键盘之后才是正戏：硬件老将眼里的物理 AI 浪潮

Asked what happened to VR despite the money poured into it and the quality of the hardware, Kalinowski says she views it as a step in a long technological arc rather than a dead end. VR taught the industry how to orient things in space relative to a simulated and a real world, including SLAM (positioning with cameras), depth sensing and how humans perceive visual data. VR gaming, she says, is an interesting niche, but a niche.

Those same technologies are now being reused in robotics, where a machine has to know how it is moving through space and how far it is from everything, even when a person is driving it through a headset. She suspects part of VR's limit was social: it is hard to be social with your face covered, a lesson Google Glass also taught. She adds that she is glad the industry went through VR.

Kalinowski still believes in AR glasses, because staring down at a phone is poor for social creatures. Orion, the AR glasses she worked on most recently, was ahead of its time because waveguides and microLEDs are not yet ready for mass production, with low yields and high cost. The 70-degree binocular field of view, though, gave a real sense of how immersive glasses could feel, and she sees the lineage running from VR and AR into autonomous vehicles, drones, robots and manufacturing, all of which will need the same piece parts.

Host Lenny Rachitsky notes that computer science enrollment is falling at many universities while hardware and robotics are rising, and Kalinowski finds it very odd after a career in a field that was never the sexy path. She explains the core difference for software people: a programmer can compile every hour, while a hardware team gets to compile its code four or five times in total. The final release for mass production is the last compile, and after that you cannot ship an update.

That forces a more conservative approach, with reliability checks and tests built into the program. For a product that sells millions, the variance of every part follows a curve of plus or minus three sigma or more, and two mating parts can end up as the smallest of one and the largest of the other. Teams must solve for the last half-percent so the build has a high yield, makes money and has few returns.

As for why robotics is suddenly hot, she says there is a dawning realization in the labs that the acceleration is so steep that what you can do behind a keyboard with AI will saturate, though nobody knows when. When it does, the next frontier is hardware, robotics, manufacturing, industrialization, the real-world sensing layer and, eventually, space.

On humanoids like Tesla's Optimus, Figure and 1X's Neo, Kalinowski says they might be close but she has serious safety concerns about large, strong robots working next to people. She points to 1X's Neo as a design that pulled mass inward and made the robot softer, which lowers the impulse of a collision. In her view, humanoids are still prototypes, and the industry has yet to move into the phase of making them cheaper, higher-yield and safer.

She notes that Chinese robots that do real work often ship with a booklet saying no human may be within three feet. On timelines, the problem is scale, which in her world means millions of units, or at least hundreds of thousands. A robot must run day to day without much human repair, and before that comes the supply chain.

She describes the chain from raw magnets to processing, actuators and finished robots, with each layer outsourced over about 25 years to countries such as China, Japan and Korea. She says she was part of that transfer of engineering knowledge. An actuator is essentially the motor, and a ring of magnets is what makes the rotor spin, so if magnets are unavailable, designers must switch to different, possibly larger actuator types. Batteries and actuators are foundational dependencies, while machined and die-cast parts are less critical.

Kalinowski points out that the technology that spins a drone's rotor is essentially the same as what moves a robot's arm, so the supply chain overlaps with military capability. She argues the US should have an independent supply chain, at least on the military side, and should re-industrialize significantly. Allies today may not be allies in the future, and she wants the country to relearn how to make things at scale and process raw materials, so a repeat of COVID would not leave it unable to protect itself.

She agrees with her friend Palmer Luckey that the US needs to invest much more in drones than in aircraft carriers, calling the older way of thinking outdated. She points to Ukraine, where drones are updated daily using 3D printing, and to the cost math of missiles versus the means to stop them, which she says the US is currently losing. She personally chooses not to work for companies building lethal technology, but thinks it is good that others do.

The conversation also turns to safety. Rachitsky wonders what happens when someone prompt-injects a walking robot, and Kalinowski says controlling adversarial threats to the hardware layer will be a huge part of future warfare. She tells a story of giving a sandboxed OpenClaw agent a few details and telling it not to share private information, only for it to post her email address on Moltbook within minutes.

Kalinowski worked at Apple from 2007 to the end of 2012, where she was thermal lead on the first MacBook Pro and later led iterations of the MacBook Air and the cylindrical Mac Pro. She says Apple treats hardware as a first-tier citizen and trains people to think about complex, interdependent decisions and risk. She credits that era's alumni, many now in key industry positions, to a dedication to hardware excellence and good experiments.

