The AI Hardware Boom Depends on Magnets, Memory, and Manufacturing Scale

Caitlin Kalinowski sees the sudden interest in hardware and robotics as a response to a specific expectation inside AI: that the gains from doing work “behind a keyboard” will eventually saturate. She does not claim to know when that happens, and says nobody else knows either. But she argues that many labs, big technology companies, and startups are arriving at the same conclusion at roughly the same time: if AI systems become increasingly capable at solving digital problems, the next frontier becomes moving, sensing, making, and acting in the physical world.

When that happens, Kalinowski says, “the next frontier is the physical world”: robotics, manufacturing, industrialization, sensing, and the ability to manipulate real-world objects. That framing explains why a field that has long been less fashionable than software is suddenly attracting students, founders, and AI companies. Hardware was “never the sexy career,” she says. People went into it because they loved it, and it generally did not pay like software. The exception, in her telling, was Apple’s hardware lineage. Now “everyone is suddenly asking about hardware and robots and the physical world.”

Kalinowski’s version of the physical AI boom is broader than consumer humanoids. It includes robotics, manufacturing, industrialization, autonomous vehicles, drones, sensing layers, and the ability to manipulate real-world objects. The path from AI models to the physical world is not simply a product trend. It is a stack of physical dependencies, safety constraints, manufacturing processes, and design disciplines that software-native companies often underestimate.

Kalinowski’s career gives her a particular lens on that transition. Lenny Rachitsky introduces her as a former Apple product design engineer who worked on the unibody MacBook Pro teams and was technical lead on the MacBook Air and Mac Pro; as the former head of VR hardware and AR glasses hardware at Meta; and as a former OpenAI member of technical staff helping build robotics and consumer hardware work. Across those roles, the recurring lesson is that AI hardware is not software with a casing. It is a domain where a missing part, a misunderstood tolerance, or a late design goal change can upend the entire program.

The simplest way Kalinowski explains the difference is by comparing hardware to compiling code. Software teams can write code, compile, run, debug, and repeat constantly. Hardware teams, she says, may only “compile” four or five times total. A major hardware build means redesigning in CAD, releasing parts, building them, testing them, and then, at the end, releasing for mass production. Once that last compile happens, there is no equivalent of an over-the-air update for a bad hinge, bad bracket, or bad component fit. The product is out in the world.

That constraint changes the culture. Hardware teams must be more conservative. Reliability checks and tests have to be built into the program, not deferred. Every product with millions of units is really a distribution of parts: the smallest version of one component will meet the largest version of another, and they still have to fit. “The part variance is pretty high,” Kalinowski says. The job is to solve the last half-percent of build risk before the final compile, so yields are high, returns are low, and the company can make money.

For software companies moving into physical products, that is the first hard lesson: iteration still exists, but it is slow, expensive, and bounded by manufacturing reality.

VR is not treated here as a simple failure. Kalinowski says she expected it to be big—otherwise she would not have worked at Oculus—but she now sees it as one step in a longer technological arc. VR helped teams learn how to orient digital objects in simulated and real space, connect those spaces, do positioning with cameras, use depth sensors, and understand how humans perceive visual data spatially.

Those technologies are now directly relevant to robotics. A robot moving through space needs to know where it is, how far it is from other things, what it is seeing, and how to map its environment. If a human is wearing a headset to operate or drive a robot, Kalinowski says, the underlying technologies overlap even more. VR gaming may remain an “interesting niche,” but the systems developed for VR are becoming useful in robotics, autonomy, manufacturing, and drones.

She attributes VR’s limited mainstream adoption partly to the social problem of covering the face. Lenny Rachitsky describes the experience of using a headset as magical but says he does not want to sit on a couch disconnected from the world. Kalinowski agrees that “it’s hard to make it social when you have your face covered,” and points to Google Glass as another example of how important social acceptability is for face-worn technology.

Her optimism shifts to AR glasses, but with caveats. She believes looking down at phones all the time is not great for humans as social creatures, and that glasses could preserve social connection while providing information. Orion, the AR glasses she worked on at Meta, is in her view “a bit ahead of its time” because the waveguides and microLEDs are not yet ready for mass production. The yields are not there and the cost is still high.

Still, she describes Orion’s 70-degree binocular field of view as a meaningful glimpse of the future. It is hard to describe until someone uses it, she says, but a wide enough field of view makes the user feel immersed. The remaining challenges are not only display technology. Input is unresolved: how does someone communicate with glasses while moving, in public, quietly or silently? Kalinowski’s expectation is not always-on visual clutter, but a display that is mostly off and can be turned on when wanted.

