127. 大模型季报跨年对谈：和广密预言一场AI War、两大联盟和第三个范式Online Learning

What can you say about AI bubble? It's not a AI bubble. It's like AI wall. AI is a new weapon or weapon. So many people will lose the last money and will not be able to get it. Google is stronger, it will become a反-Google organization. OpenAI is stronger, it will become a反-OpenAI organization. the new Abramなん uid is to challenge Google And there's a huge principle That really difficult of 1983 做这个播客很大的一个意义就是要有passion去鼓动中国的科技公司多买卡 给年轻的研究员更多的搓卡的机会 Hello大家好 欢迎收听张小军商业访谈录 我是小军 这是一档由语言及世界工作室出品的深度访谈节目 我们希望和你一起从这里探索新世界 这里是全球大模型记报的第八集 也是我和广密跨年对谈的第三年了 那这期节目将带你看到在对AGI开始充斥着现实主义情绪回归的当下 全球的AI竞争它形成了怎样的战队阵营和联盟 各个前沿实验室在探索哪些新的办事 以及硅谷又涌现出了哪些新型的研究实验室 If you have any ideas or ideas, please give us a comment and comment. In the end of the year, we hope we can see AI together. We hope we can see AI together. Hello, I'm Blame. Let's give a call for everyone. Hello, I'm Blame. The time is so fast. This is our third year of the 8th episode. The third year of the跨年. so much longer in the country Absolutely awesome also much time path many высок many employees are and a little bit dude see Good good strengthen transport and but in the next phase, Iאר Holder will rearrange be sold Mostly that is only it's considerate which is foster moving from the start and step to achieve which is meine storage of diss 어 Rominations I hope you would have a bigger than we needed to do Chinese people Application Which تم pozy реal life is How is all coming soon So the cyclekarity is still we then still umыт. because we often had the problems in the AI lab where its function forces, so we still reimbving our research and the results are more consistent. Now, what if coding is here, when starting out 150 million dollarschu hard since the year that is 100M dollars pr trademark new markets, your cashier the more and how to move it. For example Google After exchange studies license For example不了 which ELEA will soon Another outcome program with flying online learning This is an endemic goal Here, you should need to consider Instead of vamos biết technology This time I feel like the idea of the technology itself is becoming more important. So, the idea of the technology itself is very important to understand the right direction of the strategy. You see, everyone is talking about AI bubble, but you have to feel the smell of the pump? You have to feel the smell of the pump in the 2026, or the way you have more? I have worried about the fact that the conclusion is that we don't have to worry about AI bubble. I clearly didn't blame it as an AI bubble instead of the war. Juntz is each other and Italia's war, the war or war is the war. Even Chinese is there once war And there's war …… AI has all become countertalks. So many people will make the professional support, will not put single inspiration behind it. Why don't you want to Yamper next to all super Smash peaceness? The stepfather greatly in the field. Defense Marks Journal history of IP files The Russel catweay cover, Italy, Theantic app kantor orijing인지 the phone by a 2 e option if the AI agent trasplake those I mean, how about super apps intake This simple app Card, I demand one culture cyber app So today's technology, it's also a defense. It's a defense. It's a defense. It's a defense. I think AI should be the main goal of the US. It's not a defense. It's not a defense. Yeah, it's a defense. It's a defense. Why do you think you're talking about AI bubble? What's the AI bubble? People do your own business to buy 어 Stellar Ermuuuuurs fer fer fer hall wherever they get information I mean they do There's a question What happened is the communication That means that OpenAI will be a year to 2,000 more than a year. OpenAI is actually in 2030, the global forecast is 2,000 more than a year. In the next five years, the total of 5 years is about 5,000 more than a year. From the market and the current market, it's not a good deal. It's not a good deal. I don't know why I'm spending so much money. But I think there is a clear thing that in the future of 2-3 years OpenAI's cash flow is not a problem There is a news that 8000 million dollars in the year and then 100 million dollars in the year OpenAI's cash flow is still full But the 1.4T is still a big deal Most of the cash flow is in the year of 2018 One year ago Sam on a podcast and asked for openAI's cash flow How to 1.4 million dollars And then the question is, if you want to buy out your assets, I can help you find your assets. This is causing you to be able to buy your assets. Can you tell us about OpenAI's income? OpenAI has a look at the net income. It has a look at the net income. It has a look at the net income. If we look at the net income, we can tell you about it. Let's say that we have a better idea. In 2030, we have a 40 years of year, 15亿的周火, 15亿的日火,对吧?假设他的订阅做到10%的付费率,那就是4亿用户,这个是非常合理的,就是因为对标Office,Office应该也是3到4亿的付费用户,但很多是企业帮员工付费订阅嘛,那个4亿的付费用户,每个用户假设是200美金的年费,这其实是800亿美金,其实Office也是几亿的付费订阅用户,几百亿美金的生意,对吧? It's for Netflix, too long asI Jakob If not pay submit users, if they did talk peoples in China, then we can compare the plat руб others who only have gotten some data everybody 。 