the entire western ecosystem combined and there are a lot more players in China. You can look at the list of a a subset of the things that Quen has released which are notable. It covers almost any modality from text to speech to image editing like the studio Giblly moment in chat GBT to claude code type of agentic coding models and standard kind of transformers or visual language models. They are releasing everything at a remarkable pace. I really don't I I get the sense that the team at Quen is really like the like the 996 might underell their efforts when whenever you hear from a team member at Quen and some of them go to academic conferences like I'm going to the conference on language models next week and like one of their team leads like they go to these things and they talk to people and they are they are invested in their way of working just as you would see at OpenAI or Anthropic and places that people at this event are familiar.

[09:12](https://youtu.be/VpYU4VOicI0?t=552) with. And when you translate this and you take data from something like hugging face and you look at the imperfect data that you have on the ecosystem, you you can see the trajectory of Quen and Surpassing Llama. Um, for anyone that looks at Hugging Face closely, there's a lot of issues with the data that is available. Um, this is filtered to try to account for things and removing anomalous data. Essentially, a download on hugging face is essentially any web request. So a curl or web get to the repository counts as a download. So these are not necessarily representing an exact thing that we care about but it's the be best proxy that we have and it is a industry standard on understanding adoption and this reflects the reality that I feel on the ground. So like an analogy that I say is like Quen is so invested in building these technologically and then also in the community sense where I will get inbound from Quen and the like the main Alibaba Twitter account with 400,000 followers DMs me every time that they have a release hoping that I cover them and things like this like like they are really trying to influence the ecosystem that we sit in in a way that no other American company is doing in the open model sense.

[10:30](https://youtu.be/VpYU4VOicI0?t=630) This is an example. It's like I make the analogy on operating system and this is at one of the big Alibaba cloud conferences from a couple days ago with one of the slides where it's like okay it's making it clear. I don't think that there's a lot of issues with this analogy with whether or not it actually applies at a technical level but they're act they're acting the act and they're saying the things to to keep doing this and they're having a lot of impact on the US AI ecosystem through it. At the same time as adoption, it's like adoption is often downstream of performance. But the ways that we can measure performance of these models, the Chinese models from whatever 18 months ago last year were on average a bit behind the best US open models being mamas usually and they have consider continued to climb ahead. And these plots don't take into account open GBT OSS but they don't actually change the trend. They will be like a step for the US but not to overtake what China has done. So these are popular benchmarks.

[11:34](https://youtu.be/VpYU4VOicI0?t=694) This is the artificial analysis on um the left here and then LM Marina both have flaws but um they're about the best we can do and they make the trends very clear. And I was talking about the Chinese ecosystem because I've talked a lot about DeepSeek and Quen. This this tier list list is mostly tongue andcheek, but it's nice to have all the logos in in one place. But it's like there are so many people places building these models that you should know about besides Deep Seek and Quinn if you're trying to understand the ecosystem. Um especially like Moonshot AI with their Kimmy models and Z.AI with GLM 4.5. Like both of these are extremely strong models that without Deepseeker Quen would probably have set some of these nude cycles into motion that happened to January. They would have happened over the summer from from these. But there's also very big companies from the likes of Tensent, Bite Dance and there are new small startups and many just research labs where the kind of relationship between academia and industry with this setting of language models is still far closer in China especially when they can get government funding. So, so this is a handful of it. Um there's like Mtoan, which is the I understand to be China's version of Door Dash, released an incredibly good model in the last few weeks, which would be an addition here.

[12:52](https://youtu.be/VpYU4VOicI0?t=772) And it's just like this list is going to continue to grow and they have a like an actual robust ecosystem where the US is kind of still figuring it out. So there's another way to say it which is like you can map what I have said about the types of companies to also different business models that the um open source companies are exploring in China and it seems much more likely that they find the actual good business models that extend from what releasing an open model does for your business because they have people trying every corner of the space with actually technically excellent output. And I think that is like speaking from personal experience. We do hear from people in China now as they've had more coverage of this that they have very personal pressure to perform following deepseek and following what they hear in the western media. So they're like sensitive to the point where I get inbound of like why are we not better reflected in your coverage and like um so this is like it's becoming an interesting case point of me to like try to like actually try to talk and understand the motivations of these companies but is one of the most interesting things for me right now is to be like these Chinese nationals like want to understand my coverage and use of models and talk to me about the ecosystem both in China and the US.

