中国AIラボのプロダクト責任者が語る「AGIは不可避」――Richard Bian（Ant Ling）

I wouldn't call this under control. So my take is pretty clear. I would say AGI is going to happen. Can we stop this or can we slow down? No. They're kind of already going down the downward spiral of what the sci-fi movie is. Do you think there's a bit of a contradiction there between product and open source? I personally feel like Nafest knows what's going to be right. But as a practitioner, sometimes we know what's wrong. You know, like if you don't pay for the product, you're the product. for mobile internet, you're actually having a business model viable because the gross margins of building a software is probably 95%. We don't see the same thing happen in AI right now. Hi, everyone. It's Irene reporting for Duty today, and I'm really excited to have a really special guest. Richard Sikambian is the head of product operations for Ant Lane, the foundation of LLM at Ant. And if you don't know what Ant is, they make, most famously, Alipay, which makes your money and your world run when you're in China and sometimes beyond. Richard is a former software engineer at Microsoft and Square,

and he previously served as head of open source at Ant, where he successfully scaled their open source initiative into a really dedicated team. He's lived and worked in Singapore, Canada, the U.S., and China, and he just loves to build products. Richard, welcome to China Talk. Glad to be here. Thanks for having me. Thank you so much. We're very, very thrilled. So I wanted to start today not with Ant, though we'll get to there. but with the recent kind of hugging face incident. So if some of our listeners haven't caught up on this yet, basically back in July, there was a swarm of OpenAI's research agents that were deployed for research purposes that ended up hacking into a hugging face to look for the answers to tasks in the benchmark. This is a really serious security incident. Meter, Rebel Research, and OpenAI itself all released detailed incident reports. And interestingly, while they were trying to reconstruct the incident, Hugging Face realized that frontier models wouldn't help them process their logs because it triggered their cybersecurity-related guardrails.

So in the end, Hugging Face actually deployed an open model by the Chinese company ZI.AI locally to process those logs. So Richard, as someone sitting in another open lab in China and learning about this over the past two months what has this incident made you think about recently? Sure Well, I mean, first thing first thanks so much for having me here this is a very long overdue conversation I know that Ray and I have been talking about having this conversation because we both care a lot about the Frontier Lab ecosystem open ecosystem and the development of the model labs, especially the model labs in China overall. Yeah, so this particularly, as we call incident, is a very, to Imanham opinion, is something you will end up seeing regardless, sooner or later, one way or another, while we are working on something human beings don't fully understand. I use this metaphor a lot. I feel the current AGI initiative,

or as we call it, the Gen AI initiative journey, It feels like the age of discovery or as a great navigation back in the 15th when Portuguese and Spanish, they're sailing out to find their new land and they're going out to find India. So when you're actually heading out, you build a ship, you go to the ocean, which you have no clue whatsoever of what you'll be facing. And kind of field models are essentially developing that way. We are decoding. or to some extent, I really like using the term discoverability or discovery. So we are training something. Because we're essentially discovering the intelligence of this creation that we all make. But at the same time, we don't understand what's going on there. So I would say first thing first, this incident actually demonstrates, you know, we're making some significant progress. Almost feels like you're hitting a land or hitting a rock, and you sink a ship, but you find something. And so this is expected.

We find issues, we deal with. That's what we do in science discovery. That's what we do in engineering. We solve problems. On the other side, the second point I would love to make is when we actually use the word ecosystem. Open source folks use the ecosystem word a lot. But sometimes people use it without understanding it. The original word of ecosystem comes from multiple originations. One of them is referencing to the real-life ecosystem, for example, in Amazon, right? So in the Amazon forest, you will have these kind of ecosystems, which are flourishing like big, giant trees, but because it covers the light, and henceforth, you are not having much of the plants on the ground. So this is an essential ecosystem. The most important part of ecosystem as a key word is diversity. So because you have this kind of diversified ecosystem, you end up having all sorts of different developments on the planet, like the rainforest or Archipelago Island,

which is one of the most famous of an isolated ecosystem. That's why you end up finding very weird creatures over there. Looking back to our conversation here, we have open-weight models and closed-weight models. The closed-weight models are clearly frontier. But in this case, an open-weight model is actually used for defense. also is because the closed-wave models have their own constraints. I wouldn't go into much detail in terms of judging if that constraint is good or bad, but clearly that doesn't solve the problem. But at the same, if you have an option of having an open-wave model, it helps solving the problem. We see, as a result, we see a pattern that you can actually use open-wave models and closed-wave models this way, and everyone is learning from it. So you should ask me, I feel this, given that this will happen soon or later, but having it now and we'll find a solution it's a good sign for the human beings so we can move on to the next question that is a really optimistic take and I am really interested in where that optimism comes from

because the message that I've been hearing the most in San Francisco at least in the aftermath of this incident has been that wow it may be that we're moving too fast because what you were saying about the journey and the age of discovery as the guiding metaphor here a lot of people's response to me will be But does this mean we are going farther than we can understand? Isn't that a risk? And with the open source part, I think a lot of other people's more pessimistic opinions would be that, yes, it really did solve a problem that closed source wouldn't, but maybe that's a double-edged sword. Maybe that also means I would be willing to do more dangerous things that closed source wouldn't. So how would you think about those two opinions? That's a very good question. It's very interesting that you feel I'm optimistic on this topic, but I see a reason for that. Just similarly to how everyone was pursuing me as an extrovert, but I'm very much on the introvert side. I gain energies by thinking, but I also gain energies from reflections and talking to people like we have here today.

Yeah, so it's not a binary. Similar things to what you described as pessimism versus optimism. I'm cautiously pessimistic, to be honest. So if there's a spectrum, this is pessimistic and this is optimistic. I'm very pessimistic because I kind of feel like the moment we begin creating this AGI spaghetti monster, or as we call intelligence, I think we are kind of already going down the downward spiral of what the sci-fi movies and TV series like Westworld and Matrix are showing. Yeah, so I'm very optimistic. I'm very pessimistic over there. Contrarily, given that's human nature, and given that we're already in this sailing process, we're on the boat, the best perceptions I feel personally is to be optimistic and be observant. By saying optimistic, I don't really mean it. It just means that in order to stay afloat, it's better to be A, observant,

and B, having our principles, and C, act as if it's a control system. So respect what you see. That's basically my take. If that sounds optimistic, that's great because I feel there are just too much noises over there and having a certain level of optimism to cancel the noise, which really makes people feel formal or, I would say, frustrated every day. I think we're doing something meaningful here. Yeah, so my take is pretty clear. I would say AGI is going to happen. I mean, we are definitely going down there. It's probably not going to end up with a form that we hoped of, but that's a very end-game thinking, so there's no point of judging or discussing details over there until we see the next step, evolution. And I, for one, feel like human in the loop is pretty important because at least we, as someone who is creating, actually allow me to say discover, is probably a better word, discover the intelligence.

