{
  "text": "We're leading China. We're leading China by a lot. China knows that.\nThe race between America and China for AI dominance.\nThere's only going to be one winner here, and that's probably going to be the US or China.\nWho's winning? America or China?\nChina just beat America in the AI race. And the craziest part?\nThey're not only beating American AI globally, but also inside America itself.\nChinese AI models now account for nearly 60% of AI usage by US companies.\nCompanies like Pinterest, Airbnb and Coinbase are already using them.\nAnd it's easy to see why.\nRecently, Kimi K3 dropped.\nIt's the world's largest open source AI model.\nIt even beat GPT 5.6 and Fable 5 on some benchmarks.\nAnd demand exploded so fast that new signups had to be paused.\nNow, Alibaba is releasing an equally powerful QEN 3.8.\nAnd you can use them all for free.\nBut that's not even the biggest surprise.\nJust a few years ago, the United States blocked China from buying the world's most advanced AI chips and machines.\nNVIDIA is trying to unlock those advanced chip sales in China,\nbut the Chinese government wants to encourage domestic production.\nThe country with the best AI doesn't just build better chatbots.\nIt also builds better weapons, better cyber security, better scientific research,\nand eventually a stronger economy.\nWhich means AI is no longer just a technology race.\nIt's a race for global power.\nAnd when the world's biggest superpower feels the next best country is catching up,\nit doesn't just compete, it changes the rules of the game.\nThe main priority is to strategically become independent of NVIDIA and other Western manufactured chips.\nIn October 2022, the United States blocked China from buying the world's most advanced AI chips from NVIDIA.\nThe logic was simple. No advanced chips meant weaker AI models,\nand weaker AI models meant China would fall behind.\nBut this did not stop China or even slow them down.\nA few short years later, today, China is not only building some of the world's best AI models,\nit's giving most of them to the public for free.\nSo how did China manage this?\nHow did it turn such a setback into one of the biggest AI comeback stories ever?\nBefore we begin, I put a cheat sheet of all the best Chinese AI tools\nand how to use them for free in my WhatsApp community called Staying Ahead.\nGrab it from the link in the description.\nNow let's dive into the video.\nWhen ChatGPT launched in late 2022, it kicked off a global AI race.\nChatGPT. ChatGPT. ChatGPT.\nIt's been held as a game changer.\nEvery country suddenly wanted to build the smartest AI model, but there was one problem.\nTraining AI isn't like running an app on your laptop.\nAn AI model has to process billions of pieces of data until the model gradually learns patterns.\nThat requires enormous computing power which comes from thousands of specialized AI chips working together inside massive data centers.\nTo simply explain chips for those who don't know, they are parts of a computer that do all the thinking and calculations for AI.\nAnd there were just three companies that controlled almost everything needed to make those chips.\nFirst was Nvidia which designs these chips.\nTSMC in Taiwan manufactured them.\nAnd ASML, a Dutch company, built the incredibly complex machines needed to manufacture those chips.\nAnd all these were controlled by the US.\nAt this point, you might be wondering, if two of these companies aren't even American, how could the US control them?\nThe answer lies in something called the Foreign Direct Product Rule.\nIt was started in 1959 and essentially says that if a product was made using American technology,\nthe US government has the power to stop it from being sold.\nand this includes products made in a foreign country.\nAnd since TSMC and ASML rely on American technology, software and intellectual property\nto build their products, the US has the leverage to cut them off from the tools they need to\nsurvive and shut their companies down. So in simple words, America doesn't just make the\nworld's best AI chips, it controls who could buy them. And in 2022, it decided China shouldn't be\nallowed to buy them. First, Huawei and dozens of other Chinese companies were placed on a banned\nlist, making it extremely difficult for them to buy American technology. Then the US banned NVIDIA\nfrom exporting its most powerful AI chip then, the H100, to China. It also stopped companies from\nselling China the advanced machines needed to manufacture those chips. China wasn't just\nblocked from buying the world's best AI hardware, it was also blocked from building it. The situation\nwas so bad that Nvidia had to create a slightly weaker chip. And we return now to breaking news\nfrom Nvidia. Nvidia reportedly plans to release a downgraded version of H20 that could legally\nbe sold to China. So as you see, by every logical measure, China should have fallen behind in the\nAI. Most experts agree that China is still a little behind the US. If there's anything that\nyou think people could agree on, it would be that the most advanced AI technology does not go to\nChina. It is very plausible that if we get this policy wrong, China could be overtaking the United\nStates. But instead, they turned the situation around. Let me show you how. Problem one, chips.\nThe first problem was obvious. China could no longer buy the world's best AI chips from not\nonly Nvidia, but also microchips for phones from Samsung, MediaTek, and Qualcomm. And these chips\nwere being used to bring phones into the 5G era. Everyone thought Huawei and Chinese companies\nwould be stuck in 4G era because the ultra expensive machines were only with US and US allies.\nBut in 2023, Huawei made the Ascent chips, put them in its smartphone Mate 60 Pro and shocked\nthe world by connecting their phones to 5G. To show you how significant that is, even Apple took\nvery long to make its own chips to stop paying Qualcomm. Now, behind that achievement was SMIC,\nChina's largest semiconductor manufacturer. Even without access to the world's best machines,\nSMIC, found ways to manufacture advanced chips using older equipment. It wasn't as efficient\nor as cheap as TSMC, but it worked. China wasn't trying to build the world's best chip overnight.\nIt was trying to make sure it no longer depended on someone else's. Problem 2, make every chip\nwork harder. But building domestic chips was only part of the solution. China's chips were still\nslower than Nvidia's and US export restrictions made those chips very difficult to access.\nChina was in an unfair position. So first, they tried to bypass the restrictions through\nsmuggling from places such as Malaysia, Japan and Hong Kong. Some even used human couriers\nto smuggle in chips. A student was paid $100 for each of the six chips he carried from\nSingapore to China in his luggage. But eventually, all this barely scratched the surface in helping\nChina grow its AI industry. So Chinese engineers asked a different question instead of just\nbuilding bigger computers with better chips. What if we build smarter AI? This is where\ncompanies like DeepSeek completely changed the conversation. DeepSeek used a technique\ncalled mixture of experts. Let me explain this using an oversimplified example. Say you\nhave a company of 200 people and on a particular day when you want to solve a marketing problem,\nyou don't call everyone into the meeting. You only call the marketing team, right?\nPrevious AI models often worked differently. Every time you asked a question, the AI would\ncall the whole team. Basically, the whole model would calculate the answer. That required huge\namounts of power and made each response more expensive. To address this, American companies\nlike OpenAI explored the use of a mixture of experts, which basically divided the model\ninto multiple specialized expert clusters.\nHowever, DeepSeq engineers took this concept to the next level.\nInstead of dividing the model into dozens of expert clusters,\nInstead of dividing the model into dozens of expert clusters, they sliced it into 256 tiny\nhyper-specialized experts. So when you ask DeepSeq to solve a coding problem, an ultra-efficient route\ninstantly kicks in and activates just eight of those tiny experts. The rest stayed asleep.\nThat means DeepSeek can produce strong answers while using far cheaper costs. Look at this.\nWhile the best models of GPT and Claude cost $5 and $10 per million tokens in input costs,\nrespectively, the best model of DeepSeek costs only $0.4. And while GPT and Claude cost $30\n$50 respectively in output costs, DeepSeek only cost $0.87. But that wasn't all. The\nChinese didn't stop with their innovation there. They also invented a technique called\nMulti-Head Latent Attention or MLA. To understand what this does, let's take another example.\nImagine you're talking to an AI about a 100-page document. Every time you ask a new question,\nthe AI needs to remember the important details from everything it has already read. So it\nIt stores all this information in its temporary memory.\nEngineers call this the key value cache, but you can simply think of it as the AI's short-term\nmemory.\nThe longer the conversation gets, the more memory the AI needs, which makes it expensive\nto run.\nDeepSeq's technique, MLA, compresses that memory.\nInstead of storing every detail separately, it creates a much smaller summary of the important\ninformation and uses that to answer future questions.\nIt is similar to replacing 100 pages of information with a few pages of well-organized notes.\nThis reduces the model's short-term memory requirements by over 90%, allowing it to handle\nlonger conversations while using far less computing power.\nThat makes the model cheaper and more efficient to run.\nSo with all these incredible tricks, DeepSeek managed to build a world-class AI model for\nreportedly just under 6 million dollars while Anthropic and other large companies were spending\nbillions of dollars to train their models. And if you look at the comparison of its V4 Pro Max\nmodel with comparable launches back then like Opus 4.6 Max, GPT 5.4 and Gemini 3.1 Pro, you'll see\nthat its performance is comparable. So that was all about creating efficient software but this\nonly solves half the problem because even the best software needs enough hardware from chips\nto data centers to train and run ai models on for those who don't know a data center is a physical\nfacility that contains the servers and ai chips that run websites apps cloud services and ai\nmodels along with the storage networking backup power and cooling systems needed to keep everything\nrunning. And so China was not only short of cutting-edge AI chips, they also found that\nthe newly built data centers were actually underused. So China decided to treat its AI\ninfrastructure in the same way a country treats roads or electricity, as public infrastructure\nthat everyone can use. The government is building something called the National Integrated Computing\nPower Network by 2030, or even earlier. To explain it simply, it built many large data centers in\nIn regions where land was cheaper, renewable energy was more abundant and there was space\nfor huge facilities.\nChina then connected these centers to businesses and cities through high-speed fiber networks,\nallowing businesses, universities and researchers to rent computing power when they need it.\nHere is how it works using an example.\nImagine a small AI startup in Shanghai wants to train a new AI model.\nBuying hundreds of GPUs and building its own data center could cost a lot of money.\nInstead, it can rent computing power online.\nThe startup uploads its code and data and chooses how much power it needs.\nAnd the platform finds available chips in connected data centers possibly thousands\nof kilometers away.\nIts engineers can monitor the training remotely and download the finished model once the work\nis complete.\nChina has not built one fully connected national computing grid yet, but that is the goal.\ncomputing power easier and cheaper to rent across the country.\nThis brings us to Problem Number 4 Training the Models\nBecause even the smartest AI models are useless without enough data to train them. And for\nthe next generation of AI, text alone is not enough.\nVideo models need to understand how people move, how objects interact, and how the physical\nworld behaves. You cannot teach an AI what running looks like by only describing it in\nwords. It needs to watch millions of examples of different people, camera angles, lighting\nconditions, environments and movements. It needs to see how clothes move, how shadows\nchange and what happens when someone steps into a puddle. And this is where China has\na major advantage. China has one of the largest digital ecosystems in the world. Platforms\nlike Douyin, WeChat, TikTok and other Chinese apps generate huge amounts of text, images,\naudio and video every day. For companies like ByteDance, which owns the popular AI platform\nC-Dance, this is not just content, it is valuable training data. ByteDance owns Dooyan and TikTok,\nalong with several tools for creating, editing and recommending videos. This gives the company\naccess to more than just the video itself. It can also study how people react to that\nvideo, which clips do people keep watching and at what moment do viewers lose interest.