Key Takeaway

Choosing robots was not a whim: first, the hardware supply-chain expertise accumulated over nine years at Caper could be reused; second, in 2023-2024 large models began turning toward the physical world (the rise of ChatGPT, the emergence of the Mobile ALOHA / UMI papers), giving robots for the first time a chance to escape the customized mode of traditional production-line automation and move toward generalization and platformization.

MotivationAnother vertical company would not mean much

Caper was already a company doing well in the retail vertical. A second startup in another vertical would not mean much; the team wanted to bring nine years of software-hardware experience into a bigger race. York himself has a lasting interest in tangible, touchable things — switching from electronics to software as an undergrad at Zhejiang University was purely a job-prospect decision, and the desire to build real products with both software and hardware never went away.

Personally, I am actually especially interested in physical things, and not that interested in purely virtual ones. [00:04:01]

Timing2023-2024: Large models begin turning toward the physical world

When ChatGPT took off in 2023, the team's first reaction was that this thing would definitely eventually be usable in the physical world. In 2024, papers like Mobile ALOHA and UMI gradually appeared, creating early landing opportunities. What really attracted them: model breakthroughs might let robots escape the exhausting past model of heavy customization per client and become a generalized, platformized business.

Without this model breakthrough, we would feel the previous wave of robotics business was exhausting — every client needed heavy customization, which is nothing like the internet, where you build a platform and all clients use the same thing. [00:06:50]

Key Takeaway

The ChatGPT moment is essentially the combination of a model (software) + phone and computer channels that have existed for 20 years + clear demand. Even if models are extremely intelligent today, the next day you still cannot deploy them onto all kinds of hardware — because robots have no channels. So Dyna has done deployment from the company's very first day: one, to build channels; two, to get real customer demand and form a positive feedback loop, rather than only imagining in the lab what the data flywheel should collect.

ArgumentAll explosive growth happens inside existing channels

Large models are pure software; the distribution channels (phones + computers) have accumulated over 20 years and are the world's strongest millisecond-level reachable channels. Autonomous driving's channel is also relatively fixed (cars are already on the road, and people actively choose camera- and lidar-equipped ones when replacing). Only robots: lacking intelligence, they can do nothing in most scenarios; the previous generation of robots was stuck on fixed production lines — which precisely do not need the generalization embodied intelligence brings.

Look at all of human history: up to today, all explosive growth has actually only happened to products inside internet-related channels. [00:22:43]

Suppose our models are already extremely intelligent today… the next day you still cannot deploy your model onto all kinds of hardware devices, because you don't have this channel. [00:08:29]

Entry PointFolding: a small scenario, a big channel

Folding napkins and towels seems to replace only one or two workers in a store, but once a robot enters a commercial service channel like a hotel laundry room and the robot body is relatively general-purpose, housekeeping, cleaning, and food prep all become a software game — the hardware channel already exists, and more features unlock as model capability stacks up. This contrasts with the shared-bike counterexample: there isn't much you can do on a bike, but a general-purpose robot body is different.

Once you're inside this channel and your robot is standing there… what follows is a software game — as model capability stacks up, you can unlock more and more software features. [00:24:50]

StickinessHardware + B2B switching costs are extremely high

Responding to the host's challenge — if the robot body isn't the endgame, will the channels be built in vain: software switching costs are extremely low (York himself switched back and forth between Codex and GPT within days), so there is no channel moat; but hardware plus B2B businesses are far stickier than consumer and software. Back then supermarkets wouldn't even replace POS systems from the 90s; Square could enter restaurants and coffee shops but not supermarkets, precisely because of the heavy integration overhead of inventory, tag systems, and the like. Once a robot is deeply bound to a hotel's entire operating workflow and management software, it cannot simply be swapped out by whoever has better and cheaper hardware.

Hardware plus B2B is actually a very hard thing… customer acquisition costs are far higher than in the consumer market. [00:28:52]

Key Takeaway

In an early stage where neither the market nor the supply chain has converged, optimizing any single link too early means you don't know what upstream and downstream expect of you, creating massive uncertainty. Today China has no mature, cheap, deployment-ready hardware available, so building your own hardware is actually cheaper and more efficient — full-stack flexibility is the core advantage of this stage; once hardware converges or the robot's ChatGPT-scale model appears, it won't be too late to adjust strategy.