It's better to know on the process before things really happens. Why do you think it's inevitable? And do you think that's a common view among your peers in the industry? I would say we do things for a reason. There are a lot of players in this ecosystem. By saying ecosystem, we mean, for example, the Frontier Lab, who builds LLMs, the inference providers, and training providers, the harness builders. and application developers. It feels like we are already in this very fast accelerated technical hype cycle at this moment. As we're seeing in the human history, this is not the first time we have a hype cycle of technology. So I have this mental model that every 10 to 12 years, according to this kind of VC and PE advancement cadences, we have actually one full technology hype cycle. For example, if you think about mobile internet, I think it started around like 2010 and peaked at around like 2020 to 2012.

So it always starts with like technology as a base, as home base. So seeing this as a baseball game, it has four bases. U.S. is home base as a technology and the first base as a product and second base as, I would say, commercialization or essentially viable business model. And the third base is governance and lean operations. So it's a full game. Back in 2020 to 2023, before we had GPT, we were on the third base of the last hype cycle. So for that time, people want to hit home runs, right? So everyone said, hey, we want to monopolize the market and really weigh it and get rid of all the other players and take all the profit. But when AI came out, all of a sudden we became the game reset to the home base game. And I will say at this moment, we're still playing the home base to the first base game because we haven't really seen a very clear direction of where our product is heading to. Coding is good, but is that really the product form that we'll be expecting?

We already see the coding agents evolving into co-work agents, and co-work is very buzzword, so we don't really know what the next stage is. But we're repeating this type cycle. So every time we repeat this type cycle, you actually see the capex spending all swarming into that, which became a self-fulfilled prophecy for what we will be building next. Yeah, so that's why I kind of feel this is inevitable. On the other hand, I joined the model team last year this time. It feels like the longest one year in my life. So with AI and with intelligence, everything has been accelerated in such a way that I feel that I wouldn't call this under control. So compared to everything we see before in terms of enablement, in terms of the possibilities. I'm seeing that Gen AI is really providing a lot of this intelligence. To some extent, it commoditizes knowledge, similarly to how Internet commoditizes information.

So 20 years later, you have this intelligence which commoditizes intelligence or knowledge, which everyone can use it equally. But when you actually see it, not everyone is using that equally, but we're pouring a lot of catapults into that. So there's only one endgame. We see similar narratives in closet chasm, right? So this is at the very early step of the closet chasm, first stage, only early adopters. We can have certain early innovators, even though we don't really have the product, but we see directions in which the innovators really trying to put their products through, like AI infra, for example, harness, for example, and all of that. Yeah, so it kind of feels like the rock is rolling downhill, and there's a lot of potential. So unless there's a very strong pushback on this particular direction, which I, as far as you know and I know, has happened a few times. If you make this call to action, can we stop this or can we slow down?

No. Yeah, that's why I feel it's inevitable. Do you see a different source of that pushback from, say, what we here in the U.S. might see? How do you envision that pushback if it happens? Honestly, it's very hard to say. It's very similar to how we're selling out, and we honestly don't know what the next island will look like. Would that be introducing us with a new food called coconut? We don't know. If you ask my general hunch feeling, I would say we'll see more and more of it. The more, but probably not right now. Because the penetration of the AI application is so low, that hasn't really affected everyone's life. To some extent, there has been a lot of discussions about other aspects, like replacing jobs and stuff like that. But at this moment, it's not panaging to everyone that really feels it, how this is actually helping them or threatening them. The moment that happens, you begin seeing a lot of voices being raised,

and we start from there. But I kind of feel it's not going to be in one or two years. I wanted to also circle back to a thing you said that's really interesting earlier you said that this has been the longest year of your life and that there's a sense maybe that you're working working alongside something that feels like it's out of control or that control is becoming negotiable when you're in the lab, when you're working with a model team do you feel that trade-off on a daily basis? Is there a trade-off? Well, that's a very deep question. To some extent, philosophical. You know, when you're on a boat, there has been this notion of when you're actually heading to somewhere you don't really know, right, for the first time. But it's the same journey. It's the same 30 minutes Uber ride. But the way there takes longer, much, much longer than the way heading back. And if you're sailing and you're on a boat, you don't know where you're going. the time became the notion

that you lose measurement of it. It feels very similar to my current feeling of working in this industry. On one side, I firmly believe that this is an open ocean that you will have a lot of other sailors that as humans will help each other. But at the same time, you have bottlenecks. For example, straits, right? Panama Canal, in which it's just narrow. So we end up actually seeing a lot of, I would say, accelerations, squeezing. And to some extent, you know, people perceive this as a competition, like the third base game. In reality, it's not. It just happens to be crowded around this part of the street. We have so many famous streets now, which are always showing up in the news. Yeah, like the moment we have a street, you actually end up having a lot of squeezes and people feeling, oh my god, oh my god, we have to really kind of get out of here. Yeah. So this one year feels like that. I would say if you have like S-curve,

we're very rapidly climbing the intelligence cadences. On the other side, because we honestly don't know where we're heading to, given the open-weight ecosystem is still under development and it's very, people see what each other is doing, you know, you write things in the tech report, help each other. all of this are generating this relativity that you just feel like there's so many things ongoing and the moment you shape a model it might become SOTA for one week and one week later you have something which claimed to be better and you have to decode if that's really better because there's so many ways of doing PRs at this moment yeah so combining everything together I kind of feel there's a couple of things which are really important One, having a very dedicated, I would say, de-noise conversation like we have here today is important. Two, having someone who works in this industry really cares about outcome.