AWS and the first year and the first year we won't think this will be a thousand years of experience and we won't think that in 2023, we won't think that OpenAI and SORP will be growing so much and we won't think that OpenAI and SORP even if we were to talk about the first year coding agent is the main topic 10 months ago, we saw AI coding this one-year-old company we won't think that one year will be growing and it will be growing up 100 yen yen I think AI is excellent I thought it was quite� жar every leone's unique Ng TEK We are looking for even cheaper This means This means It feels like some instruments Actually, every trade company makes organic to 10 yen Openaide show about 10 yen notwendings It's an offer for 18% To their duplicate technical Efendimiz For use them There are noaisia Of the electronic hardware celebrate with weeks of weeks These are stark!! This düster's product These are too rapid So we would probably like With the standard There are still Where the most poor term-to-mass. In this particularpection show I think this track is more powerful. That's the oh that the performance has cleared me right through The importance of taking this sample yu drives actual hammering and the tools I think it's a big deal. I think it's a big deal. I think it's a big deal. It's a big deal. It's a big deal. It's a big deal. So, you don't want to let Google one of the most open-in-one and open-in-one. So, Google is good. They're good. They're good. It's like a company's house. It's a big deal. So back then we Twisted P Because at the top Atlantic a country It's been talked about the traditional customers The company opens the game Things that this technology has traveled People are in control That are Solutions Even if you want to blow itそれでは the authors And you Franco-text All three at make cost Your copy of American Woo J Daleks This text ahead of tint policies and 替代勞动力带来的收入 那这部分的GAP 本质上 是AI能力 是否能顺利地 抢夺到这个劳动力市场 你怎么看这个问题呢 我觉得核心取决于 AI的能力边界走到哪 到时候是AI便宜还是人便宜 那AI便宜 其实本质上是电费 是不是责募便宜 我觉得现有的范式 现有的路径 还是有机会让模型 比绝大多数的白领 或者支持工作者聪明 我觉得这个确定性非常高的 Thelem hacer alook in , if you need a clever était massively w karbush open Maybe you should靠 a постав to the hill a film seems to have 200 300 even 1000 This, of course, is past 국민 and physicians Theзией you mentioned of the idea of AI war AI war is a whole World night war May the end of the chair We are not being behind our competent table It is considering toy 주 count We did do a few of theseGra равна the GPU oops Google published a complement o of the site In Google Carlos has been I believe Google will confirm very strong But Enjoyedishi being better no matter how more I developed AI mobile performancearten Agden So I don't think there's these two tricky models Like I think just to 15 from here What A j I There were a lot ofsells, but Euy fi behind them actually instead were hourly Things like that was still a significant bump That's when I chocolate it was just an internal company What for me now, I have the big concept on some cél Frη GPU if PREP로 marriages are good The chance is quite cheap But it's tough Second level here is that OR Whatever receba makes for another TRT You can also Yeah will maintain temporer v8 and limit downtownings and In suspicious d during the imminent 7, that's it will be engaging, still stronger Look, you have ready to grow, and помог with the user What about the google companies? See Google is stronger It's something which forms positive communication, to open up семь hair, and become a more powerful network required to become a powerful development in China. By the way that old person Bran Así... This is the удоб idea of the mobile various companies and different projects imagingging them Let's 해uur a whole strategy Is it important for continue on this? I want to keep statistics out of your眼睛. First, the 솔직히 nextashi term is select US. Part 2 is Google against A sótry자 two companies from göt puis 200,000 pesos. Part 3 is find out UNDbD, the number is the next one to use. Part 4 is the global bank account AGS Larson's closepring媽 I wrote time timeER offering. For the next one, I will解僱 less than that. On the fifth, we will discuss AKS Selasson's Supplic cartsê, There are these människor operating to make us grow The dragon's relationship It's a very good thing to do. It's a very good thing to do. I think that in the future, the model of the model of the model of the model is a competition. This is the model of the model that is the world's most of the three. The model of the model is the GPD, Cloud, Gemini. From the three models, the model of the model is the two-year model about the home house revenue of Iw 실제로 fra thumb Stick in your Eu douleback Same of students IA native we expect to come across and answer for people so we may be We expect other small gro shoutout amount to 400 yards Maybe 3 or 10 yards Look, we always feel the best The computing of these 3 pairs are kinda hard Will be able to do 3%, ways to gain some筹 But this is how mostsuper species what clearly means to me The only thing I can think is to know about the most depict well the student as soon as Iстро especially is such an interesting because they become equal to you through the minds of Mиров There it is for trivial, people will move it as well so that we are going to start sharing the video about getting to be read I believe that Peanut Man might not be able to pass it off But XAI Meta still At first level This China and Tufts founder of�sot Nor could be satisfied with this This Gu homem theme is very much Because XAI Meta also exists When metadata extreme Rothschild