\nNow this does not mean ByteDance uses every private video to train its AI.\nWe do not know exactly what data each company uses, but owning some of the world's largest\nvideo platforms still gives Chinese companies a powerful advantage.\nBecause if chips are the engine of AI, data is the fuel.\nAnd China did not just use that fuel to build competitive AI models, it also started releasing\nmany of them for free.\nToday, models like DeepSeek, Quen by Alibaba, Kimi by Moonshot AI, and GLM models from Z.AI\nare available completely free to download and run.\nAnd this raises an obvious question.\nRunning AI companies can cost billions of dollars.\nIn fact, many major AI companies are still spending far more money than they earn.\nSo how can Chinese companies offer powerful AI models for free?\nIf users aren't paying these companies, who is?\nThe answer is that Chinese AI companies aren't really selling AI models.\nThe model is only the entry point.\nThey're selling something much bigger.\n1. They're selling the ecosystem, not the model.\nThink about Google.\nGoogle gives away Chrome for free.\nNot because browsers are cheap to build,\nbut because Chrome brings people into Google's ecosystem,\nwhere they eventually use advertising and paid services like Google Cloud.\nChinese AI companies are following the same playbook.\nWhile Western companies like OpenAI and Anthropic primarily make money by selling access to their intelligence directly.\nYou pay for a subscription or an API call or how many tokens you consume, right?\nChinese companies often make money later in the user journey.\nFor example, Alibaba not only wants you to chat with their AI model, Quen,\nit wants companies to build and run their applications on Alibaba Cloud.\nTencent, the company behind the super app WeChat, not only wants businesses to test its\nAI model, Hanuan.\nIt wants them to use Tencent Cloud, its enterprise software and the rest of its business ecosystem.\nIt needs GPU, servers, storage, security, monitoring, technical support and engineers\nwho can connect the AI to its existing systems.\nThat is where the real money can be made, whether it's cloud hosting, consulting or\neven custom AI solutions. So right now the model may be cheap or even free to download\nbut everything required to operate it reliably at scale is not. The model is not always the\nfinal product, it is the customer acquisition strategy.\nthese models become paid later, because by then many users may have no other option but\nto pay.\nReason 2 Win with distribution Remember, China couldn't compete with America on\nthe top models that Anthropic and OpenAI were constantly innovating, using better chips\nand coming out ahead. So China instead decided to win with distribution and make their products\nfree and open weight. This basically means that developers can download the trained model,\nrun it on their own computers or servers, modify it and fine tune it for their own use. For example,\na retail company could take the model and customize it by teaching it its internal policies,\nproduct catalog and more so it can accurately handle customer support queries for it.\nIt's a win for companies but it's also a win for the Chinese companies. If millions of developers\ncan freely experiment with your model, they begin building apps on top of it. And every time a\ndeveloper chooses your model instead of a competitor's, they invest time learning how it works,\nwriting code around it and integrating it into their systems. That makes them less likely to\nswitch. It's like if you invested time and effort on Twitter and built a 100,000 following,\nyou're less likely to switch to threads. Similarly, it works for developers using these\nmodels. Secondly, researchers will be able to effortlessly study these models and publish\nimprovements or techniques that others can reuse. Over time, this creates a network effect.\nThe more people use the model, the more tools, tutorials and expertise exist around it. It\nThat is the same basic reason Android became the world's largest mobile operating system.\nGoogle allowed thousands of manufacturers and millions of developers to build around it.\nAnd once enough, people are building on your platform, your ecosystem can grow faster than\nany single competitor.\nThat strategy is already working for Chinese companies.\nLook at this chart that's showing the world's top 50 most used AI models.\nYou'll see that Chinese models are rapidly gaining market share.\nAnd when it comes to the top 20 AI models in the world, between May and June of 2026,\nthe monthly token usage of Chinese AI models has surpassed the US models.\nBut let's look at more specific examples.\nAlibaba's QEN models surpassed 1 billion total downloads this year.\nMany large US companies also use it.\nFor example, Airbnb's AI customer service system uses them.\nEven Pinterest has been customizing QEN models for their usage.\nPinterest actually says it has reduced costs by as much as 90% in some workloads due to\nsuch open-source models.\nAnd Cursor, one of America's fastest-growing AI coding companies built its coding model,\nComposer2, on the base model of Chinese company Moonshot AI's Kimi K2.5.\nListen to that again.\nA top US AI company built its proprietary product on top of a Chinese foundation model.\nChinese models are no longer merely cheaper alternatives to American AI.\nare becoming part of the infrastructure on which American companies themselves are building.\nReason 3 And this brings us to one of the biggest\nreasons that Chinese models are succeeding so well. It's because China made AI a national\nmission. In 2017, China's State Council released the New Generation Artificial Intelligence\nDevelopment Plan. Its ambition was enormous. By 2030, China wanted to become the world's\nleading center for AI innovation. Universities expanded AI and computer science programs.\nGovernments funded laboratories and research projects. Cities and provinces created AI\ninnovation zones where companies could access funding, infrastructure and local policy support.\nThe plan even reached across borders. Chinese researchers who had built their careers in labs\noverseas were actively encouraged to come back home, pulled in with grants, funding and a real\npath to do their work in China. But the best thing China did was to give major companies\nspecific roles in building national AI platforms. In 2017, the Ministry of Science and Technology\nselected Baidu to develop an open innovation platform for autonomous driving, Alibaba for\nsmart cities, Tencent for medical imaging, and iFlytech for intelligent voice technology.\nThese companies were still private businesses, and they continued competing fiercely with one\nanother. Yet they were all working in the same direction for the growth of China. So this brings\nus to the last part of the case study. What can the world learn from China's AI journey? Let's\njust clarify one thing first. China did everything in its power to get ahead in the AI race. Some\nAmerican companies claim Chinese labs use distillation. This means using the answers of a stronger AI\nmodel to help train a smaller one. Then we know that there was smuggling done of restricted\nNvidia chips through black market routes. The US charges a co-founder of Supermicro with illegally\ndiverting billions of dollars of Nvidia powered servers to China. Some companies may have broken\nrules and that should be investigated but a few smuggled chips cannot build thousands of researchers,\nlarge data centers, new engineering techniques, cloud platforms and an ecosystem used by developers\naround the world. Also China did not succeed because of one shortcut. It also invested heavily\nin research and found ways to make weaker hardware perform better. The point is that\nwe should judge both sides by the same standard and not dismiss all of China's progress as\ntheft.\nSecondly, people think using Chinese AI is a threat to privacy, they will steal data\nand the models are censored by Beijing. But the West is not far behind. ChatGPT trains\non your conversations by default. Gemini merges your prompts with your Gmail, your YouTube\nand your entire Google history. Claude quietly changed its privacy policy and opted users\ninto 5 years of data retention. The idea that Chinese AI is a threat and Western AI is the\nsafe choice is completely wrong.\nWith this in mind, I have two major lessons that countries need to learn from China's\nAI race. First, every country needs some control over its AI future. Have you heard of sovereign\nAI? It does not mean every country must build every chip and every model by itself. That\nwould be too expensive and unrealistic for you.\nleast, a country should have enough of its own computing power, research talent, data,\nmodels and infrastructure that it is not completely dependent on another country. Today, most\nadvanced AI technology comes from either the United States or China.\nBut what happens if the US decides that only American companies can use its best models?\nWhat happens if China makes the same decision?\nCountries such as India could suddenly lose access to important technology because of\na political fight between two other countries.\nAnd AI is no longer just a chatbot, it is becoming part of education, healthcare, defense,\nscience, coding, manufacturing and almost every modern business. A company with better\nAI can work faster, reduce costs and compete more effectively. A country without access\nto advanced AI may slowly fall behind. The biggest lesson from China is not that every\ncountry should copy China's political system or its exact strategy. The lesson is that\na country that depends completely on foreign AI will always be vulnerable to foreign decisions.\nIndia and other countries need to build enough of their own AI ecosystem so they are not\ntrapped between the US and China.\nBecause the next major geopolitical blockade may not stop oil, weapons or physical goods,\nit may stop access to intelligence itself.\nAnd second, AI leadership is not about one great chatbot, it's about creating a whole\necosystem.\nWhen people talk about the AI race, they usually compare models.\nIs ChatGPT better?\nIs DeepSeek better?\nwhich model scores higher on a benchmark. But a model is only the visible part of a\nmuch larger system. To build strong AI, a country also needs chips to perform calculations,\nelectricity to power those chips, and data centers to keep them running. It needs universities\nto train engineers, research labs to invent new techniques, cloud companies to provide\ncomputing power, and businesses that can turn AI into useful products. China worked on all\nthese areas at the same time. So now, if Europe, India and other developing countries want\nto secure their futures, they need to step up and build such ecosystems too.\nThat was all about why China is winning the AI race and what we can learn from them. For\nmore case studies that help you understand where AI is going, who is gaining power and\nwhere the next major opportunities will emerge, subscribe to the channel and watch the video\non your screen next because it explains another major shift happening in AI that most people\nare still completely missing.\nI'll see you there.",
  "language": "en",
  "duration": 1504.725313,
  "segments": [
    {
      "start": 0.0,
      "end": 3.54,
      "text": "We're leading China. We're leading China by a lot. China knows that.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 3.76,
      "end": 7.04,
      "text": "The race between America and China for AI dominance.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 7.14,
      "end": 11.22,
      "text": "There's only going to be one winner here, and that's probably going to be the US or China.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 11.38,
      "end": 13.38,
      "text": "Who's winning? America or China?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 13.96,
      "end": 17.74,
      "text": "China just beat America in the AI race. And the craziest part?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 17.9,
      "end": 22.56,
      "text": "They're not only beating American AI globally, but also inside America itself.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 22.78,
      "end": 27.86,
      "text": "Chinese AI models now account for nearly 60% of AI usage by US companies.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.11207644414093534,
      "no_speech_prob": 1.259480367221677e-12,
      "compression_ratio": 1.752851711026616
    },
    {
      "start": 27.86,
      "end": 32.22,
      "text": "Companies like Pinterest, Airbnb and Coinbase are already using them.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 32.5,
      "end": 33.72,
      "text": "And it's easy to see why.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 34.04,
      "end": 35.68,
      "text": "Recently, Kimi K3 dropped.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 36.04,
      "end": 38.64,
      "text": "It's the world's largest open source AI model.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 38.96,
      "end": 43.58,
      "text": "It even beat GPT 5.6 and Fable 5 on some benchmarks.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 43.78,