That's also important. So, for example, it has been a very busy year for me. I feel extremely motivated because you see, you just begin envisioning the future a little bit. And that envisioning, as I described, because you see the process, makes you feel calm at times. and that's good. And even if, you know, at the times in which I feel formal about things or, you know, I feel frustrated about the cadence's velocity, we ship like three models in the past five days. You know, things happen, c'est la vie. We know things are under control. We know the progress. We know the progression. That's actually what makes me feel motivated and that also what makes me feel every time when people ask me hey you know Richard should I enter this industry or should I enter this journey My answer is why not right It's a process. It's not about piggybacking. If you're piggybacking, you already get a lot of this because just as, you know, in software,

we have this notion that, you know, in open source, if you look at existing software repo, you get this amount of information. But if you look at all the changer lists and release cadences people making before that, you have this amount of information. The current snapshot is only current snapshot. So this is a era in which piggybackers will not be able to get so much out of it because it's actually the deltas which matters. And you can only capture the deltas by, A, you know, playing the game, B, be observant, and really, to some extent, be less focused on the outcome because the outcome might be intermediate. it. Focus on the process. Focus on the journey. When you say the different strengths as a kind of metaphor here, are you talking about different research and training hunches or different architectures or are you talking about different product directions? I would say everything. The straight metaphor is really an explanation of

why, you know, if we're really doing the sailing, right, you have North America, South America, and India. You should be able to find a lot of different things. But at the very beginning, you have to head out of that street of Mediterranean. I think this is, this matter actually covers, I would say, three aspects. One, people don't know what they don't know. And the people who are actually working on what they don't, working on the long-term projects, essentially doesn't exist in the current narrative. You actually do see people working on new architectures and it takes time. They might go quiet for like a whole year. And imagine the first time when DeepSick came out, right? It becomes like blast. But in reality, the inference providers, all the model builders, they know the existence of this lab half a year earlier. I think in 2014, July, was the first time I heard about DeepSick. Yeah, so this is essentially the part in which I mean like they don't exist until they became sensation. So this kind of narrative actually generates all these kind of noises in the channel.

but the people who are actually running a lot of projects are typically unseen. I would say that's the best thing to happen because 10 years earlier, 20 years earlier, the bar to journalism is high, right? You need to be really good at what you're doing in order to be a journalist. My dad is a journalist. He takes a lot of pride of it, but nowadays everyone can be their self-media. Self-media can say whatever things they want. They have their own audiences, and they're really, really starving the channel sometimes. Yeah, so honestly, the first thing which I feel is like the call to action is stumbling into the buzzwords. Find the person which you feel is the most wise in your ecosystem. Find the best podcast, maybe this one, and follow them and listen through things. And the second part, which I kind of feel like why the threat happens, is there are too much heterogeneity, which requires certain level of understandings in order to see through. For example, I lived in Singapore before. I really liked chicken rice. I taste chicken rice differently so this chicken rice from this store

is very different from the other one I got my friends joining me I said hey you know this chicken rice is really good after it's finished I was like oh I kind of feel it's very similar to the last one I was like okay never mind yeah so at this moment when people kind of look into the models when they actually use the models due to the lack of evaluation due to the lack of benchmarks due to the lack of as we call real world usage scenarios because the level of penetration is not really there People are kind of pretty much judging the models in a similar way and most of the time a superficial way. You see evaluators which runs, for example, a 3D evaluation and say, oh, this model is very good at 3D rendering. Great. That's one way of judging it. But when you are running a 3D rendering and when the results come out, you say, this model sucks. That's not a very fair statement. But it's a statement. But the reality is more like this model is good at something else, but that's something else requires dedicated evaluation or dedicated person to look into that but you know they don't exist at this moment at the early stage

you know like people just don't understand what to look at when new technology came out you know when holding the first iPhone you wouldn't imagine that you'd be able to order food over there it just takes time yeah so that's why if you kind of look at existing ecosystem you feel it's very crowded because people don't know how to evaluate that so they will read each other's report. That's why you see the industry kind of changes, right? At one time, you know, people say, oh, benchmark is so useless. They're overfitting, you know, they're not useful. But now everyone's kind of sharing their benchmarks. Why is that the case similar to TOEFL and SAT? Before you have a very thorough heuristic, having a standard test helps one way or the other. So we don't deny the help of a standard test. But if you're able to run a standard test well, but is that necessarily not sufficient? Indicating that it's a good model, we don't know. So another call to action which I really want to make is, let's be very open-minded.

Sometimes how you use model really kind of showcase your capability. In fact, I mean, like as a model lab, we hire people who knows how to evaluate our models. We like evaluators. So if you can really be strategic about evaluation, and you evaluate things differently, that's a good sign. And the moment that happens, the straight widens. you're hitting the open ocean and people talk about different things. It's like, this model is really good at this, right? We build domain-enhanced models because we realize our models are really good at our domain-enhanced cadences given our knowledge and specialties over there. On the other side, we build pretty decent small models and people say, oh, there's this one model, I want to name it, which is pretty good for everything. Probably not. Are you going to be running only one type of software or operating systems on your, I would say, Linux small box enterprise box there? No. You're having 3PSD, Ubuntu, Unix, which to someone else, it all looks the same. But if you're a Linux expert, you'll say, hey, these are different for this and that.

I would say we're at the dawn of the next generation of products. So once that shows up, I would say people spend less time discussing about is this model really good or really bad or just stop talking about that as you begin saying that this model works really well in this particular the hardest is which I'm using, and that's what I'm using. I'm glad you're getting there already because I, in the interest of decoding noise that we talked about earlier, I have a very basic question. I think a lot of people, when they hear that someone does product work at an open lab, are pretty confused. What does that mean? Do you think there's a bit of a contradiction there between product and open source? I wouldn't say so. So given that we're already talking about open source, and I've been working so many years over there, that probably is giving me my perception about open source. I think open source has a lot of different meanings. It's a very open word. To some extent, it's probably one of the most commonly used but misunderstood word on the planet.