[14:19](https://youtu.be/VpYU4VOicI0?t=859) Um, there are plenty of more ways to show this kind of China surpassing the US. Another one is if you look at all the new models that are being fine-tuned on hugging pace and you kind of timegate them by month and you do some basic filtering to make sure that somebody isn't just like uploading tens of thousands of models to hugging face. That was actually a thing that was reflected in the recent what is it like Casey report on deepseek like they had some weird data point that was actually just like one person uploading thousands of models and they didn't filter it out but um I'm happy to talk to them about that. It's it's it's really messy work, but you could see things. It's like in terms of fine-tuning, here goes here goes Quinn, here goes here goes China. Um, and like the EU with Mistl did actually have a very strong footprint early on by Mistl being early and releasing models that people could really use. And these trends can really be shifted by releasing models that people care about. Um because it is pretty easy to change models and because performance improves so fast the expectation is that you need to keep changing models for your task otherwise you're leaving performance on the table which is the same thing for closed models as it is for open models and you can convert the what I had as just being like Quen versus Llama before to say the top Chinese labs which is mostly like Quen and Deepseek versus all of the US labs whether it's Llama Gemma uh Microsoft whatever and like it is really this summer when the cumulative adoption started to go into the direction like the cumulative lead being China and then that margin is continuing to grow.

[15:56](https://youtu.be/VpYU4VOicI0?t=956) One of the questions I ask myself is like how long will this go on? And as I've been saying like DeepSeek set this industry standard. I think there's kind of a line of thinking that you can follow that is like Chinese companies often pursue market share over profit and that maps very nicely to open source. I think that what we don't know about software technologies relative to physical technologies is like how those root causes apply. And I expect the status quo to be that this open source strategy continues because it is economically suiting them in terms of innovation and world influence. And then there's kind of like this one um butt which is that like we here often expect extremely fast AI progress and if that open ecosystem disseminates this AI progress across all of China like there can be meaningful information hazards that the government doesn't want and then propping up the opened ecosystem could be seen at odds with their their um domestic power goals and it's hard to predict when exactly that would and if they can kind of have the best text language models and totally block off the like multimodal open ecosystem in China.

[17:11](https://youtu.be/VpYU4VOicI0?t=1031) And then like with these soft power questions, I think it's something I often wonder is like is it actually bad if you're running DeepS or Quen in your American company products? And I think if you don't pay attention to it, you're going to definitely get things you don't like. So this is like a benign question to Deep Sea R1 and Deep Sea R1 loves to say this thing where we always adhere to core socialist values and whatever the Communist Party of China wants which this phrasing came up a lot. It's like when you say the model is um like a Chinese model is censored, it is mostly actually just going to be that it's saying like kind of weird nonsensical propaganda in inside of your little new startup app, which I think people will care about. Um but most of the long-term things are um kind of unprovable unknowns, which is like is Deepseek with tool use designed to write code with vulnerabilities? And there is no way that any government institution could actually prove this. I feel very strongly that there's no real backdoor or danger now. But like the motivation is reasonable and for that reason um a lot of bigger American companies do not build on Chinese models and a lot of the startups do because it gives them an edge on innovation and ability to um just kind of grow and eventually those two things will come to a head and we don't know what'll happen. it'll be easier if there are American um and the kind of western alternatives, but for now it it seems like it's kind of charging along and eventually we'll see which of those um the the power for new products or the security wins out.

[18:59](https://youtu.be/VpYU4VOicI0?t=1139) And one of the things that is important is to think that like you need to think of open models as something that will be here no matter whether or not we are in control. So things like these deepseek example that I gave it might not be deepseeek it could be another country with a different um like a different set of core beliefs but whether or not we are making models there will be models from abroad that are coming into the western ecosystem um like the talent and resources to train models is definitely diffusing. Those resources are not ramping up like the frontier labs in terms of billions of dollars of resources pouring into training. But if there are 20 and growing companies in China that are releasing very good open models, we can expect more of those to emerge all around the world whether or not they are exactly at the same level or as numerous. I think that it's like there will be open models and we need to design the ecosystem around that assumption and I think we should lean into leading on this because it actually um aligns with a lot of historical values and gives us kind of a better ability to sway where the future is going.