      "end": 47.34,
      "text": "And demand exploded so fast that new signups had to be paused.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 47.48,
      "end": 51.42,
      "text": "Now, Alibaba is releasing an equally powerful QEN 3.8.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 51.5,
      "end": 53.1,
      "text": "And you can use them all for free.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 53.24,
      "end": 55.12,
      "text": "But that's not even the biggest surprise.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1515836408061366,
      "no_speech_prob": 2.6043945645926048e-12,
      "compression_ratio": 1.4734982332155477
    },
    {
      "start": 55.12,
      "end": 61.96,
      "text": "Just a few years ago, the United States blocked China from buying the world's most advanced AI chips and machines.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 62.06,
      "end": 65.68,
      "text": "NVIDIA is trying to unlock those advanced chip sales in China,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 66.0,
      "end": 68.9,
      "text": "but the Chinese government wants to encourage domestic production.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 69.32,
      "end": 72.6,
      "text": "The country with the best AI doesn't just build better chatbots.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 72.8,
      "end": 77.28,
      "text": "It also builds better weapons, better cyber security, better scientific research,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 77.58,
      "end": 79.56,
      "text": "and eventually a stronger economy.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 79.84,
      "end": 82.72,
      "text": "Which means AI is no longer just a technology race.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 82.72,
      "end": 84.54,
      "text": "It's a race for global power.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.15343231714072347,
      "no_speech_prob": 4.8645028977167826e-12,
      "compression_ratio": 1.6178343949044587
    },
    {
      "start": 84.54,
      "end": 88.62,
      "text": "And when the world's biggest superpower feels the next best country is catching up,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 88.88,
      "end": 91.88,
      "text": "it doesn't just compete, it changes the rules of the game.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 92.14,
      "end": 99.62,
      "text": "The main priority is to strategically become independent of NVIDIA and other Western manufactured chips.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 99.82,
      "end": 106.38,
      "text": "In October 2022, the United States blocked China from buying the world's most advanced AI chips from NVIDIA.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 106.7,
      "end": 110.6,
      "text": "The logic was simple. No advanced chips meant weaker AI models,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 110.92,
      "end": 113.48,
      "text": "and weaker AI models meant China would fall behind.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1403803472165708,
      "no_speech_prob": 1.883299722768572e-12,
      "compression_ratio": 1.6219931271477663
    },
    {
      "start": 113.48,
      "end": 116.92,
      "text": "But this did not stop China or even slow them down.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 117.16,
      "end": 122.12,
      "text": "A few short years later, today, China is not only building some of the world's best AI models,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 122.28,
      "end": 124.52,
      "text": "it's giving most of them to the public for free.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 124.76,
      "end": 126.26,
      "text": "So how did China manage this?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 126.52,
      "end": 131.0,
      "text": "How did it turn such a setback into one of the biggest AI comeback stories ever?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 131.3,
      "end": 135.2,
      "text": "Before we begin, I put a cheat sheet of all the best Chinese AI tools",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 135.2,
      "end": 139.06,
      "text": "and how to use them for free in my WhatsApp community called Staying Ahead.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 139.36,
      "end": 140.96,
      "text": "Grab it from the link in the description.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 141.34,
      "end": 142.58,
      "text": "Now let's dive into the video.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09782475062779018,
      "no_speech_prob": 3.278587904637087e-12,
      "compression_ratio": 1.650943396226415
    },
    {
      "start": 143.48,
      "end": 150.68,
      "text": "When ChatGPT launched in late 2022, it kicked off a global AI race.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 150.68,
      "end": 153.68,
      "text": "ChatGPT. ChatGPT. ChatGPT.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 153.68,
      "end": 155.48,
      "text": "It's been held as a game changer.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 155.48,
      "end": 160.92,
      "text": "Every country suddenly wanted to build the smartest AI model, but there was one problem.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 161.26,
      "end": 163.86,
      "text": "Training AI isn't like running an app on your laptop.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 164.22,
      "end": 170.0,
      "text": "An AI model has to process billions of pieces of data until the model gradually learns patterns.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.22210772986550933,
      "no_speech_prob": 3.318030962831675e-12,
      "compression_ratio": 1.5397489539748954
    },
    {
      "start": 170.0,
      "end": 178.38,
      "text": "That requires enormous computing power which comes from thousands of specialized AI chips working together inside massive data centers.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.14662681831108346,
      "no_speech_prob": 2.739432595696978e-12,
      "compression_ratio": 1.5770609318996416
    },
    {
      "start": 178.72,
      "end": 185.34,
      "text": "To simply explain chips for those who don't know, they are parts of a computer that do all the thinking and calculations for AI.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.14662681831108346,
      "no_speech_prob": 2.739432595696978e-12,
      "compression_ratio": 1.5770609318996416
    },
    {
      "start": 185.74,
      "end": 190.38,
      "text": "And there were just three companies that controlled almost everything needed to make those chips.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.14662681831108346,
      "no_speech_prob": 2.739432595696978e-12,
      "compression_ratio": 1.5770609318996416
    },
    {
      "start": 190.58,
      "end": 193.18,
      "text": "First was Nvidia which designs these chips.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.14662681831108346,
      "no_speech_prob": 2.739432595696978e-12,
      "compression_ratio": 1.5770609318996416
    },
    {
      "start": 193.7,
      "end": 195.86,
      "text": "TSMC in Taiwan manufactured them.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.14662681831108346,
      "no_speech_prob": 2.739432595696978e-12,
      "compression_ratio": 1.5770609318996416
    },
    {
      "start": 195.86,
      "end": 202.58,
      "text": "And ASML, a Dutch company, built the incredibly complex machines needed to manufacture those chips.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 202.9,
      "end": 205.02,
      "text": "And all these were controlled by the US.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 205.22,
      "end": 211.12,
      "text": "At this point, you might be wondering, if two of these companies aren't even American, how could the US control them?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 211.62,
      "end": 215.04,
      "text": "The answer lies in something called the Foreign Direct Product Rule.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 215.32,
      "end": 221.56,
      "text": "It was started in 1959 and essentially says that if a product was made using American technology,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 221.56,
      "end": 225.14,
      "text": "the US government has the power to stop it from being sold.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1157249759983372,
      "no_speech_prob": 1.9129812751234043e-12,
      "compression_ratio": 1.5695792880258899
    },
    {
      "start": 225.14,
      "end": 228.14,
      "text": "and this includes products made in a foreign country.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08239986419677735,
      "no_speech_prob": 2.1508569664902e-12,
      "compression_ratio": 1.5477941176470589
    },
    {
      "start": 228.52,
      "end": 234.42,
      "text": "And since TSMC and ASML rely on American technology, software and intellectual property",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08239986419677735,
      "no_speech_prob": 2.1508569664902e-12,
      "compression_ratio": 1.5477941176470589
    },
    {
      "start": 234.42,
      "end": 240.0,
      "text": "to build their products, the US has the leverage to cut them off from the tools they need to",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08239986419677735,
      "no_speech_prob": 2.1508569664902e-12,
      "compression_ratio": 1.5477941176470589
    },
    {
      "start": 240.0,
      "end": 245.02,
      "text": "survive and shut their companies down. So in simple words, America doesn't just make the",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08239986419677735,
      "no_speech_prob": 2.1508569664902e-12,
      "compression_ratio": 1.5477941176470589
    },
    {
      "start": 245.02,
      "end": 252.0,
      "text": "world's best AI chips, it controls who could buy them. And in 2022, it decided China shouldn't be",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08239986419677735,
      "no_speech_prob": 2.1508569664902e-12,
      "compression_ratio": 1.5477941176470589
    },
    {
      "start": 252.0,
      "end": 257.7,
      "text": "allowed to buy them. First, Huawei and dozens of other Chinese companies were placed on a banned",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.059780566491813306,
      "no_speech_prob": 2.109167441394222e-12,
      "compression_ratio": 1.6241610738255035
    },
    {
      "start": 257.7,
      "end": 263.4,
      "text": "list, making it extremely difficult for them to buy American technology. Then the US banned NVIDIA",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.059780566491813306,
      "no_speech_prob": 2.109167441394222e-12,
      "compression_ratio": 1.6241610738255035
    },
    {
      "start": 263.4,
      "end": 270.18,
      "text": "from exporting its most powerful AI chip then, the H100, to China. It also stopped companies from",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.059780566491813306,
      "no_speech_prob": 2.109167441394222e-12,
      "compression_ratio": 1.6241610738255035
    },
    {
      "start": 270.18,
      "end": 275.3,
      "text": "selling China the advanced machines needed to manufacture those chips. China wasn't just",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.059780566491813306,
      "no_speech_prob": 2.109167441394222e-12,
      "compression_ratio": 1.6241610738255035
    },
    {
      "start": 275.3,
      "end": 280.52,
      "text": "blocked from buying the world's best AI hardware, it was also blocked from building it. The situation",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.059780566491813306,
      "no_speech_prob": 2.109167441394222e-12,
      "compression_ratio": 1.6241610738255035
    },
    {
      "start": 280.52,
      "end": 285.56,
      "text": "was so bad that Nvidia had to create a slightly weaker chip. And we return now to breaking news",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.058178209625514204,
      "no_speech_prob": 2.4086334055734016e-12,
      "compression_ratio": 1.6114864864864864
    },
    {
      "start": 285.56,
      "end": 291.54,
      "text": "from Nvidia. Nvidia reportedly plans to release a downgraded version of H20 that could legally",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.058178209625514204,
      "no_speech_prob": 2.4086334055734016e-12,
      "compression_ratio": 1.6114864864864864
    },
    {
      "start": 291.54,
      "end": 297.08,
      "text": "be sold to China. So as you see, by every logical measure, China should have fallen behind in the",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.058178209625514204,
      "no_speech_prob": 2.4086334055734016e-12,
      "compression_ratio": 1.6114864864864864
    },
    {
      "start": 297.08,
      "end": 302.22,
      "text": "AI. Most experts agree that China is still a little behind the US. If there's anything that",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.058178209625514204,
      "no_speech_prob": 2.4086334055734016e-12,
      "compression_ratio": 1.6114864864864864
    },
    {
      "start": 302.22,