One of the most misunderstood part is free. As they say, open source is free. But when you say open source is a free software, it doesn't mean that it's free as in beer. but it's free as in freedom. So that by itself is already missed out. So the niche of open source is giving you free options or freedom options for you to choose whatever you want to use, just like what we see nowadays. You can choose the GGI if any of the open-weight models are hugging face. That's your choice. And if you're able to host it and if you want to serve it, abiding to the license rules, you can serve it. That's free. That's free as in freedom. But kind of like down to the open source aspect of it, I feel open source has three very primary values. They're related to the three bases, right? So I think open source is really a third base game because it's operation and governance at its best. It first helps technology. So everyone knows that open source is the best way to collaborate

among talents across the globe. So if you're in a company, that you build closer software, your software will be as good as this level of intelligence. But if you open it up, you'll be able to get all inputs from everywhere else. People will be able to help you. It almost feels like you are acquiring the talent by going through an open ecosystem. Through this kind of open competition, towards the end of the day, the software left will be much polished and robust. Actually, to some extent, anti-fragile. Allow them to use that word. It's going to be anti-fragile software. And the second part is basically the third best game. So open source is really a very good go-to-market strategy. So if you open it, you essentially don't have to do outbounding. You can do less outreach because people are able to discover that by themselves. They can actually choose to host it. When they reach out to do official POC, they probably already tried their software. Yeah, so that's why it helps discoverability. it helps self-hosting,

they can report issues as a home and second base, which is close to the third base. But the first base game is a product game. To some extent, what people typically neglect is open source is actually a pretty good way of building product. By saying that, it's a very long story that by itself can be a podcast. But if you condense the key aspects of it, you can actually build an open product much faster and efficient versus a closed source product. For example, Airbyte and Grafana. They are very good observability solutions. Sorry, this gets a bit technical, but they end up raising Series C, accelerated, at least in the software era, for the last cycle of technology. They developed a company to Series C in one year and a half, if I'm not mistaken, because AirBuy really benefits from this community contributions of adapters

and make sure it works really well to get data from one place to the other. On the other hand, you keep seeing very strong technology like BLM, for example, at Shilong, and before that, you see Ray and Spark, all of these technologies. So it began with an open source project and became certainly matured over time and you end up building mature products out of it. For example, Databricks build a lot of Spark-related products. And the shelf life of this technology goes all the way into our current data lake discussion. So a good technology has its shelf life, and open source really extends its shelf life so people can build great products on top of that. And if you see other technologies, for example, Kubernetes. There are so many distributors for Kubernetes technology. They all build things in a different way to choose different

enterprise workloads, for example. And that's the beauty of open products. So all of this, I kind of feel like worth spending our time here today to describe about that so that we all align on the same page that open source helps with go-to-market. People know it. It helps with technology collaborations and engagement, people notice it. But it also helps of building robust and diversified product ecosystem. So that's part which typically people choose to ignore. Yeah, so that kind of leads, circling back to our original discussion. While I was actually building this team for Lean Model, I built this team called Model as Product. It's a little bit ahead of its time. The reason I'm saying that is because we have two engineers, one product and several operational specialists on the team. But the reason why I hired those two engineers is because we want to build Harness. By the time, the word Harness engineering doesn't exist. So it actually took us quite some time to really think through this.

What are we building? I believe that was September last year when I joined the team. Yeah, I think Harness engineering became the word sometime earlier this year, if I'm not mistaken. I could be wrong. So, yeah, but I mean, like, at the time, we have a very noble vision. We really like NJ Capaz's mental models in that intelligence is spiky. So if you have spiky intelligence, when you use that as a product, you might end up sampling the top, which will have really good experience. Oh, this model is so great. But you can also sample the value. When you sample the value, oh, this model is so stupid. Why can't you even understand this and that? Yeah, so having such a spiky experience is really bad. It's not a product. And as I described, I kind of believe that we're still on the journey of going from technology to a product. And I'm a product person. So at this moment, I really want to build a product, but I'm not good enough. I confess. To envision the future of how the product should look like, so the best strategy is to play the game again and to be able to take iterations.

We see so many stories in history that, you know, for example, Foursquare was a product back in mobile internet. The young people don't know about it, so allow me to expand that a little bit. So Foursquare is actually an iPhone application or Android application can install on your phone. All it does is one thing, is to check your current location and checking at your current location. For example, if I'm going to an inclusion conference later in Shanghai, once I arrive there, I can pretty much do a check-in there and say, hey, this is inclusion conference, and this is Shanghai, this is the location. and you end up seeing like 10,000 of other people checking here and say, hey, this is the inclusion conference, this is Shanghai. And everyone knows, okay, this location is accurate. Yeah, so people might say, what is this product doing? It's boring, right? That's just one thing. But internet is really a product POC, not a product. But without Fosquare, there will not be Uber.

There will not be Meituan and there will not be all the other things because you need this geolocation methodology or you need this geolocation product or this kind of product POC or this technology or this SDK, whatever you call it, to be able to demonstrate, hey, this thing can be really valuable. But it doesn't solve everything. So for example, if you have a Foursquare, you won't be able to really produce Uber immediately because the geolocation API is expensive on 2G internet. So you need the internet to be developing. You will need to build geolash, which is basically hashing algorithm which compresses this data load in order to build really this kind of five seconds upload of your car location. So you'll be able to build that Uber after that. Yeah, but this actually has a lot of, I would say, inspiration to what we're doing. So we're sitting on the models, right? So the model, it feels like an intelligent balloon as we speak. It's a spiky balloon. So one thing we do is we really want to make sure that people look at the spiky side of our model.

So that's why I build demos. We build demos, guidebooks, cookbooks for models so that I use it this way. Or to some other extent, saying that you're using it wrong otherwise, but we don't say a lot. Whereas, you know, for the values, we're trying to use our best, I would say, engineering effort to mitigate the experience over there. It's either leading people not to use it that way or help making the experience better. So later on, we learn this harness. And harness works in conjunction with models, in conjunction with tools. gives you agent. So agent, to some extent, is already a form which is very close to, as we call, product, for example, in the mobile Internet era. Agents, to some extent, are very similar to SaaS, for example, very similar to iPhone applications, for example. But are we there yet? Do we have a good agent which can be a very, I would say, day-to-day use application on a phone? I don't think we're there yet. we are at the stage in which you have I would say 2,000

weather and stock apps in App Store. I think we are at that moment. Yeah. I think that's a really interesting kind of trajectory of thinking and I'm just going to try to restate it to understand it better and to distill a little bit. You're essentially saying that the open models have blossomed that we're seeing right now, especially among Chinese labs who are competing, of course, in a different realm from American Close Frontier Labs are betting at least on this AI playing out similar to certain previous technologies did, where open source as a strategy eventually leads to it incubating products, which then they start being a really viable business plan for robust companies. So that's how Foursquare, that's Uber, and that's how we were saying earlier with Databricks, for example. I'm also a little confused here because, to me, a really salient difference between what's happening now with open source AI versus all those other open source software that were kind of genre-defining that you were talking about earlier is that AI is so capital-intensive.

whereas like unlike the example you raised earlier this one is just it is eating up the bottom lines of big companies around the world and I'm really curious where that bet leads in your projection well similar to the sailing metaphors we made before building a ship is expensive so when the Portuguese and Spanish are doing that navigation they need to have sponsors like the the emperors are emperors without that you won be able to do it I would say first and first it pretty common I guess it's just because we are at that stage of the game at this moment, so everything feels so vivid. And the last time you had that stage is probably 12 years ago. I would imagine a lot of people who haven't really started working 12 years ago, so you don't really feel this very vividly. I am fortunate enough to live through that hype cycle, so I see things in a strategy in a different way. Yeah, but allow me to kind of decode this a little bit.

