OpenAI Built Its Own Chip in 9 Months, Fable 5 Is Coming Back, AI Is Trying to Kill the Common Cold

Host: Welcome Back to today's AI news, brought to you by 9X Productions.

Co-host: Glad to be back. We've got, uh, a lot to cover today.

Host: We really do. But first, imagine trying to win a Formula one race, but you have to buy your engine from the exact same guy. Selling engines to all your competitors.

Co-host: Right. You're never going to get an edge that way.

Host: Exactly. To really pull ahead, you eventually have to build your own engine. And today we're looking at how the biggest name in AI just fired up theirs.

Co-host: It's a massive shift in the landscape for sure.

Host: Yeah. So today's mission for this deep dive is simple. We are skipping the fluff and diving straight into the Latest developments from OpenAI's custom silicon and a, uh, half billion dollar swing at curing the common cold to some quick hits and a deeply moving community workflow you won't want to miss.

Co-host: Yeah, it's just a wild convergence of like physical, political and biological tech frontiers all at once.

Host: Definitely. So let's start with that physical frontier, the hardware. Anyone following this knows Nvidia has essentially been the, you know, the bottleneck.

Co-host: Right. They make the chips that make the AI happen.

Host: But OpenAI is aggressively trying to control its own destiny. Now, they just designed their first custom chip, it's codenamed Jalapeno and they co built it with Broadcom. Jalapeno. Gotta love the codenames.

Co-host: Right. But the timeline is what really caught my eye. They went from design to factory ready in just nine months.

Host: Okay, let's unpack this. Because isn't chip design famously slow?

Co-host: Oh, absolutely. I mean, a nine month turnaround in semiconductor manufacturing is basically light speed. Traditional stuff usually takes like two or three years of agonizing development.

Host: So how do they do it so fast? Is OpenAI basically using AI to build better brains for the AI creating this like massive feedback loop?

Co-host: Yeah, that is exactly what they're doing. They actually use their own AI models to help design and optimize the architecture of the chip itself.

Host: That is wild.

Co-host: It really is. But, uh, we should probably clarify something important here because this isn't meant to totally replace Nvidia.

Host: Right. Because there are different phases of AI development.

Co-host: Right, Exactly. You've got training and you've got inference. Training is the heavy lifting. It's when you feed the model the entire Internet and teach it at a reason.

Host: Yeah.

Co-host: Nvidia is still completely anchoring that training phase.

Host: Okay, so Nvidia is still the go to for the brute force stuff.

Co-host: Yeah. But Jalapeno is an asic and application specific integrated circuit. And it's built strictly for inference.

Host: Inference meaning like when you or I actually type a prompt into ChatGPT and it generates an answer.

Co-host: Right. It's the model applying what it learned. And owning that silicon lets OpenAI tune every single layer, the hardware, the models, the products, for, you know, maximum cost and speed efficiency.

Host: Which they desperately need. Right. Because the testing on Jalapeno shows the performance per watt is substantially better than the current state of the art.

Co-host: Oh, y. Yeah. And the energy costs for inference are astronomical when you're serving millions of users. Their larger goal here is to power 10 gigawatts of compute with custom chips by 2029.

Host: 10 gigawatts. That's. I mean, that's city level power.

Co-host: Oh, easily.

Host: Yeah.

Co-host: It just shows how physical this digital industry has actually become.

Host: Yeah. But, you know, while OpenAI has taken control of their hardware constraints, their biggest rival, Anthropic, is battling a very different constraint. The government.

Co-host: Yeah, this is a totally different kind of bottleneck.

Host: So Anthropic's Fable and Mythos models are currently offline due to a U.S. government order. They've been on ice for nearly a month now.

Co-host: Right. Which is an eternity in AI time.

Host: Exactly. But it looks like a thaw might be coming. There's a new Claude code update log that shows references to Fable usage, and they actually removed the purchased separately language,

Co-host: which is usually a pretty strong technical signal that things are moving behind the scenes.

Host: Yeah, and Wired just reported that the Trump administration is, quote, happier in their talks with Anthropic right now, largely because they are dealing with co founder Tom Brown instead of CEO Dario Emodei.

Co-host: Oh, really? What was the context there?

Host: Well, the Wired quote from an insider literally says, Thom Browne is not being a weirdo like Dario and can actually engage.

Co-host: Wow. Okay, that's, uh, pretty blunt.

Host: It is. And just to be absolutely clear for you listening, we. We're remaining completely neutral on all the political figures, the lawsuits and the US Order here.

Co-host: Right. Strictly impartial. We are just reporting what Wired and the politicians and the lawsuits are claiming without endorsing anyone's viewpoint.