      "end": 307.06,
      "text": "you think people could agree on, it would be that the most advanced AI technology does not go to",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.058178209625514204,
      "no_speech_prob": 2.4086334055734016e-12,
      "compression_ratio": 1.6114864864864864
    },
    {
      "start": 307.06,
      "end": 312.7,
      "text": "China. It is very plausible that if we get this policy wrong, China could be overtaking the United",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06597369292686725,
      "no_speech_prob": 2.8160130477461465e-12,
      "compression_ratio": 1.586092715231788
    },
    {
      "start": 312.7,
      "end": 319.14,
      "text": "States. But instead, they turned the situation around. Let me show you how. Problem one, chips.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06597369292686725,
      "no_speech_prob": 2.8160130477461465e-12,
      "compression_ratio": 1.586092715231788
    },
    {
      "start": 319.46,
      "end": 324.74,
      "text": "The first problem was obvious. China could no longer buy the world's best AI chips from not",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06597369292686725,
      "no_speech_prob": 2.8160130477461465e-12,
      "compression_ratio": 1.586092715231788
    },
    {
      "start": 324.74,
      "end": 331.04,
      "text": "only Nvidia, but also microchips for phones from Samsung, MediaTek, and Qualcomm. And these chips",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06597369292686725,
      "no_speech_prob": 2.8160130477461465e-12,
      "compression_ratio": 1.586092715231788
    },
    {
      "start": 331.04,
      "end": 336.68,
      "text": "were being used to bring phones into the 5G era. Everyone thought Huawei and Chinese companies",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06597369292686725,
      "no_speech_prob": 2.8160130477461465e-12,
      "compression_ratio": 1.586092715231788
    },
    {
      "start": 336.68,
      "end": 342.94,
      "text": "would be stuck in 4G era because the ultra expensive machines were only with US and US allies.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1027382992683573,
      "no_speech_prob": 2.7294991353926656e-12,
      "compression_ratio": 1.4767441860465116
    },
    {
      "start": 343.34,
      "end": 349.88,
      "text": "But in 2023, Huawei made the Ascent chips, put them in its smartphone Mate 60 Pro and shocked",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1027382992683573,
      "no_speech_prob": 2.7294991353926656e-12,
      "compression_ratio": 1.4767441860465116
    },
    {
      "start": 349.88,
      "end": 355.5,
      "text": "the world by connecting their phones to 5G. To show you how significant that is, even Apple took",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1027382992683573,
      "no_speech_prob": 2.7294991353926656e-12,
      "compression_ratio": 1.4767441860465116
    },
    {
      "start": 355.5,
      "end": 361.18,
      "text": "very long to make its own chips to stop paying Qualcomm. Now, behind that achievement was SMIC,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.1027382992683573,
      "no_speech_prob": 2.7294991353926656e-12,
      "compression_ratio": 1.4767441860465116
    },
    {
      "start": 361.18,
      "end": 366.04,
      "text": "China's largest semiconductor manufacturer. Even without access to the world's best machines,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09252214431762695,
      "no_speech_prob": 2.4274267491908308e-12,
      "compression_ratio": 1.6585365853658536
    },
    {
      "start": 366.22,
      "end": 372.34,
      "text": "SMIC, found ways to manufacture advanced chips using older equipment. It wasn't as efficient",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09252214431762695,
      "no_speech_prob": 2.4274267491908308e-12,
      "compression_ratio": 1.6585365853658536
    },
    {
      "start": 372.34,
      "end": 378.22,
      "text": "or as cheap as TSMC, but it worked. China wasn't trying to build the world's best chip overnight.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09252214431762695,
      "no_speech_prob": 2.4274267491908308e-12,
      "compression_ratio": 1.6585365853658536
    },
    {
      "start": 378.58,
      "end": 383.96,
      "text": "It was trying to make sure it no longer depended on someone else's. Problem 2, make every chip",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09252214431762695,
      "no_speech_prob": 2.4274267491908308e-12,
      "compression_ratio": 1.6585365853658536
    },
    {
      "start": 383.96,
      "end": 389.72,
      "text": "work harder. But building domestic chips was only part of the solution. China's chips were still",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.09252214431762695,
      "no_speech_prob": 2.4274267491908308e-12,
      "compression_ratio": 1.6585365853658536
    },
    {
      "start": 389.72,
      "end": 395.32,
      "text": "slower than Nvidia's and US export restrictions made those chips very difficult to access.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.12473367487342613,
      "no_speech_prob": 1.8325163435306546e-12,
      "compression_ratio": 1.5813148788927336
    },
    {
      "start": 395.32,
      "end": 400.4,
      "text": "China was in an unfair position. So first, they tried to bypass the restrictions through",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.12473367487342613,
      "no_speech_prob": 1.8325163435306546e-12,
      "compression_ratio": 1.5813148788927336
    },
    {
      "start": 400.4,
      "end": 406.12,
      "text": "smuggling from places such as Malaysia, Japan and Hong Kong. Some even used human couriers",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.12473367487342613,
      "no_speech_prob": 1.8325163435306546e-12,
      "compression_ratio": 1.5813148788927336
    },
    {
      "start": 406.12,
      "end": 411.82,
      "text": "to smuggle in chips. A student was paid $100 for each of the six chips he carried from",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.12473367487342613,
      "no_speech_prob": 1.8325163435306546e-12,
      "compression_ratio": 1.5813148788927336
    },
    {
      "start": 411.82,
      "end": 417.74,
      "text": "Singapore to China in his luggage. But eventually, all this barely scratched the surface in helping",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.12473367487342613,
      "no_speech_prob": 1.8325163435306546e-12,
      "compression_ratio": 1.5813148788927336
    },
    {
      "start": 417.74,
      "end": 423.52,
      "text": "China grow its AI industry. So Chinese engineers asked a different question instead of just",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08679641723632812,
      "no_speech_prob": 2.20176437845665e-12,
      "compression_ratio": 1.5551724137931036
    },
    {
      "start": 423.52,
      "end": 428.98,
      "text": "building bigger computers with better chips. What if we build smarter AI? This is where",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08679641723632812,
      "no_speech_prob": 2.20176437845665e-12,
      "compression_ratio": 1.5551724137931036
    },
    {
      "start": 428.98,
      "end": 434.02,
      "text": "companies like DeepSeek completely changed the conversation. DeepSeek used a technique",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08679641723632812,
      "no_speech_prob": 2.20176437845665e-12,
      "compression_ratio": 1.5551724137931036
    },
    {
      "start": 434.02,
      "end": 439.72,
      "text": "called mixture of experts. Let me explain this using an oversimplified example. Say you",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08679641723632812,
      "no_speech_prob": 2.20176437845665e-12,
      "compression_ratio": 1.5551724137931036
    },
    {
      "start": 439.72,
      "end": 444.66,
      "text": "have a company of 200 people and on a particular day when you want to solve a marketing problem,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.08679641723632812,
      "no_speech_prob": 2.20176437845665e-12,
      "compression_ratio": 1.5551724137931036
    },
    {
      "start": 444.66,
      "end": 448.52,
      "text": "you don't call everyone into the meeting. You only call the marketing team, right?",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06759779381029533,
      "no_speech_prob": 3.3835026794215572e-12,
      "compression_ratio": 1.641025641025641
    },
    {
      "start": 448.86,
      "end": 453.98,
      "text": "Previous AI models often worked differently. Every time you asked a question, the AI would",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06759779381029533,
      "no_speech_prob": 3.3835026794215572e-12,
      "compression_ratio": 1.641025641025641
    },
    {
      "start": 453.98,
      "end": 459.18,
      "text": "call the whole team. Basically, the whole model would calculate the answer. That required huge",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06759779381029533,
      "no_speech_prob": 3.3835026794215572e-12,
      "compression_ratio": 1.641025641025641
    },
    {
      "start": 459.18,
      "end": 464.74,
      "text": "amounts of power and made each response more expensive. To address this, American companies",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06759779381029533,
      "no_speech_prob": 3.3835026794215572e-12,
      "compression_ratio": 1.641025641025641
    },
    {
      "start": 464.74,
      "end": 470.18,
      "text": "like OpenAI explored the use of a mixture of experts, which basically divided the model",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.06759779381029533,
      "no_speech_prob": 3.3835026794215572e-12,
      "compression_ratio": 1.641025641025641
    },
    {
      "start": 470.18,
      "end": 472.82,
      "text": "into multiple specialized expert clusters.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.31751705470838043,
      "no_speech_prob": 3.4361966397278287e-12,
      "compression_ratio": 1.3658536585365855
    },
    {
      "start": 473.22,
      "end": 476.82,
      "text": "However, DeepSeq engineers took this concept to the next level.",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.31751705470838043,
      "no_speech_prob": 3.4361966397278287e-12,
      "compression_ratio": 1.3658536585365855
    },
    {
      "start": 477.16,
      "end": 480.0,
      "text": "Instead of dividing the model into dozens of expert clusters,",
      "chunk": 0,
      "language": "en",
      "avg_logprob": -0.31751705470838043,
      "no_speech_prob": 3.4361966397278287e-12,
      "compression_ratio": 1.3658536585365855
    },
    {
      "start": 480.0,
      "end": 484.12,
      "text": "Instead of dividing the model into dozens of expert clusters, they sliced it into 256 tiny",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09596133232116699,
      "no_speech_prob": 1.5310030803891705e-12,
      "compression_ratio": 1.6118143459915613
    },
    {
      "start": 484.12,
      "end": 490.92,
      "text": "hyper-specialized experts. So when you ask DeepSeq to solve a coding problem, an ultra-efficient route",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09596133232116699,
      "no_speech_prob": 1.5310030803891705e-12,
      "compression_ratio": 1.6118143459915613
    },
    {
      "start": 490.92,
      "end": 496.9,
      "text": "instantly kicks in and activates just eight of those tiny experts. The rest stayed asleep.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.20527793482730264,
      "no_speech_prob": 2.729476150306609e-12,
      "compression_ratio": 1.58008658008658
    },
    {
      "start": 496.9,
      "end": 502.92,
      "text": "That means DeepSeek can produce strong answers while using far cheaper costs. Look at this.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.20527793482730264,
      "no_speech_prob": 2.729476150306609e-12,
      "compression_ratio": 1.58008658008658
    },
    {
      "start": 502.92,
      "end": 509.36,
      "text": "While the best models of GPT and Claude cost $5 and $10 per million tokens in input costs,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.20527793482730264,
      "no_speech_prob": 2.729476150306609e-12,
      "compression_ratio": 1.58008658008658
    },
    {
      "start": 509.36,
      "end": 516.08,
      "text": "respectively, the best model of DeepSeek costs only $0.4. And while GPT and Claude cost $30",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.20527793482730264,
      "no_speech_prob": 2.729476150306609e-12,
      "compression_ratio": 1.58008658008658
    },
    {
      "start": 516.08,
      "end": 522.92,
      "text": "$50 respectively in output costs, DeepSeek only cost $0.87. But that wasn't all. The",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1387905188366375,
      "no_speech_prob": 1.8541147350087384e-12,
      "compression_ratio": 1.4769736842105263
    },
    {
      "start": 522.92,
      "end": 527.44,
      "text": "Chinese didn't stop with their innovation there. They also invented a technique called",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1387905188366375,
      "no_speech_prob": 1.8541147350087384e-12,
      "compression_ratio": 1.4769736842105263