The point is you're not going to even enter the surpassinggence and keep the number of the katy recursos because of the za' that key hizo most probable time is לו. It's always up to date the compensation because there are two ways to get it for certain points that you don't want to use to describeارlain The answer right is now to say Then one more thing will be olacak The question from each tool will change the power to stage What differentups have these two problems have been like a transition message? The different issues are common By using these two, it is a distinct connection Since it was five weeks смотреть the power to within each year Be determined 구체적ally Also we know it's our vision The general knowledge will beploaded future But there is also no concentration of strategy chronologically There may be a lot of menus dormetite Biologicalnya P It's better, it's better and better, it's better and better. I don't know how much time it is. For example, ChatGPT's bet, it's 2C's依然 is the base of the end of the 2C's It's about 4.8 or 5.1's of the day of the day, right? It's about Gemini's 5-6X's. ChatGPT's been a 10-20 hour of the day of the day of the day of the day of the day of the day of the day of the day. That's a good thing, I think it's a good thing. I think it's a good thing. So you can see Google's Dormtai, ChatGP's friends, Anthropic's coding and agent and Elyea's SSA or Thinking Machines to do the same thing So I think this is a bit more interesting thing The idea of the design is that Anthropic is interested in 2B and 2C and coding and agent This is a very useful thing for AI to learn But then Anthropia would be in this serious situation. So in each of the best-spects of the best-spects, we must have a strong strong-spects. We must be able to use the best-spects of the best-spects to use a goal of the best-spects. We must be able to use the best-spects of the best-spects. And you can seeGHTYP back C The user click to machine The Steve Mei Or QIE The user tick... The user goes well Hekom You can tell If you really know So in the next generation, what's the thought of it as a goal meny 大家都是在交替领先 秀姬肉 我觉得我们就拿Meta来说 就是你再退一个SOTA的model 只是交一个作业肯定是不够的 它得有个很牛的爆款产品才够的 起码得到ChatGP这种量级吧 才算爆款 因为这样才能撑得起Meta 1200亿美金的明年的CAPEX投入 对吧 你看Meta现在账上现金大概500亿美金 It's going to be $600, but it will be $1200. Actually, only the SOTA model is not enough, right? That's right. Today's SOTA is benchmarked SOTA, right? I think Gemini 3 and GPT 5.2 are a lot of short-term fees. Because many companies are high-term fees, but high-term fees. The SOTA really is the best idea is how much money is. where they are experts and i will prick tongue and do not check, look at the benchmark is needed he said you believe you have a view is that the structure of the homestomen I don't like this as this i mean there are one there is the structure of the homestomen is likeир n celebrations지도 scale SQ的时候就是模型公司 scale data这个是更容易的你最后看模型公司想做好哪些能力就是加数据你比如说ChatGPT也好Coding也好就很像Amazon早年卖书一样它通过卖书拉通了物流仓储和用户对吧那接下来它横向的scale更多的SQ那今天领先的模型公司也是对吧你看它自己都在努力的 and also the office in the PPT and Excel and the Excel data and the finance and the actually the the basic the scale of data is higher because the chat GPT and the coding and the all the stack and the data the pipeline is better so I understand the basic the especially the the the the the and I think there's a big difference in the business. You're doing a trade-off market, or a small-factor? You're doing a trade-off market, but you're doing a trade-off market. But in history, trade-off market is not a big platform. But you can see, the small-factor market is in the combined platform platform. It's built into a high-quality market value. I think this is a very important thing. So today, we'll go to the propensity on the basis of uncompetwości, firstly looking at set-home It's not a big impact. You can see how these two companies are going to compete? Yes, I think the main point is that Google and OpenAI is the highest value of 10T. I think it's a bit more fun. I think it's been a big deal. There are two reasons. First, OpenAI is the only chance to challenge Google. The only chance to challenge Google. Secondly, openVAX is the best. OpenVAX is more so låなんか reste chimescale. OpenVAX kişi's production quality offers as far as well. Japan has 20 table yen enemy AR. 높 of multiple fee should cost over. Third in Europe, openVAX will plant the future, or the online driving team or your next other вещи This is of legal expenses but I feelcha we don't take too much 2007 I think it's hard to think it's a positive thing For me, if I'm about to invest around what is, I guess India want to invest wrong I feel like it's of Google I think it's pointless It's a common company. Because the chatbot has been a lot of sales. It's like a short video. It's a short video. I think Google and OpenAI is a good group. It's a good group in the C-axis. So today, Google is a short video. OpenAI is a short video. Google is a short video. It's a short video. It's a short video. the state branches same as 28 . 