By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
a16zWhat iPhone Packaging Reveals About the Robotics Opportunity
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 1:07
Global supply chains and the “what is the robot for?” framing
The conversation opens by grounding robotics in today’s interdependent global supply chain and the push to build industrial capacity. The hosts set a key critique: many robotics pitches start with models and data rather than clearly defining the end job the robot must accomplish.
- •Josh Gruenstein says that in 2026 the US and China remain tightly interconnected via a global supply chain, and both are thinking hard about long-term industrial strategy.
- •Erik Torenberg says they see many robot pitches, but few focus on what the robot is actually supposed to do in the real world.
- •The episode frames robotics as tied to real operational needs in manufacturing and logistics rather than purely technical demos.
- 1:07 – 2:50
Tutor Intelligence overview: deploying robots to create the data flywheel
Josh Gruenstein introduces Tutor Intelligence’s origin thesis: deploy robots first to generate the real-world data needed for robot learning, while simultaneously building manufacturing and customer adoption. He describes the company’s scale, footprint, and emphasis on deployment as the foundation for a durable robotics platform.
- •Josh Gruenstein says he started Tutor about five and a half years ago with the thesis that deployment enables datasets, models, and supply chain needed to proliferate robots.
- •He describes Tutor as about 160 people based primarily in Boston, training frontier robot AI models, manufacturing robots at scale, and deploying into large supply-chain businesses.
- •Martin Casado highlights that Tutor is among the few robotics companies deploying at scale, not just pitching.
- •Josh says Tutor robots have helped manufacture and ship over a quarter of a billion products to American consumers.
- 2:50 – 6:46
Start with the job, then build full-stack: robotics as a systems platform
Josh explains why starting from the job-to-be-done forces a full-stack approach across hardware, software, manufacturing, supply chain, and go-to-market. He argues robotics progress comes from making trade-offs across layers rather than optimizing one layer in isolation.
- •Erik Torenberg contrasts typical robotics pitches (model-first) with Tutor’s job-first approach and asks where that insight came from.
- •Josh describes a thought experiment: even with “infinite” human intelligence connected to robots, the hard problems remain manufacturing robots at scale and building real customers and software.
- •Josh says robotics success requires combining intelligence, hardware, and customers—plus broader business layers like supply chain, product, and go-to-market.
- •He argues ignoring a layer eliminates most of the trade-off space needed to make forward progress.
- 6:46 – 9:08
From verticalization to platform: why humanoids can be the “clean abstraction”
The group discusses why robotics companies often end up going vertical, and whether robotics will ever horizontalize like computing. Josh argues verticalization is a market-maturity issue, and that humanoid-like form factors can become a general-purpose abstraction analogous to CPUs—followed later by re-specialization.
- •Martin Casado observes that many robot companies start horizontal and then go vertical, asking whether that is temporary or inherent.
- •Josh says it reflects market maturity and compares it to early computing and specialization-to-generalization cycles.
- •He argues humanoids provide a general abstraction similar to CPUs: not perfectly optimized for every task, but broadly usable.
- •Josh extends the analogy to compute diversity (CPUs, GPUs, ASICs) to explain why different robot form factors will still exist.
- 9:08 – 12:24
Why it’s called Tutor: teleoperators, “internet of robot training data,” and tutor economics
Josh explains the company name and the core mechanism: connecting robots in real customer sites to human teleoperators (“tutors”) to teach tasks and generate training data. The conversation then turns to whether tutors go away, framed as a pragmatic unit-economics and trade-off question rather than an ideological goal.
- •Josh says digital work is naturally networked into data (text, emails, social media), while physical work is not recorded in the same way.
- •He describes connecting robots with human teleoperators—called tutors—to teach new tasks and create an “internet of robot training data.”
- •Erik says customers often don’t notice or care when tutors vs models are operating, which he credits to seamless execution.
- •Josh frames tutor usage as an operational complexity and unit economics question, comparing it to the cost and utilization of compute for model serving.
- 12:24 – 16:57
How frontier models and simulation accelerate robotics—while real-world data remains central
The discussion explores whether robotics can match the “distribution in, distribution out” economic magic of generative AI. Josh argues economically valuable data comes from robots doing paid work in real sites, and describes using simulation and other data sources to widen distributions; he also explains how improved spatial reasoning in frontier models is a major shift.