    },
    {
      "start": 527.44,
      "end": 533.98,
      "text": "Multi-Head Latent Attention or MLA. To understand what this does, let's take another example.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1387905188366375,
      "no_speech_prob": 1.8541147350087384e-12,
      "compression_ratio": 1.4769736842105263
    },
    {
      "start": 533.98,
      "end": 539.96,
      "text": "Imagine you're talking to an AI about a 100-page document. Every time you ask a new question,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1387905188366375,
      "no_speech_prob": 1.8541147350087384e-12,
      "compression_ratio": 1.4769736842105263
    },
    {
      "start": 539.96,
      "end": 545.02,
      "text": "the AI needs to remember the important details from everything it has already read. So it",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1387905188366375,
      "no_speech_prob": 1.8541147350087384e-12,
      "compression_ratio": 1.4769736842105263
    },
    {
      "start": 545.02,
      "end": 548.4,
      "text": "It stores all this information in its temporary memory.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 548.4,
      "end": 553.18,
      "text": "Engineers call this the key value cache, but you can simply think of it as the AI's short-term",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 553.18,
      "end": 554.18,
      "text": "memory.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 554.18,
      "end": 558.68,
      "text": "The longer the conversation gets, the more memory the AI needs, which makes it expensive",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 558.68,
      "end": 559.68,
      "text": "to run.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 559.68,
      "end": 563.18,
      "text": "DeepSeq's technique, MLA, compresses that memory.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 563.18,
      "end": 568.26,
      "text": "Instead of storing every detail separately, it creates a much smaller summary of the important",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 568.26,
      "end": 572.2,
      "text": "information and uses that to answer future questions.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.12870687192624752,
      "no_speech_prob": 2.4559813818075815e-12,
      "compression_ratio": 1.6272401433691757
    },
    {
      "start": 572.2,
      "end": 578.3,
      "text": "It is similar to replacing 100 pages of information with a few pages of well-organized notes.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.10921555625067816,
      "no_speech_prob": 8.10036933363012e-13,
      "compression_ratio": 1.5077519379844961
    },
    {
      "start": 578.3,
      "end": 584.2,
      "text": "This reduces the model's short-term memory requirements by over 90%, allowing it to handle",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.10921555625067816,
      "no_speech_prob": 8.10036933363012e-13,
      "compression_ratio": 1.5077519379844961
    },
    {
      "start": 584.2,
      "end": 588.26,
      "text": "longer conversations while using far less computing power.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.10921555625067816,
      "no_speech_prob": 8.10036933363012e-13,
      "compression_ratio": 1.5077519379844961
    },
    {
      "start": 588.26,
      "end": 591.56,
      "text": "That makes the model cheaper and more efficient to run.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.10921555625067816,
      "no_speech_prob": 8.10036933363012e-13,
      "compression_ratio": 1.5077519379844961
    },
    {
      "start": 591.56,
      "end": 596.86,
      "text": "So with all these incredible tricks, DeepSeek managed to build a world-class AI model for",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.10921555625067816,
      "no_speech_prob": 8.10036933363012e-13,
      "compression_ratio": 1.5077519379844961
    },
    {
      "start": 596.86,
      "end": 602.54,
      "text": "reportedly just under 6 million dollars while Anthropic and other large companies were spending",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.15180121297421661,
      "no_speech_prob": 8.966276614404134e-13,
      "compression_ratio": 1.5298804780876494
    },
    {
      "start": 602.54,
      "end": 608.46,
      "text": "billions of dollars to train their models. And if you look at the comparison of its V4 Pro Max",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.15180121297421661,
      "no_speech_prob": 8.966276614404134e-13,
      "compression_ratio": 1.5298804780876494
    },
    {
      "start": 608.46,
      "end": 617.66,
      "text": "model with comparable launches back then like Opus 4.6 Max, GPT 5.4 and Gemini 3.1 Pro, you'll see",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.15180121297421661,
      "no_speech_prob": 8.966276614404134e-13,
      "compression_ratio": 1.5298804780876494
    },
    {
      "start": 617.66,
      "end": 623.34,
      "text": "that its performance is comparable. So that was all about creating efficient software but this",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.15180121297421661,
      "no_speech_prob": 8.966276614404134e-13,
      "compression_ratio": 1.5298804780876494
    },
    {
      "start": 623.34,
      "end": 628.94,
      "text": "only solves half the problem because even the best software needs enough hardware from chips",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.042091699746938854,
      "no_speech_prob": 2.4658183481185425e-12,
      "compression_ratio": 1.6916299559471366
    },
    {
      "start": 628.94,
      "end": 635.34,
      "text": "to data centers to train and run ai models on for those who don't know a data center is a physical",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.042091699746938854,
      "no_speech_prob": 2.4658183481185425e-12,
      "compression_ratio": 1.6916299559471366
    },
    {
      "start": 635.34,
      "end": 642.22,
      "text": "facility that contains the servers and ai chips that run websites apps cloud services and ai",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.042091699746938854,
      "no_speech_prob": 2.4658183481185425e-12,
      "compression_ratio": 1.6916299559471366
    },
    {
      "start": 642.22,
      "end": 649.02,
      "text": "models along with the storage networking backup power and cooling systems needed to keep everything",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.042091699746938854,
      "no_speech_prob": 2.4658183481185425e-12,
      "compression_ratio": 1.6916299559471366
    },
    {
      "start": 649.02,
      "end": 654.62,
      "text": "running. And so China was not only short of cutting-edge AI chips, they also found that",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09718439625758751,
      "no_speech_prob": 1.748606684795484e-12,
      "compression_ratio": 1.5918367346938775
    },
    {
      "start": 654.62,
      "end": 659.98,
      "text": "the newly built data centers were actually underused. So China decided to treat its AI",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09718439625758751,
      "no_speech_prob": 1.748606684795484e-12,
      "compression_ratio": 1.5918367346938775
    },
    {
      "start": 659.98,
      "end": 665.66,
      "text": "infrastructure in the same way a country treats roads or electricity, as public infrastructure",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09718439625758751,
      "no_speech_prob": 1.748606684795484e-12,
      "compression_ratio": 1.5918367346938775
    },
    {
      "start": 665.66,
      "end": 671.04,
      "text": "that everyone can use. The government is building something called the National Integrated Computing",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09718439625758751,
      "no_speech_prob": 1.748606684795484e-12,
      "compression_ratio": 1.5918367346938775
    },
    {
      "start": 671.04,
      "end": 678.04,
      "text": "Power Network by 2030, or even earlier. To explain it simply, it built many large data centers in",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.09718439625758751,
      "no_speech_prob": 1.748606684795484e-12,
      "compression_ratio": 1.5918367346938775
    },
    {
      "start": 678.04,
      "end": 683.04,
      "text": "In regions where land was cheaper, renewable energy was more abundant and there was space",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 683.04,
      "end": 684.5,
      "text": "for huge facilities.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 684.5,
      "end": 690.64,
      "text": "China then connected these centers to businesses and cities through high-speed fiber networks,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 690.64,
      "end": 696.36,
      "text": "allowing businesses, universities and researchers to rent computing power when they need it.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 696.36,
      "end": 698.38,
      "text": "Here is how it works using an example.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 698.38,
      "end": 703.24,
      "text": "Imagine a small AI startup in Shanghai wants to train a new AI model.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1681988855426231,
      "no_speech_prob": 2.04428206133922e-12,
      "compression_ratio": 1.608695652173913
    },
    {
      "start": 703.24,
      "end": 708.36,
      "text": "Buying hundreds of GPUs and building its own data center could cost a lot of money.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 708.36,
      "end": 711.06,
      "text": "Instead, it can rent computing power online.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 711.06,
      "end": 715.4,
      "text": "The startup uploads its code and data and chooses how much power it needs.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 715.4,
      "end": 720.24,
      "text": "And the platform finds available chips in connected data centers possibly thousands",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 720.24,
      "end": 721.6,
      "text": "of kilometers away.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 721.6,
      "end": 726.2,
      "text": "Its engineers can monitor the training remotely and download the finished model once the work",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 726.2,
      "end": 727.2,
      "text": "is complete.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 727.2,
      "end": 732.34,
      "text": "China has not built one fully connected national computing grid yet, but that is the goal.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11824512481689453,
      "no_speech_prob": 1.8325495201171327e-12,
      "compression_ratio": 1.6721854304635762
    },
    {
      "start": 732.34,
      "end": 736.28,
      "text": "computing power easier and cheaper to rent across the country.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 736.28,
      "end": 739.6,
      "text": "This brings us to Problem Number 4 Training the Models",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 739.6,
      "end": 745.18,
      "text": "Because even the smartest AI models are useless without enough data to train them. And for",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 745.18,
      "end": 748.94,
      "text": "the next generation of AI, text alone is not enough.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 748.94,
      "end": 754.14,
      "text": "Video models need to understand how people move, how objects interact, and how the physical",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 754.14,
      "end": 759.38,
      "text": "world behaves. You cannot teach an AI what running looks like by only describing it in",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17236592071224946,
      "no_speech_prob": 1.741314124142912e-12,
      "compression_ratio": 1.5714285714285714
    },
    {
      "start": 759.38,
      "end": 764.56,
      "text": "words. It needs to watch millions of examples of different people, camera angles, lighting",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11180431063812558,
      "no_speech_prob": 2.1340466286462467e-12,
      "compression_ratio": 1.605072463768116
    },
    {
      "start": 764.56,
      "end": 769.88,
      "text": "conditions, environments and movements. It needs to see how clothes move, how shadows",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11180431063812558,
      "no_speech_prob": 2.1340466286462467e-12,
      "compression_ratio": 1.605072463768116
    },
    {
      "start": 769.88,
      "end": 774.94,
      "text": "change and what happens when someone steps into a puddle. And this is where China has",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11180431063812558,
      "no_speech_prob": 2.1340466286462467e-12,
      "compression_ratio": 1.605072463768116
    },
    {
      "start": 774.94,
      "end": 780.5,
      "text": "a major advantage. China has one of the largest digital ecosystems in the world. Platforms",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11180431063812558,
      "no_speech_prob": 2.1340466286462467e-12,
      "compression_ratio": 1.605072463768116
    },
    {
      "start": 780.5,
      "end": 787.06,