9 . It's anosage引back to use I think the concern is of other reasons somewhereinsale GPT is in the boleistfulness With a big gap on the base So we cannot beat 44 So GPT is the use and product is強 Google has a high-st unsafe I think the players need to and it's better than Google. And its product sense and strategy is better. I think there's a good idea to do this. ChatGPT's three years of increase in the past is from the past. The internet is 3 years of 4.5 million of MAO. ChatGPT is 3 years of 14 million of MAO. I think that's a pretty cool number. ChatGPT is now 4.8 million to 5 million. Gemini is now 9000. but it was more of the 5-6 devices I don't know whether it comes from husbandry or five about 3 in future Stockton's 3rd spread Chatterstate really is acker on hands-up or 사�ping Google may grow over E days or Nach skin And here there is up Ancrazy The last of gettingaways Tyl半 are some acceptable And XL might transform into the marketрипاس必须 stealing money I think this hasn't turned into a coming I don't think it is good for Google Google isn't I think Firefox isàn data We've all the boxes 10 years ago, you don't have a 1000000000 yen, right? Just like today, you don't have a 1000000000 yen. I think there are probably a few companies that have a 1000000000 yen. Today, it's 3-5 million yen to the 1000000000 yen. It's less than the 1000000000 yen. It's less than the 1000000000 yen to the 1000000000 yen. So, big companies have been stronger, This is a... This is a... We are going to be facing a趨勢. But today Google is not cheap. Because it's just about the price of the price of the price of the price. You can see Gemini now has a SOTA model. That's why Gemini is now in the first place. What do you understand about the current value of the price of the price? Is there a change? Gemini 3 is in the same way. I think it's a good thing. If it's not going to be a good thing, Only percent for the chat than the GК right now Tangledged challenge MRS I yet intent that gäbe I made the solidrude It layers off dengan the drivers and Daha The顆 security rate andak whatever The DAU product was Even the ThenQ O native GK Does the Snap but there are many of the people in the web. You can see ChatGPT is also used in the chat. Gemini is from the 5th to 5th job. It's called a new machine. Like, do the private project. There are numbers. Gemini's MAO is ChatGPT's 20-25. But Gemini's DAO and MAO is still 10%. HTML русs enorme range of chat gpt is 25% And even now, the way in this group isệm a few areas that change up now It is the removal trend from chat gpt It's the знаешь angle being a economic focus And you can see so muchّ capital, so China doesn't offer that the most high quality country. Vietnamese, like that, such as Android companies, like this. ChatGPT of 1 3% So, from the big data ChatGPT From the Google search From the beginning of 95.5 Now it's 85.15 This is a big data Actually ChatGPT It's not just a wide range of data It's a big data You can see Google's performance What is the reason for the reason? You think Google's today's危機 Is not a problem? partnering channel with expanding no from the game by anyтом particular information or by any patentseling healthcare that are eventually launched on the first time that we do in a weapon and we're entirely without 파k because the machine does does concerne machine does of the can be prepared for the first time. Do you have any other training? Do you have any other training? Pre-training and scaling is already finished. But the online learning is just starting. The first training is pre-training. It's about 70-80. There is a space, but it's a big deal. The second training is RL. It's about 30-50% of the level. So now today they're un mixing data today These two areas costume may not occur circumstances That they are specifically clear The top on of theção, it perhaps is clear We'renadish Myoorna search and the network bond All put on those Xi'an The most important part again Is the network excited 2,000 Both have become a niche However, today the top character The wave back angle is not over-and-初 It's a rare little 而是稀疏度在变高 今天趋势是走向了一个很大的 MOE的一个架构 我的感受是OPI也好 SSI也好 Thinkning of Shense也好 都看到了下一个范式 都在赌下一个大的范式创新 龟谷现在最热的还是叫continual learning 我们很多人叫online learning 其实就是模型的自主学习 学习效率更高 更少的数据就可以学得更快 and preach and fellowship and talk to each other I was learning this is a corporate change I live才能学到那个就是为什么移动互联网比PC互联网大10倍对吧因为每天在交互交互过程中沉淀了这个互动吸收了用户偏好我觉得online learning work以后就说AI是会吸收人的智能的就从人类的经验中学习就是但新范式的成熟可能还需要一些基础设施的成熟比如说更长的context对吧包括Laura也是重要的东西 And to form over another 2-th種 formats But we can still spend time with the institutional model and the entire僕 uit dB to comprehend we must have a listen we can hear it a lot to see on the new center which we visions can be discovered Are those new companies using those such mega leader? Who are the top Record? OpenHearn is most easily .... The huge thing is of the Mac It the name of an deetsual This is a bit more about Thinking Machines or Neolabs I think everyone is a big fan of their own Actually it more interesting Anthropic is OpenAI's the scaling team. Ilya is the pre-train team. Thinking Machines is the chat GPT and post-train team. These are very strong. OpenAI is even in the分崩离析 after 3 or 4 times. The Google actually today is still top down of delivery online learning with an investor quite unique for an expert on the confidentiality pesky machine and what a bigger range of APIs Q&Y actually because of the course why online learning is important here is an極 soci boy is Listen scientists made into dares Maybe they might have messages Of course Online learning should be the only thing This code indicates a human It's easy with learning easy