Robots don't stall because they're slow—they stall because no one refills them
Dyna's napkin-folding robot achieved 99%-plus success rates with paying customers, but the company halted scaling—completing a task isn't managing a workflow, and robots need constant human resupply like a babysitter.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Founder called herself a cockroach to survive a death spiral
Kaber, Linden's previous venture, built autonomous shopping carts. At Y Combinator graduation, she ranked as the cohort's weakest fundraiser with only $300,000 raised. She and three cofounders then moved into shared housing on zero salary, spending just $7,000 monthly for two and a half years. Pitching over 1,000 angel investors yielded only two commitments, each for roughly $50,000. She sustained the team with a survival mantra: ‘我们是蟑螂, 我们死不了’ (we are cockroaches, we cannot die). This lean period eventually led to a $2 million seed round.
— LindenTeam split over research vs. commercialization—workflow became the bridge
Last year-end, Dyna's team fractured over a strategic identity crisis: some members pushed for commercialization and revenue, while others advocated continuing pure research. Rather than choosing sides, Linden extended the timeline to find a path that could do both. The solution was reframing the product around ‘workflow’—unifying all steps from resupply through transport, folding, and storage. This goal simultaneously forced capability improvements and created a genuine business closure.
— LindenNinety-nine percent success isn't enough: robots need a babysitter that doesn't exist
Dyna One, the napkin-folding robot, achieved over 99% success rates and secured more than 10 paying customers. Yet Linden chose not to scale. The core issue: after folding 3-to-4 buckets of napkins—roughly 30 to 60 minutes—the robot stops because it has no fresh material to process. Staff members must set phone alarms to manually refill it. This dependency Linden defines as the fundamental, unsolved product problem—the robot needs a ‘机器人保姆’ (robot babysitter). Scaling would only multiply this defect across all deployed units.
— LindenProduct roadmap upgraded from task completion to entire workflow management
Dyna redefined its product goal from completing a single task to managing an entire workflow. A complete cycle requires the robot to: locate material stock, retrieve supplies, transport them to the work station, complete folding, place finished goods on the designated shelf, retrieve the empty container, and begin again. Each step carries its own success rate; the combined chain is exponentially harder. Linden set a hard metric: one robot's 24-hour output must match one person's 8-hour output. The team has already achieved and surpassed this target.
— LindenEgocentric data revealed Scaling Laws the industry thought didn't exist
The industry long believed that scarce high-quality data was the ceiling for robot intelligence, so Dyna initially pursued that path. But while developing the workflow, the team shifted focus to egocentric—first-person—data from human operators. Using roughly 1 million hours of pretraining footage, they discovered remarkably strong Scaling Laws. This suggested robot data ceilings are far higher than previously assumed. Linden challenges the field's consensus that high-quality data should come first, arguing it's not a first-principles position.
— LindenModel leaps will obsolete some optimizations—but not if you stay maximally general
When asked whether a leap in model capabilities might render Dyna's scene-specific fine-tuning obsolete—as happened with many Agent applications this year—Linden acknowledged yes, it could. But she views this as compatible with her approach. She points to Anthropic's experience: after the Coding Agent's breakthrough, the same training framework scaled directly to Excel and other domains without rework. Dyna maintains maximum generality in pretraining, with workflow-specific tuning only in the fine-tune stage. This means scene-specific experience survives model leaps and can be reapplied.
— LindenThe brain sets whether it's possible; the body sets how good it gets
From day one, Linden insisted on building a 40-to-50-person hardware and supply-chain team in China—reasoning that embodied AI cannot produce world-class robots without supply chains. Her calculus inverts common thinking: the brain (AI) sets the lower bound—what's theoretically possible; the body (hardware and supply chains) sets the upper bound—how good the robot actually becomes. If the field moves toward a cost-and-quality competition within a decade, supply chains become more critical than AI capabilities. Dyna's China team is already operating at 100-unit production scale.
— LindenDropping the perfectionist burden revealed the path she'd been too anxious to see
During last year's strategic confusion, Linden discovered her deepest constraint: the burden of ‘不能犯错’ (cannot make mistakes). Because she had already sold one company, she felt pressure to prove the success wasn't luck, and that burden pushed her toward non-first-principles decisions. Once she set it down and clarified direction, her first strong emotion was gratitude—for the cofounders, team, and family who had carried her through. She frames entrepreneurship as a marathon with no visible finish line, where 修心 (cultivating the heart-mind) is what sustains you.
— LindenIn their own words · checked verbatim
There's no contradiction. I'm a very idealistic person. I'm not doing robots to make money. I think making money is one path to survival. My dream is the stars and the vast ocean, but to get there, my path must be grounded in reality.
没有分歧 我是非常理想主义的人 我来做机器人不是为了赚钱的 我觉得闻得到钱是生存的一条路 我的梦想是星辰大海 可是我的路径是需要脚踏实地
Linden0:00
I told my team then: we are cockroaches. You are cockroaches. We cannot die.
当时我就跟我团队说我们是蟑螂 你们是蟑螂 我们死不了
Linden29:28
But folding is different. If I fold it wrong, I can fold it again. So it's about loss rate. Loss rate and quality requirements aren't inherently very high.
可是在折叠不一样 折叠我折坏了可以重折 所以说就是损失率 损失率和对质量的要求本质上不是特别的高
Linden48:45
It's actually not that robots are too slow. It's that robots themselves need a babysitter. So the fact that robots need a babysitter is a very big product problem.
那其实就不是机器人太慢 而是机器人本身需要一个保姆 所以机器人需要保姆这件事情 是一个非常大的一个产品问题
Linden59:59
I think demand for general-purpose robots in B2B will be higher than B2C. Why? Because human desire is unlimited.
我觉得B2B的通用机器人的demand会比B2C更高 为什么 因为人的欲望是无限的
Linden1:16:16
We discovered that using egocentric human data, we could find remarkably strong Scaling Laws at 1 million hours of pretraining. I think this is an extremely important inflection point for the industry.
我们发现了其实在用人体第一视角数据的上面 是可以在1 million hours的一个预训练集下面 我们发现了非常强的Scaling Load本身 我觉得这是对行业来说应该是一个非常重要的一个节点
Linden1:28:38
I think the brain is the robot's lower bound. The body is the robot's upper bound.
我觉得大脑是机器人的下限 我觉得本体是机器人上限
Linden2:08:23
I was terrified of making mistakes then. Because I'd already been a successful founder once, and my burden was that I couldn't afford to make mistakes.
我当时非常怕犯错 因为其实我也算是成功过一次的创始人了 所以其实我当时对自己的包袱其实就是我不能犯错
Linden2:37:51
Figures
| Kaber sale price | $350 million | 19:18 |
| Kaber seed-stage burn rate | $7,000 monthly for 2.5 years | 14:14 |
| First seed round funding | $2 million | 14:14 |
| Dyna One daily output vs. customer demand | 850 per 24 hours vs. 1,300 needed | 54:51 |
| US unemployment rate cited | 4.5% | 1:16:16 |
| US hotel count | 100,000 | 1:19:20 |
Glossary
- product research
- Guiding research choices by whether they form a complete product cycle, rather than pursuing research without limits.
- egocentric data
- First-person video of human operations, used to train robots to replicate human movements and actions.
- Scaling Law
- The predictable relationship where model capability scales proportionally with increases in data or compute.
How to listen
Founders, investors, and embodied AI engineers concerned with diverging paths in the robotics industry and product scaling decisions.
Minutes 3:00–32:00: founder's early business ventures and investment banking experience—loosely connected to robotics decisions, can skip.