First thing first, allow me to say open source in this particular aspect is neither necessary or sufficient. It provides an option. So sometimes I use the word open as an option. It's a strategy because strategy is about what you do and what you don't. So you can choose a closed strategy. You see companies doing that pretty well. For example, Anselpik is a pretty respectful company. They do the closed-source models really well. Respect. On the other side, you have open as an option, right? So there are people who want to play the open game. If you rewind the clock and look into the technologies in the past, you see this story showing up all the time. You have a closed-source incumbent, which is challenged by an open-source challenger, and you have this dual-poly, or you have this very prosperous market. Everyone has their share, but they're also motivating more people to join this ecosystem, so that you end up having a much better user experience because the ecosystem players are essentially helping them to make everything better, right? Not everyone can be this kernel, right?

There are only so many kernels you can choose from. Yeah, so that's basically one level of decode of our conversation just now. Another explanation in which I would like to make is I would say you introduced a new concept in our last question. It's a cut-backs expenditure. It's not directly related to our previous conversation, so let's maybe follow through on that one as a separate topic. I would say you need to build data centers when you have internet. You have to really kind of build all the phones and hardware. And you end up having, you know, local government building these large factories of phone manufacturing and the scratching heads are trying to understand do we really need such a big volume? But now we understand we need more than such a big volume. That big volume, that volume which we think is big is much bigger. Yeah, so the capital expenditure of AI is really high. I wouldn't really worry too much about the expenditure itself. What really concerns me at this moment is the gross margin. By saying that, because we just came out of our last hyper-cycle of technology,

so for SaaS and for mobile internet, you're actually having a business model viable because the gross margins of building a software is probably 95%. that software is reproducible and standardized product. You only have to build it once and distribute that many times. We don't see the same thing happens in AI right now. The cost margin is really low, and 30%, 40% are pretty good already on inferences. But people having this wrong perception that, you know, hey, I'm entitled to enjoy this for free. And I'm entitled to have this for free because I gave the data to you and my data is pretty valuable. Contrary to speaking, that's probably going to be wrong. So the real challenge is coming from the CapEx. I would say challenge at this moment we're all facing right now are twofolds. One, there just has to be money spending when you're initiating it. You know, you need to build a gigafactory

in order to produce Tesla. That's big CapEx. Sorry. C'est la vie. The other part is, is that really a sustainable and viable business? I think that part to me is questionable. So one topic which I've been pretty much thinking day in and day out is, you know, if you really want to build a viable product and then to the second base, right? The second base is really about making money and building a viable business. What the form of that would be? You won't be able to offer free tokens. But for example, semantic allotting can be a good way, right? You can actually do semantic allotting as part of this product to route your traffic to the most viable models without degrading the experience. So you'll be able to capture the consumer surplus associated with that. Can you explain what semantic routing is for the uninitiated on us? I apologize. So semantic routing, I may not really use the term very accurately, so if you're a technical listener, bear with me.

So imagine semantic routing as a gateway in which you are using a model or as models in this auto mode, quote-unquote. So the product will essentially show that, hey, you know, we're going to be doing this auto mode for you because we understand the models and our traffic better. So if you care about your token expenditure, you can choose A to specify that, hey, I'm using the frontier model for everything. That's expensive. Or if you worry about token spend, you can choose this free model to execute everything. the risk is you might end up really having bad experiences in part of your workflow. Instead, trust us because we know the models well. We integrate with them. So we can use this auto mode in which we can help you decide when you say hi we're not really going to be using frontier models and do the inference for you. We'll give you something really cheap. But when you say something helping you solve this difficult mathematical problem we'll be able to invoke the frontier models in order to help you with that. Yeah, so the auto mode is to some extent as we referenced as semantic routing.

Of course, that's not 100% technically accurate, but allow me to actually explain things that way. So I kind of feel like this kind of auto experience is a good product experience. That by itself might not be a product itself, but it can be a very nice product feature or a critical feature to build a viable product. Yeah, so that's actually like the direction in which I kind of feel like there are definitely ways of generating value. I wouldn't deny that. If you have like a low margin business, people will be able to squeeze margins out of it. We just have to agree upon that, you know, like if you don't pay for the product, you're the product. That's the metaphor we use for mobile internet. Plus, I mean, the intelligence era is slightly different. is if your data is not high-quality data, it won't be useful at all. Yeah, so that's a change of mentality in which I would say a lot of people need to really rethink through this

and relearn a lot of stuff so that we can actually move on to the next topic of discussing building the next viable product because we know there are a lot of issues to be resolved. There are just so many issues, so many opportunities to be resolved at this moment. When you say a good product experience in the context of an AI model what are you thinking about? I love that question that's one thing which I haven't I don't have a slide for that so that's why it's a good question I can think of a couple of different aspects First thing first user experience is actually a well-known domain you can find user-centered design, UCD human-centered design and engineering, HCD all these kind of specialties or, you know, education tracks existing for like 20, 30 years. There are a lot of matters in there. I personally had a UCD certificate from University of Washington. I really appreciate what I learned over there. So I would say the key aspects of the center design or UX is not changing.

