Tech Brew Ride Home - Wed. 07/16 – Jensen Gets Bullish

Episode Date: July 16, 2025

Jensen is feeling his oats after the reprieve on China, spilling lots of tea about where he see the AI industry. OpenAI is going after the office and also the storefront, with interesting new integrations. Why is it so hard to create LLM’s in other languages? And a first person account of what its like to work at OpenAI, the culture, the pressure, etc. Links: Nvidia Boss Expects US to Move Fast on First H20 China Licenses (Bloomberg) OpenAI Preps ChatGPT Agents in Challenge to Microsoft Excel and PowerPoint (The Information) OpenAI to take cut of ChatGPT shopping sales in hunt for revenues (Financial Times) Fed up with ChatGPT, Latin America is building its own (Rest Of World) Reflections on OpenAI (Calvin French-Owen) Raiders, Rulers, and Traders: The Horse and the Rise of Empires Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Wednesday, July 16th, 2025. I'm Brian McCullough today. Jensen is feeling his oats after the reprieve on China, spilling lots of tea about where he sees the AI industry. Open AI is going after the office and also the storefront with interesting new integrations. Why is it so hard to create LLMs in other languages and a first person account of what it's like to actually work at Open AI, the culture, the pressure, etc. Here's what you miss today in the world of tech. After seemingly winning back that huge China business, Jensen Wong is feeling his oats. He's running around giving wide-ranging sort of interviews and comments on the state of the AI industry. Quoting Bloomberg, Wang is touring Beijing just days after meeting with U.S. President Donald Trump, who he said wished him a great trip. The NVIDIA CEO turned on the charm during high-profile events from the summit to widely covered meetings with the likes of Jaume's Lejeune and vice-premier-Haye Lefe. the key negotiator for China in trade discussions. At that meeting, they discussed Beijing's

commitment to remain open to foreign investment, Huang said. Huang made it a point to wow his audience, donning a modernized Tang jacket posing with Lai, locally celebrated for creating a smartphone to compete with Apple. On social media, lauding household names from Tencent to Alibaba to Bite Bight Dance. Models like Deepseek, Alibaba, Tencent, Minimax, and Baidu Ernie Bot, our world class, Wong told conference attendees. China's open. open source AI is a catalyst for global progress, giving every country and industry a chance to join the AI revolution, end quote. And quoting CNBC, Wong on Wednesday also praised Chinese companies for taking an open source approach to AI, meaning developers can access the underlying code for free.

Notably, Open AI in the U.S. has not yet taken this approach. Alibaba-backed startup moonshot last week released a new open source model called Kimi K-2 that claims to beat OpenAI's chat GPT and Anthropics clawed on certain coding metrics. Wong added that open-source technology is also key for AI safety and enables international cooperation on standards. Wong also described how AI powers China's consumer techs such as Tencent's WeChat, social media app, Alibaba's Taubo shopping app, bite dances, do yen, short video app, and Maytwan's super convenient delivery, end quote. And quoting the times about the regulatory about face that seemingly has put wind in a sales, quote, I don't think I changed his mind, Mr. Wong said of Mr. Trump.

It's my job to inform the president about what I know very well, which is the technology industry, artificial intelligence, the developments of AI around the world. Nvidia has found itself facing Chinese export controls around a rare earth metal called dysprosium that it uses in many of its chips, but in small quantities. Beijing put controls on the metal, which is refined almost exclusively in China in April. But Mr. Wong said he hadn't discussed that issue with Chinese officials in their meetings this week and suggested that enough dysprosium remains available for Nvidia's needs. needs. The volume we use is not that high in the grand scheme of things. I think the amount of overall

inventory around the world is sufficient for us, he said. Asked whether he had discussed China's rare earth or battery technology restrictions on Wednesday with Chinese officials. Mr. Wong replied with a laconic no, end quote. Wong also said, NVIDIA will, quote, accelerate the recovery of its China chip sales as he expects its U.S. export licenses to, quote, come through very shortly, though also worth noting that U.S. Commerce Secretary Howard Lutnik did admit, NVIDIA's plan to resume its H20 AI chip sales to China is part of U.S. trade negotiations over rare earths and magnets. We put that in the trade deal with the magnets, Lutnik told Reuters, referring to an agreement President Trump made to restart rare earth shipments to U.S. manufacturers.