      "text": "like Douyin, WeChat, TikTok and other Chinese apps generate huge amounts of text, images,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.11180431063812558,
      "no_speech_prob": 2.1340466286462467e-12,
      "compression_ratio": 1.605072463768116
    },
    {
      "start": 787.06,
      "end": 792.5,
      "text": "audio and video every day. For companies like ByteDance, which owns the popular AI platform",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17772633320576436,
      "no_speech_prob": 2.27172989621105e-12,
      "compression_ratio": 1.6392857142857142
    },
    {
      "start": 792.5,
      "end": 799.5,
      "text": "C-Dance, this is not just content, it is valuable training data. ByteDance owns Dooyan and TikTok,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17772633320576436,
      "no_speech_prob": 2.27172989621105e-12,
      "compression_ratio": 1.6392857142857142
    },
    {
      "start": 799.5,
      "end": 804.8,
      "text": "along with several tools for creating, editing and recommending videos. This gives the company",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17772633320576436,
      "no_speech_prob": 2.27172989621105e-12,
      "compression_ratio": 1.6392857142857142
    },
    {
      "start": 804.8,
      "end": 809.9,
      "text": "access to more than just the video itself. It can also study how people react to that",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17772633320576436,
      "no_speech_prob": 2.27172989621105e-12,
      "compression_ratio": 1.6392857142857142
    },
    {
      "start": 809.9,
      "end": 814.72,
      "text": "video, which clips do people keep watching and at what moment do viewers lose interest.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17772633320576436,
      "no_speech_prob": 2.27172989621105e-12,
      "compression_ratio": 1.6392857142857142
    },
    {
      "start": 814.72,
      "end": 819.18,
      "text": "Now this does not mean ByteDance uses every private video to train its AI.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 819.18,
      "end": 824.48,
      "text": "We do not know exactly what data each company uses, but owning some of the world's largest",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 824.48,
      "end": 828.76,
      "text": "video platforms still gives Chinese companies a powerful advantage.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 828.76,
      "end": 832.3,
      "text": "Because if chips are the engine of AI, data is the fuel.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 832.3,
      "end": 837.72,
      "text": "And China did not just use that fuel to build competitive AI models, it also started releasing",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 837.72,
      "end": 839.22,
      "text": "many of them for free.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13108702380247791,
      "no_speech_prob": 1.2595244942500972e-12,
      "compression_ratio": 1.5632183908045978
    },
    {
      "start": 839.22,
      "end": 846.42,
      "text": "Today, models like DeepSeek, Quen by Alibaba, Kimi by Moonshot AI, and GLM models from Z.AI",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 846.42,
      "end": 849.56,
      "text": "are available completely free to download and run.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 849.68,
      "end": 851.24,
      "text": "And this raises an obvious question.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 851.56,
      "end": 854.04,
      "text": "Running AI companies can cost billions of dollars.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 854.36,
      "end": 859.12,
      "text": "In fact, many major AI companies are still spending far more money than they earn.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 859.12,
      "end": 863.26,
      "text": "So how can Chinese companies offer powerful AI models for free?",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 863.52,
      "end": 865.94,
      "text": "If users aren't paying these companies, who is?",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13739595242909022,
      "no_speech_prob": 1.5674170945542687e-12,
      "compression_ratio": 1.5625
    },
    {
      "start": 865.94,
      "end": 870.1,
      "text": "The answer is that Chinese AI companies aren't really selling AI models.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 870.46,
      "end": 871.88,
      "text": "The model is only the entry point.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 872.0,
      "end": 873.18,
      "text": "They're selling something much bigger.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 873.52,
      "end": 876.34,
      "text": "1. They're selling the ecosystem, not the model.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 876.74,
      "end": 877.48,
      "text": "Think about Google.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 877.84,
      "end": 879.5,
      "text": "Google gives away Chrome for free.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 879.82,
      "end": 881.62,
      "text": "Not because browsers are cheap to build,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 881.9,
      "end": 884.64,
      "text": "but because Chrome brings people into Google's ecosystem,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 885.0,
      "end": 889.1,
      "text": "where they eventually use advertising and paid services like Google Cloud.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 889.46,
      "end": 892.32,
      "text": "Chinese AI companies are following the same playbook.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.1291790426823131,
      "no_speech_prob": 2.1422089362815866e-12,
      "compression_ratio": 1.7445255474452555
    },
    {
      "start": 892.32,
      "end": 899.5,
      "text": "While Western companies like OpenAI and Anthropic primarily make money by selling access to their intelligence directly.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13055339455604553,
      "no_speech_prob": 2.871476927762684e-12,
      "compression_ratio": 1.5807692307692307
    },
    {
      "start": 899.84,
      "end": 904.04,
      "text": "You pay for a subscription or an API call or how many tokens you consume, right?",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13055339455604553,
      "no_speech_prob": 2.871476927762684e-12,
      "compression_ratio": 1.5807692307692307
    },
    {
      "start": 904.44,
      "end": 907.88,
      "text": "Chinese companies often make money later in the user journey.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13055339455604553,
      "no_speech_prob": 2.871476927762684e-12,
      "compression_ratio": 1.5807692307692307
    },
    {
      "start": 908.24,
      "end": 912.38,
      "text": "For example, Alibaba not only wants you to chat with their AI model, Quen,",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13055339455604553,
      "no_speech_prob": 2.871476927762684e-12,
      "compression_ratio": 1.5807692307692307
    },
    {
      "start": 912.78,
      "end": 916.56,
      "text": "it wants companies to build and run their applications on Alibaba Cloud.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.13055339455604553,
      "no_speech_prob": 2.871476927762684e-12,
      "compression_ratio": 1.5807692307692307
    },
    {
      "start": 916.56,
      "end": 922.6,
      "text": "Tencent, the company behind the super app WeChat, not only wants businesses to test its",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 922.6,
      "end": 924.12,
      "text": "AI model, Hanuan.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 924.12,
      "end": 930.18,
      "text": "It wants them to use Tencent Cloud, its enterprise software and the rest of its business ecosystem.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 930.18,
      "end": 936.88,
      "text": "It needs GPU, servers, storage, security, monitoring, technical support and engineers",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 936.88,
      "end": 940.22,
      "text": "who can connect the AI to its existing systems.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 940.22,
      "end": 944.8,
      "text": "That is where the real money can be made, whether it's cloud hosting, consulting or",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.14937579516068245,
      "no_speech_prob": 1.883239441127782e-12,
      "compression_ratio": 1.6145038167938932
    },
    {
      "start": 944.8,
      "end": 950.86,
      "text": "even custom AI solutions. So right now the model may be cheap or even free to download",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17202190183243662,
      "no_speech_prob": 9.143178835574473e-13,
      "compression_ratio": 1.4625
    },
    {
      "start": 950.86,
      "end": 956.56,
      "text": "but everything required to operate it reliably at scale is not. The model is not always the",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17202190183243662,
      "no_speech_prob": 9.143178835574473e-13,
      "compression_ratio": 1.4625
    },
    {
      "start": 956.56,
      "end": 960.0,
      "text": "final product, it is the customer acquisition strategy.",
      "chunk": 1,
      "language": "en",
      "avg_logprob": -0.17202190183243662,
      "no_speech_prob": 9.143178835574473e-13,
      "compression_ratio": 1.4625
    },
    {
      "start": 961.82,
      "end": 966.68,
      "text": "these models become paid later, because by then many users may have no other option but",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1778410556269627,
      "no_speech_prob": 1.4840629810122796e-12,
      "compression_ratio": 1.602189781021898
    },
    {
      "start": 966.68,
      "end": 967.68,
      "text": "to pay.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1778410556269627,
      "no_speech_prob": 1.4840629810122796e-12,
      "compression_ratio": 1.602189781021898
    },
    {
      "start": 967.68,
      "end": 972.08,
      "text": "Reason 2 Win with distribution Remember, China couldn't compete with America on",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1778410556269627,
      "no_speech_prob": 1.4840629810122796e-12,
      "compression_ratio": 1.602189781021898
    },
    {
      "start": 972.08,
      "end": 977.76,
      "text": "the top models that Anthropic and OpenAI were constantly innovating, using better chips",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10872350954541973,
      "no_speech_prob": 2.0126545829252107e-12,
      "compression_ratio": 1.6082474226804124
    },
    {
      "start": 977.76,
      "end": 982.96,
      "text": "and coming out ahead. So China instead decided to win with distribution and make their products",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10872350954541973,
      "no_speech_prob": 2.0126545829252107e-12,
      "compression_ratio": 1.6082474226804124
    },
    {
      "start": 982.96,
      "end": 988.16,
      "text": "free and open weight. This basically means that developers can download the trained model,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10872350954541973,
      "no_speech_prob": 2.0126545829252107e-12,
      "compression_ratio": 1.6082474226804124
    },
    {
      "start": 988.16,
      "end": 994.72,
      "text": "run it on their own computers or servers, modify it and fine tune it for their own use. For example,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10872350954541973,
      "no_speech_prob": 2.0126545829252107e-12,
      "compression_ratio": 1.6082474226804124
    },
    {
      "start": 994.72,
      "end": 999.92,
      "text": "a retail company could take the model and customize it by teaching it its internal policies,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10872350954541973,
      "no_speech_prob": 2.0126545829252107e-12,
      "compression_ratio": 1.6082474226804124
    },
    {
      "start": 999.92,
      "end": 1004.8,
      "text": "product catalog and more so it can accurately handle customer support queries for it.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11250621832690193,
      "no_speech_prob": 2.0845842413352855e-12,
      "compression_ratio": 1.6798561151079137
    },
    {
      "start": 1004.8,
      "end": 1010.72,
      "text": "It's a win for companies but it's also a win for the Chinese companies. If millions of developers",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11250621832690193,
      "no_speech_prob": 2.0845842413352855e-12,
      "compression_ratio": 1.6798561151079137
    },
    {
      "start": 1010.72,
      "end": 1016.16,
      "text": "can freely experiment with your model, they begin building apps on top of it. And every time a",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11250621832690193,
      "no_speech_prob": 2.0845842413352855e-12,
      "compression_ratio": 1.6798561151079137
    },
    {
      "start": 1016.16,
      "end": 1021.68,
      "text": "developer chooses your model instead of a competitor's, they invest time learning how it works,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11250621832690193,
      "no_speech_prob": 2.0845842413352855e-12,
      "compression_ratio": 1.6798561151079137
    },
    {
      "start": 1021.68,
      "end": 1026.56,
