---
id: "lib-2026-10-15-dyna-s-founder-on-product-research-task-vs"
type: "library"
title: "Dyna's founder on product research, task vs. workflow"
title_original: "Lindon Gao / Dyna: Finding Clarity in the Fog of Robotics (张小珺商业访谈录)"
lang: "ZH"
org: "张小珺商业访谈录 Zhang Xiaojun Podcast"
source_type: "Podcast"
date: "2026-10-15"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-15-dyna-s-founder-on-product-research-task-vs/"
original_url: "https://www.youtube.com/watch?v=iDaxRNIddME"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Long-form interview with Dyna Robotics founder Lindon Gao that gives rare deployment numbers and explains why a working product was not scaled; all figures are the founder's own and unverified."
summary_basis: "transcript"
original_minutes: "171"
tags: ["ai technology", "robotics deployment", "workflow vs task"]
---

# Dyna's founder on product research, task vs. workflow

Original title: Lindon Gao / Dyna: Finding Clarity in the Fog of Robotics (张小珺商业访谈录)

Tier B (Useful, limited): Long-form interview with Dyna Robotics founder Lindon Gao that gives rare deployment numbers and explains why a working product was not scaled; all figures are the founder's own and unverified.

Summary based on the full transcript.

## In short

A Mandarin episode (15 October 2026) of Zhang Xiaojun's business interview podcast with Lindon Gao, founder of Dyna Robotics and earlier of Caper, the AI shopping-cart company sold to Instacart. Most of the value is in the deployment story of DYNA-1, a napkin-folding robot that worked at a 99%+ success rate for 24 hours, had more than ten customers, and was still not scaled. Gao's reason is the difference between automating a task and automating a workflow: the robot needed a person to keep feeding it.

## Key numbers

- 1,300 vs 850: napkins a day the restaurant customer needed versus what DYNA-1 folded in 24 hours
- 99%: success-rate bar for commercial use, at 50-60% of human speed
- 10+: customers Dyna had when it chose not to scale the napkin-folding product
- 2 min vs 20 s: recovery time after grabbing two napkins at once, versus time per fold

## What it reports

- Gao's rule from Caper: ROI defines product-market fit. At Caper, 150 supermarket owners did not want Amazon Go-style checkout even for free; the value of the cart turned out to be bigger baskets and in-cart advertising, not saved cashier labour.
- The 2025 commercial bar was a 99% success rate at 50-60% of human speed. DYNA-1 folded napkins for 24 hours at 99%+.
- A large restaurant customer needed 1,300 napkins a day; DYNA-1 managed 850 in 24 hours, so the team worked on speed. A double grab cost about 2 minutes to recover against about 20 seconds per fold.
- Every 30-60 minutes the robot emptied its stack and waited until staff refilled it. Customers called it "too slow"; the data showed idle gaps. In his words, the robot needed a babysitter.
- The next product covers the whole workflow (find stock, carry, fold, deliver, fetch a new bin) on a mobile platform, first in hotel laundries. Per-step success rates multiply, so workflow reliability becomes a pretraining problem.
- Data: he calls his 2025 bet on high-quality data a mistake at this stage; diversity and scale come first. Dyna reports strong scaling on about a million hours of egocentric human video (unverified).
- He expects scalable product-market fit in specific workflows within one to two years.

## How they got it

A 2 hour 51 minute interview. The detailed summary covers the first 1 hour 50 minutes; later segments on the China supply-chain team, competitors and personal practice are covered at a high level only.

## Limits

- All company figures are the founder's claims; no customer is named and nothing is independently checked.
- The interview format favours the founder's narrative, including his account of an internal strategy dispute.
- Robotics-specific; the economics of hardware differ from software agents.

## Why it matters for companies

The task-versus-workflow point applies to software agents as much as robots: an agent that completes a step but needs a person to feed it inputs and move its outputs does not remove the role around it. Gao's evidence came from idle-time data, not from model accuracy, which is where many pilots stop measuring. Buyers can ask the same question of any automation vendor: who babysits it, and how often.

## Key passages in translation

> 我觉得ROI不成立就不是一个产品 就没有product-market fit

I think that if the ROI does not hold, it is not a product. There is no product-market fit.

> 一天要叠1300张餐巾 我们就会发现说 DYNA-1其实那时候24小时只能折850张

They needed 1,300 napkins folded a day, and we found that DYNA-1 at the time could only fold 850 in 24 hours.

> 其实就不是机器人太慢 而是机器人本身需要个保姆 ... 所以说当时我选择不scale

In fact the robot was not too slow; the robot itself needed a babysitter ... so at that point I chose not to scale.

Original: https://www.youtube.com/watch?v=iDaxRNIddME