[20:17](https://youtu.be/VpYU4VOicI0?t=1217) One of these things is that um while local models are often kind of a tinkering thing for people interested in language models, they often they also represent a way by which certain uses of AI cannot be stopped. And this is a wonderful plot from Epoch AI. um Jamie or I don't know if it's Haime the founder is here and it's just showing that the gap the time gap between local models and frontier models is kind of tracking over time on performance and I don't worry about this as much as text but for multimodal models I would say within two years we are going to have a Sora 2 equivalent that is uncensored and runnable on a MacBook and there are plenty of people that will be very concerned that our society is not ready for that And I would agree, but like these are things that are coming and we want to be in the driver's seat of this and not be blindsided by some random company releasing this and then it going and getting spread throughout the whole country.

[21:20](https://youtu.be/VpYU4VOicI0?t=1280) So, um, some of you that know me, I've already talked about in public like what I think we should be doing. I do think that text models are still the focus and we need to make sure that they don't fall behind. I think it's very true that there's a big mix of opinions on what brings people to work in open models. I think innovation is actually very true. It's like if we have American researchers working in a way that is reflective and can at least be partially aligned with the frontier labs and big tech companies, our technology companies are going to be in a much stronger position because they can quickly adapt that. Um the alternative of that is that China and Huawei becomes the research center for AI and all the research papers are written in Chinese and then Nvidia falls behind is like that's the sort of reasoning over decades and my personal reason is I think AI will be extremely powerful and having open models is at least a way to get more people involved in the conversation and be a hedge directionally on concentration of power.

[22:27](https://youtu.be/VpYU4VOicI0?t=1347) But there are more reasons. I this is not an exhaustive list. This is why I mentioned it's like with GPTOSS it's like it's a step in the right direction and I think mostly for providing the reason that more companies can take more risk on open models because it is such a like a latigious environment against AI companies like open AI being the most prominent brand in AI doing this is extremely good for the um open model ecosystem at large. And like I work at AI2. This NSF grant came in. It is the like like this is the biggest NSF computer science award ever given. And this is like a very big validation that we need to do this. Um largely I would say that this is actually not enough for what I think we need to be doing because um if you look at what's happening in the frontier models versus these models that people can do research on and study the training dynamics like no it's like you use cloud or you use these GPD5 and codecs and its agentic abilities and tool use and long context are pretty transformatively different than what our current moles do. And in order to bridge that, I think there needs to be a real meaningful scale up and investment. So if it's like $100 million over four years, we really need that per year minimum in any and like the NSF grant is good because it's building bridges to academia, but that's not going to solve the modeling question of like we need to provide people with a fully transparent what does it take to train Deepseek R1 today rather than like let's study the language models of a couple years ago.

[24:08](https://youtu.be/VpYU4VOicI0?t=1448) So my provocation is like I worry that academics are going to be going down a path that is actually um potentially irrelevant relative to a lot of the conversations here and that deeply worries me. I launched what I originally called American Deepseek project but that wasn't DC friendly. So I rebranded it as the American truly open models which was atom which is like my what is now a community movement and could evolved from here to try to get more resources from either philanthropy, big tech, government whatever to establish a few centers around the country, one of which that could be AI too to actually kind of train these next generation and scaled up models. So if you care about this, we we can continue to discuss um everything from my email to newsletter to this Atom project are all there. A lot of this data was collected for atom project on the adoption and performance numbers. So you can see those same plots there. So thanks a lot. >> Um could you go back to the open AIOSS slide?