He did not provide additional detail. So there are still factors here outside of Jensen's control. Open AI is not standing still. Two interesting new initiatives from them. First, sources say Open AI is preparing chat cheapy tuesday. agents to let users create files compatible with things like PowerPoint or Excel, allow them to generate reports and handle tasks involving websites. The better to go after enterprise users, right? Quoting the information. The company has designed buttons below the chat GPT search

bar that direct users to create a spreadsheet or presentation, the person said. It isn't clear when Open AI plans to release the features, but they imply that ChatGPT customers would also be able to download and open these Excel and PowerPoint files using a variety of applications made by companies other than Microsoft. Anyone can create files that are compatible with Excel and PowerPoint because Microsoft has made the formats for those files open source, so OpenAI doesn't need Microsoft's permission to do so. The new chat GPT agents will also help customers generate reports based on corporate or public data or handle repetitive tasks involving websites, such as scheduling and booking appointments. Taken together, these productivity features could

make ChatGPT even more attractive as a business tool, posing a threat to productivity suites sold by the likes of Microsoft and Google. Ironically, ChadGPT runs almost entirely on Microsoft servers due to the company's financial arrangements. Hundreds of millions of people use ChatGPT, including tens of millions of paying subscribers, and OpenAI wants to make the app a gateway through which consumers and enterprises use online services or get work done, end quote. And maybe just as interestingly, the FT says OpenAI is aiming to add a checkout system inside of ChatGPT to ensure users complete transactions within the platform with merchants paying a commission. Quote, the San Francisco-based company currently displays products on the platform

with an option to click-through links to online retailers. It also announced a partnership with Payments Group Shopify in April. According to multiple people familiar with their proposals, it now aims to integrate a checkout system into ChatGBTGBT, which ensures users complete transactions within the platform. Merchants that receive and fulfill orders in this way will pay a commission to OpenAI. The e-commerce pushmarks a strategic shift for the loss-making startup valued at $300 billion, which has made revenue primarily from subscriptions to premium services. Taking a cut of sales from ChatGPT would allow the company to make money from users of its free

version, a so-far untapped source of revenue. OpenAI's move also represents a further threat to Google. business model as consumers increasingly move to AI chatbots to conduct searches and discover products. The feature is still in development, so the details may change. However, OpenAI and partners such as Shopify have been presenting early versions to brands and discussing financial terms, these people added. Shopify offers checkout technology that can be integrated into other online services. It already works with social media platforms, for instance, underpinning TikTok's shopping feature. ChatGPT's product recommendations are currently generated based on whether they are relevant to the

user's query and other available contexts such as memory or instructions like a specified budget. OpenAI has recently enhanced its memory, which allows the model to remember user preferences and provide more personalized responses. However, when a user clicks on a product, OpenAI may show a list of merchants offering it according to its website. This list is generated based on merchant and product metadata we receive from third-party providers. Currently, the order in which we display merchants is predominantly determined by these providers, it adds. Open AI does not factor in price or shipping into these merchant options but expects this to evolve as we continue to improve the shopping experience. Brands and advertising agencies have been experimenting with how to promote their products in chatbot search results.

For example, by posting content they believe will be more likely to be picked up by the models. The practice, similar to so-called search engine optimization or SEO, has become known in the industry as AIO. It starts to pose big and difficult questions around what preferences AI shows. and its results, one advertising chief executive said, this can potentially destroy the idea of paid search via traditional platforms, and also, of course, disintermediate the way advertising agencies operate today. As recently as December OpenAI, which is currently restructuring into a for-profit company, said it had no active plans to pursue advertising, end quote. One of the things that is percolating in the background with AI, at least in this LLM moment, is actually a very cultural thing.

Basically, you need to localize models to specific regions and cultures for best results. Rest of World has a look at the Chile-led Latam GPT project, which involves more than 30 Latin American and Caribbean institutions, collaborating to release an open-source LLM in September. Quote, while large language models including GPT and Meta Islam are trained on a wide range of data in languages other than English, their capability in those languages remains limited, particularly in dialects and locales. idioms to address these shortcomings which have led to inaccuracies and hallucinations or fabrications. A group of over 30 institutions across Latin America has spent the last two years developing Latam-GPT. The Chile-led Latam GPT project is, quote, building AI in Latin America for Latin Americans.

Hector Bravo, lead of disruptive technologies at Sanda, a Chilean IT firm that is not involved in the project, told rest of world. It means redefining success metrics, not just accuracy or speed, but cultural representation. representation, social impact, and accessibility. Latam GPT is being designed for deep multilingualism and includes indigenous languages such as Noado, Quecha, and Mapodongan, as well as dialect variants, including some from the Caribbean, said Bravo. Latin America is following the lead of other regions. Southeast Asia's Sea Lion is a family of open source LLMs trained in nearly a dozen regional languages besides English. In Africa, users can interact with Ulyza Lama in at least five different languages, including Sosha and Zulu, while in India, Barat GPT supports over 14 regional

languages with the government recently announcing that it was building its own LLM as well. Latin America has been slow to adopt AI. It is beginning to catch up, however, with Chile leading in terms of regulation and institutional development, according to the Atlas of Artificial Intelligence for Latin America and the Caribbean, a 2025 report from the United Nations Development Program. Chile's National Center for Artificial Intelligence, was founded in 2021. The idea for Latam GPT emerged shortly after. Although LLM, such as GPT and LAMATU, support multilingual capabilities, including Spanish, many of the data sets they are trained on are from Spain, or translated from text originally written in English, limiting their ability to understand

cultural and linguistic nuances. Latam GPT, which is being trained with data from schools, businesses, libraries, and historical texts, quote, helps the model better understand the context and needs of Latin American users. Omar Flores, Latam GPT's technical lead for pre-training told rest of world. There is increasing demand for generative AI platforms in the region. Brazil has the highest number of users of chat GPT after the U.S. and India, according to Demand Sage, a sales analytics platform and Lama downloads have also surged in Latin America. Teachers and students use them in classrooms, while business owners turn to them to offer customer support. Even government offices employ them to reduce processing times. In Buenos Aires, for example, the courts use chat GPT to draft legal decisions.

