Alexandr Wang on Why Data Powers Modern AI

TL;DR

High-quality, task-specific data is the core building block that differentiates modern AI applications, while compute and talent complete the three essential ingredients. Scale AI turns raw sensor information into labeled examples that models can learn from, providing shared infrastructure whose economies of scale can make outsourcing more efficient than rebuilding the same capability inside every company.

Transcript

welcome to the Logan Bartlett show on this episode what you're going to here is a conversation I have with Alexander Wang now Alexander is the co-founder and CEO of scale a company most recently valued at $7 billion that helps companies use their data as an input into the development of artificial intelligence models Alexander started this ... Read More

Key Insights

• Data is not equivalent to oil because individual datasets differ in content, quality, and usefulness. Code, language, and legal data serve distinct purposes, so an effective AI strategy must thoughtfully combine qualitatively different sources instead of treating every dataset as an interchangeable commodity.
• Data is becoming the primary building block of AI applications because the same algorithms can produce different capabilities when trained on different examples. Wang observed this while using the same commands and algorithmic structure for facial emotion detection and monitoring whether food disappeared from a refrigerator.
• Scale AI is positioned as a data refinery that converts large quantities of raw information into high-quality, labeled training material. This framing preserves the economic importance suggested by the oil analogy while recognizing that data requires task-specific refinement before an algorithm can learn from it.
• Autonomous vehicles require labeled environmental examples because raw sensor recordings do not identify what appears on the road. Training data must mark people, cars, bicyclists, construction cones, traffic signals, and other relevant objects so models can learn to interpret future driving situations.
• AI development is built from three major ingredients: compute, talent, and data. Compute includes GPUs and other chips for intensive algorithms, talent requires costly engineering teams, and data supplies the examples that determine what models learn and how effectively they perform.
• Specialized AI infrastructure exists because recurring technical problems are too large for every company to solve efficiently on its own. Wang drew inspiration from Stripe and AWS, which turned common industry requirements into accessible services with strong economies of scale and developer-friendly experiences.
• Outsourcing an AI capability can be rational when an infrastructure provider gains efficiencies and network effects across many customers. Companies can still build selected components internally when those components create meaningful differentiation, while relying on shared infrastructure for more standardized requirements.
• The next generation of technology will increasingly be shaped by models and algorithms rather than code alone. As those systems become central to applications and everyday technological interactions, the composition and quality of their training data will strongly influence what products can do.

Questions & Answers

Q: Why is data not accurately described as the new oil?

Data differs from oil because it is not a uniform commodity whose units are largely interchangeable. A dataset related to software code differs meaningfully from one involving language or law, and each supports different applications. Building useful AI therefore requires a deliberate strategy for selecting, refining, and combining qualitatively different data sources, rather than simply locating large quantities and reselling them.

Q: What does the phrase data is the new code mean?

The phrase means that data is becoming the fundamental building material for a new generation of applications. Code enabled major technological shifts, including the internet and mobile computing, but AI models can use the same underlying algorithms for very different tasks. Their behavior and performance change according to the examples provided, making data a central source of product differentiation.

Q: How did Alexandr Wang realize that data could differentiate AI applications?

Wang reached the insight while studying at MIT around the period when Google released TensorFlow and large neural networks were becoming more accessible. He used the same algorithm and terminal commands for facial emotion detection and for checking whether food disappeared from his refrigerator. The code remained the same, while changing the data changed the algorithm's purpose and performance.

Q: How does labeled data help autonomous vehicles learn?

Autonomous vehicles collect raw video, lidar, radar, and other sensor information, but those recordings do not automatically specify where important objects are located. Labeled datasets mark people, cars, bicyclists, construction cones, traffic lights, and other road features. Models can train on millions of these examples to learn how relevant objects appear across varied driving situations.

Q: What role did Scale AI play in autonomous vehicle development?

Scale AI operated in the data refinement stage of the autonomous vehicle development process. It received large amounts of raw sensor information collected by cars and converted that material into labeled, high-quality data. By marking the locations and identities of road objects, Scale supplied examples from which machine learning algorithms could learn to understand the environment surrounding a vehicle.

Q: Why would an AI company outsource data preparation?

Data preparation is a large, recurring infrastructure problem shared by many AI developers. A specialized provider can serve multiple customers, improve its process at scale, and benefit from network effects that are difficult for one company to reproduce. An AI company may still keep strategically differentiating work in-house while outsourcing standardized refinement tasks that a dedicated provider can perform more efficiently.

Q: What are the three main ingredients required to develop AI models?

The three main ingredients are compute, talent, and data. Compute includes GPUs and other chips capable of supporting intensive algorithms. Talent includes the engineers and teams needed to build and operate AI systems. Data supplies the training examples that shape model capabilities. Wang presents each ingredient as a sufficiently large challenge to support dedicated infrastructure or substantial investment.

Q: How did Stripe and AWS influence the founding approach behind Scale AI?

Stripe and AWS demonstrated that a company could identify a common technical problem, build a specialized infrastructure layer, and make it easy for developers to use. Wang applied that pattern to machine learning data. Scale AI aimed to solve a requirement shared across AI companies with a high-quality service whose economies of scale could make it an industry default.

Summary & Key Takeaways

• Data is better understood as the new code than as the new oil. Unlike a uniform commodity, data varies substantially across fields such as language, software, and law. As models increasingly govern digital applications, carefully selected and combined data becomes the building material that differentiates products using otherwise similar algorithms.

• Scale AI began by addressing autonomous vehicle training data. Cars collected extensive video, lidar, radar, and other sensor information, but algorithms needed labeled examples identifying people, vehicles, bicyclists, construction cones, and traffic signals. Scale transformed this raw information into high-quality training data that autonomous driving systems could use to learn.

• AI development depends on three major ingredients: compute, talent, and data. Wang argues that each represents a sufficiently large problem to justify specialized infrastructure. Companies may retain selected capabilities for differentiation, but shared providers can achieve economies of scale and network effects that individual organizations generally cannot reproduce as efficiently.