Host: Exactly. Speaking of the lawsuits, Legion, which is a legal tech firm, they filed a suit calling the US Order unlawful. Um, retaliation.

Co-host: It's getting messy. And didn't some lawmakers step in too?

Host: Yeah, representatives Sam Locardo, Jay Obernault, and Ted Lieu actually gave the Commerce department a firm June 26 deadline to explain how the public gets the model back.

Co-host: Okay, so the pressure is really mounting

Host: from multiple angles But I have to ask, is sidelining an outspoken CEO for a more, you know, diplomatic co founder becoming a standard playbook for tech companies when politics get messy?

Co-host: Well, I mean, it makes sense, right? Frontier AI is no longer just a technical raise in a lab. It's now deeply entangled in legal, diplomatic, and public relations maneuvering.

Host: So you need diplomats just as much as you need engineers now.

Co-host: Exactly. The skill sets required to navigate Washington are very different from the skill sets required to build a neural network.

Host: That makes total sense. Okay, so while Anthropic is dealing with governmental roadblocks, a whole coalition of these AI giants is trying to bypass traditional roadblocks in a totally different massive system. Global healthcare.

Co-host: This is one of the most fascinating swings I've seen in a while.

Host: Yeah, stripe anthropic, the OpenAI foundation, and a few others just formed intercept. It's a $500 million nonprofit fund.

Co-host: Half a billion dollars? That is not small change.

Host: No, it's huge. And it's aimed at funding respiratory virus prevention and cleaner indoor air tech. Basically, they want to make routine sicknesses like the common cold and the flu a thing of the past.

Co-host: Which sounds like science fiction, honestly.

Host: Right. The money is funding shots, nasal sprays, pills, air cleaning tech for offices and schools. And the economic reality behind this is staggering. People lose, like, 15 to 25 days a year to respiratory infections.

Co-host: Yeah, which causes around $600 billion in global productivity losses every single year.

Host: But I have to push back a little here. Why are AI companies suddenly playing doctor? I mean, is this just classic tech hubris, like, we fixed search algorithms, now we'll cure the cold?

Co-host: Well, it definitely looks like that on the surface, but you have to look at the genius of this specific targeted philanthropy. The 500 million isn't actually meant to bring these products to market.

Host: Wait, really? Then what's it for?

Co-host: It's designed to fund what researchers call the awkward middle of development. See, traditional pharma companies won't touch early,

Host: risky research because it's too expensive if it fails.

Co-host: Exactly. They want a sure thing before they run those massive, costly final clinical trials. By funding that specific risky middle phase, these AI giants are using their deep pockets to de risk the science.

Host: Oh, I see. So they get it far enough along that big pharma, and investors are finally willing to take over for that costly final stretch.

Co-host: Right, they are essentially getting a universal problem over the hump that traditional capital markets just ignore.

Host: Okay, that actually makes a ton of sense. So, from solving the common cold on a macro level, let's zoom in to how these models are solving everyday headaches. On a micro level, let's get into our quick hits and a really clever way to snag highly coveted tickets.

Co-host: Yes, the agentic workflows. This is where the rubber meets the road for daily use.

Host: Exactly. So There's a new AI training workflow going around for using ChatGPT agent mode as a World Cup 2026 ticket scout.

Co-host: Oh man. Getting those tickets is usually a nightmare.

Host: A total nightmare of refreshing tabs. But here's the prompt setup. You type agent and you detail your location, your party size, the venues you want, the dates and your budget.

Co-host: You're basically giving it a super specific man.

Host: Yeah. You instruct it to check Ticketmaster seatgeek, Vivid seats, all of them. And then you have it return a ranked table match, date, venue, ticket, total transport costs, and an all in total

Co-host: with the link, the transport costs too. That's really smart.

Host: Right? And then you just set it as a recurring task every morning at 8:00am oh, and the pro tip here, use the Codex desktop app for longer scouting runs across even more sources.

Co-host: That is brilliant. It really highlights how we're moving away from AI just being a chatbot.

Host: Totally. Speaking of moving away from chatbots, let's hit some other rapid fire tools in the news. First up, Claude Tag. You can now tag Laude as a shared teammate right in your Slack channel.

Co-host: Oh, that's huge for team collaboration. Just bringing the AI into the room where the conversation is already happening.

Host: Exactly. Next is OpenArt Director. It's a vibe directing tool for making five minute films from chat.

Co-host: A Vibe director. How does that work?

Host: Basically it maintains the consistency of the aesthetic and the narrative over longer generations, which has always been the hardest part of AI video.

Co-host: Right. Stopping characters from suddenly sprouting an extra arm or changing clothes between shots.

Host: Exactly. Also, Crea2 just launched. It's an open weights in house image model, which is great for the open source community. And Mistral released OCR4.

