Lindon | 张小珺Jùn｜商业访谈录 Summary

1. Key Themes

"Product Research": Research Bounded by a Product Guardrail

Lindon positions Dyna as neither a pure "deployment company" nor a pure open-ended research lab. Research priorities are reverse-engineered from the end-state product, and commercial loops are used as an instrument to discover where the model's intelligence ceiling is. This is the core framing of the company.

He defines it: "Product research differs from traditional research... traditional research teams say 'I want to find AGI' and explore divergently... With product research, within the boundaries of research I have guardrails." 00:44:01 He adds: "I'm using product loops to probe which direction our research should go next." 01:07:38 And on the cost of divergence: "Our team's energy is limited, so we only bet on the most important things." 00:44:22 He is explicit that the "deployment company" label is wrong: "The most inaccurate thing is that we're a deployment company. From day one we never built this company around a deployment-first view." 02:35:15

From Task to Workflow: The Napkin Robot That Needed a "Nanny"

The pivotal operating lesson: Dyna-1 folded napkins at 99%+ success but still failed commercially, because it solved a task rather than the workflow. The missing loading/unloading meant a human had to babysit it.

Lindon: "Every half hour to an hour we'd finish folding a stack of 100 napkins, and then the robot would stop because there were no napkins left... so the robot needed a nanny. That a robot needs a nanny is a huge product problem." 00:59:23 He concluded: "If you scale an incomplete product, that product becomes a liability for you in the future." 01:00:31 That is why he stopped scaling despite having 10+ customers: "I chose not to continue scaling... because the problem I solved wasn't fundamental enough. I only did the task, I didn't do the entire workflow." 00:58:35

ROI Is the Definition of Product-Market Fit in Hardware Robotics

Lindon repeatedly collapses PMF into ROI, a lesson carried from his retail-tech startup Caper, where 150 supermarket owners rejected Amazon-Go-style tech even for free.

On Caper: "Nobody wanted it. I asked, what if I give it to you free? He said I still don't want it... they said the cost-performance isn't high, I have no ROI." 00:10:15 On robotics: "In hardware robotics, if ROI doesn't hold, it's not a product, there's no product-market fit." 00:23:00 He also warns that companies finding "PMF" through customization often fall far short of the real value: "Many robotics companies found so-called product-market fit but found the customization... was far from the product form customers need." 00:23:00

Egocentric Human Data and the First Scaling Law

Dyna reversed its own earlier conclusion about data quality and now bets on egocentric (first-person human) video as the terminal data form, claiming a scaling law at 1 million hours.

Lindon: "We found a very strong scaling law at a 1 million hours pre-training set using human first-person perspective data." 01:28:33 On the reversal: "In the first Dyna-1 we used the highest quality data... [later] the most important thing is whether we can give the model the most diverse, broadest human experience. Once we reach internet-scale data, then ablation and distillation make sense." 01:29:42 On why egocentric: "I think we'd bet egocentric or exocentric becomes the final form of data... betting on teleop or UMI is in some sense betting that robot hands won't work." 01:31:34 He believes the industry belief is that "robot hands are an engineering problem, not a research problem." 01:31:56

Pre-training Drives Robustness; Post-training Datasets Shrink

Dyna runs its own pre-training and has shifted compute from 50/50 pre/post to heavily pre-training. The key finding: most robustness comes from pre-training, and post-training data requirements keep falling.

Lindon: "A model's vast majority of robustness comes from pre-training, not post-training." 00:57:19 On trend: "The most fundamental change is that the data needed in post-training is getting less and less." 00:52:39 He identifies what pre-training can and can't fix: grabbing two napkins at once and collapsing throughput "is a problem pre-training can't solve," whereas neatly placing napkins in a bin "is entirely something pre-training can solve through language." 01:42:59

Three-System Architecture

Dyna-2 is organized into System 2 (embodied reasoning), System 1 (a world-action model), and System 0 (whole-body control via RL).

Lindon: "System 2 is the embodied reasoning layer... System 1 is a world-action model... and System 0 is how we use reinforcement learning on motion control to get actions that are fast, smooth and accurate." 01:08:02 Generalization is measured on two axes: "how many demonstrations until you learn" and "language steering capability." 01:09:30

The Brain Sets the Floor, the Body Sets the Ceiling

Lindon inverts a common framing and argues for full-stack ownership with the brain driving hardware requirements.

