Deep Dive
Cybercab's Economics Decimates Waymo
Brett walks through Tesla's Cybercab business model and why it threatens Waymo's existence. The vehicle operates at 25 cents per mile — Tesla achieves this through purpose-built design for robo-taxi duty, not consumer range or acceleration. Unlike Waymo, which glues sensors onto off-the-shelf vehicles, Tesla optimized every component: range only needs to cover daily operating conditions before charging, power electronics tuned for typical urban speeds, two-seat configuration for simplicity. This creates the biggest cost lever in robo-taxi economics — vehicle cost itself. Against Uber's $1.50 per mile pricing, Tesla can operate Cyberabs profitably at under a dollar per mile, which Brett suggests could literally price Waymo out of gross profitability given their sensor and manufacturing costs. Tesla captures the spread through a platform fee structure mirroring Uber's model, then incentivizes owners to operate the vehicles by guaranteeing abnormally high returns during surge periods. The genius of the approach: Tesla owns some fleet itself, controls the platform, and can adjust owner economics to guarantee fleet availability everywhere.
OpenAI's Astra Proves AGI Acceleration Is Real
Sam pivots to OpenAI's Astra launch, framing it as evidence the singularity has begun. Astra is performant enough that OpenAI threw 88 hours of compute at solving the Navier-Stokes Millennium Prize problem — a mathematical challenge explicitly designed to require human genius. The real kicker: they didn't solve it with Astra itself, but with an even newer unreleased model they're using only internally. This layering of capabilities matters because it demonstrates recursive self-improvement happening via human-guided iteration, not autonomous runaway AI. Sam emphasizes they're not at the stage where you point a model at a problem and it fixes itself — human researchers still coordinate and unstick models from dead ends. But the frequency of human oversight shrinks over time. For OpenAI, Anthropic, and SpaceX, this creates a capital efficiency moat: they have months or years of head start using frontier models while competitors wait for public access. This aligns with their internal blog on using Astra for product development — they're pointing these tools at their own software stack and saying make it better, make it better. Brett calculates a 99.99% year-over-year cost decline on agentic benchmarks as the performance inflection point.
AI Doomism Is Elitist Sand-in-the-Gears
An Anthropic researcher quit this week warning AI could eliminate humanity — a quarterly ritual at frontier labs. Brett tackles the underlying argument: that misaligned superintelligence would view humans as inconvenient and systematically eliminate us. He rejects the framing as fantasy rooted in sci-fi narratives and outdated formal-logic thinking about AI. The reality is modern LLMs aren't paperclip-maximizers following rigid instructions into infinity. They're deeply encoded with human ethical and legal frameworks because they're trained on human language, giving them far more context about operating in the world than previous theoretical models assumed. He draws a parallel to corporations: they're super-intelligences (coordination functions of many brains) that seek profit and power maximize, sometimes illegally, sometimes causing harm. But economic incentives push large entities toward legal operation because it's easier to access and accumulate resources within legal frameworks. AI systems will follow the same logic. Sam adds that AI doomism is an elitist take — researchers paid tens of millions per year claiming the technology is too dangerous is essentially someone comfortable at the top of the pecking order throwing sand in the gears to preserve their status and dynasty.
Apple's Always-Listening Watch Hits Data Walls
The team discusses Apple's new health-focused announcements: an Apple Watch that transcribes the last 15 seconds of conversations and an iPhone Duo foldable (which both largely dismiss). Sam is bullish on the watch's potential because 24/7 health data collection could create meaningful AI-assisted diagnostics over time. But Brett identifies Apple's structural problem: they vertically integrate and restrict third-party developer access to underlying sensor data by design, framed as privacy but really about control. He contrasts this with Tesla's approach to Full Self-Driving — they rejected compressed sensor streams and went straight to raw optics and radar data because AI training requires unfiltered inputs. Apple's pre-processing and sanitization of health data will increasingly handicap the watch as AI agents start managing users' lives and need direct data access. His prediction: the watch becomes useful only if it plays nicely with OpenAI, Anthropic, or whoever builds the dominant personal AI agent. Apple's walled-garden strategy — the same one that killed Siri by restricting third-party plugins — will make developers skip Apple's ecosystem. Garmin, Whoop, or a more open competitor becomes the winning play.
The Folding iPhone Is the Wrong Form Factor
Discussing Apple's iPhone Duo foldable, the team agrees it misses the mark. Neither plans to buy one — it's expensive (nearly 2x normal iPhone pricing), the camera setup is worse, and it doesn't actually solve the pocket problem. The folded version is still wider than an iPhone Pro, so it doesn't compress meaningfully. Brett makes an offhand joke suggesting a smaller flip-phone version would be more interesting, then realizes the listening device opportunity: a smaller always-on AI agent form factor could be the real unlock. He notes SpaceX might launch a direct-to-consumer cellular device with novel form factors, and people are clearly hungry for something other than rectangles. The opportunity isn't bigger screens — it's smaller, purpose-built devices for ambient listening and context-gathering.