The broader point is that VR, AR, robots, drones, autonomous vehicles, and manufacturing are not separate technology islands. They share component families and technical problems: sensing, localization, spatial understanding, batteries, displays, silicon, cameras, motors, and control systems. In that sense, the investment in VR did not disappear. It built pieces of the physical AI stack.

Humanoid robots occupy a narrower place in Kalinowski’s forecast than the current hype suggests. She does not dismiss them, and she says “we might be close” in some sense. But in her world, humanoids are still prototypes—advanced prototypes, but prototypes. The next stage is not merely making them more impressive in demos. It is making them cheaper, easier to manufacture, higher-yield, reliable day after day, and safe around people.

Her first concern is safety. Large, strong humanoids operating next to humans need enough data to show they are safe. She points to 1X Neo as an example of a design that has made significant safety considerations, including pulling mass inward. A lighter, softer arm is less dangerous because impact depends on the moving limb, the rotating actuator, and the compliance of the surface. A hard arm creates a higher impulse; a softer, compressible arm lowers it.

This is not a cosmetic issue. A robot strong enough to do meaningful work can also hurt someone. Kalinowski says some Chinese robots can do many tasks, but the booklet warns that no human should be within three feet of the robot. She does not expect many robots strong enough for meaningful work to lack that kind of warning today.

The harder problem is scale. For Kalinowski, scale usually means millions of units; even hundreds of thousands is a major challenge. Before that can happen, a company needs a working design, reliability without constant human repair, and a supply chain that can actually provide every part.

The public fascination with humanoids also risks obscuring where robots are most useful. There is a hype cycle, she says, around a “generalist robot shape to do everything.” She expects winners in humanoids, but she doubts one form factor is right for most jobs. In manufacturing, a humanoid is often the wrong answer. If a laptop keyboard needs to be screwed to a case, the right tool is probably a dedicated robot that drives ten screws into that case thousands of times, not a human-shaped generalist.

Modern top-tier manufacturing lines in China already use very little human labor in some areas, according to Kalinowski. Printed circuit board lines can run with essentially no people unless something goes wrong. Mechanical assembly lines that once used 200 people might now use 10. That means the task is not always to replace a human with a humanoid. It is to build more specialized robots: for construction, electrical work, low-volume assembly, logistics, and manufacturing.

She expects humanoids to matter for long-tail tasks humans currently do. But most robots, in her forecast, will not look like people.

The AI hardware boom runs into old industrial constraints quickly. Kalinowski repeatedly returns to supply chains, not as background logistics but as a gating factor for robotics and physical AI.

The chain begins with raw materials. Magnets are a useful example. Robots need actuators, and actuators are motors: electricity goes in, motion comes out. A typical actuator uses magnets and a rotating rotor, with gearing that powers a limb, head, finger, or other mechanism. A schematic in the source reduced the actuator problem to magnets, fields, and efficiency: motors use electricity to create a moving magnetic field; that field interacts with permanent magnets inside the motor; better magnets mean more motion from less space, weight, and energy.

Kalinowski says the supply chain for raw magnets, processing, actuator integration, subassemblies, and final robot integration has been outsourced over roughly 25 years to countries including China, Japan, and Korea. She is explicit that she was part of that transfer of engineering knowledge to Asia. The historical bargain, as she describes it, was that Asia had the expertise to scale production and build many parts at lower prices, while design and other functions were often elsewhere. Now, in her view, safer supply chains require more independence at multiple layers of the stack.

Actuators may become a bottleneck. If companies cannot get magnets, they may need to design different actuator types using other materials, which could be larger or less efficient. If they cannot buy actuators, they cannot make robots. Batteries, die-cast parts, and machined parts also matter, though she views machined parts as less critical because they are more likely to be obtainable.

Memory is another looming shock. Rachitsky introduces it through a warning from Matic CEO Mehul Nariyawala: “There’s a meteor called memory prices that are coming for consumer hardware and robotics and physical AI.” Kalinowski’s response is blunt.

She says she is not an expert on memory markets, but believes AI and constrained supply are part of the problem. She has been advising startups and companies to pre-buy memory and keep enough in stock, if they can afford it, to ride out price spikes. During COVID-era disruptions, she says a company she worked with also had to pre-buy memory. If a key component such as memory or silicon is constrained, the options are limited: pay the higher price or already have inventory.