Lindon: "The brain is the robot's floor, the body is the robot's ceiling." 02:08:32 "The brain determines whether this thing can be done. The hardware determines how well it can be done." 02:08:44 On why owning the body matters: "Two different motion-control algorithms can train two very different models... I can hardly imagine, for B2B deployment, that if you don't own the robot body, you can reach their quality and speed efficiency." 02:09:42 And the spec flows from research: "The hardware spec is always given by the research team. Our camera FPS, global vs rolling shutter, camera angle... all defined by research." 02:10:38 On why not wait: "Hardware needs at least three versions before real production... If you do hardware three years from now, you lose three years of iteration." 02:11:56

Deployment Is a Flywheel That Takes Years

Lindon argues the deployment stack (observability, support, maintenance) takes 2–3 years to mature, so starting late means you're structurally behind.

Lindon: "Your hardware has problems, software has problems, model has problems, motion control has problems... one day your throughput was supposed to be 1,300 napkins but you only did 800 and you don't know why. So you need to build an observability stack, a support and maintenance stack." 01:58:01 And: "When you've finished the intelligence and then go build the deployment loop, you'll find you're two or three years behind." 01:58:31 He ties this to his Caper experience, where scaling took years even after the product worked. 01:57:34

The Brain-Camp vs. Body-Camp Divide (US vs. China)

Lindon describes a structural geographic split in robotics strategy and says Dyna is "full-stack."

Lindon: "The US thinks the brain is the only problem and the body will solve itself. China thinks the supply chain is the only problem and the brain will solve itself." 00:02:33 On his ten-year view: "If you make me look ten years out, I think supply chain may be more important... the brain determines your product's floor, the body determines your ceiling... in the long term this becomes a cost and quality war." 02:04:17

Founder Inner Work as a Competitive Input

A substantial through-line is that mental clarity, discipline, and "cultivating the mind" (修心) function as operating infrastructure. Lindon reads Meditations, Wang Yangming, and Buddhist and Christian texts.

Lindon: "The lower the trough, the more you need to cultivate your mind." 00:17:47 On decision quality: "Most of a CEO's major decisions... finding people and money don't require much brainpower. The most important is finding direction. Each decision needs to be essential and high quality." 00:27:44 On the pivot out of the strategic crisis: "I dropped the baggage of being a CEO who doesn't make mistakes... making decisions to avoid mistakes is essentially not a good decision." 02:38:36

2. Contrarian Perspectives

Scaling the Wrong Product Is Worse Than Not Scaling

Most founders with 10+ paying customers would keep scaling. Lindon deliberately stopped. His logic: an incomplete product creates compounding debt. "If you scale an incomplete product, that product becomes a liability for you in the future." 01:00:31 He chose not to produce the previous-generation robot at all: "We chose not to mass produce the previous generation, because I think it fundamentally wouldn't work." 02:00:50 This runs against the typical "get revenue, iterate later" instinct.

Intelligence Gains Don't Equal a Business, and a Robot That Zero-Shots Everything May Still Not Be a Product

Many believe the physical-AGI brain is the whole bottleneck. Lindon says even perfect generalization on a task doesn't create a commercial loop. "I can have any task zero-shot into any new environment... it's indeed a rise in intelligence, but I think it's not a commercial loop." 01:03:02 And on consumers: "Even if a robot can do things in every home, it doesn't mean this product holds... consumer is a very subjective market." 01:13:30

B2B General-Purpose Robots Will Out-Demand B2C, Because Human Desire Is Infinite

Contrary to the home-humanoid-as-crown-jewel narrative, he argues B2B demand exceeds B2C. "I think B2B general-purpose robot demand will be higher than B2C... because human desire is infinite... if labor cost becomes negligible, consumer products get cheaper, and people want to buy more." 01:16:23 And: "If the home robot is three people to a household, America might only need 100 million robots. But B2B... America will definitely need more B2B robots versus B2C." 01:17:09 He also argues the humanoid form factor is purely an ROI question: "Do these two legs bring higher value that far exceeds the cost of the two legs?" 01:49:27