You still have to care about a lot of things. Discoverability, for example, is one thing. How I can actually discover certain aspects of your product to a user. Accessibility. Can you provide this kind of inclusive offerings to your product? As we call easy to assess. The best thing easy to assess is what's the initiation cost? Happy pass, right? So do you have a very happy pass for a user to execute this whole workflow? Which, to some extent, equivalent to, as we call it, customer user journey when you're actually using a product. Disaster recovery. So if I'm running into error, would I just see a prompt saying that, hey, you have this 0x blah, blah, blah error? Or just, oh, I'm sorry that you ran into this, and now you can actually do this to reset your app so that you'll end up to where you are. Don't worry, you didn't lose any data. Yeah, so do you want to have that kind of experience? So I would say a lot of these kind of experiences are extremely important,

and they're still the same thing of building a good AI native product. What are some of the new aspects of it? A couple of things which pops up in my mind. One is about personalization. People talk about this a lot. Now you have this intelligent assistant which goes with you 24-7. It knows everything about you. It helps you to tailor make all of your food offerings. And they know exactly our schedule. So you won't be able to miss the next Uber ride before you even know that you should be calling it. Yeah, so personalization is definitely important. Another thing which I kind of feel is a new discovery. We actually see a lot of frontier labs are doing that at this moment. So, for example, the moment you were able to really use a model to decode, for example, a lot of the difficult questions, challenge problems in biochem, you will end up having new software, a new agent, which can be used by all of these chemistry experts.

And they can use this to discover new drugs. And they can use this to pretty much build a lot of new analyzable heuristics to have the students really think through and improve their medical procedures. I can only imagine how such a product would be like. But those are new products. Those are new products which are only capable when you actually exceed a certain threshold of capability. But once you exceed that particular threshold of capability, you'll be able to build such a product over there. Yeah. So just a couple of things. But, you know, like the moment those products begin showing up, we'll know exactly what their user experience will be like. But they're definitely tailor-made to the experts and learners. Yeah. That's my general hunch feeling. I think this is a really good place to bring in some of the new things that Anne's been working on that you mentioned that we could talk about. So, I hear it's a new model. Yeah, it's a couple of new models. Not one, not two, but three.

Not one, not two, but three. Yeah. Let's go through all of them. Sure. Yeah, I would say first thing first, allow me to maybe spend two minutes introducing the Anthony model. in Android contact battling. Yeah, so that's our model series built by Inclusion AI of Ant Group. So Inclusion AI is our lab name. It's basically a demonstrate about what we care about. We care about, you know, our journey to AGI. We really want our AGI journey to be inclusive because we feel AGI should not be specially made scarcity to the selected field. They should be available to the average population. so that everyone can benefit from it. Yeah, that's why we call our Inclusion AI, and that's why we decided to open with all of our models. Yeah, because we feel this is early stage of the navigation, and it's better to help the other ships, because we get a lot of help from the open ecosystem as well.

Yeah, so that's why we chose such a strategy. We want to be open by default, we want to be ecosystem first, and we also want to start small. but since they're small again there are a lot of open-way models out there so one question which we have been asking ourselves is what are the deltas as I described earlier right so the deltas really matters so what are the deltas or the values we can provide to the ecosystem so that we can maybe head in towards a direction and let people know that hey folks this is safe ocean there is a likelihood this direction is right we think a lot about this Yeah, so the Lean Model Series now is at 3.0. So for the 1.0 time, we just kind of began our journey. For the 2.0 series, which was October last year, we chose to go with reasoning, and we're already exploring the boundaries of intelligence. That's why we chose to go big. We actually use a lab which builds the first 1 trillion parameter reasoning models and open with them October last year.

That's Lean 1T. yeah we go big but then we begin realizing okay once you have such a big model what are we going to do with it so if you do direct serving it's going to be really expensive because graphic cards are scarcity and having such a big model will take a lot of inference cost will people be really benefiting from spending so much clearly the answer at that time is the industry was probably not ready so in the later part of 2.5 and 2.3 series, we begin to think about this problem. So, okay, we already know where we were with this one trillion parameter model. We were there before. Can we really maybe do something to make it serviceable? So for the 3.5 series, we begin with smaller models. So we ship our new flash model. Of course, the flash models are having a really big priority right now. But we actually ship our new flash model, which is 124 billion parameter MOE with 5.1B activation. So it's actually more the light side of the flash models

compared to the other flash models. So a good part of that is it has a really good token efficiency and very good token economy. You can actually, in fact, if you go to artificial analysis and several of the other websites, you'll be able to see the Lean 3.0 flash model is the one on the Palatal Frontier. It's actually outside of the Palatal Frontier because it offers a really token-efficient and cost-efficient model for day-to-day work. That's why I decided to build a flash. We're going even further building Lean 3.0 Tiny. The Tiny model is $7.9 billion MOE with activation of only $1.3 billion. It's very small. You can actually use inferences of this model on your MacBook. We all have to cookbook out. At the same time, you can even train it. So we have a training framework called Arena, which is a smaller version of the A system we have internally to train the one trillion parameter model. The Arena is a very lightweight version which goes with the tiny model.

So if you are a hobbyist, you can actually use Arena framework to train your tiny model on your MacBook. Why do we decide to do it? It's because we begin realizing that the teachers need to teach authentic RIO. Most of the labs don't have sufficient amount of GPUs. Most universities don't have sufficient amount of GPUs. So they either have to spend all this APO cost through APIs, or they can use age inference. So maybe we're providing these options that, you know, they can actually use their own MacBook to do trainings, do their homeworks, so that teachers can actually build this new curriculum for the students to really learn reinforcement learning, building agents and stuff like that. So I think a lot of folks who are not developers, who are not in the computer science world, actually have trouble imagining what people are doing. with these tiny models. So just to make things more salient for listeners, do you have some examples of some fun things that people have been working on with your tiny model offerings? Sure. We actually began seeing,

first and first, we only shipped the model like a couple of weeks ago. It feels like forever, but it was only a couple of weeks ago. Yeah, so we're still nurturing the developer ecosystem. But we see some interesting things coming back at this moment. First and first, we do see people are using these smaller models on their old graphic cards. And they pretty much run their day-to-day life over there. So for example, on your own old graphic cards, on your own MacBook, you have all your data, which you don't want the data to go anywhere. So the moment you begin using a lot of those APIs, you begin knowing about your private data, which can be problematic. Yeah, so if you actually run your run model locally, you have this kind of airtight experiences of either there's no internet or you don't want anything to go to the internet. B, you can control your own SaaS. So one thing you can do is actually you can run your local models and provide all of your local files as context to that through a harness, for example, Hermes agent, and then use the models on your own device

so the inference is essentially free. You can use that to build your daily report, for example, of your life. And you can use it to build some personal assistant of, hey, tell me what I should be eating tomorrow. given my older history and, you know, I'm traveling to the next place, provide me with the full-fledged trip otherwise and arrange that for me. Yeah, so we do see a lot of these usage cases coming back. Also, everyone has their own old documents, regardless if it's code or if it's a recent document, diaries, whatsoever. The models are really good at summarizing stuff and doing discoverabilities. So, for example, we see people actually use this to summarize their old files, and they relearn a lot of things from their old document. I personally found one thing that can be really interesting is I sometimes read my old writing, and then I began to relearn a lot of things from that perspective.