      "text": "writing code around it and integrating it into their systems. That makes them less likely to",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11250621832690193,
      "no_speech_prob": 2.0845842413352855e-12,
      "compression_ratio": 1.6798561151079137
    },
    {
      "start": 1026.56,
      "end": 1032.64,
      "text": "switch. It's like if you invested time and effort on Twitter and built a 100,000 following,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.12437495998307771,
      "no_speech_prob": 1.3095323436740358e-12,
      "compression_ratio": 1.6317689530685922
    },
    {
      "start": 1032.64,
      "end": 1037.3,
      "text": "you're less likely to switch to threads. Similarly, it works for developers using these",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.12437495998307771,
      "no_speech_prob": 1.3095323436740358e-12,
      "compression_ratio": 1.6317689530685922
    },
    {
      "start": 1037.3,
      "end": 1042.82,
      "text": "models. Secondly, researchers will be able to effortlessly study these models and publish",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.12437495998307771,
      "no_speech_prob": 1.3095323436740358e-12,
      "compression_ratio": 1.6317689530685922
    },
    {
      "start": 1042.82,
      "end": 1048.66,
      "text": "improvements or techniques that others can reuse. Over time, this creates a network effect.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.12437495998307771,
      "no_speech_prob": 1.3095323436740358e-12,
      "compression_ratio": 1.6317689530685922
    },
    {
      "start": 1048.66,
      "end": 1054.24,
      "text": "The more people use the model, the more tools, tutorials and expertise exist around it. It",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.12437495998307771,
      "no_speech_prob": 1.3095323436740358e-12,
      "compression_ratio": 1.6317689530685922
    },
    {
      "start": 1054.24,
      "end": 1059.18,
      "text": "That is the same basic reason Android became the world's largest mobile operating system.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1059.18,
      "end": 1064.44,
      "text": "Google allowed thousands of manufacturers and millions of developers to build around it.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1064.44,
      "end": 1068.78,
      "text": "And once enough, people are building on your platform, your ecosystem can grow faster than",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1068.78,
      "end": 1070.3,
      "text": "any single competitor.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1070.3,
      "end": 1073.28,
      "text": "That strategy is already working for Chinese companies.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1073.28,
      "end": 1077.08,
      "text": "Look at this chart that's showing the world's top 50 most used AI models.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1077.08,
      "end": 1080.88,
      "text": "You'll see that Chinese models are rapidly gaining market share.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.13305438813709078,
      "no_speech_prob": 1.3148297554888e-12,
      "compression_ratio": 1.6125827814569536
    },
    {
      "start": 1080.88,
      "end": 1085.96,
      "text": "And when it comes to the top 20 AI models in the world, between May and June of 2026,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1085.96,
      "end": 1090.76,
      "text": "the monthly token usage of Chinese AI models has surpassed the US models.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1090.76,
      "end": 1092.96,
      "text": "But let's look at more specific examples.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1092.96,
      "end": 1097.58,
      "text": "Alibaba's QEN models surpassed 1 billion total downloads this year.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1097.58,
      "end": 1099.9,
      "text": "Many large US companies also use it.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1099.9,
      "end": 1103.9,
      "text": "For example, Airbnb's AI customer service system uses them.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1103.9,
      "end": 1107.44,
      "text": "Even Pinterest has been customizing QEN models for their usage.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14379454088640642,
      "no_speech_prob": 1.5674231660864346e-12,
      "compression_ratio": 1.5693430656934306
    },
    {
      "start": 1107.44,
      "end": 1113.48,
      "text": "Pinterest actually says it has reduced costs by as much as 90% in some workloads due to",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1113.48,
      "end": 1114.96,
      "text": "such open-source models.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1114.96,
      "end": 1120.38,
      "text": "And Cursor, one of America's fastest-growing AI coding companies built its coding model,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1120.38,
      "end": 1125.94,
      "text": "Composer2, on the base model of Chinese company Moonshot AI's Kimi K2.5.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1125.94,
      "end": 1127.04,
      "text": "Listen to that again.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1127.04,
      "end": 1132.62,
      "text": "A top US AI company built its proprietary product on top of a Chinese foundation model.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1132.62,
      "end": 1136.46,
      "text": "Chinese models are no longer merely cheaper alternatives to American AI.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.1912954783035537,
      "no_speech_prob": 1.5859525064748636e-12,
      "compression_ratio": 1.5758620689655172
    },
    {
      "start": 1136.46,
      "end": 1141.86,
      "text": "are becoming part of the infrastructure on which American companies themselves are building.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14461893385106867,
      "no_speech_prob": 1.5192192122370773e-12,
      "compression_ratio": 1.5353159851301115
    },
    {
      "start": 1141.86,
      "end": 1144.26,
      "text": "Reason 3 And this brings us to one of the biggest",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14461893385106867,
      "no_speech_prob": 1.5192192122370773e-12,
      "compression_ratio": 1.5353159851301115
    },
    {
      "start": 1144.26,
      "end": 1149.96,
      "text": "reasons that Chinese models are succeeding so well. It's because China made AI a national",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14461893385106867,
      "no_speech_prob": 1.5192192122370773e-12,
      "compression_ratio": 1.5353159851301115
    },
    {
      "start": 1149.96,
      "end": 1155.58,
      "text": "mission. In 2017, China's State Council released the New Generation Artificial Intelligence",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14461893385106867,
      "no_speech_prob": 1.5192192122370773e-12,
      "compression_ratio": 1.5353159851301115
    },
    {
      "start": 1155.58,
      "end": 1161.54,
      "text": "Development Plan. Its ambition was enormous. By 2030, China wanted to become the world's",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.14461893385106867,
      "no_speech_prob": 1.5192192122370773e-12,
      "compression_ratio": 1.5353159851301115
    },
    {
      "start": 1161.54,
      "end": 1167.78,
      "text": "leading center for AI innovation. Universities expanded AI and computer science programs.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10015209154649214,
      "no_speech_prob": 1.75528634595995e-12,
      "compression_ratio": 1.6464285714285714
    },
    {
      "start": 1167.78,
      "end": 1173.14,
      "text": "Governments funded laboratories and research projects. Cities and provinces created AI",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10015209154649214,
      "no_speech_prob": 1.75528634595995e-12,
      "compression_ratio": 1.6464285714285714
    },
    {
      "start": 1173.14,
      "end": 1179.06,
      "text": "innovation zones where companies could access funding, infrastructure and local policy support.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10015209154649214,
      "no_speech_prob": 1.75528634595995e-12,
      "compression_ratio": 1.6464285714285714
    },
    {
      "start": 1179.06,
      "end": 1184.34,
      "text": "The plan even reached across borders. Chinese researchers who had built their careers in labs",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10015209154649214,
      "no_speech_prob": 1.75528634595995e-12,
      "compression_ratio": 1.6464285714285714
    },
    {
      "start": 1184.34,
      "end": 1190.34,
      "text": "overseas were actively encouraged to come back home, pulled in with grants, funding and a real",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10015209154649214,
      "no_speech_prob": 1.75528634595995e-12,
      "compression_ratio": 1.6464285714285714
    },
    {
      "start": 1190.34,
      "end": 1195.18,
      "text": "path to do their work in China. But the best thing China did was to give major companies",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.09437121391296387,
      "no_speech_prob": 1.2893643401221122e-12,
      "compression_ratio": 1.5807560137457044
    },
    {
      "start": 1195.18,
      "end": 1201.62,
      "text": "specific roles in building national AI platforms. In 2017, the Ministry of Science and Technology",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.09437121391296387,
      "no_speech_prob": 1.2893643401221122e-12,
      "compression_ratio": 1.5807560137457044
    },
    {
      "start": 1201.62,
      "end": 1207.6,
      "text": "selected Baidu to develop an open innovation platform for autonomous driving, Alibaba for",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.09437121391296387,
      "no_speech_prob": 1.2893643401221122e-12,
      "compression_ratio": 1.5807560137457044
    },
    {
      "start": 1207.6,
      "end": 1213.18,
      "text": "smart cities, Tencent for medical imaging, and iFlytech for intelligent voice technology.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.09437121391296387,
      "no_speech_prob": 1.2893643401221122e-12,
      "compression_ratio": 1.5807560137457044
    },
    {
      "start": 1213.5,
      "end": 1218.16,
      "text": "These companies were still private businesses, and they continued competing fiercely with one",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.09437121391296387,
      "no_speech_prob": 1.2893643401221122e-12,
      "compression_ratio": 1.5807560137457044
    },
    {
      "start": 1218.16,
      "end": 1223.1,
      "text": "another. Yet they were all working in the same direction for the growth of China. So this brings",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.05906537148804791,
      "no_speech_prob": 5.279938812552398e-12,
      "compression_ratio": 1.6372881355932203
    },
    {
      "start": 1223.1,
      "end": 1227.78,
      "text": "us to the last part of the case study. What can the world learn from China's AI journey? Let's",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.05906537148804791,
      "no_speech_prob": 5.279938812552398e-12,
      "compression_ratio": 1.6372881355932203
    },
    {
      "start": 1227.78,
      "end": 1232.98,
      "text": "just clarify one thing first. China did everything in its power to get ahead in the AI race. Some",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.05906537148804791,
      "no_speech_prob": 5.279938812552398e-12,
      "compression_ratio": 1.6372881355932203
    },
    {
      "start": 1232.98,
      "end": 1238.38,
      "text": "American companies claim Chinese labs use distillation. This means using the answers of a stronger AI",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.05906537148804791,
      "no_speech_prob": 5.279938812552398e-12,
      "compression_ratio": 1.6372881355932203
    },
    {
      "start": 1238.38,
      "end": 1242.82,
      "text": "model to help train a smaller one. Then we know that there was smuggling done of restricted",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.05906537148804791,
      "no_speech_prob": 5.279938812552398e-12,
      "compression_ratio": 1.6372881355932203
    },
    {
      "start": 1242.82,
      "end": 1248.74,
      "text": "Nvidia chips through black market routes. The US charges a co-founder of Supermicro with illegally",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11984812979604684,
      "no_speech_prob": 3.47728985362894e-12,
      "compression_ratio": 1.5922330097087378
    },
    {
      "start": 1248.74,
      "end": 1254.62,
      "text": "diverting billions of dollars of Nvidia powered servers to China. Some companies may have broken",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11984812979604684,
      "no_speech_prob": 3.47728985362894e-12,
      "compression_ratio": 1.5922330097087378
    },
    {
      "start": 1254.62,
      "end": 1259.56,
      "text": "rules and that should be investigated but a few smuggled chips cannot build thousands of researchers,",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11984812979604684,
      "no_speech_prob": 3.47728985362894e-12,