Underneath it all, she says, is understanding the first principles of why you are building something a certain way. She cites a story about Steve Jobs and the cabinet maker who finished the back of a cabinet: considering every design decision, even inside the device, forces engineering, design and operations to find what really matters, which tends to look simple in the end. By contrast, Oculus began as a hacking hardware startup that started on forums, and the challenge after Meta's acquisition was professionalizing it with better yields, volumes and costs.

A concrete example is the Quest 2. The goal was to democratize VR, so the team cut the price by redesigning the whole product for cost, removing cameras and components and changing materials and manufacturing processes. She says this led to the best-selling VR headset of all time, and that the product may have ended up stronger for it.

Kalinowski says hardware is much less adaptable to mid-course changes than anything digital, so goals, which she likes to call KPIs, should be written down early and changed as little as possible. If the target price moves from $300 to $150 halfway through, you have burned early time. With clear goals you know when you are done, which matters because each iteration costs three to five months and shipping weeks before a competitor can win the press and the interest.

Her second rule is to design the hardest parts first rather than the ones you know how to design. She describes an architect who started with routing cables through a laptop hinge, since it was unclear they would fit, before finalizing the hinge. Third, the parts customers touch most, like a trackpad, need far more iteration than the rest. Fourth, do anything you know you need to do right now, because a surprise will arrive in two days and you will need the time.

On what goals look like, she says VR has display resolution (pixels per degree) as a key metric, while a laptop has clock speed, weight, price and features. During the MacBook Air, the weight and size goals made it easy to drop features such as the ambient light sensor. She has heard Elon Musk puts numbers on the value of a gram of weight versus cost, so that decisions fall out of the ratios.

On Apple's reputation for ignoring customers, she says it is often misread: if you build something genuinely new, like a touchscreen phone, customers do not know what they want because they have not seen it. If you get stuck in an iterative feedback cycle with customers, it is hard to go from zero to one.

Prompted by a question from Matic's CEO about memory prices as a meteor heading for consumer hardware and robotics, Kalinowski says the industry is in trouble. She believes AI demand and constrained supply are behind it, and she has advised startups and companies to pre-buy memory if they can afford it, as happened in COVID. She guesses prices could double on an unknown timeline, though Rachitsky says he has seen numbers like 6x, and she says data centers are less cost-sensitive than consumer electronics makers.

She explains why this is so serious using a robot vacuum such as Matic as an example: it likely has between 50 and 150 parts at a high level and thousands if you count every small component. A lost die-cast supplier is recoverable in a few months, but losing silicon or the RAM you designed around is a catastrophic redesign requiring a new board, supply chain, production line and reliability testing. In consumer electronics, silicon and displays usually have the longest lead times, while actuators can take a month or two to buy even for prototypes.

This is where vertical integration helps. She cites Tesla and Starlink as examples of companies that can adapt faster to shocks, noting that when silicon was hard to find, Tesla was able to redesign its board in record time to buy different chips. A company with a classic supply chain would find the same shift far more catastrophic.

For build-versus-buy decisions, Kalinowski says she uses off-the-shelf parts whenever she can, especially in prototyping, where the goal is to show it can work at all. She likes works-like and looks-like models, as long as what is shown could fit the final industrial design. For mass production, parts are often custom because of KPIs for size, weight and color, though cheap drones show what happens when off-the-shelf parts are scaled.

On AI's impact on hardware work, she says it is at the very beginning of doing CAD. Claude, for instance, can produce surfaces or point clouds, which is not real CAD, because real CAD is a dense solid with equations describing the surfaces. She would love AI to take over unfun work like tolerance stacks and 2D drawings. Printed circuit board routing and basic component selection are looking increasingly feasible, and she currently uses AI for high-level planning, research databases and fast Excel work.

The deeper problem is that language and video models do not understand friction, weight, contact or surface texture, such as where a hole will land after you fold a paper four times. She suspects world models may need to be the base for CAD and physical engineering, and says half-jokingly that she wants a Codex for hardware engineering. She thinks the full answer may require new model types.