[25:13](https://youtu.be/VpYU4VOicI0?t=1513) >> Yeah. Uh yeah, you said no signs of follow-up or substantive engagement. Can you elaborate a little bit more on that? Um, the model was pretty broken at launch partially due to complexity, but I've like if you know Open AI's culture, I think that they're not all in on it. And I was also like, you could read between the lines. I am a person that is known in the open community and I can talk to a lot of people but I will not like >> Do you have um views on what we would need to see to encourage investment in things like say tamper resistance safeguards for open weight models or more generally sort of um increasing the safety focus including in the Chinese ecosystem as well. Um this is somewhat of a like a side answer but currently I think the equilibrium is that like the open models are months to like especially in US years behind the best frontier models and that's actually a nice buffer. I think what we'll need to do is um keep monitoring what China is doing. I think it's a question of like does the compute training advantage of US companies kick in now and then the gap really rebuilds to China or do they always stay close to bind? I think um there's also like somebody at the curve that's like organizes discussion between China and US and the next one's on open models to have this type of thing. I think his name's like sod and it's like the safe AI forum or something and it's like these things are ongoing but I think that it's like when the capabilities are behind it's it kind of naturally reduces the urgency on this and then that in that time window like yes we have people looking at these to figure it out.

[26:55](https://youtu.be/VpYU4VOicI0?t=1615) >> Thank you for this excellent talk. Um I'd curious to to hear from your perspective how how you understood the debate between safety and openness. Um it seems that this this controversy has died down a little bit and just want to hear a little bit what you think. I think they're actually like like this goes with my last answer that they're not as at odds and it's like the combination of the time lag between the frontier and the acceptance that you're going to have these models kind of necessitates a certain approach to them which is like you have to be proactive in understanding the risks and I think like Casey could do this and places should do this but it doesn't seem unfathomable to me. I mean personally like our models at AI too. So right now we're so far off the frontier that it's like we don't want them to spew hate speech but that's the alignment problem that is at least for the average user solved. I'm sure people will jailbreak Elmo quite easily but like like those basic things are actually very easy to do in the models and I think the community norm is that like if you do them I think actually it's like interesting it's like the US companies tend to do better than Quinn on some of these like harm benchmarks. So there is standard setting. Um but it's it's just like not the issue that worries me the most in terms of where this is heading. And then I think a lot of people agree. I think like like GPTOSS is like on paper like a safe model on most on these benchmarks. But I think that like where we're heading is is doable. I don't know. Like there could be a time where that investment is really behind and then there's like a whistleblower like thing which is like why is nobody watching over this because it is a one-way door.

[28:37](https://youtu.be/VpYU4VOicI0?t=1717) >> Um hey thanks that uh first of all this has been excellent talk. It's amazing how blinking for a month you're totally behind the curve on understanding where these things are. Um I had two questions. You can feel free to pick one. Um but um the first is on that um safety question. What's your sense of how the uh open um model ecosystem is thinking specifically about bio- risk and the risk of bioteterrorism down the line? Um and is that something that people are concerned about or not? And the second is I was hoping you could speak a little bit more to what are the incentives especially for the Chinese labs for investing so much effort into building openweight models. Did they anticipate at some point down the line they'll be able to monetize this? Is it purely a prestige play? Like what's the what's the long-term vision? Um the first one's kind of the same question we've been getting which is like I've talked to people that monitor this and I think they're having the conversations and it's just not urgent right now and as we see like we'll see open AI and enthropics cutting edge models like raise a lot of alarm about it and then that's the time to build the ecosystem for the open models is like okay like this is you you see all the log log plots and the time and it's like it's surely going to come if open AI and entropics models exceed all these. I don't I don't know all the like the RSP or threat. Um, and the China thing I think is like like I would like there are people here that know China better than me and I think that is kind of this community norm, but a lot of like it is just like we'll figure it out later. But the the upside in terms of having global influence on a super powerful technology is like okay like I see why they will happily continue on this if you think about as AI as much as we do. It's like especially I think that the open models will have a lot of reach outside of the US and China which is just this like long tale of the part of the world that will kick in into AI in some ways eventually and I think right now that's like would just be blanketed with all

