Tech Brew Ride Home - Thu. 04/17 – Mooooaaar Models!

Episode Date: April 17, 2025

OpenAI has a new reasoning model, and more lower cost models as well. Is OpenAI about to acquire a coding startup? Is Perplexity turning to Samsung for distribution and branding? A Neuralink rival gets FDA approval. And why is Jensen Huang on a code-red mission to China at the moment? Sponsors: 1Password.com/ride Links: OpenAI's new o3 and o4-mini models are all about 'thinking with images' (Engadget) OpenAI debuts Codex CLI, an open source coding tool for terminals (TechCrunch) Vibe Check: o3 Is Here—And It’s Great (Every) OpenAI In Talks to Buy Windsurf for About $3 Billion (Bloomberg) Perplexity AI in Talks to Integrate Assistant Into Samsung, Motorola Phones (Bloomberg) Neuralink Rival Gets FDA Clearance in Path to Less Invasive Brain Implants (Bloomberg) Nvidia chief Jensen Huang flies to Beijing for talks (FT) 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 Thursday, April 17th, 2025. I'm Brian McCullough today. OpenAI has a new reasoning model and another lower cost model as well. Is OpenAI about to acquire a coding startup? Is perplexity turning to Samsung for distribution and branding? A neuralink rival gets FDA approval and why is Jetson Huang on a code red mission to China at the moment? Here's what you miss today in the world of tech. More models, everybody. OpenAI has launched Outs. O3, its most advanced AI reasoning model yet, as well as O4 Mini, a lower cost alternative that

still delivers, according to them, impressive results for chat GPT paid users. Quoting and gadget, a mere two days after announcing GPT4-1, OpenAI is releasing not one but two new models. The company today announced the public availability of O3 and O4 Mini. Of the former, OpenAI says O3 is its most advanced reasoning model yet, with it showing strong performance in coding math and science tasks. As for O3Mini, OpenAI is billing it as a lower cost alternative that still delivers impressive results across those same fields. More notably, both models offer novel capabilities not found in OpenAI's past systems. For the first time, the company's reasoning models can use and combine all of the tools available in chat GPT,

including web browsing and image generation. The company says this capability allows O3 and 04 Mini to solve challenging multi-step problems more effectively and, quote, take real steps toward acting independently. At the same time, O3 and O4 Mini can not just see images, but also interpret and think about them in a way that significantly extends their visual processing capabilities. For instance, you can upload images of whiteboards, diagrams, or sketches, even poor quality ones, and the new models will understand them. They can also adjust the images as part of how they reason. The combined power of state-of-the-art reasoning with full tool access translates into significantly stronger performance across academic benchmarks and real-world tasks, setting a new standard in both intelligence and usefulness, says OpenAI. Separately, OpenAI is releasing a new coding agent a la Claudecode named Codex CLI. It's designed to give developers a minimal interface they can use to link OpenAI's models with their local code.

out of the box it works with O3 and O4 Mini with support for GPT41 on the way, end quote. Yes, on that coding agent, quoting tech crunch, codex CLI appears to be a small step in the direction of OpenAI's broader agentic coding vision. Recently, the company's CFO, Sarah Fryer, described what she called the agentic software engineer, a set of tools OpenAI intends to build that can take a project description for an app and effectively create it, and even quality assurance test it. Codex CLI won't go that far, but it will integrate OpenAI's models, eventually including 03 and 04Many, with the clients that process code and computer commands, otherwise known as command line interfaces. It's also open source, OpenAI says. And a blog post provided to TechCrunch OpenAI added, you can get the benefits of multimodal reasoning from the

command line by passing screenshots or low fidelity sketches to the model combined with access to your code locally via Codex CLI. to spur use of Codex CLI, opening I plans to dole out $1 million in API grants to eligible software development projects. The company says it'll award $25,000 blocks of API credits to the project's chosen, end quote. But let's get back to 03 and the fact that it is a reasoning model. Dan Shipper at Every wrote up a summary of his experiences with it, and while I encourage you to read his entire thing to get a sense of the different ways he used it, I'm just going to read from his conclusion here, quote, this is the biggest one.

wow moment I've had with a new open AI model since GPT4. The company has successfully lengthened the leash that an AI gets in order to do its tasks. Now you can reliably let it work for minutes at a time to get higher quality answers with no intervention. But what makes O3 especially powerful is the integration between the model and chat GPT. It has tools like web search, reminders, memories, and code interpreter that allow users to access more of its power with less hassle. This is a very valuable strategic position for OpenAI to be in. one that will further its lead as the destination chapout of choice for this AI era. O3 is also a lesson in the seasons of change that companies go through.

For about a year after Sam Altman's firing, Open AI was slow to release new products, and analysts were prognosticating about its downfall. With this release stacked on its deep research and GPD 4.5 recently, OpenAI is back on a tear, end quote. Sticking with Open AI for a second, Bloomberg says, Open AI is in talks to acquire WindSurf, an AI coding tool formerly known as codium for around $3 billion. WindSurf was valued at $1.25 billion in a 2024 funding round. So we know that coding AI is the new hotness, and Open AI has been moving to plant their flag in this arena.