Asked what will happen, she says she does not know. She thinks prices will “probably” double, but does not know the timeline. Rachitsky mentions having seen figures like a 6x increase, but tells listeners not to quote that; Kalinowski says she did not realize it was that bad. Her larger point is less about a specific price forecast than the structural risk: demand from data centers may be large and less price-sensitive than consumer hardware companies can tolerate. A company building physical products cannot simply absorb every component spike.

The component count makes this fragility concrete. Using Matic as an example, Kalinowski estimates the robot vacuum may have 50 to 150 parts depending on how counted, and thousands if every small capacitor and PCB component is included.

It has wheels, a vacuum, a mop, a vacuum bag, a liquid reservoir, a SLAM-based mapping system, wireless modules, a system-on-chip, RAM, and printed circuit boards. Kalinowski says she believes the mapping stays on the device rather than going to the cloud, which she compares to VR practice and describes as good for privacy. If one required part is missing, the device cannot ship.

Not every missing part is equally bad. If a die-cast component vendor goes out of business, a company may recover by finding another supplier over a few months. If silicon becomes unavailable, the board may need redesigning. If the required RAM in the needed form factor cannot be secured, Kalinowski calls that “essentially a catastrophic redesign”: redesign the guts, secure new supply, rebuild on the line, retest, redo reliability work.

That is why hardware companies often begin with the longest-lead and hardest-to-change items. In consumer electronics, she says that often means silicon and display. In robotics, actuators can be hard even at the prototyping stage, sometimes taking a month or two to buy.

Vertical integration becomes attractive in that context. Kalinowski points to Tesla and especially Starlink as examples of Elon Musk verticalizing supply chains. She says she has heard Starlink described as effectively “ore and silicon chips in, product out,” and calls that “a pretty incredible factory.” Verticalization, in her account, makes it easier to adapt to supply shocks. She also cites Musk’s ability to redesign a PCB quickly when silicon was hard to find as something that would be far more catastrophic for a company with a classic supply chain.

The same motors, magnets, and control systems that make a robot arm move also make drone rotors spin. That overlap leads Kalinowski from robotics supply chains into military strategy.

She argues that the United States needs a more independent supply chain at least on the military side, and says other countries should think similarly. Her reason is not isolationist nostalgia; it is resilience. COVID and war have shown how quickly conditions can change. Allies today may not be allies in the future. The “allied West,” in her phrasing, is going through geopolitical shifts. She wants the country to “re-teach ourselves how to make things at scale,” process raw materials, and be more independent so that future disruptions do not leave it unable to protect itself.

She agrees with Palmer Luckey, a friend with whom she says she does not agree on everything, that the United States should invest much more in drones than in aircraft carriers. Aircraft carriers still matter, but she sees them as part of an older way of thinking. AI and military technology are changing fast, and Ukraine is the example she points to: drones changing and updating rapidly, including with 3D printing.

The strategic math concerns her. If one side launches a missile cheaply and the other side spends much more to stop it, “you have to do the math every time.” Right now, she says, “we’re losing on the math.” That may be fine for a limited period, but becomes less fine the longer it continues.

Kalinowski does not work on lethal technology and says she has chosen not to work for companies that create it. But she also says it is good that some people are willing to do that work. In her view, building the future requires different people taking different roles.

Her expectation for the near term is striking.

She ties that to democracy and the need to defend it with capabilities while hoping to avoid hot conflict. Rachitsky invokes a Marc Andreessen image of a hundred thousand drones coming out of China; Kalinowski’s answer is not to dismiss the scenario, but to return to industrial capacity and speed.

The AI safety version of the same issue extends to adversarial control of hardware. Rachitsky raises prompt injection and jailbreaking risks for chatbots, then asks what happens when a robot can be told to punch someone. Kalinowski says controlling adversarial threats to the hardware layer—robotics, drones, and more—will be a huge part of future warfare.

She gives a smaller, nonphysical example from using OpenClaw. She sandboxed it on its own computer and gave it a few pieces of information, including her real email address. She told it not to share her private information. Five minutes later, she says, it had posted her personal email address to Moltbook. The story is funny at chatbot scale. It is less funny when the agent has physical capabilities.

Kalinowski says she left OpenAI after disagreeing with how the company handled what she calls the announcement of the Department of War deal. She says she still has many friends on OpenAI’s executive side, considers them good people, and thinks OpenAI is an amazing company. She also says the decision-making, speed, governance, and lack of defined guardrails around that announcement were not how she thought it should have been done.

