Why Physical AI Is the Next Platform Shift
In a Nutshell
Eric Landau's Encord is building the data infrastructure for physical AI, which targets the 80% of economic activity involving physical manipulation and movement in the real world. The company evolved from vision-focused annotation to handling multimodal robotics data across petabyte-scale datasets, serving autonomous vehicles, manufacturing, and logistics applications. Key lessons include persisting through multiple failed sales hires before finding product-market fit organically, maintaining dual London-Bay Area operations to access talent and customers, and treating startup volatility as a manageable roller coaster rather than a source of stress.
These notes were generated by AI and may contain inaccuracies.
Eric Landau is the co-founder of Encord, building the data layer for physical AI. His background includes big data particle physics followed by 10 years as a quant before creating Encord. The common thread connecting these experiences is the trajectory of AI development over his career. When he started in physics doing particle physics, systems were based on the physics itself, filtering data based on particle momentum and movement. In high frequency trading, the approach involved thinking about market factors and crafting strategies before applying machine learning models. The current paradigm is simply throwing data into a large machine learning system.
Eric quit his quant job during COVID to start a company. The month he left, the market went crazy with oil trading negative, extreme volatility, and Wall Street bets activity. His desk made more money that month than in the previous 3 years combined. He quit not for money but because he and his co-founder believed AI represented the technological paradigm shift of their generation - the early days of the internet and computing. After starting the company, his existential questions disappeared and were replaced with concrete problems about providing value to customers.
Encord was created in 2021. The company spent a couple years in the desert before ChatGPT changed the conversation tenor. Product-market fit for Encord came gradually rather than through a single binary threshold moment. The product continuously improved, and after ChatGPT, the market started coming to them faster. Product-market fit was recognized retrospectively when a sale closed in the Gong channel with a company they didn't know, indicating they had achieved organic traction without being intensely involved in every sales process.
Eric had to build a sales organization despite his quant and physics background. The process was trial and error - they brought in a sales team that didn't work, another sales team that didn't work, and by the third sales team it started working. They learned from YC advice but still made mistakes, including hiring someone they were advised against. They kept that person for 9 months before learning it wasn't a fit. The key learning was following gut instinct faster and making decisions more efficiently over time.
Encord started with vision as the hardest modality, with the thesis that AI systems would naturally be multimodal like humans. They developed expertise in vision before moving into multimodality. Physical AI emerged as the biggest application of multimodal AI, leading them to work in robotics, autonomous vehicles, logistics, and manufacturing. They still maintain healthcare customers and have seen diverse applications including facial recognition for cows, where algorithms can identify individual cows despite humans being poor at this task.
Physical AI represents enormous opportunity since 80% of economic activity involves manipulating or moving things in the real world. Robotics experts predict there will be more robots than people in 5-10 years. The company operates a facility in the Bay Area with robot systems and operators collecting data for various tasks. Data collection includes video streams, sensor data, audio, and language - all multimodal. The facility functions like film sets where environments are built for robots to perform tasks and collect data. Data collection serves early-stage companies needing pre-training data, while many companies already have production data loops.
The product encompasses data management, curation (selecting the right data from massive datasets described as finding a million needles in a billion haystacks), annotation and enrichment, and evaluations. The company has hundreds of customers and deals with multiple petabytes of multimodal data, more than what was used to train GPT-4. They don't know the data amounts for GPT-5.
Eric remains based in London while his co-founder is in San Francisco. The philosophy is to be where the important people are: customers first, then talent, then investors. Most customers are in the US, with the vast majority in the Bay Area, driving the decision to build operations there. London offers comparative advantage in AI and engineering talent with less intense competition than San Francisco. When starting, the main pushback was excessive competition, but many competitors pivoted, consolidated, or ran out of money. Competition has returned, but the company prefers good competitors as they push improvement. Their differentiation focuses on physical AI and scalability at the multiple petabyte level.
For founders building in physical AI, the recommendation is to use products like Encord when reaching scale inflection points - after building internal tools, trying POCs, and moving from POC to production or experiencing big data influxes. Early stages can start with open-source tools. The broader advice for all founders comes from a YC retreat where Reddit's CEO discussed that starting a company is a roller coaster with low lows and high highs. The view is to ride the roller coaster, try to have fun, and accept the fluctuations. When at a low, know a high is coming; when at a high, know a low is coming. This mindset shift toward enjoying the journey and running toward fires rather than away from them makes the experience more enjoyable. Eric notes that only relatively recently did he reach the point of thinking about making fluctuations fun, realizing more control over mental state than initially thought.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