Clearly, the resources behind chat GPT dwarf those of Latam GPT, which will be text only for the foreseeable future. It will also lag behind on general questions and those not related to Latin America, said Soto. Latam GPT requires ultra-high capacity infrastructure, specialized talent, and relevant data sets, three areas where gaps still exist in the region. Carlos Honorado, chief executive officer of Orion, a Chilean, AI company told rest of world, end quote. Finally today, something that has gotten a lot of chatter online overnight is this post from Calvin French Owen, a former OpenAI engineer who detailed his experience working at OpenAI, including its culture, engineering practices, rapid growth,

the launch of Codex, their coding platform. He said that when he joined, Open AI was just over a thousand people, but by the time he left, it had tripled in size. That kind of growth creates a lot of strain, obviously, communication, hiring, shipping products, even how teams organize. And apparently, the culture at OpenAI isn't uniform. Some teams run flat-out sprints, others move steadily, and research, applied work, and go-to-market-all operate on different time horizons. A surprising detail was that almost everything happens on Slack. French Owens said they got maybe 10 emails their whole time there. It can be overwhelming if you're not careful, but if you curate your Slack channels, apparently it works.

the company is incredibly bottoms up. Early on, he asked what the next quarter's roadmap look like and was told there isn't one. It's very meritocratic, apparently. Ideas matter more than politics. Leaders tend to rise because they consistently deliver, not because they're great at all hands or internal maneuvering. And there's a bias toward action. People build prototypes without asking permission, and if something shows promise, teams form around it. Researchers are treated like many executives who chase what interests them. He emphasized how quickly Open AI pivots, even at its size. The company will abandon a plan overnight if new data suggests a better direction. At the same time, the scrutiny is intense. He mentioned seeing press stories before internal announcements and said secrecy is taken seriously, even internally. Access is restricted, but under that secrecy is a strong sense of mission. The stakes feel enormous to OpenAI errs, apparently. They really believe in the fact that they might be building AGI, serving hundreds of millions of users competing against giants like Google and peers like Anthropic, all under the watch of governments and the public. He liked how Open AI shares its models.

Cutting-edge tools aren't locked behind enterprise contracts. Anyone can try chat GPT or sign up for the API. That openness, he said, is still core to the company's DNA. Safety work is another area they highlighted. Contrary to some online speculation, a lot of people there focus on practical risks like abuse, misinformation, or prompt injection. More theoretical risks are also considered, but less of a day-to-day focus. They also dove into the engineering site. Open AI uses a giant Python mono repo with some rust and go, and code quality varies wildly. Production-grade systems sit alongside quick Jupiter notebook experiments. Everything runs on Azure, though a lot of in-house tools fill gaps where Azure is weaker. He noted a strong influx of former meta-engineers, and much of

the infrastructure feels, according to him, inspired by meta's early days. Decisions are made, by whoever is willing to do the work. So duplicate systems do exist, but the velocity is impressive. The highlight of his time there was the Codex launch. After returning from paternity leave, he joined a frantic seven-week sprint that merged two teams and built a coding agent from scratch, container runtime, Git integrations, fine-tuned models, a new interface. They worked late nights, early mornings and weekends, the night before launch. A small group stayed up until 4 a.m. deploying, then showing up for the 8 a.m. launch and watch traffic pour in. They said it was the hardest they'd worked in a decade in one of the most rewarding projects of their career.

He closed by reflecting on why he joined in the first place. He wanted to understand how models are built, to work alongside brilliant people, and to ship something meaningful. In his words, all three boxes were checked. His advice to founders was interesting. If your own startup feels stalled, either take bigger swings or consider joining one of the major labs. So some late-night insomnia-induced reading learnings for you, sort of tech history. I'm reading a book about the history of the horse as a technology and its impact on history. A couple things I learned. First, did you ever wonder why horses tend to pull things like carts and carriages and stagecoaches in pairs?

Like, historically, you'd see two or four or more horses in front of a stagecoach, but rarely just one. It's because horses are such social animals. It's hard to convince one singular horse to pull something, but if it senses another horse, horse riding alongside it, it will be more likely to be like, I guess this is what we're doing. If you only have one thing to pull a thing, historically people used oxen instead, not horses, oxen are apparently willing to do solo work. The other thing I learned is that the chariot actually came hundreds of years before people started actually riding horses. So the, you know, Egyptian chariot armies came centuries before the like Huns and the cavalry of Alexander the Great

interesting that the cart literally came before the horse. The book is called Raiders, Rulers and Traders. The Horse and the Rise of Empires, by the way. And it's very readable, actually. So if you're curious, link to that in the bottom of the show notes. Talk to you tomorrow.