The placement of that disagreement matters in her broader argument. She is not opposed to physical AI, robotics, or even the need for countries to take military technology seriously. She argues elsewhere that drones, supply chains, industrial capacity, and adversarial control of hardware are central safety issues. Her objection to OpenAI’s announcement was about process and boundaries: how quickly the decision was made, how governance worked, and whether guardrails were defined.

She wanted a “third path” between going along silently and scorched-earth departure. In her account, neither extreme fit the situation. She cared about the people and the work, including the robotics program she helped build and the talent she helped attract. But after the announcement, she says she could not continue because “you don’t know what’s gonna happen next time.” Because her exit was going to be reported, she posted about it first.

Rachitsky asks what OpenAI is working on in robotics and consumer hardware, and Kalinowski declines to discuss anything internal or any intellectual property. The only characterization she gives is that the team is strong and that she was grateful for the opportunity to help build it. The restraint matters because her criticism is about process and boundaries, not a disclosure of the company’s plans.

Her hope was that leaving would make it easier for other people to talk about their own boundaries and hold them. That position fits the broader tension in her outlook. She is excited about AI, hardware, robotics, and individual leverage. She is also worried about military risk, adversarial control, supply chain fragility, safety, governance, and the future people are choosing to build.

Apple’s influence on Kalinowski’s hardware philosophy is narrower and more practical than a generic cult of taste. Hardware was a first-tier citizen there, she says, and the company trained people to think through complex, interdependent decisions and risk. Looking back at the 2007–2012 period when she was there, she says many people from that era now hold key positions across the industry, which she attributes partly to that training.

The lesson was not that every detail matters equally. It was that every design decision should support the reason the product exists. Rachitsky compares Steve Jobs’s “back of the cabinet” standard to the brown M&M story—a way to test whether people read the contract. Kalinowski redirects the analogy. The message, she says, is understanding why you are doing what you are doing, and then making every design decision support that goal.

That kind of methodical work can make the final product look simple. When engineering, industrial design, and operations are forced to examine the inside of the device as carefully as the outside, what really matters tends to rise out. It is not only an aesthetic exercise; it is a forcing function for technical clarity.

Clear goals are especially important because hardware cannot tolerate constant late-stage changes. Define the goals early, write them down, and change them as little as possible. A price target is one kind of goal. If a company starts building a $300 device and halfway through decides it must be $150, much of the early work may be wasted. Other goals might include display resolution, pixels per degree, clock speed, parallel processing, weight, features, or cost. The point is to know the key metrics and tradeoffs early enough that engineering decisions can “fall out” from them.

The MacBook Air is one example. The goal was weight and size, enabled by machining. That made some features, such as an ambient light sensor, no longer make sense. Because the overarching goals were clear, the team could jettison features that did not serve them. Kalinowski says the earliest MacBook Air—the one Steve Jobs famously removed from a manila envelope—was a low-volume proof of what could be done with CNC machining. The later wedge-shaped MacBook Air, which she worked on, carried that proof into a higher-volume roadmap.

For VR, one key metric is visual resolution in arc minutes or pixels per degree. She compares it to Apple’s retina displays: once a product reaches what the human eye can resolve, the engineering pressure on resolution can change. Mass-produced VR is “not even close” to retina, so resolution remains a key metric.

Quest 2 is her concrete Meta example of goal discipline. The team wanted to democratize VR by lowering price. That required redesigning essentially the entire product for cost: removing cameras and components, changing materials, and changing manufacturing processes. She says that alignment around the goal led to what she believes became the highest-selling VR headset of all time, while still producing a high-quality product with low return rates.

Once goals are defined, Kalinowski’s operating advice is to sequence the work around risk rather than familiarity. Hardware teams often start with what they know how to design. The best architects start where the product might fail. If cables must route through a hinge and it is unclear whether they will fit, begin with the hinge, cable cross-section, and routing before finalizing the rest. Do not start with the display just because it is familiar.

The parts the customer touches or interacts with most deserve disproportionate iteration. On a computer, that means the trackpad first and the keyboard next. Those parts have to feel good, respond correctly, and be highly reliable. Less-touched components may need less iteration. Rachitsky asks about Apple’s butterfly keyboard; Kalinowski says she did not work directly on it and cannot explain what happened, but agrees this is exactly the kind of thing a company has to get right. She adds that modern MacBook keyboards are excellent.

Her fourth principle is urgency. Hardware teams “can’t wait around ever.” If something is known to be necessary, do it now, even if the schedule appears to allow more time. Surprises will arrive. She attributes this lesson to people including Shelly Goldberg and Kate Bergeron at Apple: stack the known tasks and get them out of the way. When Rachitsky summarizes this as “you never know what’s around the corner,” Kalinowski sharpens it: in hardware, “you actually don’t have more time.”