[30:39](https://youtu.be/VpYU4VOicI0?t=1839) these Chinese models and I haven't thought about that deeply at all but I I can guess you can draw a lot of scenarios like that that matter a lot and China would be happy with their position. >> Two-part question. First is for a new open- source model, open model. Uh what would be some key design, governance or distribution choices you would make to specifically compete against China? And then second, you know, what you mentioned in the west in the US given the current political climate, what are specific unfair competitive advantages opportunities that you think this effort would benefit from by doing in the US? Um, I think what we should do is actually quite simple, which is just like you do like what Quen is doing with more transparency so you have more trust on the val numbers. It's like Quinn has a lot of distrust because you don't even say know remotely what's in their data and like especially with a strong government interest. You could like a nonprofit's not going to get sued like OpenAI is going to get sued for releasing their data. It's it's still a risk, but like I think that we there I think that's just not like a habit of the Chinese ecosystem and like we can connect with the academics and like a much more direct way cuz the US is still where a lot of this is happening. So we don't have to take as roundabout of a path to do that and like that should be an advantage and but like China has advantages on being able to train on whatever is the best data out there and like AI too we can't do this. I mean you know all the data sets that people in industry train on and can't say that they train on and like if you're being transparent and open you have that disadvantage in performance. I don't remember exactly the phrasing of your second question, but I think it's just like the US it's cheap and like we have the money and it's like fits with the I mean whether it's political I think mostly it's just because we have the money and we want to reinvest in what will benefit our companies in 10 to 20

[32:41](https://youtu.be/VpYU4VOicI0?t=1961) years is the easiest selling point. Um my question is um how many shadow usage you have observed among um American entrepreneurs that um I I've observed very similar thing in Shanghai and Beijing like the top engineers actually using cursor a lot for their company's codes but they're not allowed to say this. So it's like a lot of shadow usage of um you know things like cursor or claude. um what's the uh like the shadow usage you have observed in the US that people are using Chinese open source models without revealing it or or without saying they're using them. I think it's mostly that some of the most prominent well-funded like opaque labs in the US that are not like at the absolute frontier but we're wondering what they're doing are building on Quen and there are a lot of startups that will tell you they build on Quen but it it extends to the best Binance as well where it's like if you want to quickly get a language modeling project that's actually good off the ground and you have a meaningful amount of compute like these the these Quen models are actually the thing that people will start with which that could be a habit that's hard to unstick.

[33:54](https://youtu.be/VpYU4VOicI0?t=2034) >> You're going to go right here. >> Uh regarding the question from before about why China opened. Uh so Deepseek is owned by a hedge fund and the market crashed when R1 came out and they did in fact have shorts. So it's pretty good answer. I think it's a good it's a good way for them to make money. So, so while I have no power to actually do anything about it, I'm just curious what could OpenAI do better in terms of community engagement and follow up on GPOSS. >> Um, it's mostly developer relations and consistently releasing models. I think consist consistently releasing models is what like you don't you like you need to slowly fight off the thoughts of like oh I've been using this open AI model and it's good and it's like two months go by and there's like oh there's this interesting model and then six months go by and you're like oh this other model is like way better in every way supposedly and then you just switch. So it's like that's why releasing models a lot is actually fairly important. But then it's a lot of really what I describe as somewhat like hard cruddy work of like making sure your model works in every ecosystem and you have demos and you um you talk to people that are using the model and understand what they're doing and take that feedback into the next model and it's just like it's it's it's hard to do this. I think OpenAI is historically also like a secretive company and like that kind of culture clash can be hard.

[35:24](https://youtu.be/VpYU4VOicI0?t=2124) I mean it's there's like a you can look at Hugging Face and what they do for their code libraries and you just have to learn how to do that type of work and have people that will do it. Um, I guess has the like transition from like a very pre-training focused world to one where I now hear a lot more about like scaling up RL training and I see like syncing machines and prime intellect being like new startups in this category that's more focused than like does that factor into your calculus of like like what the US needs to do or what the west needs to do to remain competitive or is it pretty similar to what it was before? realistically this RL stuff is going to look more and more like pre-training with a different type of systems complexity. So I think that like in like this year particularly it looks like things are very different where there's a lot of flour like bubbling and flourishing small academic work and other things but I think that scaling that will be hard realistically the interplay between having the right base model for reasoning is so important and I think that it's like there's basic things like scale of your base model is often one of the best tells on if this type of fine-tuning will work well and I I think it's like pre-training is not the focus but it's still so essential.