- •Martin Casado questions whether robot tasks will always be slightly out-of-distribution relative to training data, making tutor removal difficult.
- •Josh argues the key is data source alignment: paid deployments guarantee the data is economically valuable, unlike speculative data collection.
- •Josh says the company is not purist and uses multiple tools, including simulation and internet-scale data, to broaden training distributions.
- •Josh says frontier models have undergone a “sea change” as labs post-train on 3D spatial data, enabling robots to leverage general intelligence that previously couldn’t express the needed spatial reasoning.
- •He says architectural tricks matter mainly insofar as they accelerate the deployment-data-manufacturing flywheel.
- 16:57 – 18:36
Building robots in-house: why legacy industrial robots don’t fit, and what Tutor manufactures
Josh details what it means for Tutor to “build robots,” arguing existing industrial robot supply chains are too small and optimized for different trade-offs. He explains Tutor sources some components from OEMs but designs and manufactures the complete robot systems at scale in Watertown, Massachusetts.
- •Martin Casado asks where Tutor sits on the spectrum from buying off-the-shelf robots to machining parts end-to-end.
- •Josh cites low deployment volume in the US and argues current industrial robots are designed for systems integrators and repetitive motion, not co-designed with frontier models.
- •He says Tutor has been forced to design and manufacture its own hardware to deploy at scale.
- •Josh says Tutor is not building actuators or cameras, but is building the full robot system (mobile base with arm) at scale in Watertown, Massachusetts.
- 18:36 – 20:42
Two robot lines: Cassie for heavy-duty work and Sonny for dexterous bimanual tasks
Josh introduces Tutor’s two robots and the design trade-offs behind each. Cassie targets heavy-duty pallet/case handling with suction and single-arm design, while Sonny is a wheeled semi-humanoid designed for long-term factory operation, higher payloads, and safer collaboration around humans.
- •Josh describes Cassie as a heavy-duty robot with a single arm and suction cups, powered by foundation models in the 200–400 million parameter range.
- •He says Cassie is co-designed around manufacturability, model capability (e.g., predicting grasps/affordances), and market demand driven by pallet and box handling.
- •Josh describes Sonny as an industrial semi-humanoid designed to run for years in factories with reliability; it is wheeled (not legged) to match indoor floors.
- •He says Sonny has higher payload capacity, moves faster, and meets collaborative safety standards to operate around humans.
- •Josh says Tutor operates 100 Sonnys in Boston in a “data factory,” and has a growing real-world fleet doing more dexterous bimanual manipulation.
- 20:42 – 23:17
What the robots actually do today—and what they can’t yet do (outside and home)
Josh challenges the popular mental image of fully automated factories by describing how much handwork still exists in large facilities. He explains the physical tasks Cassie and Sonny target, then outlines the next frontiers: outdoor environments and, eventually, homes—both constrained by hardware and economics.
- •Josh says many people’s view of US manufacturing comes from highly automated examples like “How It’s Made,” which represent a tiny fraction of reality.
- •He describes large facilities where people still do manual work at conveyors and tables, and frames Cassie as a fit for backbreaking tasks with difficult hiring and turnover challenges.
- •Josh says many consumer goods still involve extensive human boxing/packaging, even at very high scale, and positions Sonny for more dexterous manipulation work.
- •He says Cassie and Sonny can’t operate outside due to wheeled hardware limitations (e.g., gravel), pointing to opportunities in construction and landfill material reclaiming that may require legged robots.
- •He says homes are possible long-term but current robots are too industrial, and economics and safety make home deployment harder today.
- 23:17 – 25:51
Selling to first-time robot buyers: pricing, speed of deployment, and greenfield adoption
Josh explains that most customers have never bought a robot, so Tutor sells a new proposition: robots priced like labor, with minimal customer risk and rapid rollout. He shares timelines from first call to contract and from contract to on-site production, emphasizing standardized robots rather than bespoke integration.
- •Josh says most Tutor customers have never purchased a robot before because most US factories and warehouses have no robots, making the market largely greenfield.
- •He says Tutor doesn’t lead with AI or models, but with a simple proposition such as a robot for an hourly rate.
- •Josh says Tutor’s approach enables fast sales and deployment cycles: 40–60 days from first call to contract, about 40 days from contract to robots on-site, and production stand-up often within hours.