But what does this startup have that Open AI doesn't think they can build on their own? Quoting from Bloomberg. The deal would be OpenAI's largest acquisition to date and could help the company take on rising competition in the market for AI-driven coding assistance, systems capable of tasks like writing code based on natural language prompting. WindSurf, formerly called ExaFunction, had recently been in talks with investors, including Kleiner Perkins and General Catalysts to raise funding at a $3 billion valuation. The company was valued at $1.25 billion in a deal led by General Catalyst last year.

If the deal closes, Open AI would compete more directly with companies like Anthropos, Microsoft-owned GitHub and AnySphere, which offers services in the fast-growing field of AI coding. The deal would also be a signal that, as more companies pile into the sector, the number of mergers and acquisitions will rise. OpenAI has acquired other startups before, including Vector Database Company Rockset and Multi, a remote collaboration platform. Investors have recently been enthusiastically backing AI coding tools, including AnySphere, the startup behind Cursor, which was in talks with investors to raise money at a

evaluation of almost $10 billion earlier this year, end quote. Three interesting tweets on this. First, quoting Austin Al-Read. If OpenAI buys WinServe, does it get data as to all of the input prompts and then outputs for all of the different models? I think it might. And you'd be able to see whether the output is accepted or rejected, end quote. Then Zach Weinberg.

I'm sure everyone deep in the industry already knew this, but very clear the core model companies realize they need to own the customer slash application layer to actually make money. Tons of M&A likely coming to fight for the most mature application layer companies in the biggest categories, coding, marketing, maybe customer service will be very tough to compete with the OpenAI or Google or Microsoft application bundle in the long run, end quote. Then at semi-analysis, cursor and Anthropic had a mutually beneficial relationship, but labs realize controlling the main application of a model is as valuable as owning the model itself. With this acquisition, Open AI gains greater ecosystem control and can build better products, end quote.

The race is real, folks. Bloomberg sources say perplexity is in discussions to integrate its AI assistant on Samsung's devices and has reached an agreement to preload perplexity on Motorola's phones. Quote, the details of the Samsung deal are still being hashed out, the people said, the terms could involve offering perplexity as a default AI assistant option or preloading the startup's Android app on phones. Samsung, the world smartphone market leader, also could promote the assistant heavily as an option within the Galaxy store where users download apps. A deal with perplexity is complicated by Samsung's wide-ranging partnership with Alphabet, which powers many of the AI features on Samsung devices and provides the default search engine.

But Samsung has been fostering a relationship with perplexity for months. South Korea-based company's investment arm called Next, backed the startup last year. Samsung is discussing making another investment in perplexity in the near future, the people said. The startup has been in broader talks with investors for a new round of funding, which would double its valuation to $18 billion. It is discussed raising between $500 million and $1 billion. Bloomberg News reported in March. Founded in 2022, Perplexity has gained popularity for real-time results and search engine-like user interfaces. The assistant has the ability to service in-depth research, and perplexity can tap into models beyond its own, including those

from OpenAI and Anthropic PBC. The Motorola deal will include perplexity being preloaded as an alternative assistant to Google's Gemini. There will also be a tailored user interface for Motorola's updated razor-foldable phones, and marketing from the device maker will encourage customers to try perplexity. Today, Motorola offers its own homegrown AI system and access to Google Gemini. Samsung takes a similar approach, providing its Bixby service alongside Google's AI platform. For San Francisco-based perplexity, securing Motorola and Samsung deals would be a major boon to building its brand and gaining popularity in the U.S. where ChatGPT and Gemini are more established. While Motorola's market share is slim in the smartphone industry, the company is making

strides. Samsung, for its part, has a share of about 20% globally, according to data from IDC, end quote. The U.S. Food and Drug Administration has cleared Neurrelink rival precision neuroscience to offer a less invasive brain implant. Quoting Bloomberg. One of the co-founders of Elon Musk's Neurolake Corp is building a different kind of brain implant. That device made by the startup precision neuroscience corp went in 63-year-old Tim Fisher earlier this month. In an operating room in Philadelphia, a surgeon placed a sliver of film thinner than a human hair and embedded with over a thousand electrodes on Fisher's brain. After a little training when Fisher moved his left hand, so did a nearby robotic appendage, guided by signals from his brain and the new brain patch.