Co-host: Ah, uh, Mistral's OCR4 that is actually layout aware document understanding. Right?

Host: Yeah. It doesn't just read the text, it understands the structure of the document. But the biggest news in this agent space might be from Google. DeepMind just added computer use functionality directly into Gemini 3.5 Flash.

Co-host: Oh wow. So what can it actually do?

Host: It allows the agents to navigate browsers, apps and desktops natively.

Co-host: So it's literally moving the cursor and clicking around for you.

Host: Exactly. And the talent war to build all this is getting intense. We just saw some major brain Drain Google, AI researchers Jonas Adler, Alexander Pritzell and Arthur Conmee all left to join Anthropic.

Co-host: The talent shuffle in this industry is just non stop. Everyone is trying to build the ultimate agent, right?

Host: And speaking of ultimate agents, there's a new lab called Mirandil. It was launched by former employees from Anthropic, OpenAI, Google and XAI.

Co-host: Okay, an all star team. What are they building?

Host: They are building AI that runs its own R and D for self improving science research. Like it hypothesizes, tests and improves itself without human intervention.

Co-host: That is simultaneously, uh, incredible and slightly terrifying.

Host: It really is. But here's the analogy I want to make. We're moving from AI as an encyclopedia to AI as a tireless intern.

Co-host: That's a great way to put it.

Host: But the contrast between these high minded science labs like Marindel and the messy reality of open systems is hilarious to me. For example, xai's grok.

Co-host: Oh, I saw this report.

Host: Yeah, the information just reported that Grok is heavily driven by NSFW generations. Employees are saying that well over half of its traffic comes from explicit content.

Co-host: It's just so predictably human. Right. We build these hyper advanced tools capable of like self directed scientific R and D. And what do people use them for?

Host: Bypassing safety filters for explicit images. It's ridiculous.

Co-host: It really is. But it shows the massive spectrum of utility. I mean the ChatGPT, Ticket Scout and Gemini 3.5, Flash's desktop control, they represent the exact same massive trend. AI is natively taking over multi step cross application workflows.

Host: Exactly. And all this talk of autonomous agents and massive scale can feel a bit, you know, abstract.

Co-host: Oh, definitely. It gets very technical very fast.

Host: Right, so we always like to finish off by covering community AI workflows to really ground this tech in real human impact.

Co-host: These are always my favorite part of the deep doc.

Host: Yeah, me too. So a reader named Steve C. From Troy, New York reached out. He actually built a voice based practice app from scratch using Claude.

Co-host: From scratch? Was he a developer?

Host: Not at all. He just had a specific problem to solve. His young family member has selective mutism.

Co-host: Oh, okay. So comfortable at home, but speaking is really hard in social situations.

Host: Exactly. The anxiety just takes over. So Steve used Claude to build an app that lets her choose who to practice talking with. Like a peer, a teacher or a coach.

Co-host: That's a great way to handle exposure therapy.

Host: Right, and she can choose a setting like a school cafeteria. The app even includes realistic background audio for that setting. The AI character asks her age Appropriate questions out loud. She responds with her actual voice. And then Claude gives her warm, honest feedback.

Co-host: Wow, that is incredibly thoughtful. And he built that over a weekend?

Host: Yeah. He even made it so parents can adjust the feedback style and the number of questions. Steve said, quote, this was not about building something complex. It was about using AI to solve a real problem for someone I care about.

Co-host: I mean, that just gives me chills. That right there is the true value of AI. It isn't just in raw compute power. You know, like we were talking about with the jalapeno chip.

Host: Right, the 10 gigawatts of power.

Co-host: Exactly. The true value is in hyper personalization for these edge cases. Traditional software companies would never bother to address this because the market is just too small. But now a dad in New York can build a custom therapeutic tool for his kid in a weekend.

Host: I am just in genuine awe, uh, of how accessible this technology has become. It's profound.

Co-host: It really is. And, you know, I think it leaves us with a really important final thought to mull over.

Host: Go for it.

Co-host: We started today with OpenAI using AI to design the great chips that power it. And we ended with a father using AI to help a child find her voice in a noisy world. It raises the ultimate. As AI gets better at optimizing itself, how do we ensure it continues optimizing for us?

Host: Man, that is a fantastic question to leave off on. You know, these tools are powerful, but it's really up to us how we point them exactly.

Co-host: Right?

Host: Well, that is all for today's deep dive. If you want to keep optimizing your own feed and stay updated on all things AI, make sure you hit that subscribe button.

Co-host: Yeah, please do.

Host: And, hey, if you enjoyed the conversation, we'd love it if you could rate the show five stars. It really helps us out. We will see you back here for tomorrow's episode.