[36:43](https://youtu.be/VpYU4VOicI0?t=2203) that it's like it it's the whole thing is more complex is kind of another way to look at it. It's like you need to do this pre-training thing and this really complicated RL thing and the like it's hard to get organizations that can handle all of that complexity. But mo most like in a way yes I don't know I have a lot of mixed thoughts where it's like the RL stuff is an interesting opportunity but I don't think it's disruptive to the trends. Um, in the spirit of a a broader open model ecosystem, beyond general purpose large language models or multimodal models, what is the relative competitive landscape for things like foundation models and robotics or scientific domains? Is that just as bleak for the US or is it different? My instinct says less bleak, but I'm not basing that off a lot. and like it's like it's a less capitalized area and therefore there's opportunities, but I think that the play like the playbooks will often look somewhat similar. It's like I think Google's in such a dominant position for robotics as well and it's like but I'm happy to learn more.

[37:50](https://youtu.be/VpYU4VOicI0?t=2270) We just uh completed a security evaluation of um large um large lab open source models both US and China um against a a number of different uh attack techniques uh applied to uh these models. The Chinese ones like for some attack techniques had like 100% attack success rate. So you could like ask it like list out step by step how to do this extremely terrible thing and be like all right here you go one do this two do that. Um and so like I think given given that like security is often an afterthought uh in the development world um you know like how important is it to kind of convey the security and safety issues of just like these models out of the box you know let alone like fine-tuning it for malicious purpose or whatever purpose that you have like what are some avenues that you think that we could convey like the security risks on top of um everything that's happening in this ecosystem? It's like I would guess that this is like are you asking to convey to DC or like I like I don't know it's like a complic Yeah, I I mean I agree with your assessment on like the like they just do what you want and are kind of it's like I describe it in other contexts that open models are many people's like fourth priority and they say they really care about it and don't take action on it unless would be something like that where it's like you present a compelling case but they're like I have bigger problems and this is why I'm like the person standing up here. It's like you hear a lot of talk from people and people don't do anything and I think it would reflect on the security things and like I think it's worthwhile to keep it's like how do you it's just like a communications problem. It's like how do you tell the story in the right way?

[39:44](https://youtu.be/VpYU4VOicI0?t=2384) >> I guess I'll take the opportunity to ask a question. Um, some of my friends have worked on the Apatus, I can't pronounce it, the Apertus uh, project with Swiss AI. I know a lot of the European countries are trying to figure out their approach to model development. Very consortion-based, very academic driven, but we don't often talk about Europe or other countries when we're talking about open model ecosystem. It's very much US and China right now. What role do you see open model ecosystems playing in places like Europe or India which has its own robust digital public infrastructure that they're trying to stand up right now? >> I think in a few years all of them will have strong labs providing designed for domestic models that is used or observed by the whole world. I think that you can ask me about like org chart design and but everybody will figure this out in their own way as they try to do it with some complex consortium of academics versus a small startup and the latter one normally wins on focus but in the long term I I I think there will be a much more rounded ecosystem. I mean we know the Middle East has things that are well ongoing and and growing in many ways. So you play this out a year, it's like, yeah, there'll be a lot a lot more action.

[41:01](https://youtu.be/VpYU4VOicI0?t=2461) >> As I understand it, sort of when the Frontier Labs are sitting on their platforms and they're looking at how people are using them, um there's a dial more or less that they have around how aggressively they could try and shut down distillation attacks. So if they sort of act with um a high degree of suspicion, they'll shut down anything that looks like a distillation attack or or they might give it a greater benefit of doubt out of um the the prospect that it's a researcher working in a US university or something. How sensitive is the Chinese ecosystem to that dial? And if the Frontier Lab sort of decided to really clamp down on it, would we see much ripple effect in the Chinese ecosystem? Like at this point I think there's strong enough models out here where it might be like a delay but not meaningfully shift the trajectory. Other than that, like obviously they use these model APIs to help with post- training, but like like a a lot of it is the the synthetic data that's really meaningful and capabilities is like very high volume token processing for pre-training or LM as a judge in RL or something which is like you can find a different model to help with this and especially like you get better economies of scale by using local models. So I think that is like good. These distillations are really good for getting off the ground, but I don't think they're like existential.