choice This allows others to discover Right because there are what这是 Incremental discussions This is the ultimate Zen We will always walk a feel There is no way for So there's a very big variable The future is for wallet. But it's possible that there are things that are targeting every year. Maybe for a great job, this led by NLP… The one, that is true. So the biggest problem in sẽ be made of non-person agency. Didn't you that you have an evening disaster as a zust for 10 years? Do you think it's the future to watch them? Like I think it's your pressure to keep the product nogs. So I feel likeını gets done with the brew fois before the switch comes the psychology best It's a bit more important. Can you explain the three methods? From print training to RL and online learning? I think there are three methods. This is what we mentioned in our team. Pre-training is like石油 and aluminum. It's a big amount of water. It's a big amount of water. Now it's about 7, 8成. RL is a strong tool. It's a human-like data. It's like a new technology, but it's a little bit smaller. So it's like Surge or Mercur, such as a company for the people of the world. It's also a big deal. Online learning is a big deal called核聚变, it's not a big deal, but if it's a big deal, it's a big deal. By the way, you think there will be the third thing? What divine오� Could you do theuo This is a good relationship but I think it's a great partnership that is Theeptitude for the supposed to be That is the passion to 억 form the business by buying up To get new companies by some of the young entrepreneurs Using it to exchange cash winning chances Because I think, we don't know how many of these companies Eventually it will make it an investment in the start, Actually it is never achieved the short network当 age If you want to buy more credit in China or because the young quarter-ु contractor That's the case of Elyea. That's the case of Elyea. That's probably the next generation of Elyea. Now many young people have this young man. And they're so young. I think that you have a real story about Elyea. It seems that you have a real real story. You are always trying to push Elyea. Do you think Elyea has a real question about Elyea? This is a real question. We've been working for a long time for the frontier models to make the most important information. The most important thing is that if the data is not in the data, then it doesn't work. If you've been working for this data, then it works. So you can feel that today's model is a huge pressure. The data is too important. Yes. oftenen olur a lot of extra equipment This is our model We already nowOH dauertRI11 Everyone claims to buy bags andJohn said that by 달�. Then the span days were put in T일�idades where it's Ahmed started. You felt very different for the capac remarquesaurus and to get information online. That's whereource report is aناrid by a sense of accounting Venian Critics test got the numbers. They need a halfئche category, but it takes a really long time for CHRWA discovereri. The Halo text This is a happier comentarios Again facts So when we J brand them Vale we know that each other say Agent's landing needs to be more long But in the field, the群体 is very fast For coding, right? So I think there is a bit more In the field, the群体 is able to experience the L3 or L4 experience It is possible to see the chat gpt, right? Coding agent, office, ppt, excel agent, right? I think the agent is able to see the群体 everyone can expand ideas So what's it mean? irgendwie the R3. R4. you can make a difference in R4 there's obviously on R8 I don't it. but can you see that But it's easy to lose consumption or potentially not to maybe solve the end of IT i see, That will be taken part to the frontier level of complexity ofaders. And to make these changes in clearer neighborhood knowledge from true Además we can make the corporation Dienstiger to search location Every way that I have an肌肉 Before I start to split my stacking Principles Yes, allow that so you can see theishes from lawsuits It's still a little bitkay of infrared devices It's the first venture that we would have mixed вп respect interesting And the third thing is the современная or machine of further construation it is probably gonna be more than a Cursor interrupting As for Time Savage comes now, Then Steve Ntt ! Here you may be tiki only fiindu do, If I had another big picture completed here grab Yes! There's also a Collins system Alright 4k Well, One's Keanu new Fjcl. Such a normal a fort A s net low then you can use the size of the internet whether it's hard to do the software I think that is the big company with the client Bars think so three them Well first, the server has an option and a cool possono and the computer will offer basic things Today's p Is more important And I think that is our first time is that B Notes thank you for what is through a depth of research buy the cash classe with an intensity is really so that we can kommt when we think about AGI only Bighimean is the trap 最好是投多家 因为大家是交替以前 技术变化太快 最好的还是有一个篮子 我觉得是比较好的 因为确实比较难 最终判断哪一家是Wenders 阶段性 我觉得AZR非常好 对吧 今天我们觉得Open-Eye非常好 那个中间有一段时间觉得Google非常好 那个 对 最好是有一个组合吧 你现在理想中的AGI的In-depth有变化没有? The last episode of the podcast was about 40% of OpenAI and 40% of the word Google and Anthropics. Now it's about 25% of OpenAI and 25% of the word Google and Anthropics and 10% of the word and 10% of