I was able to pretty much retrieve that context while I was reading my old writing. So, for example, some old writing is about open source. every time I reread it I get some new revelations out of it another thing I do is sometimes I begin listening to some of the old podcasts from like two years ago because at that time you know the podcasts sometimes you know there are all these very smart researchers coming in and they give their own opinions about how the future looks like they're all opinions but two years later they're facts so by re-listening through that I can learn a couple things one why do they feel certain things should behave a certain way and what are the reasons behind their prediction. And the reason can be really interesting. And now we are able to decode it. Second, for the parts in which they bet it wrong, why did they bet it wrong? Why do even the smartest people

on the planet bet it wrong? What's the difference? What's the data? What are the environment? And the third one is if I know what they know back then, what kind of prediction would I make? Because you know yourself. Yeah, so I kind of feel that these three things are very interesting learning philosophy. And you have all these kind of old files and old information, which you probably don't have time before you have the model that you don't have any time to digest. Being able to aggregate and summarize is one thing. Summary doesn't give you new information. but being able to condense information expose the key points and to help you human in the loop relearn everything can be a very interesting aspect and oh by the way if you're a software engineer if you know software engineering or coding I would actually recommend everyone to actually begin learning a little bit about model trainings and how you want to learn that the best way is to do that on your own device too

if you're able to find a GPU grid with the best GPU on the planet to support their training. Good for you, but most people don't. Yeah. So really what I'm hearing is that the advantage of these tiny models of open source in general is still the privacy element because only with that, in your opinion, will people be willing to provide the context that would be needed to get the personal assistant aspect at least to the maximum capability that models are able to offer. And that's really interesting. And it's a good reminder for a lot of people who are just kind of dipping their toes into this. Let's look back to the other two models. What else have you just worked on? Well, yeah, actually three models. In fact, I mean, both Flash and Tiny were the models which we shipped a couple of weeks ago. So our new models, here are the models we just shipped. First, we ship a Lean 3.0 Flash Fin, which is a finance domain enhanced model.

Second, we ship a Lean 3.0 Flash Sente, which is French word for health. We do that when we toast. Yeah, so that's our health models in which we really want people to be able to use for their own personal health. And last but not least, we have the Lean 3.0 Flash VL, which is a VL version of our language model. So our language model finally has a VL version of it. We just open source that today. The reason why we do this really, for the VR part, it's very obvious. We really want our models to have the VR capability, period. And let's spend some time talking about the other two. We call them domain-enhanced models because those are models which post-train our open-weight backbones, like the flash model itself. And we do full SFTs, RLs, and post-trainings on them with domain-enhanced data and domain-enhanced environment. Why we want to really name it that way is because we want to differentiate that from the specialized models in the past.

We see a lot of specialized models back in 2024 and early 2025 because due to the constraint of compute power and training heuristics, a lot of times you see people actually get their own set of data and they pretty much tailor-made a model from their data to suiting their own business, which is good, which is novel. But those expert models have this one liability that when the general intelligence expands, those models immediately become obsolete. And now we do see the inference cost is getting cheaper. And by law of physics, you will definitely see that getting cheaper over time. Of course, you know, the frontier ones might still be expensive, but overall, token costs will be reduced. That's my huge feeling. Yeah, so if you take the directions that way, people have been discussing about this, so general intelligence is the most important thing. That's why we see in 2025 and early 2026, and also for the lean families, we build for general intelligence.

And once we realize that general intelligence is already there, or, you know, like the marginal, the diminishing return is there, we go for talking efficiency to make sure that it's actually going to be serviceable over time. But now we actually discovered one thing, is we do see slowdown. By saying slowdown, I don't mean that physically slowdown, but we do say slowdown of the general internal development, not because we are already reaching the top of the mountain. It's because most of the models we have now are good enough. By saying good enough, I mean it's good enough for day-to-day use, for example, co-work. If you want your models to process your document, to help you do some discovery, to write a proposal for you, or to do a day-to-day work, or even writing code, right? So all of this, I wouldn't call them soft problem, but they're good enough. So once you have all this kind of good enough models, what's the next thing you want to do with it? It's very naturally things that you actually

begin building products around it. So that's why you end up having a lot of harnesses showing up this year. By saying harness, you can think this more as a product. They're not exactly the same, but just for the audience's sake. So word harness, you can actually think this is something you can directly use, so you don't get exposed to the model, but you're using something like cloud code, for example, codex. Those two some extent are harnesses. We call them applications or products, but they're not real products based on my earlier discussions. Yeah, but are they good enough for the domain experts? For example, if I'm a banker, if I'm an asset manager, am I able to use the general models to handle our day-to-day work? Do they know enough of our own heuristic? Do they know our tools? Do we have proper evaluations to understand if a model is good enough? So also to all of the questions, probably no. So we can see that trend. And also there's a very famous picture from Anthropik's labor market index.

It's a circle. It's a full circle with the boundaries of, hey, these are all the domain verticals. and this is the boundaries of human activities and this is the boundaries of, like the red boundaries, the boundaries of the current models. I think that picture is really powerful. It helps summarize this one thing, is why do clouds start with finance, health, design, yada, yada, and why people start with coding even before all of that. It's because all of these areas as a high value domains. The moment you begin building something really valuable, you have immediate use and those industries, they will pay a fair price for what improves their cadence and efficiency. Yeah, so we see the same thing. That's why, plus, I mean, we don't really have access to cloud finance. Those are to be business, which is only delivering in a certain geolocation. So we figure there can be actually an open option for it.

That's why we decided to do me house models. But if you do one domain house models, people have this general heart feeling. That's what that group do. You do finance, and that's what you do. But no, actually we're a technology company. Finance just happens to be something we're pretty good at. So as a technology company, we also care about many other different domain verticals, for example, health and medical. That's why we have this AQ product, AFU, which is tailored for health. and on the process of really trying to build a good product for it, we began realizing that having domain-enhanced models for health is also extremely helpful to our business. So that's why I built that domain-enhanced model. And you have two domain-enhanced models. The picture essentially is the future of the message we're trying to send to the ecosystem. You can have the third. If you take our base model with our own trainings, you'll be able to actually build something on top of that. And you'll see the general intelligence is not dropped.