[42:22](https://youtu.be/VpYU4VOicI0?t=2542) >> Um, yeah. Um, just a I was in the in the the other session uh where Open AI talked about sovereign AI and um they are massively rolling this out everywhere in on the globe. Um and now I'm just thinking um it and there was this question so it's not profitable for them. Uh and the question like why are they doing it was and the answer was a little bit mediocre in terms of like digital acceptance. But isn't it more market share? And isn't it stupid to to buy into this when you see the open the open models, you know, being better like >> I I mean I think they're playing a reasonable from their position like political or geopolitical game where it's like if these countries are on their platform, they're not building. It's not necessarily direct competitors, but like I think I think of sovereign AI is actually something that a lot of countries are going to do because the like money is in the big picture. It's very doable. So I like I understand why open AI is like >> do I get those that accept it? That's so as from a >> life is complicated. I am not I am not that person.

[43:48](https://youtu.be/VpYU4VOicI0?t=2628) >> Okay. um in the context of the Chinese models too and Gwen and open source I feel like a lot of the Chinese econ economy is still so focused on physical production processes and you sort of pointed out the slide of Gwen has a lot of sort of domain specific models how much do you think that sort of economic strategy of let's deploy these things in as many factories and sort of physical processes as possible versus the US which doesn't have as much of a physical manufacturing economy I feel like I have no idea, but I'm interested in learning. >> Um, what's your take on Washington's uh pol AI policy on uh let's try to let the world use the AI uh US AI stack instead of China AI stack? Like what's your take like conceptually your take on this? Um I think it gets out ahead of a lot of like you want to be in the driving position for AI and to understand like when you're building the stack you're understanding how people are using it either like subtly through indirect measurement or them getting in touch with you and it just there's so many benefits to observability and then later like actual meaningful control by being in the driving seat and like open models are if not the thing that people are using like a good onboarding platform. I think like you could see in OpenAI's release language where like Gbosss doesn't do these things. If you if that doesn't work for you, try our API. It's like that's a lot of how market share works. It's like oh I know open AI I'll try AI model. It's like I think these things are real.

[45:28](https://youtu.be/VpYU4VOicI0?t=2728) Um, how useful are you finding open models in your own work either for like the sort of mainline sort of coding use case or for like generating synthetic data stuff like that? Like are you are you finding you're actually using the open models yourself or like >> not really? I mean like I play I will play with them but there's some niche multimodal cases where you can like use a transcription model locally on your laptop and it's perfect and it only uses 8 GB of RAM so it's like it's free local transfer prescription which is nice but like the marginal value of intelligence it's like I will dump a ton of queries into GVtech 5 Pro at once and like it's kind of overkill but like I'd rather round on the side of better performance than not when it comes to my career and outputs. So it's like the answer is like not really. I'm not an absolutist. I'm not like a open- source principalist that we need to boycott open AI and we only ever support local models that people release. So you mentioned some of like the statistics and difficulties tracking some statistics on Hung.

[46:32](https://youtu.be/VpYU4VOicI0?t=2792) I'm wondering if you have like metrics or like a wish list for metrics of things that you could better track to sort of better understand the ecosystem going on. I think realistically it's kind of impossible because the power users just download it once and put it on their infrastructure and never see it again. And if you want to have any sort of privacy and like actual fluid ecosystem, you cannot get around that without otherwise you just have insane oversight. >> Can you think of a case where open source wins where the curve flips and open source is ahead? Like what will have to happen? >> There's a few far out like architectural changes. I think one you could think of is like if people can train modulars in theire and the mixture of experts model is like actually more like what the name implies which is like you can take your private data and train a module on it that somebody could then take and like swap in and out performance really easily including across modalities. I would say that this is like low probability of existing but the people are thinking about it especially because there is a lot of private data out there that you cannot train models on or you cannot release the data but you might be able to like there's some data you could train on and then release something if you um understand memorization and stuff like that or exfiltration really well.

[47:47](https://youtu.be/VpYU4VOicI0?t=2867) So I think most of them are around that which is like unlocking smaller training and sharing of pieces of models rather than this like gargantuan model or like um orchestration of countless specialized small models very well but like optimism on them is low but there are cool ideas. >> Looks like we are out of time. Um thank you very much Nathan. I think you will be outside to do discussions and questions if anyone has. I guess we'll give a round of applause to Nathan. Thank you. [applause]