Robots Needn't Be Faster Than Humans, Just Run Longer

Rather than chase human speed, he re-defines the value metric. "The most important value metric is not doing it faster than a person, but doing it longer than a person." 01:11:45 The team target: "A robot working 24 hours matches a human working 8 hours" 01:11:57, and he says they've already exceeded it. 01:12:18

The Last Mile of Intelligence Isn't Required for Universal Adoption

He claims robots can reach "everyone" without achieving general intelligence: "I think robots don't need intelligence to reach generality to be usable by everyone." 01:45:13 He frames this as a product-definition issue, analogizing GPT-3 (API form) to ChatGPT (chat form): "GPT-3 to ChatGPT's essential change is a product change." 01:46:25 He teases a next release where the form factor is "not a robot form but a product for how to use robots." 01:48:00

3. Companies Identified

Dyna Robotics

Lindon's company. Full-stack robot foundation model plus hardware, with a US research team (~200) and a 40–50 person China hardware and supply chain team. Mentioned as the subject. Notable claims: "the industry's largest dataset" and a scaling law at 1 million egocentric hours. 01:27:33 "We're at initial mass production, hundred-unit level." 02:00:37 Stated cash position: "Quite a lot... fully" able to fund model R&D. 02:15:29 First commercial demo: "Dyna-1... we were the first company to show a commercializable demo... folding napkins, 24 hours, 99%+ success rate." 00:47:37

Caper (Lindon's prior startup)

Smart shopping carts with vision and sensor fusion for grocery retail, acquired in 2021 for $3.5 hundred million (3.5亿). Mentioned as the source of his product and deployment lessons. "One of the few survivors in this industry" of dozens of unmanned retail companies. 00:10:25 Pitched as: "we were the few survivors"; key insight: value was in raising basket size and ads, not cutting cashiers. 00:21:55 Funding path: only $300K post-YC, burned $7,000/month for 2.5 years, Seed (~$2M, First Round Capital) and Series A (>$10M, Lux Capital). 00:14:23

Y Combinator

Accelerator that admitted Caper on the first attempt with Sam Altman and Jeff Ralston as mentors. Caper came out near the bottom of the batch on fundraising. 00:12:45

First Round Capital

Seed investor in Caper, the single institutional yes out of 100+ pitches. "Rejected by all except one, which was First Round Capital." 00:16:12

Lux Capital

Led Caper's Series A (~$10M+) after 60+ investors passed. 00:17:14

Amazon Go

Cited as the trigger for the 2017 unmanned-retail wave, and as the consensus approach Caper contrarianly avoided. 00:10:00

Instacart

Acquired Caper; Lindon spent three years there learning management and org-building from execs who came from Uber, Facebook, and Google. 00:28:32

Physical Intelligence (Pi)

Cited as a research-oriented robotics lab. Lindon declined to characterize their strategy but said there's "no essential difference" in the brain research goals. 01:49:38

Generalist

Named alongside Pi as a research-oriented robotics company. 00:44:54

Google DeepMind

Mentioned as a comparison for robot-brain research. 01:49:38

Unitree (宇树)

Praised for hardware: "They've made the sim-to-real gap this small, which is essentially a very strong breakthrough." 02:05:11

AgiBot (智元)

Described as having its own intelligence and hardware, with ample capital enabling a broad product line and divergent brain research. 02:05:40

OpenAI, Anthropic, Google (as data-strategy analogies)

Used to illustrate that pre-training objectives reflect founder choices: "OpenAI was consumer application all the way... Anthropic very early decided 'we're going to bet on coding agent'... Google, multimodal." 01:33:00 He also cites Anthropic's coding agent as proof that deep pre/post-training enables expansion beyond one use case (e.g., Excel). 01:41:10

Jasper

Cited as an example of a company that turned the GPT-3 API into a successful product. 01:46:25

4. People Identified

Sam Altman

Former YC president; Caper's YC mentor. His advice: "launch, launch, launch and iterate." Lindon said he lacked the capability then to resonate at Sam's frequency: "I think my capability wasn't strong enough to be at the same frequency as him." 00:24:10

