Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Why the next AI boom is physical AI | Caitlin Kalinowski (ex-OpenAI, Meta, Apple)
TL;DR
The next AI boom is physical, not just digital — Caitlin Kalinowski says labs increasingly see a ceiling on “what you can do behind a keyboard,” which is why talent is flowing from computer science into robotics, manufacturing, drones, autonomy, and the sensing layer of the real world.
VR didn’t fail so much as become the training ground for robotics — work on SLAM, depth sensing, spatial perception, and headset-driven teleoperation at Oculus/Meta now maps directly onto robots, drones, and autonomous systems, while AR glasses like Meta’s Orion still look real but are blocked by yields and cost of waveguides and microLEDs.
Hardware is brutal because you only get to ‘compile’ a few times — unlike software, where you can ship fixes constantly, hardware gets maybe four or five major builds total, so teams must lock goals early, solve the hardest constraints first, and obsess over tolerances, yields, and reliability before mass production.
Humanoid robots are still advanced prototypes, and safety is the real gating factor — Kalinowski argues that today’s strong humanoids aren’t ready for humans at scale, pointing to issues like impact energy, arm compliance, and even manuals that warn people to stay 3 feet away, while praising softer designs like 1X Neo as directionally safer.
The real bottleneck may be supply chain, not model intelligence — from magnets and actuators to RAM and silicon, one missing part can trigger a catastrophic redesign, and she warns memory prices could double as AI infrastructure absorbs supply, which is why she’s advising companies to pre-buy key components.
The biggest near-term disruption may be warfare, not consumer gadgets — citing Ukraine, Palmer Luckey, and Marc Andreessen’s “100,000 drones” scenario, she argues the U.S. needs to reindustrialize, build independent supply chains, and invest far more in drones and robotics than legacy systems like aircraft carriers.
The Breakdown
VR’s real legacy: it taught machines to understand space
Lenny opens with the obvious question: after billions from Meta and a magical-but-still-niche Vision Pro, what happened to VR? Kalinowski’s answer is refreshingly unsentimental — VR mattered because it taught the industry SLAM, depth sensing, and spatial perception, and those same capabilities now power robotics, drones, and autonomy. She still believes in AR glasses, especially as a healthier alternative to staring down at phones, but says products like Meta’s Orion are ahead of manufacturing reality because waveguides and microLED yields still aren’t there.
Why hardware suddenly became the sexy thing
She says it’s “very odd” to watch hardware and robotics become hot after years of being the unglamorous path you chose only if you loved it. The reason, in her view, is that AI progress is going so vertical that people can already see a future where purely digital work saturates — and once that happens, the next frontier is the physical world. That means robotics, manufacturing, industrialization, and eventually space.
Hardware is harder because you only get a few shots
Her clearest analogy lands here: software engineers can compile constantly, but hardware teams only get to “compile” four or five times total. That constraint changes everything — you have to define KPIs early, avoid late changes, start with the hardest failure points, and pour extra iteration into the parts people touch most, like a trackpad or keyboard. She also shares the Apple-style discipline she absorbed: understand why you’re building something, then let every design decision support that goal.
Humanoids are exciting, but they’re still not ready
Kalinowski likes the ambition around Tesla Optimus, Figure, and 1X Neo, but she’s blunt that current humanoids are still prototypes. The limiting factor isn’t just intelligence — it’s safety around humans, from the mass of a moving arm to whether the surface is soft enough to reduce impulse on impact. She notes that some robots today still come with warnings that no human should be within 3 feet, which is a pretty good reality check against the hype.
The bottleneck under everything: magnets, actuators, memory, and China
This section turns from cool demos to industrial anxiety. She walks through the stack from raw magnets to processed materials to actuators to subassemblies, arguing that the U.S. outsourced too much of it over the last 25 years and now needs to re-industrialize for both economic and military safety. She also flags a “meteor” coming for robotics and consumer hardware: memory prices, which she believes could double as AI demand soaks up supply, and reminds listeners that if just one key part disappears, your whole product may need a catastrophic redesign.
Drones, war, and why old military math no longer works
The podcast gets darker here, but she doesn’t duck it. Watching Ukraine convinced her that drones, rapid iteration, 3D printing, and AI-updated systems are changing war faster than consumer electronics are changing everyday life, and she agrees with Palmer Luckey that the U.S. should invest much more in drones than aircraft carriers. Her point is simple and stark: when cheap autonomous systems attack and expensive legacy systems defend, “we’re losing on the math.”
Apple, Meta, and how to build teams for a new era
Kalinowski credits Apple with teaching a generation of hardware leaders how to think through complex interdependent decisions, with Steve Jobs setting an uncompromising quality bar and even caring about the “back of the cabinet.” At Meta, she admired how clearly reviews, decision-making, and technical tradeoffs were run under Mark Zuckerberg and Andrew Bosworth. Her hiring playbook today is to mix hardcore specialists with adaptable generalists and “AI native” young engineers, because in fast-moving fields like robotics, nobody has done the exact job before.
What makes robots feel less creepy — and why she left OpenAI
One of the most human parts of the conversation is her explanation of robot behavior: a machine that just stares blankly is creepy, but one that acknowledges you, signals intent, and moves with softness feels safer. She says Pixar and Disney may be the best in the world at this kind of emotional design. Near the end, she also explains her OpenAI exit: she cared deeply about the people there, but disagreed with the speed, governance, and lack of guardrails around the company’s Department of Defense-related announcement, and left to make her own boundaries clear without going “scorched earth.”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