It was an amazing experience, said Fisher, a retail worker who lives in Westchester, Pennsylvania, and has Parkinson's disease. He said he's drawn to these kinds of implants because his tremors make everyday tasks like typing or opening a jar near impossible. The brain device that Fisher used was cleared by the U.S. Food and Drug Administration late last month for limited uses, bringing precision one step closer to a commercially available technology that one day might allow people to control computers or other devices with just their thoughts. That could help people with paralysis, electronically communicate with others, use computer software, and get a job. The device is already being used to help doctors map brain activity in more detail, to plan

procedures, and could also enable researchers to study the brain more closely and accurately. Precision was started in 2021 by Michael Magar and Benjamin Rapaport, a brain surgeon who previously helped co-found Neurrelink, in contrast with some other experimental systems like Neurilinks, which stick into the brain tissue, precision's strip is designed to sit on top, Rappaport said. The idea is to build something that can be inserted and removed without leaving a trace on the brain. Fisher's Patch, for instance, was only in place temporarily during the procedure. Around the size of a U.S. postage stamp, the precision device contains 1,024 tiny metal electrodes that are each a fraction of a millimeter in diameter.

It's been studied in at least 37 patients across five medical centers and about one-third of U.S. medical centers that do major brain surgeries have expressed interest in it, according to the company. We've been amazed by the demand for it, Rappaport said. The strip was greenlit by the FDA to record, monitor, and stimulate brain activity and can be used for less than 30 days. The regulator cleared the device based on a submission that included testing, but not the human studies. At least one other company, BlackRock NeuroTech, has received a similar clearance for its Neuroport Array device, end quote. I know, I know, I promise not to do this, but tech stocks were down, significantly yesterday. What's worrying people now? Well, that's what I want to talk about. It's

about that wrinkle around chips and tariffs and export controls, but also specifically about NVIDIA. All the big chip players, AMD, NVIDIA, and ASML were down 7% yesterday. And NVIDIA CEO Jensen Huang apparently arrived in China yesterday and visited Beijing today, a source says, to meet officials to discuss Trump's new chip export rules. quoting the FTE. According to two people familiar with his travel schedule, Huang met NVIDIA clients, including the founder of Generative Artificial Intelligence Startup Deepseek to discuss new chip designs for Chinese customers. He then held separate talks with Chinese premier He Lai Fang, according to one person familiar with the meeting. Huang said China

was a very important market for NVIDIA and expressed hope that his company could, quote, continue cooperating with the country, according to state broadcaster CCTV. On Tuesday, NVIDIA said it expected a $5.5 billion hit to earnings from new U.S. export restrictions on its H-20 chip, a lower-powered model that had already been designed to comply with Joe Biden-era controls limiting exports to China. Wong's talks indicate that NVIDIA is not willing to give up on the Chinese market and is considering designing yet another chip for it, even though its previous efforts have been banned by Washington. Plans for the NVIDIA chief's visit to Beijing were finalized after U.S. President Donald Trump's unexpected move to ban the H-20 chip. Invidia reported $17 billion in sales from China last year,

but faced growing threats to its business from Beijing, even before Trump interceded, end quote. In a separate Financial Times piece, they report that NVIDIA is hustling now because they were actually caught flat-footed. After a meeting with Trump at his Florida residents at Mara Lago earlier this month, Nvidia executives were left with the impression they could escape tougher enforcement of any curbs, the people said, adding the company's plan to invest $500 billion in the U.S. had also impressed the president. This led NVIDIA to tell Chinese clients, including tech giants, Alibaba, bite dance, and Tensit, that orders of H20s would not be affected, the people said. NVIDIA was then blindsided as Trump decided to clamp down on the export of H20, a product

that Chinese tech groups have relied upon in their efforts to challenge their global peers to develop large language models. Frustrated Chinese tech companies have complained about not having enough warning about such a major policy change, but our understanding that the shift is beyond the control of Nvidia, according to the people with knowledge of recent discussions. AI demand jumped in China after DeepSeek's successful launch of its low-cost reasoning model led local companies to put almost $17 billion in orders for H20 chips this year, according to one of the people.

While Nvidia typically takes more than six months to deliver such chips, most of this year's orders from its Chinese clients are yet to be filled and will probably be affected by the latest U.S. restrictions. The $5.5 billion hit-to-earnings NVIDIA announced are mostly the cost of materials to be used to produce such orders and related penalties and operational costs for not delivering based on agreed terms. The actual affected revenue from China could be more than $10 billion, the person estimated. China's tech giants are racing to find a replacement to the H-20, while Trump's new export controls could significantly help the sales of domestic manufacturers, led by Huawei, which has been pushing to produce more AI processors. It also remains unclear how Chinese groups can apply for a license to obtain H-20s and on what basis they would be issued. Intel told its Chinese clients last week that chips would require a license for exporting to China if they have a total D-RAM bandwidth of 1,400 gigabytes per second or more,

IO bandwidth of 1,100 gigabytes per second or more, according to a company email. Intel's Gaudy series as well as Nvidia's H20 far exceed these requirements, end quote. So yeah, sounds like Jensen is on a code red mission. As soon as I hit publish on this, we are going to take a tour of the Stanley Hotel, which is what the hotel and the Shining is based on. Apparently Stephen King came here in the 70s right when the Stanley was shutting down for the season. And boom, inspiration for the Shining. after he visited the hotel bar and met with the bartender whose name coincidentally was

Lloyd. Talk to you tomorrow.