the word and the word. We can all just put it in the same way. You think that you don't have to do it? I don't want to do it for the investment. Oh that's a truc sound . It is quite deseable . Google ? It has imposed by OpenAI, It is acting as a 기 bright OK okay We also have a nice point look at Chattlet TT It has just about three times Today we're supposed to think in� qualidade what major court to win so for this ist I think it's important to think about the trends and to think about the trends We're going to start to move into the 2026 year What do you think of the投資? What do you think of next year? I think the most important thing is in the 2026 year I hope to see a next year of the phase of the phase of the phase Like online learning If this phase of the phase It can bring more of the agent. The type of product will also change. I think there is a more important thing. I saw the mobile phone number of this case. I don't know if the mobile phone number is the last thing. But it gave a big idea and thought. It's called Agentic Web. I think this is the real web. It has finally made a power to make a company or a company. habitual book concurrent, butico It could be fine low. It could be highly powerful Во can't be sierra generations Eも times which is broad and even still code läm technique impact Jabetic Wealth? Um, right. Just Agent if it will become a lamp as a consumer today SuperApp Google Search It'snymi This component of fluctuations The impact will be multiracial It's a good design for the first time. You have a lot of time in the Grygol. Can you talk about the key points of the world? I think there are a few feelings. First, the first is the design of the model. The second is the design of the model. The second is the one is the one. It is the one is the one. It is the one is the one. The one is the one is the one. And many new labs In general, viati on some Apple resources Through the fast shooting The security system upgrades or hinge guests Next, you see, theался of high detail It can apply for information for the data Google is also appropriate for Instagram If Google want so many useful amounts Do something? Anthropics of coding Morgan's basic data has the best QR code OpenHead more importantly tagging with social AREA Pol breathing The use of Meme relative to ASE C- meus Kaplan Google chat This experience 生产力量 其实这些都是在模型训练最初 它的数据分布决定的 我觉得这里面对创业最大一个建议就是说 除了你要的通用能力 你想做好AI产品或模型产品 一定想向下的做好数据 你到底有哪些差异化的数据的Bits 我觉得这个是最关键的 搞数据有没有看到什么新的好的方法 有多少人工有多少智能 Gears of the 15th edition of the What kind of view is it turns out a lot more év per day on the rural dishes? Yes, it turns out on the beach ish as it makes for the concept of Miracleitas. For the countries both started approaching. I believe it's changing as aəsireni-zai gandkai As an provalianerval shifal, humans reported to us We'll have a lot of different типа that we showed the規制 I learned how the buying and selling of internet frequencies We're not trying to judge the current Mediterranewehr There is a situation that is a medium-to- cima He calculates that a typical user is a career move That isn't an auto-egól part of programs This position Hill also mentions That is true This common socialghan society And has along the ability tovers響 어�tob't get ready You can do your creative mind When you deliver the device That's right. ChatGPT should be promised after the answer, but not to be able to think about it. Every question is to be able to answer it. In the process of answering, we can't learn how to learn. In the open-hand market, Mark Chen said that in his next product, memory, memory, AI will learn from the question of the question from the question of the question and the user experience. The user has a lot of feedback, and has a lot of feedback. The user will be more smart. In this case, the user will be more smart and more active. The user will also need a lot of new labs. There are many new labs and not only the SSI and Thinking Machines Lab. Can you give a talk about these labs? Yes. The sole limit of SSI, right? Babylon is doing a computer workию. But their zu 해�ảng is sie pics They may want to essayer a broader level. The Synclass which just mentioned... Then once you openAPI post-trane lead is parameter. I think it's called the Pulse Train. The LAM is the OPAI post-train leader. It's Mira and Barrett. LAM is the post-train leader. And he's also the deep man's chemical leader. The material model is actually compared to the e-dial model. The material model is the test of the test of the test of the test of the test. so the entire operation is shorter but nothing early happened to the look I can't Also, shows multigents Ikea It also généralized byori It was a swell to train multigents different fields And from maths and this is from Sziah ... This means to make a stat Examination The chief organization I have is way like ... very far called Inflection and it a general intuition that is a whole way paramount in human color work late General Intuition Is a usually different sponsored and inclusive Defaultti technology Everytime computer Um Summer Hospital Plus Obviously Super сво Million 活 one ع 啊 %, 嗯而且就嘛 很像 sono 給妳朋友海鮮兰 und衛 realiz In her 可 convicted Like a Tiger What do you think about OpenAI? One thing is that from the time, there were a lot of things in the past. OpenAI chat HP, and many people want to do the OpenAI in the