Jeff Ralston

YC mentor to Caper and later YC president/CEO. 00:12:45

Yang Shiyuan (杨世远, also referred to as York)

Lindon's cofounder of 10+ years; former rock musician. "Extremely strong empathy, structural thinking, execution, technology, management. A rare technical genius who also has management-level cognition." 00:38:20

Jason (Dyna research cofounder)

Chosen for research taste visible in his papers: "In PhD years one and two, many papers were fancy... as he progressed his papers got simpler and cleaner, with very clear core concepts. I think this is very good taste." 00:39:11 Also pragmatism about landing impact. Lindon names Jason's Eureka paper as the most influential robotics paper and how he met him: "Eureka... redefined the form of how robots self-learn in the future." 02:49:59

Huang Yiming (黄一鸣)

Another cofounder, through whom Lindon met York at the previous company. 00:38:49

Lindon's father

Source of Lindon's discipline habits: "I learned this from my dad, who is extremely self-disciplined." 00:27:02

Wang Yangming / Marcus Aurelius

Philosophical influences on Lindon's inner cultivation: "Western Meditations, China's Wang Yangming, even Buddhist and Christian systems... the core concept is cultivating the mind." 02:29:19

5. Operating Insights

Prioritize Research by Reverse-Engineering the End Product, Not by Diverging

Lindon sequences research with explicit stage-gating rather than parallel exploration. "We need to do this thing first at this stage, then another thing at the next stage, rather than diverging simultaneously." 00:44:22 Tactic: write the end-state product, then derive the research agenda backward.

Resolve Team Identity Splits by Lengthening the Time Horizon

When the team split between "ship now" and "research," Lindon found the overlap by pushing the target further out: "We lengthened the front line... the commercial bottleneck is stuck on intelligence... so I gave the team a farther goal." 00:41:49 The unifying goal ("one employee" doing the whole workflow) forced both pre-training gains and deployment infrastructure.

Lower the Skill Floor of Deployment to Free Researchers

Dyna built infrastructure so non-engineers can train models: "We've reached the point where any material staff [data-ops staff] can train a model... all deployment videos are made by the data-ops team, not the engineering and research team." 01:51:39 This keeps ~80–90% of capacity on intelligence.

Hire the Strongest People for the Most Important Work; Keep Orgs Flat

Using his Caper and Instacart lessons: "Hire the strongest people to do the most important thing." 02:21:21 Most ICs are former managers who chose IC roles. "Large companies have 10 to 11 levels... in the current AI state the org has become very flat," allowing CEO and execs to steer execution directly. 02:20:09 Fewer people means less management cost and easier "one heart."

Treat Each Rejection as Reinforcement Learning, but Fix the Listening Problem

Lindon's 100+ rejections taught him he was "only speaking, not listening," and had no investor mental model. 00:18:55 He now classifies investors by what they care about: "financials, technology, hardware, end-state, founders." 02:28:43

6. Overlooked Insights

Rejected Mass-Production of the Previous Generation Reveals Capital Discipline

In a throwaway line, Lindon says: "We're at hundred-unit level... workflow robot. The previous generation we chose not to mass-produce." 02:00:46 A company with real customers voluntarily orphaning a shipped product is rare and signals a willingness to take a revenue haircut to avoid technical debt. Investors should read this as unusually rigorous product discipline, likely to produce better unit economics later.

Language-Following Is the "Glue" for Workflows, and VLA Models Were Weak at It

Briefly noted: "We discovered the VLA's language-following ability wasn't very strong," and this was attributed entirely to pre-training. 01:05:51 Chaining long workflows depends on language commands from the reasoning layer to the action layer. Because workflow success compounds multiplicatively ("99 times 99 times 99... exponential" 01:05:08), language steerability is arguably the critical enabling capability for long-horizon robotics, and Dyna names it one of its two pre-training goals alongside demonstration efficiency. 01:35:00 It's an under-discussed bottleneck relative to hand dexterity, which Lindon dismisses as "not important for now" (gripper-to-dexterous-hand swap will happen "when dexterous hands are ready"). 02:12:39