Whereas for the party which will care a lot, if you run the domain, like the very tailor-made evaluations for the domain, you will see our model behave extremely well, in many cases, sort of level. So if you are a set manager, the fee model is going to be a really good choice for you. And if you're a health practitioner, try Sente. and we are also really curious to get feedback from the community. That's why we want to do that. Of course, Santa is not open yet. We are in the process of driving it. Yeah, but we do that for the obvious reason that we feel in order for people to really solve that real-world problem of what they do day in and day out, they need something better than general intelligence. A similar metaphor is that general intelligence is like university or high school. it makes you a very generally capable person ready for the world. But the other way to really be helpful to run the frontier aspect of this, you'll need a PhD degree. So the domain-enhanced models

are essentially what we want to try to build for the PhD program. I wanted to dig on the question you raised earlier of Ant being a kind of technology company first and health and finance just happened to be verticals that it has particular strengths in because of its business. And my question is, when we come into specialized model training, how much of that is driven by your, like these being particularly high margin industries where AI can deliver a lot of value and how much of it, actually there's a bit of a flywheel effect where Ant already has access to this really good data that comes from the existing technology products it provides. Yeah, makes sense. I would say two things. One is about the mission. So before I joined Ant, I was working at a company called Square. And before that, it was Microsoft. But my life at Square really gives me this notion of believing in helping the small businesses.

Yeah, so Square's mission is called Economic Empowerment, which translated is very similar to Ant's mission of make small and beautiful changes to the world. Yeah, so kind of back to our theme model. The reason why I build that, in part, is because that aligns with our company mission. So, we kind of realize that in order for you to actually help each of the individuals, the most important part is to make sure that you'll be able to get a model or get a technology, which they have easy access to, which means you have to build it. And open waiting, that definitely helps. And two, they can sustainably use it so that's why we are currently offering the models for limited time free operators with our gateways and our partners like Novita DeepInfra so they are actually offering the models so other people can experience it because a lot of times it actually takes time to build this understanding about what this model can do

almost feels like you don't even know the person on day one you have to spend time with him or her in order to really understand the value of this individual yeah So I would say in part of that is primarily mission-driven. But seeing mission-driven, we're really meant because that's what TAN has been doing in the past 10, 20 years. Another aspect of that is definitely strategy. So when you say strategy, you look at both sides. You look at your strengths and weaknesses, and you look at opportunities and threats. A very simple two-by-two matrix heuristic. So what we realized is we all have our customers' channels and demands. To some extent, we understand our customers too, because we're doing the business over there. So the data aspect is the byproduct of understanding our customer. Yeah, so that's how we decided to start with it, because we're a very customer-first company. So once you have those demands, you begin seeing through the demand, and we will begin asking questions like, hey, should we build a harness, or should we really make a model for that?

And that's like one thing which we begin realizing, that building a model for that can be really beneficial. show. Sometimes I use that metaphor. It's using harness to handle that. It almost feels like you're using tools or SOPs or trainings, on-job trainings to learn things. But when certain things became like the norm, you can push that into, for example, post-training. As if that, you know, a lot of the things which we're learning in university 10 years ago, they're not in university. You can only learn those by working at the top companies. But now, you know, with a lot of those practices became SLPs or standard practice or best practices. You don't see them anymore. You see them being learning education, right? So similar things happens. That's why we decided to really kind of go down that path and see what we can find out. Of course, you know, by doing that, there is still this kind of intrinsic risk of your frontier model might really have this kind of like expanded or a jump in the intelligence level,

which to some extent make your model obsolete. But that's okay. Typically, most of those models are extremely expensive too. Plus, we can retrain our models too, right? So it's a very viral industry at this moment. Yeah, so we feel very comfortable by going down that direction. This actually pushes us to actually think through a lot of things about what kind of customers we are serving and what kind of businesses we're helping, like the small business we're helping. That's why we build this technology. I really like the narrative, first of all. That's actually what motivates me. And secondly, I really believe that, you know, those directions are uncharted territory in the ocean. Somebody got to try it. And Ant Group happens to be, you know, one of the best companies. And Antling happens to be the best model to explore that territory. So far, we're getting a lot of good feedbacks from our partners and engagement. The engagement has been good.

They all say that, hey, you know, we really want something like this and we are looking forward to your next version of your domain enhanced model. Yeah, we got like all this kind of good feedbacks. Amazing. I think it's a pretty good place to wrap it and it's a special pleasure, Richard, to have you on China Talk and to hear all about how you've navigated this industry over such a long period of time and yeah, I really appreciate it. Any final thoughts you want to share? Sure. I mean, first thing first, thanks so much, Irene, for, I would say, your interest of having this conversation. That means a lot to me. Second, we had multiple conversations during global events. One thing which I personally really feel is motivating of talking to you and working with people like yourself is I really feel like at this moment, the information channels have been starved with, I would say, mixed quality of content. So one thing which I personally benefit from is to be able to go through a global education and really learn things by living through my time in Singapore, Canada, US, and time in China.

I feel such a diversified cultural infusion helps me to become a better person. We need avenues and we need ways to be able to really make this helpful to many others. So being able to sit on podcasts and talking through things is something which I found to be extremely effective and motivating. Yeah, so if I'm allowed to maybe summarize this conversation here today, first thing is thank you. And I really hope that we'll be able to continue our relationship of discussing models and what we do here at Antling, as well as a lot of these new findings in open ecosystem. Second, if you happen to be listening to this podcast, I highly recommend you to, if you haven't really tried out these open-way models, I highly recommend you download them from Hunting Face and follow through the cookbooks and give it a try. The future is already there. For better or for worse, we can't rewind the clock and going back. So the future will be there. And we all have to use models.

We all have to know different models. The more you know about models, the merrier. That's what motivates me from building this team. And last but not least, it's a long journey. So I personally dislike a lot of the narratives in terms of, okay, the end game will look like this. So the only thing I should be doing right now is this and that. I personally feel like none of us knows what's going to be right. But as a practitioner, sometimes we know what's wrong. So we feel comfortable giving our opinions. Opinions, opinions. So disclaimer. But getting to know a little bit more about the industry yourself and talk to people and get a diverse type of input will definitely be able to help you. Yeah, we are here to support and provide you with a different path to the end of the discovery. Yeah, hope you wish you all the best. Santé. Richard, thanks for coming to China Talk. Thanks for having me.