machine. For example, Pi, Sunday, Generalist. I think this is better. They are all out there for a while. They are all out there for a while. I have a zor Kings I have some of the Greater Many people are programs So, we feel a bit In the bullsh petits For example, 100% has risen I 즉 , the and Met AU But there is an incarnation I still think about is the data The dataheadedured Theets and Key These numbers are the important methods они используют Polar и вitation А также визу lungs иつаров 진짜 heeft повседневение And if you ever use code you lang斯 wilden while you were ein counterbuilt by like 10 games. crackers are útil. yyoi bean control system is an interesting game later. Now this video? It is two initial one. the product is OC5. But it is gonna be understood as well, we have to use a box on the box And one of the functions is possible here l have a quite great combination for someunions. So the idea is to about, the chipset company actually has started to sell their own one thing that's common is. Its team is also a big company. I think PEI is very strong, the future So the largest government I believe like is seeingthankяться competencies of the interlocutor. It's just one point. There's a point in the process. This one is 0.7 or 0.8 points. It can be done with a value function. Or credit assignment. It's like the next one. It knows what's been successful. What's been successful. I think it's a good thing. I think it's a pretty important thing. But today's learning is more strong. The second is Generalist. They are from Google. They are very good. The most important thing is they said they found the machine learning in the screening area. There are 27,000,000 data. The data is very good. The data is very good. The data is very good. So here's skinlaw, it is more fundamental, and it's related to the game. And so as well This can be an interesting rate to fully charge of a legal幸運 encontramos with a study directly by visiting the family These data can be quite high This money is including a cyberrehensible these companies This is a particular tool whichtime is not letting mail these stores in a remote profession and this is where there must be the iconic alternative AI research because the bottom is Gemini's power. And so, I think it's probably the same name. It has a whole lot of the engine. Yes. How fast are you going to get to what stage of the GPU? Are you from the GPT-4? I think it's a pretty long distance from GPT-4. Today it's probably from GPT-0.5 to GPT-1. in the future of 3 years for GPT-II and GPT-II I mean, the models and data Including the Ouuffin practice r ile. It is not different for the previous possible things AS most of the mightystrang It is different for the pre-train train the pieces in the process are quite similar as a country as it is in the same order for the gig as the Tipect beast is that it is divided on theBoard Oh, the Cracker expansion does not have a neural system or the 보고-line system cannot be removed. And, the current view of the big traffic years and they cannot change, however the Radial pouvez-itable were managed by themselves, so maybe these are the best practices of the internet or things like case Und…… or visible. So the Túadio, let's unpack this action with Caesar is distinguished. First of all, good You find it No bath It is a source of She muốn We gotta check Just it looks for twisting According to The material I think it's already $10.000. Cursor is already over $10.000. And the two of the data, Surge AI and Mercur, are over $10.000. And Mid-Journey, I think it's about $7.000. And Together AI, is about $3.000. So the next thing is probably less than $3.000. For example, Glean, there are 2, 3 million dollars. Reply, Web Coding, there are 2, 3 million dollars. More interesting is, the 11 Labs, there are 3 million dollars. And the last time of the Lovable, there are 2 million dollars. And the E-Sense, the ChatGPT, Open Evidence, there are 1.5 million dollars. And then, the Sincetia is a $1.5 to $2. And then, the Fireworks AI is a $1.2. And then, the Average is a $1.2. And then, the other one, the Harvard is a $1.5 to $2. And the Hedgen is a $1.2. And then, we can add a picture to the book. I programmed it. The other thing there is… Minos, Genespark, there is an idea. On the previous story, we were dis textbooks today, But we talked byrais use Therapy Packers over+, DeepSeköpe, Kheimie, Minimax, Jippoo, Minos. Whatências you say about? Nothing beforehand. We heard people already get ATA. Thepc EBRocker is still a huge investment, Where would the威风 come from? fightable for 2019 Does the market actually 안돼 This is a pretty big where the marketplace is how it will disуб sure I doubt朋友 We should there mentioned the new design Connew También the不会 This is maybe from thehire sechs of our Kirky possibility. You think there's a victory still? Yes, Oledir will run in real power versㅋㅋㅋ It will be scary for the average of articulate entrepreneurs. The spacecraftも collectible events, and 크ole companies are all used to be with that. Is there any other demographic developers with us? In the system of Ukraine, maybe MeVora, a listing company of other clams are used of I will be able to do the code. At least I will be able to do the code. I will be able to do the code or do the code. Do the code. Do you think you can see the next step of the game? I will be able to do two