CounterintuitiveBuilding your own hardware is cheaper and more efficient

Many believe that since China's hardware industry is developed, there will eventually be many Unitree-like companies supplying hardware. York believes that will happen, but it hasn't yet — leveraging the social division of labor doesn't yet deliver marginal cost reduction, because things change too much. Dyna had one hardware R&D person from day one; today it has grown into a full hardware R&D and manufacturing supply-chain department in Shanghai, continuing Caper's hardware-made-in-China structure.

In embodied AI right now, it's a very counterintuitive thing: building your own hardware is actually cheaper and more efficient, rather than relying on the division of labor to have each factory make one component. [00:15:29]

ModelFoundation model and post-training climb in alternation

Dyna's earliest public demo was folding napkins for 24 hours straight, because deployment scenarios require continuous, uninterrupted work. On the data strategy: first accumulate high-quality robot-body data (rather than early low-quality teleoperation/human data), and this year begin gradually ramping in lower-quality data for pre-training. As upstream data diversity grows, the amount of post-training data a single downstream scenario needs will shrink — piling up only towel-folding data overfits to tricks.

The foundation model goes up a bit first, then your post-training climbs a bit more… so it's this alternating upward pattern. [00:20:44]

EvolutionFrom VLA toward video models as the backbone

Models are fully in-house: early on they borrowed from the π series but later adjusted everything themselves; recently they've started building a world model — more precisely, a model architecture with a video model as the backbone — gradually shifting away from the VLA route. On hardware they currently use off-the-shelf arms plus self-designed frames and shells for fast iteration, aiming to de-risk the model layer first — hardware is a long game, and being best on day one doesn't guarantee long-term leadership.

Key Takeaway

The value of the 10 months isn't how much money was made, but the lessons learned: staying one week is completely different from staying ten months. Three big categories of problems had to be solved: hardware (several arms replaced), operations (dispatching people on-site vs. a remote observability system), and changing requirements (updating the model daily vs. describing behavior changes in language). As for a domestic company mass-deploying in-home robot services, York's take: entering scenarios to gain know-how is reasonable; scaling up before the product is mature harms the company.

PracticeThree lessons from the laundry room

First, hardware: when the arms bought from suppliers early on failed, they could only swap them, not dig in and change the design — which is exactly what later drove them to build their own hardware department. Second, operations: when something breaks, do you send someone over, or monitor and fix it remotely with an observability system? Third, changing requirements: when clients have special demands on the finished product, do you update the model every day, or describe the behavior change to the robot in language?

These ten months, for us, aren't about how much money we made from this client — it's really about how many lessons learned this client brought us. [00:33:15]

LessonThe two-year cost of forcing the first-generation cart to mass production

Caper's deepest lesson: when the first-generation cart finished EVT/DVT/PVT and was heading to mass production, several hundred units deployed with too many problems, trapping more than half the team in firefighting — P0 issues forced everyone to seek the fastest, simplest fixes, nobody built long-term systems, and the product saw zero evolution for two or three years; the user interface didn't change at all in two years. The mature product eventually deployed at scale was the third generation.

Scaling up while your product is immature is extremely difficult, and it is harmful to the company… everyone is doing fire fighting — when a client has a problem you go fix it, and again, and again. [00:36:38]

Key Takeaway

What matters most about the three gates is that they cannot drift too far apart: you may start from any one of them, but you must not max out any single gate, or external expectations and internal reality misalign and the company gets crushed. Also beware two popular fallacies: equating sold with ROI (the metaverse is the counterexample), and hoping to raise big money at IPO to ride out the winter (a company that raises a billion burns money differently from one that raises a hundred million — investors won't let you save). Technically, demand is the essence of every engineering discipline — look at technology from demand, not demand from technology.

FrameworkThe three gates: the art of not drifting too far apart

While writing, York found little evidence of a fixed order — success and failure cases are quite random in sequence (some companies raised little money early yet ran long and survived). So the conclusion converges to: exactly how far apart is very hard to measure — this is art, not something a formula can measure.