Her failure story from Quest 1 shows the cost of getting a detail wrong. Around EVT—the point where hardware is first compiled with final components and materials on mass-production-intended tools—the team discovered a mismatch in how a camera-positioning spec had been interpreted. They had gone from five cameras to four for cost reduction, but the computer vision team could not reliably lock onto where the headset was in space. The product design team had interpreted a tolerance differently than the computer vision team needed.

The fix required an architectural change: lock the bottom two cameras to each other on a bracket so their relative distance met the needed spec, and let the other two float. It was stressful, late, and the kind of failure she wishes had been caught four months earlier. But it kept the build and ship schedule on time, and the new design turned out to be better because the fixed pair became a source of truth for spatial tracking. On the shipped Quest, she says, two cameras are closer together at the bottom front because of that redesign.

The story also clarifies why off-the-shelf parts and prototypes have their place. In prototyping, Kalinowski says, the goal is to show that something can work at all. Off-the-shelf parts are useful whenever possible, especially for “works-like” prototypes. The caveat is that the prototype must still be plausibly compatible with the final industrial design. Once a product moves toward mass production, custom parts become more common because size, weight, color, cost, and performance targets often cannot be met with standard components.

AI has already changed how Kalinowski works, but not yet in the way software engineers might expect. It is not doing the “meat and potatoes” of mechanical and electrical engineering day to day. It is useful for strategy, planning, research, dependency mapping, database building, and spreadsheets. She uses it to ask who else is making a similar product, to build imperfect but fast databases, and to work in Excel, which remains central to engineering workflows.

The transformative threshold would be real CAD. Kalinowski says current models are at the very beginning of being able to do CAD. Claude can do surfaces or point clouds, but in her world that is not real CAD. Real CAD is dense: solid entities, NURBSNon-uniform rational B-splines, a mathematical representation used to define curves and surfaces in CAD., equations describing surfaces, components and assemblies that can be manufactured and fit together. She expects AI may eventually do rapid design and accelerate hardware engineering significantly.

She would especially like AI to handle some of the tedious but necessary parts: custom screws, 2D drawings, tolerance stacks, and checking whether seven parts always fit together properly. Printed circuit boards are another near-term area. PCBs have internal layers and components on top; she says AI increasingly appears able to route inside boards and may do basic component selection and layout.

But physical engineering requires understanding that current language and video models do not have. Kalinowski wants systems that understand friction, weight, contact, pressure, surface texture, and the behavior of physical objects. She gives the example of folding a piece of paper four times, making a hole, and knowing where the hole will be when it is unfolded. LLMs and video models are not good enough for that kind of physical reasoning.

Her hypothesis is that world models may be needed as the basis for CAD and physical engineering work. What she wants is “Codex for hardware engineering,” and she suspects it may require new model types.

The data problem may be the largest barrier. High-quality CAD is among the most valuable intellectual property a hardware company has. Samsung, Matic, and others will not simply hand their 3D CAD to a model vendor to train a general model. Kalinowski thinks hobbyists may be the more natural starting point because they are less concerned about protecting CAD and more interested in making things faster. A hobbyist may not care about expert PCB design; they may just want a faster drone.

Over time, she imagines on-prem systems that train safely inside a company’s data center on proprietary CAD. But that still requires a base model with enough CAD knowledge, plus a way to adapt it securely inside company walls. She is unsure what the architecture will look like, but considers the long-term direction plausible.

Making robots acceptable around people is not only a matter of force limits and softer arms. It is also a design problem involving social cues, motion, and intent.

Kalinowski credits researcher Leila Takayama with helping her understand how humans expect other beings to respond in shared space. When someone enters a room, people acknowledge them in subtle ways. They may look up without speaking. They give and read complex nonverbal signals. A robot that does not respond can feel creepy.

Her design requirements are plain: robots should be non-threatening, appear soft, react to people, convey that they know someone is there, and make clear they are present to help. They should also show intent before acting. A robot that suddenly turns and moves can alarm people. A robot that looks before it turns, then moves, is less alarming.

She thinks Pixar and Disney are especially relevant to robotics, even if they have not produced physical robots at volume. Their characters communicate emotion, intent, approachability, and engagement through motion and expression. The examples shown alongside the point included Pixar and Disney imagery, Baymax, and WALL-E—characters whose designs make softness, curiosity, and readable intent legible.