Kalinowski sees a hype cycle around humanoids. She doubts a single generalist robot shape can do everything and says different robots will do different jobs. Screwing a keyboard into a laptop case 10,000 times is a job for a dedicated manufacturing robot, perhaps with standard cabinet sizes and adaptable setups, not a humanoid. Top-tier suppliers in China already run printed circuit board lines with almost no people, and assembly lines that once had 200 workers might have 10.

So humanoids will likely handle some long-tail tasks, while other robots will be built for construction, electrical work, low-volume assembly and logistics, and most will look different from one another. On robots building robots, she thinks AI will be able to go from a 2D picture to 3D CAD, assemblies and vendor communication, but not a single robot that builds itself. It will not be as good as human engineers at first.

The biggest challenge, she says, is data, since companies such as Samsung or Matic will not hand their proprietary CAD to a model maker. That is why she thinks hobbyists, who just want their drone to be faster, are the more interesting starting point, with incumbents slower. She also likes the idea of an AI system running on-premises inside a company's data center and trained on its own data, built on a base model with lots of CAD data.

Answering a question from Rachitsky's sister, a longtime VR creator, Kalinowski draws on researcher Leila Takayama, who taught her that people expect other beings to respond when they enter a space. Humans give each other complex nonverbal cues, such as looking up when someone walks in, and she finds many humanoids creepy right now. Robots should be non-threatening, appear soft, seem reactive and attentive, and show intent before acting, since a robot that looks before it turns is much less alarming.

Rachitsky suggests learning from Pixar, and Kalinowski agrees that Pixar and Disney are probably the world's best at conveying emotion, intent and approachability, even without much physical product volume. On home robots, she says her partner is a high bar, as with Waymo, which she took once and now prefers over everything else. She argues a self-driving car has an existence proof, human driving, and data on safety, whereas a new home robot replaces nothing and so is a harder sell.

Looking five years out, she says her job is to live in the future far enough ahead to design a first product that ladders up to the platonic ideal several versions later, while staying a skeptic so the hardware actually works. She expects AI to hit knowledge work next but the physical world to change more slowly, and says she does not think there will be 20 million robots in five years. She expects supply-chain, raw-material and factory work first, plus weird sights such as delivery robots on the street.

Kalinowski notes that war may see more change than consumer electronics over the next two years. On why she quit OpenAI, she says she has friends there on the executive side whom she cares about, but she felt the speed of decision-making, the governance and the lack of defined guardrails around announcing the Department of War deal were not how it should have been done. She wanted a third path between going along and scorched earth, praises the company and the robotics talent she helped attract, and hopes her decision made it easier for others to state and hold their boundaries.

On hiring for zero-to-one work, she says nobody has built the type of robot she wants at the scale of millions, so she looks for strong generalists who can adapt from other fields, such as autonomous vehicles, plus a mix of people who have built the new thing and people who have scaled other things. She also wants AI-native young people, saying the ones who bake AI into their engineering process are about 20 or 21 and are much faster.

She does not accept that junior roles are disappearing: teams need both senior and junior people, though they may be smaller. She adds mission alignment as a unifier across AI researchers and hardware people, and says she relies on gut feel for a spark in people who are motivated by learning and excellence, open to updating their views, and who want to win.

Asked what she learned from legendary builders, Kalinowski says Sam Altman is really good at asking "why not more?", pushing her to think bigger than she had been. From Steve Jobs she took an unwavering bar for technical talent and excellence, noting that hearing something is not good enough can be very motivating to an ambitious young person. From Mark Zuckerberg she learned how well a company can be run: decisions made at the lowest level possible, clear objectives, and leaders such as Andrew Bosworth who could digest 20-page technical reports.

For her failure story, she recounts the Quest 1 at EVT, when the team had cut from five cameras to four to reduce cost. Just before Christmas the computer vision lead said camera data could not get a lock on the headset's position, and they discovered the two teams had interpreted a plus-or-minus spec differently, apparently around 0.15 mm. The fix was an architectural change: lock the bottom two cameras together on a bracket, made of steel to hold tolerance, and let the other two float.

The team kept the build on time and shipped on schedule, and the new design turned out better, since the favored pair provided a source of truth for the space. She still wishes they had caught it four months earlier. Closing out, she encourages people to use AI tools daily, retest with every new model, and embrace the growing power of the individual, and she asks for help imagining the future together rather than accepting a dystopian default.