      "compression_ratio": 1.5922330097087378
    },
    {
      "start": 1259.94,
      "end": 1265.3,
      "text": "large data centers, new engineering techniques, cloud platforms and an ecosystem used by developers",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11984812979604684,
      "no_speech_prob": 3.47728985362894e-12,
      "compression_ratio": 1.5922330097087378
    },
    {
      "start": 1265.3,
      "end": 1270.74,
      "text": "around the world. Also China did not succeed because of one shortcut. It also invested heavily",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.11984812979604684,
      "no_speech_prob": 3.47728985362894e-12,
      "compression_ratio": 1.5922330097087378
    },
    {
      "start": 1270.74,
      "end": 1275.44,
      "text": "in research and found ways to make weaker hardware perform better. The point is that",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1275.44,
      "end": 1279.98,
      "text": "we should judge both sides by the same standard and not dismiss all of China's progress as",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1279.98,
      "end": 1280.98,
      "text": "theft.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1280.98,
      "end": 1285.54,
      "text": "Secondly, people think using Chinese AI is a threat to privacy, they will steal data",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1285.54,
      "end": 1290.08,
      "text": "and the models are censored by Beijing. But the West is not far behind. ChatGPT trains",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1290.08,
      "end": 1294.76,
      "text": "on your conversations by default. Gemini merges your prompts with your Gmail, your YouTube",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1294.76,
      "end": 1299.82,
      "text": "and your entire Google history. Claude quietly changed its privacy policy and opted users",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.10542420196533203,
      "no_speech_prob": 2.0440981806507663e-12,
      "compression_ratio": 1.5828402366863905
    },
    {
      "start": 1299.82,
      "end": 1304.58,
      "text": "into 5 years of data retention. The idea that Chinese AI is a threat and Western AI is the",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.15604128586618524,
      "no_speech_prob": 2.1508279098719774e-12,
      "compression_ratio": 1.6056910569105691
    },
    {
      "start": 1304.58,
      "end": 1306.44,
      "text": "safe choice is completely wrong.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.15604128586618524,
      "no_speech_prob": 2.1508279098719774e-12,
      "compression_ratio": 1.6056910569105691
    },
    {
      "start": 1306.44,
      "end": 1310.48,
      "text": "With this in mind, I have two major lessons that countries need to learn from China's",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.15604128586618524,
      "no_speech_prob": 2.1508279098719774e-12,
      "compression_ratio": 1.6056910569105691
    },
    {
      "start": 1310.48,
      "end": 1316.98,
      "text": "AI race. First, every country needs some control over its AI future. Have you heard of sovereign",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.15604128586618524,
      "no_speech_prob": 2.1508279098719774e-12,
      "compression_ratio": 1.6056910569105691
    },
    {
      "start": 1316.98,
      "end": 1322.7,
      "text": "AI? It does not mean every country must build every chip and every model by itself. That",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.15604128586618524,
      "no_speech_prob": 2.1508279098719774e-12,
      "compression_ratio": 1.6056910569105691
    },
    {
      "start": 1322.7,
      "end": 1324.7,
      "text": "would be too expensive and unrealistic for you.",
      "chunk": 2,
      "language": "en",
      "avg_logprob": -0.534008264541626,
      "no_speech_prob": 2.044145668705921e-12,
      "compression_ratio": 0.8545454545454545
    },
    {
      "start": 1326.045,
      "end": 1332.045,
      "text": "least, a country should have enough of its own computing power, research talent, data,",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.12448991775512695,
      "no_speech_prob": 2.8161435856877137e-12,
      "compression_ratio": 1.6481481481481481
    },
    {
      "start": 1332.045,
      "end": 1337.765,
      "text": "models and infrastructure that it is not completely dependent on another country. Today, most",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.12448991775512695,
      "no_speech_prob": 2.8161435856877137e-12,
      "compression_ratio": 1.6481481481481481
    },
    {
      "start": 1337.765,
      "end": 1342.325,
      "text": "advanced AI technology comes from either the United States or China.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1342.325,
      "end": 1347.705,
      "text": "But what happens if the US decides that only American companies can use its best models?",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1347.705,
      "end": 1350.645,
      "text": "What happens if China makes the same decision?",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1350.645,
      "end": 1354.945,
      "text": "Countries such as India could suddenly lose access to important technology because of",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1354.945,
      "end": 1357.825,
      "text": "a political fight between two other countries.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1357.825,
      "end": 1364.085,
      "text": "And AI is no longer just a chatbot, it is becoming part of education, healthcare, defense,",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.132415771484375,
      "no_speech_prob": 1.7014665501177695e-12,
      "compression_ratio": 1.6090225563909775
    },
    {
      "start": 1364.085,
      "end": 1370.025,
      "text": "science, coding, manufacturing and almost every modern business. A company with better",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.09728891953178075,
      "no_speech_prob": 1.2594798251205908e-12,
      "compression_ratio": 1.6505576208178439
    },
    {
      "start": 1370.025,
      "end": 1375.785,
      "text": "AI can work faster, reduce costs and compete more effectively. A country without access",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.09728891953178075,
      "no_speech_prob": 1.2594798251205908e-12,
      "compression_ratio": 1.6505576208178439
    },
    {
      "start": 1375.785,
      "end": 1381.425,
      "text": "to advanced AI may slowly fall behind. The biggest lesson from China is not that every",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.09728891953178075,
      "no_speech_prob": 1.2594798251205908e-12,
      "compression_ratio": 1.6505576208178439
    },
    {
      "start": 1381.425,
      "end": 1386.885,
      "text": "country should copy China's political system or its exact strategy. The lesson is that",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.09728891953178075,
      "no_speech_prob": 1.2594798251205908e-12,
      "compression_ratio": 1.6505576208178439
    },
    {
      "start": 1386.885,
      "end": 1393.165,
      "text": "a country that depends completely on foreign AI will always be vulnerable to foreign decisions.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.09728891953178075,
      "no_speech_prob": 1.2594798251205908e-12,
      "compression_ratio": 1.6505576208178439
    },
    {
      "start": 1393.165,
      "end": 1398.325,
      "text": "India and other countries need to build enough of their own AI ecosystem so they are not",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1398.325,
      "end": 1401.285,
      "text": "trapped between the US and China.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1401.285,
      "end": 1406.765,
      "text": "Because the next major geopolitical blockade may not stop oil, weapons or physical goods,",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1406.765,
      "end": 1409.425,
      "text": "it may stop access to intelligence itself.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1409.425,
      "end": 1414.665,
      "text": "And second, AI leadership is not about one great chatbot, it's about creating a whole",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1414.665,
      "end": 1415.665,
      "text": "ecosystem.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1415.665,
      "end": 1420.025,
      "text": "When people talk about the AI race, they usually compare models.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1420.025,
      "end": 1421.785,
      "text": "Is ChatGPT better?",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1421.785,
      "end": 1423.105,
      "text": "Is DeepSeek better?",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13695377773708767,
      "no_speech_prob": 1.7830649066616355e-12,
      "compression_ratio": 1.5833333333333333
    },
    {
      "start": 1423.105,
      "end": 1428.225,
      "text": "which model scores higher on a benchmark. But a model is only the visible part of a",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13320912036699117,
      "no_speech_prob": 1.5311470624376766e-12,
      "compression_ratio": 1.6231884057971016
    },
    {
      "start": 1428.225,
      "end": 1434.905,
      "text": "much larger system. To build strong AI, a country also needs chips to perform calculations,",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13320912036699117,
      "no_speech_prob": 1.5311470624376766e-12,
      "compression_ratio": 1.6231884057971016
    },
    {
      "start": 1434.905,
      "end": 1440.545,
      "text": "electricity to power those chips, and data centers to keep them running. It needs universities",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13320912036699117,
      "no_speech_prob": 1.5311470624376766e-12,
      "compression_ratio": 1.6231884057971016
    },
    {
      "start": 1440.545,
      "end": 1446.025,
      "text": "to train engineers, research labs to invent new techniques, cloud companies to provide",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13320912036699117,
      "no_speech_prob": 1.5311470624376766e-12,
      "compression_ratio": 1.6231884057971016
    },
    {
      "start": 1446.025,
      "end": 1452.385,
      "text": "computing power, and businesses that can turn AI into useful products. China worked on all",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.13320912036699117,
      "no_speech_prob": 1.5311470624376766e-12,
      "compression_ratio": 1.6231884057971016
    },
    {
      "start": 1452.385,
      "end": 1457.885,
      "text": "these areas at the same time. So now, if Europe, India and other developing countries want",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1457.885,
      "end": 1462.665,
      "text": "to secure their futures, they need to step up and build such ecosystems too.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1462.665,
      "end": 1467.385,
      "text": "That was all about why China is winning the AI race and what we can learn from them. For",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1467.385,
      "end": 1472.145,
      "text": "more case studies that help you understand where AI is going, who is gaining power and",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1472.145,
      "end": 1476.725,
      "text": "where the next major opportunities will emerge, subscribe to the channel and watch the video",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1476.725,
      "end": 1482.185,
      "text": "on your screen next because it explains another major shift happening in AI that most people",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.08443905135332527,
      "no_speech_prob": 2.8716094172681617e-12,
      "compression_ratio": 1.6847133757961783
    },
    {
      "start": 1482.185,
      "end": 1483.745,
      "text": "are still completely missing.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.4031370282173157,
      "no_speech_prob": 2.131521955467397e-12,
      "compression_ratio": 0.875
    },
    {
      "start": 1483.745,
      "end": 1484.385,
      "text": "I'll see you there.",
      "chunk": 3,
      "language": "en",
      "avg_logprob": -0.4031370282173157,
      "no_speech_prob": 2.131521955467397e-12,
      "compression_ratio": 0.875
    }
  ],
  "segmented_transcription": {
    "status": "SEGMENTED_TRANSCRIPT_OK",
    "created_at": "2026-08-04T10:50:18",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260804_104928__America_Banned_Nvidia_Chips_to_Stop_China_s_AI_-_It_Backfired/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260804_104928__America_Banned_Nvidia_Chips_to_Stop_China_s_AI_-_It_Backfired/audio_16k_mono.wav",
    "audio_sha256": "6572dec08485ce25281e0ee11b8bba87f2307f3674ee70215095a82da3477483",
    "audio_duration": 1504.725313,
    "model": "mlx-community/whisper-large-v3-turbo",
    "language_hint": "en",
    "condition_on_previous_text": false,
    "chunk_count": 4,
    "resumed_chunks": 0,
    "merged_segment_count": 319,
    "final_segment_end": 1484.385,
    "tail_repeat_run": 1,
    "tail_repeat_text": "Countries such as India could suddenly lose access to important technology because of",
    "qc_ok": true,
    "qc_errors": []
  }
}