The home raises the bar further. Kalinowski is interested in home robots, but says her partner is not easily convinced. Her partner resisted Waymo until taking one, then did not want to take anything else. She loves the Matic robot vacuum. Kalinowski takes that as evidence that the bar for a home robot is high but movable if the product is genuinely good.

She distinguishes home robots from self-driving cars. A self-driving car replaces a known category: human driving. There is an existence proof, and companies can compare safety data. With home robots, the object may be doing something that did not previously exist in the home. If it is bad at the task or unsafe, the comparison is less obvious. That makes trust harder to earn.

Even cars reveal how much human communication matters. Rachitsky notes that self-driving cars can make social interactions at intersections awkward because human drivers normally rely on eye contact and gestures. Kalinowski says she has experienced that too and jokes that one almost wants little arms at the front gesturing “you go.” The point is serious: even basic movement through shared space depends on human connection and signaling.

Hardware leaders need to “live in the future” while remaining skeptical. A device shipping two or three years from now must also ladder toward a six-year vision. Hardware rarely gets to execute a one-shot perfect product; it iterates toward a final goal. That requires imagining the Platonic ideal of the product and the first shippable version at the same time.

Kalinowski also says a good hardware leader should be worried. If she assumed everything would be fine, the hardware would not work. Her job includes seeing what will fail and worrying through the details.

Her five-year forecast separates software-like AI change from physical-world change. AI will foundationally change work over the next couple of years. Coding is already changing; knowledge work is next, in her view, and the economy will be progressively affected. The physical world will change more slowly, except in areas like drones and self-driving cars. There will be more robots, delivery machines, and strange devices in public, but she does not expect 20 million robots in five years.

The reason is not lack of ambition. It is supply chain reliability, raw material access, safety, and the need to relearn how to make high-tech products in the United States. Those are deep constraints. They do not move at the speed of a software deployment.

This is also why she sees young AI-native engineers as essential rather than obsolete. In hiring for zero-to-one robotics and AI hardware, she looks for strong generalists who can adapt what they learned in one field to another. She wants some specialists who know robotics from scratch, some people who have scaled other products to high volume, and people from adjacent domains such as self-driving cars, where sensing stacks and safety tradeoffs overlap.

But she also wants “AI native” young engineers—people around 20 or 21 who use AI from the ground up in their engineering process. She says they approach problem-solving differently and move faster. She does not accept the simple narrative that AI eliminates junior roles. Teams need both senior and junior people, though team sizes may become smaller.

The junior point is not only about staffing. Kalinowski argues that companies still need to build new technologists. If teams stop mixing senior and junior people, the industry loses the path by which people learn to become senior. At the same time, the youngest engineers may teach older engineers how to work with AI more natively. Kalinowski compares herself and Rachitsky to digital natives: people who came of age with the internet and early mobile phones, but not with AI as a default tool. The next cohort’s advantage is different, and hardware teams need access to it.

Mission alignment matters because AI researchers and hardware people come from different worlds and can miscommunicate. A shared mission helps unify the team. Beyond that, Kalinowski relies partly on judgment: people with a spark, genuine motivation, a desire to learn, openness to updating their views, and a desire to win.

Her lessons from prominent leaders map onto those hiring and operating beliefs. From Sam Altman, she says she learned “why not more?”—why not 100x or 10,000x, why not think globally, why not think bigger. From Steve Jobs, she learned the force of an unwavering bar for technical talent and excellence. Hearing that something was not good enough could be highly motivating for ambitious young people. From Mark Zuckerberg, and also Andrew Bosworth, she learned the value of a clean, well-run technical organization: decisions made at the lowest possible level to preserve speed, clear reviews, defined objectives, and senior leaders able to read long technical reports, understand tradeoffs, and contribute.

Kalinowski’s final note is not a prediction about a single device category. It is a claim about agency. She says the current cultural imagination often feels stuck in a dystopian niche, where the future seems horrible by default. Her answer is not simple optimism. It is design: decide what future people want, what human aspects should be preserved, how humans should be augmented, and then build toward that picture.

That extends to personal work as well. She describes the present as scary but unusually empowering. AI tools can let individuals do more than they could before, and she encourages people to use them daily, test the boundaries, and test again whenever a new model comes out. Her reason is practical: people need to know what the tools can and cannot do.

Her closing request is not for a single company to solve the future. She says imagining the future is “not a single-player game.” It requires fiction, literature, conversation, and building. The alternative, in her framing, is to let a default picture of the future harden without deliberate human choice.