Stewart Squared
Episode #97: How AI Is Rewriting the Rules of Computing
- AI & Agents
- Chip design
- GPU architecture
- CPU virtualization
- OpenAI's Jalapeno chip
- NVIDIA history
- disk operating systems
- digital projection
- TeamLab
Listen
Download MP3About this episode
Stewart Alsop sits down with his father, Stewart Alsop II, to unpack what chip design even means anymore, starting with OpenAI's Jalapeno chip and the wild claim that it was designed in nine months using their own LLM. They trace the CPU from its personal computer origins through GPUs, FPGAs, and the strange new world where anyone might vibe code a chip, then swing into digital projection and showrunner systems at TeamLab and Meow Wolf, autonomous vehicles and the LiDAR fight between Tesla and Waymo, Gaussian splats and world models, a detour into 3D printing and a failed IRL collectibles startup, and a closing stretch on patents, IP trolls, and whether China's open source AI push means the US proprietary model is losing ground.
Timestamps
00:00 Apple chips, CPU vs GPU and why chip design feels virtualized now.
05:00 GPUs for video games, why productivity ignored graphics, and how the CPU became a bundle of multiple cores.
10:00 Team Lab and Meow Wolf as digital-projection worlds: projectors, microcontrollers, and the showrunner idea.
15:00 Interactive exhibits as “everything at once,” then a shift into Anthropic, biotech, and the pace of AI innovation.
20:00 Real-time systems, lipsync, world models, and why LLMs struggle with space-time.
25:00 Autonomous vehicles: Cruise, Waymo, LIDAR, Tesla’s camera-only approach, and the debate over edge cases.
30:00 More on Waymo vs Tesla, safety incidents, and whether Gaussian splats matter for robotics.
35:00 Chip design, OpenAI’s “jalapeño” chip, firmware, memory shortages, and why Apple memory costs are rising.
40:00 Patents, software IP, LLMs, and how China and the US diverge on open source versus proprietary AI.
Key Insights
- The CPU has quietly become plural. What used to be a single processing unit is now many cores managing memory, disk, and networking all at once — the concept of "central processing" has essentially been virtualized from the inside out.
- Chip design may no longer require deep technical expertise. OpenAI's Jalapeno chip, reportedly designed in nine months using their own LLM, suggests that designing silicon is becoming something closer to "vibe coding" than a specialized engineering discipline.
- Digital projection systems like TeamLab and Meow Wolf run on lightweight computing, not heavy processing power. The magic comes from networked microcontrollers and a "showrunner" system, a concept borrowed from television, that keeps hundreds of projected events in sync without conflict.
- Tesla and Waymo represent two opposing bets on autonomy. Tesla relies purely on cameras and processing power, while Waymo loads its cars with LiDAR, radar, and video. Both approaches still hit real-world edge cases, from Waymo pulling cars off freeways after construction-zone incidents to a fatal Tesla crash with no clear explanation.
- World models are trying to give machines a sense of space and time. Gaussian splats, used by companies like Marble and Niantic, create detailed spatial reconstructions, but they're not yet real-time, which limits how directly they can be applied to something like robotic driving.
- Intellectual property often only reveals its value after failure. A collectibles startup pairing physical figurines with digital twins collapsed alongside the NFT market, but the conversation underscores how "IP trolls" and specialists like Nathan Myhrvold later mine failed patents for value nobody recognized the first time around.
- China's AI progress is closing the gap through an open source strategy the US mostly abandoned. Coupled with Anthropic's accusation that Alibaba scraped its codebase millions of times, the episode frames China's non-profit-driven, open approach as a real competitive threat to America's proprietary model.
Episode transcript
Everyone is wrong about AI, but if you know the history of technology, you can be wrong and still be right.
I know the history since I started in technology with the personal computer in 1980. So listen in while we explore what is true about AI.
Welcome to the Stuart Squared podcast. Today, we are going to talk about Apple and Apple chip design and how that is. Do they design GPUs?
That's all they've been doing is GPUs. Well, I mean, not all it, because they've done a bunch of accessory chips. you know, one of the issues in chip design for me is a chip's no longer a chip. Right. It used to be that you kind of divide it, design this thing that looks like a chip and it would be a CPU and it would be the core of your design for a computer. all of that seems to have changed. I'm a little unsure of what CPU and GPU mean anymore. It's kind of been virtualized to the point that, you know, we wanted to start talking about this because OpenAI announced the Jalapeno chip, which they claim they designed in nine months using their own LLM. And that was a mind blowing concept to me. I didn't know that you, that an LLM could design these massively complicated, very trans shrunken, chips. You know, that Nvidia has been doing it for years and then suddenly OpenAI shows up and says, oh, well, we designed a new chip for LLM. I could, I couldn't have anticipated that. So I'm not sure exactly what the terms mean anymore. If, you know, you're not doing discrete chips, a CPU and a gpu, which is where Nvidia, started in a computer where the CPU is managing the memory process and execution and then the GPU is putting shooting pixels up on the screen. Kind of the first application of GPUs. I don't, don't. So I'm kind of lost in there. I don't know much enough about design to be very precise about it.
Your first full transcript is free. After that, an email opens every transcript in the index — a list of readers we can write to, not a guest book.