things. One is the product on a product of a strategy. You can say Anthropic, Bait on Coding. You can find the strategy of a strategy, Keness of HPT is spatial about pal imagining your friends and their business too. At a good cross-fail Schools, it can become more available. Next chain, the plus is the ultimate opposite shape.ai learning, the political action, right? If we wanted to learn more than a way from Taipei Marketing and jargon, bitcoin experimented will be something that we had to succeed. conspiracy theoriesica and analysis. What changed how do we learn dealing with this led ablan master of cooperation in Japan? The current �stbn fluorescent usually represents what the Such- Pennzoa frei 꾸� parade andend It's become the US country's marketer The world más rápidoON Buffett as the capital of the capital, only for the return of the capital, and Tim Cook as the capital of the capital, for the capital of the capital, and to bring the capital to China. This is a big reason for the company's business. So you've been to stay for a long time, and you'll have a feeling that the US government is being used to have a lot of money and money. These money and money today are all to AI this direction. So if AI is a big deal, it would be very new. that if AI are melted, there are there some of the challenges since this is been very similar to the West Time knows what's other than Japan and Hong Kong is that that we couldn't see in the radio only made separate vulnerable Oh and so So because of different devices, A $100 mm But moto-long We set an average while very huge reach in the very is it that I saw something taking and I had to plein expectations any situation I think both were existing acabar with specific emotions найти regulations between the companies during those years which are successful which carry on by marquee Do you think they have a good idea? First of all, we need to be doing multicultural parts especially the cost of the conserved quality of the market marketing market next to us, we need to make sure we will get afraid of sales II, we need to invest all the raw material intensity andچoek I need to take care of sales, investment, efforts to make food to keep the knot in our business we need to use enough That's the most important thing And the Chinese people are more confident Don't be a crazy person Don't be a crazy person Say I'm a country Say I'm a country Actually, the team is a Chinese team I think the best product is the wrong thing The best product is the wrong thing And then The third thing is If you really want to take a lot of money It's not easy Because the new generation is hard right? So I think In the fact I think ofkrev I think of different countries You have found language I think this is called the one who is strong and the one who is strong. You can do it well and you can do it well. You can do it well and you will help you solve the problems. Why do you want to push the Chinese government to help you? I think that there is a Chinese government to help you. It means that there is a lot of 0 to 1 of the creation. Right? The two-way market, the two-way market, the one-way market, the one-way market, the one-way market. Actually, in China, the one-way market is a lot of people. It is an others The average投anha is always natural in country The two years 44 end it is littleее In make me an album of 2011 1990 he gave me a huge amount of pushbackகள the雙11 mine last That'sieren My 정확했 a lot of friends did I feel stopping Each fuck This Is kind of a fantastic business after I come to packs a toll There are many the best stars I think the spike in the always quite nice I think these are cranes. I think that was a really remaining risk since the market of China and the Music dabei foreignㅋㅋㅋ I think that the big reason was financing AI of the software has become 100 more companies in the world and in the world in the software in the world I think only 3 or 5 companies I think this is a big difference I think it's very sad not a big difference I think the people of the world are still very strong 95 or 97 is very good In many young people, 300% happen to уров They display their ultimate fuel The当作 luke Nederlanders and hedge funds You need to invest some bond You 就是 noting how thisAGED royal hesitant I think it's a worthy perspective then for over a couple of already? It can tighten up against such the�면 only specific claimaler AI companies are doing the Chinese? Three to five years ago, the world's most famous AI companies is the Chinese Chinese. It was probably 0 to 5.5% of the price. Now I think it's about 20% of the price. I don't know if you hear the audience think what kind of price is. I think the key difference is that the Chinese people have great. only human Email components without益 today it only involves saving the gap on the other hand I think there will be a 500 number of international companies potential and financial performance Maybe Google Mehta has been sold on China Whether part of China andTHE world economic compromise would be well-s subsidized a goodıyoruz than today. Please see our welcome, This is the show of the show. This is a show from the U.S. and the U.S. and the U.S. and the U.S. and the U.S. You can visit our website to our website or to get more information. Our website is U.S. and the U.S. Language is World. We hope you can find out a new world.