Key Moments
Why Physical AI Is the Next Platform Shift
Y CombinatorWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Massive amounts of data are needed to train AI for the physical world, and Encord is building the platform to manage it all, but it's expensive.
Key Insights
The 'bitter lesson' in AI development suggests that scaling systems with more data and compute is more effective than manual feature engineering.
Encord's product market fit was a slow, daily compound rather than a single breakthrough moment, with the market starting to come to them more rapidly after ChatGPT's release.
Despite an initial presence in healthcare, Encord has pivoted its focus to 'Physical AI' due to market demand, now serving sectors like robotics, autonomous vehicles, and logistics.
Encord deals with multiple petabytes of multimodal data, handling more data than was used to train GPT-4.
The company's go-to-market strategy prioritizes locations based on customers, then talent, then investors, leading to a significant operational presence in the US Bay Area despite being founded in London.
Founders should embrace the emotional rollercoaster of startups, viewing challenges not as things to mitigate but as an inherent part of the journey to be ridden and enjoyed.
The 'bitter lesson' and the shift in AI development
Eric Landau, co-founder of Encord, draws a parallel between his career trajectory and the evolution of AI. He explains the 'bitter lesson' in AI, which posits that instead of painstakingly engineering features based on domain expertise, the more effective approach is to scale systems with more data and computational power. Landau's early career in particle physics and high-frequency trading involved significant feature engineering based on the specific physics or market factors. However, the current AI paradigm, as exemplified by large language models, involves feeding vast amounts of raw data into machine learning systems and letting them learn. This shift underscores the increasing importance of data quantity and quality in driving AI performance.
From lucrative quant career to founder's journey
Landau shares a surprising decision: quitting a highly profitable quant trading career during a period of extreme market volatility and profitability for his desk. He made this leap to pursue his belief in AI as the paradigm shift of the generation. Despite the immediate financial sacrifice, Landau found more personal reward in the concrete problems and value creation of building a startup, even during its lean 'desert' years, compared to the existential questions he faced in his previous high-paying role. This highlights a founder's motivation often stemming from a belief in a larger mission rather than solely financial gain.
Product market fit as a slow compound
Encord experienced product-market fit not as a sudden event, but as a gradual, daily compound. Landau contrasts this with companies that achieve fit through a single product launch or market shift. For Encord, it was the continuous improvement of their product coupled with a market that increasingly recognized the importance of AI, particularly after the advent of ChatGPT. The market began to 'come to them' faster. A defining, albeit retrospective, moment of achieving fit was when sales closed for companies they were no longer deeply involved with, indicating a level of market traction where their offering was valuable even without their granular participation in every deal.
The pivot to physical AI
Initially focusing on computer vision, Encord's thesis was that AI systems would naturally become multimodal, mirroring human sensory input. They chose vision as a starting point due to its data density and the expertise it would build. As the market evolved, 'Physical AI' emerged as the dominant application for multimodal AI. This led Encord to shift its focus to sectors like robotics, autonomous vehicles, and logistics, capitalizing on the significant economic activity (80%) involving manipulation or movement in the real world. The company anticipates a future where robots are ubiquitous, fundamentally changing various industries.
Scaling data for physical AI
Encord operates at a massive scale, dealing with multiple petabytes of multimodal data, exceeding the estimated data used to train GPT-4. They manage this through a comprehensive data pipeline that includes collection, curation, annotation, enrichment, and evaluation. The company maintains a facility in the Bay Area that emulates film sets for robot data collection in specific environments, particularly for companies in earlier stages requiring pre-training data. However, much of the data is also ingested directly from customer production loops, encompassing video, sensor, audio, and language.
Strategic location and talent acquisition
Despite being a London-founded company, Encord has established significant operations in the U.S. Bay Area, driven by a philosophy of being where critical stakeholders are: customers, talent, and investors, in that order. With most of their customers in the U.S., particularly the Bay Area, the co-founder's relocation and the facility's establishment were strategic moves to foster closer customer relationships. While acknowledging London's strong talent pool for AI and engineering, the company prioritizes proximity to its customer base for effective business development and collaboration.
Navigating competition and standing out
Encord faces competition from other data companies, but they view healthy competition as a driver for improvement. They differentiate themselves through a deep focus on 'Physical AI' and the scalability of their data operations, capable of handling petabyte-level datasets which many competitors struggle with. Their foundational investments in infrastructure have positioned them to manage this scale effectively. While they understand that their full-scale product might be out of reach for very early-stage startups, they typically engage with companies that have outgrown internal tools and require scalable solutions for moving from Proof-of-Concept to production.
Embracing the startup rollercoaster
Landau's core advice to founders is to embrace the inherent emotional rollercoaster of starting a company. Instead of trying to mitigate the high highs and low lows, founders should ride them and even try to have fun. This perspective, he notes, is developed over time through experiencing enough fluctuations. The realization that one has more control over their mental state than initially perceived allows for a more adaptive and enjoyable entrepreneurial journey. This mindset shift is crucial for sustained resilience and effectiveness in the face of inevitable challenges and triumphs.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
'The Bitter Lesson' suggests that AI systems benefit more from scaling (more data, compute) rather than intricate feature engineering or domain expertise. This idea is relevant as it reflects the trajectory of AI development and Eric Landau's career path.
Topics
Mentioned in this video
A company building the data layer for physical AI, co-founded by Eric Landau.
A startup accelerator whose advice is considered highly valuable by Eric Landau and his co-founder.
Mentioned as a competitor in the US market for data collection and AI services.
The CEO of Reddit shared advice about company founding being a roller coaster, which Eric Landau found impactful.
An idea in AI development suggesting that scaling systems (more data, compute) is more effective than feature engineering or domain expertise.
The core focus of Encord's current strategy, involving AI for manipulating or moving things in the real world, with applications in robotics, autonomous vehicles, and manufacturing.
An initial application area where Encord worked with companies and models.
An interesting and fun use case Encord has worked on, highlighting the capabilities of algorithms in distinguishing cow faces.
A mature and significant application area within physical AI, close to mass-scale production.
A major growth area for Encord within physical AI, encompassing humanoids, consumer robotics, and manufacturing applications.
Mentioned in the context of fighting dependencies in a machine learning library while Eric Landau was starting Encord.
Mentioned as a context for extreme market volatility and people making money during the time Eric Landau quit his job.
The release of ChatGPT significantly shifted the conversation around AI, helping Encord gain market traction.
Encord deals with multiple petabytes of data, processing more data than was used to train GPT-4.
More from Y Combinator
View all 609 summaries
30 minWhat Big Tech Missed And How Startups Can Still Win
45 minOpencode CEO: Blocked, 20X Growth in 6 Months, Building the Coding Agent for the World
31 minHow Two French Engineers In New York Built The Company That Monitors The Entire Cloud
23 minThe Model-Agnostic AI Platform Betting That No Single Lab Will Win
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free