The core of the three gates is that they cannot drift too far apart… once one gate maxes out, you easily fall into a state where external expectations don't match internal reality, which makes it even easier to crush the company. [00:38:34]

CounterexampleUsage ≠ ROI: the metaverse mirror

What the company keeps defining internally: after the machines are sold, do customers actually save money long-term — not idle units brought in for PR. York himself bought a pile of AR/VR headsets that now sit in a corner at home, unsellable — sold but unused; in the end, what lets a company survive is still high-frequency usage and real profit.

It sold, but that doesn't mean it really has ROI. The real question: long after it's sold, is it used at high frequency, and does it actually make money in use? [00:41:25]

CritiqueThe paradox of raising big money to weather the winter

About quite a few companies rushing to IPO with still-unclear business models, York says outright it doesn't make much sense: raised money isn't left unspent; investors in a billion-dollar round expect you to go maximal scale and maximal aggression — try to scrimp and ride out the winter, and investors won't stand for it. By contrast, an IPO from a company like Unitree with clear profits at the hardware infrastructure layer is perfectly fine.

A company that raised a billion dollars and one that raised a hundred million burn money at different speeds… investors' expectation for your billion is that you charge. [00:40:04]

First PrinciplesDemand is the essence of every engineering discipline

Technology-driven forces bring variables (like large models), but without channels such as computers and phones, technical breakthroughs cannot turn into commercial adoption. Jobs' iPhone wasn't technology-first either — it precisely captured humanity's subconscious demand to make the phone more useful. Engineers are a tiny slice of human society — he dislikes the argument that if you push technology to the extreme you will surely win.

Every engineering discipline… exists at its core because people are too lazy — humans build something to help themselves — so starting from demand is extremely important. [00:45:00]

Key Takeaway

Caper started in 2016 with smart security tags (a YC project) and lost to B2B activation costs — a store unwilling even to swap security tags would never accept an Amazon Go-style retrofit; pivoting to high-frequency supermarket carts was the right answer. The Instacart acquisition was itself unplanned (they were raising a Series B at the time). Looking back, York admits that purely commercially Caper was a slow company: supply chain and manufacturing demanded deep technical accumulation while the deployment scenario was extremely vertical, so the input-output ratio wasn't great — the outside skepticism of selling shopping carts for that much money makes a lot of sense to him today.

Starting PointSmart security tags: losing to B2B activation costs

Four graduates with no work experience (CEO Lindon had a bit over a year in investment banking) started from the pain point that queuing wastes too much time. The first product stuck a QR code on the security tag — pay with your phone and it unlocks automatically. It looked simple, but the bottleneck was operations: no store would close for a day or two so every employee could retag all the clothes in the store. This taught them: for B2B merchants, activation must be easy — which is also why they didn't follow the Amazon Go route.

For B2B merchants, the core is showing them that getting started is easy; if activation is very complicated, it becomes an extremely complicated thing, and they just can't do it. [00:50:45]

PivotPivoting to shopping carts: a high-frequency, easy-to-enter form factor

The first product's usage frequency was too low: Black Friday comes once a year, everyday queues are under five minutes, and buying clothes is itself low-frequency. The move to grocery was because in the US a supermarket run every week or two is a high-frequency event with a clear pain point (even self-checkout machines have long lines at rush hour). There was another pivot along the way: investors from the first company were carried over to the new business, so they raised only two rounds but got heavily diluted and grew slowly.

▸2016: smart security tags get into YC — the first product

▸Lost to B2B activation costs → pivoted to smart shopping carts

▸Three generations of carts; forcing gen 1 to mass production dragged things down for two to three years

▸2021: unplanned acquisition by Instacart (while raising a Series B)

▸York spent nearly nine years at Caper, then started Dyna

Starting OverWhy keep hustling after financial freedom

One, his level of financial freedom is not fully free — there is still room to climb; two, he and Lindon are both born fidgets — in the later Instacart days, big-company resources and conflicts of interest made many things move very slowly, which York found miserable; he spent a lost year and read thirty or forty books looking for new perspectives.