- •He attributes speed to shipping a generally capable machine without customization and structuring the business model to avoid putting economic risk on the customer.
- 25:51 – 29:04
Robots as equipment—and as coworkers: Human-Robot Interaction, naming, and user adoption
The conversation turns to how people perceive robots after deployment, distinguishing between the buyer and the day-to-day user. Josh describes practical HRI tactics—like naming robots—and how worker relationships can matter even when robots are viewed as machines by management.
- •Josh says operations teams respond simply to working systems: if it works, they replicate it across sites.
- •Martin Casado asks whether robots are treated like equipment or coworkers; Josh says factory managers largely see them as machines once they work.
- •Josh notes robotics differs because the buyer (e.g., COO/GM) is not the user (factory workers), so user relationship-building matters.
- •He references the academic field of Human-Robot Interaction and says Tutor uses tactics like giving every robot a name.
- •Josh says technicians often name robots initially, and customers can name robots on repeat orders; he describes customers naming robots after kids, athletes, or pop stars, indicating attachment.
- 29:04 – 31:22
Why small and mid-sized businesses matter: product iteration vs bespoke automation traps
Erik asks why Tutor works with many local and mid-market businesses rather than only large enterprises. Josh argues this reflects where the market is, and that serving diverse customers forces a true product approach—iterative, general-purpose, and repeatable—rather than one-off custom systems that often fail.
- •Erik notes Tutor’s customer base includes many smaller/local businesses compared to robotics companies that focus on very large enterprises.
- •Josh says most of the manufacturing/logistics market is small and mid-market, so serving the real market requires reaching them.
- •He critiques a common robotics failure mode: spending years building a custom system for one big customer, then failing to iterate or losing relevance when needs change.
- •Josh says diverse customers force the “right shape” of product: general robot bodies and generally capable models that can get to work immediately.
- •He suggests fast contracting and quick deployments are signals of true product success rather than bespoke engineering.
- 31:22 – 42:04
Cost, business model, and the future physical economy: factories, geopolitics, and supply chain verticalization
The episode expands to economic convergence, factory location, and geopolitical implications. Josh argues robots could lower costs and unlock new market discontinuities, potentially increasing the number of factories; he also discusses US–China supply chain interdependence and why Tutor is investing in building more robotics supply chain capacity in the US.
- •Josh says if robots are successful, goods can get cheaper and demand can rise; he gives examples of discontinuities where small cost changes could reshape markets.
- •Erik asks how general-purpose robots might change where factories are located and how communities organize around them; Josh says more people could start and scale physical-world businesses with flexible labor.
- •Josh says robots could be a “great equalizer” between the US and China, arguing the US is building automation greenfield while China may need to tear down prior automation investments.
- •Martin Casado asks about dependency on China in robotics supply chains; Josh says in 2026 supply chains remain highly interconnected and both sides are strategizing and investing across borders.
- •Josh says as Tutor grows, OEM relationships can shift, and Tutor expects to continue verticalizing while also needing a mature US robotics supply chain; he points to building in Watertown, Massachusetts as part of that effort.
- •Josh also explains that ongoing costs (inference, teleoperation, maintenance) are a larger portion of cost than many expect, compared to traditional CapEx-plus-ROI framing.
- 42:04 – 44:56
Real-world reinforcement learning at scale: the 100-robot “Data Factory One” and hiring across the stack
In the closing segment, Josh describes Tutor’s research approach: prioritize real-world robot experience traces, enabled by a large robot fleet used as a reinforcement learning testbed. He closes by describing the interdisciplinary team structure and broad hiring needs across technical and operational roles.
- •Erik asks how Tutor approaches robot learning research amid competing theories about which models will win.
- •Josh says that, as with LLMs, diverse data matters but the “top” data is real-world reinforcement learning traces doing the target task.
- •He describes a 100-robot fleet in Boston called Data Factory One, used to explore real-world reinforcement learning at scale.
- •Josh says reinforcement learning methods don’t work well at small batch sizes, so a large fleet enables experimentation not feasible in academia or many companies.
- •Josh says Tutor is hiring across every layer of the stack (research, engineering, hardware, materials, electronics) as well as technicians, emphasizing that robotics forces engagement across disciplines.
- •He says the company has teams across the US, especially in go-to-market and field technician roles supporting nationwide customers.