There is always a return for your effort, because you lived through everything, success or failure; but if you insist on treating success as the reward, the world is brutal. [00:59:05]

Key Takeaway

York and Lindon are the most complementary pair: Lindon reasons backward from the furthest endgame, York works forward from the nearest execution; their conversations never take more than 15 minutes. In hiring they believe in the bar raiser — you don't need to be well-rounded, but you must have one exceptionally high bar that lifts the company's overall bar; misalignment on values or long-term vision is an instant veto. A revolutionary friendship forged through hard times makes the two take money, interests, and power very lightly.

PartnershipComplementary long view and near view

York, an engineer by training, is naturally wary of grand long-term visions; Lindon reasons backward from the long term — but people with long-term vision alone are a dime a dozen; Lindon's rarity is that he also understands the execution process. The hardest founder configuration is two people who are exactly alike (both long-term or both short-term), irreconcilable when ideas collide. Inside Dyna, when nobody else understands Lindon's strategy, York grasps it quickly and translates it into something tangible that engineers need, then passes it down.

He keeps pushing me to think further out, and I push him to think closer in… the two of us basically never need more than 15 minutes of talking. [01:01:09]

HiringBar raiser: one exceptionally high line

What a team fears most is everyone having exactly the same capabilities — everyone thinks the same things, and tons of duplicated, useless work gets done. Dyna's hiring core: you may be deficient in some ways or even mediocre across the board, but you must have one bar that is exceptionally high and can push the company's capability upward; the precondition is alignment on culture and long-term vision. Hiring few people with many demands — Shanghai headhunters already have plenty of negative comments — this is deliberately cautious expansion.

As long as he has one bar that's especially high… one that can push your bar upward, we really want to talk to this person. [01:05:32]

FamilyFamily and startup growing together

The founders' wives were all girlfriends from the startup days and now help the company with HR and finance — they jokingly call themselves a family business. His father's death from cancer, surgical complications, and other uncontrollable things moved York from a sense of omnipotence toward accepting that things are impermanent. His child is three now; he just plays with the kid and does no tutoring, with a Zen life philosophy: family staying intact and everyone happily living each day matter most.

Whenever I have time, my second choice is always to stand with my family… I just hope nothing goes wrong at home, and everyone can live each day happily. [01:14:02]

Key Takeaway

Rock taught him authenticity — My Chemical Romance's three albums, from screaming dissatisfaction with the world, to a look back on the edge of death, to a still-critical but sunny attitude, mirror York's own trajectory; the Buddhist book Blooming in Barren Ground (Chi Di Hua Kai) taught him to let go of ego and accept impermanence. An Instacart VP who was a natural at rallying people made him fully admit: some abilities are baked into personality and cannot change — rather than trying to do everything you want to do perfectly, push to the extreme the things you can do well.

Rock MusicMCR's three albums = York's three stages

He started listening to rock from late middle school through his first year of high school; rock's critical framework made early him lash out at whatever annoyed him — which is also why his online comments were sharp. In recent years he has mellowed a lot: what once seemed unreasonable, from that seat now seems possible to happen; reasonable does not mean endorsing.

His growth is actually very close to my own — from when I was unhappy with everything and wanted to slam them into the ground, to today when there's a lot I can let go. [01:22:55]

Turning PointThe disappearance of the sense of omnipotence

Managing fifty-some people at Instacart, York was in a strained state: he wanted to do the full CTO job well, yet saw every day that he couldn't. Observing that VP naturally enjoying communication, persuading for resources, and gathering people — some things no amount of effort can change. Since then, his view of roles at Dyna has been more flexible: fill in wherever needed, but while filling in, find someone better suited to take over.

I feel I'm no longer omnipotent; that sense of omnipotence has now disappeared — as long as I do well the things I can do well, I'm already in good shape. [01:18:32]

AdviceThe only advice for young people

He is no longer keen on giving young people advice: ten thousand pieces of advice are hard to understand without lived experience. The only thing he can offer is to avoid the omnipotent-me state — being in your early twenties thinking you are the chosen one who can do anything. High expectations for yourself are fine, but carrying high expectations too early brings enormous pressure too. People in AI and robotics will eventually find that humans become useless things, and then what matters most is figuring out what you actually want out of this life.

Ten thousand pieces of advice — when you haven't truly lived through them, they are very hard to understand… even before you hit the wall, it's hard to truly know these lessons will happen to you. [01:25:23]