Alexandr Wang: Building Scale AI, Transforming Work With Agents & Competing With China

Wang recounts leaving MIT after studying AI and applying to YC with an idea for chatbots for doctors. The team realized chatbot companies needed data and human labor, then launched an API for human tasks. Focusing on self-driving cars brought early traction, including Cruise as a major customer, though Wang says that market alone was not large enough to sustain a giant business.

Scale began working with OpenAI in 2019, and Wang says GPT-3 made scaling laws feel real; work on InstructGPT helped make generative AI a turning point for Scale. He expects companies to build specialized models from their own data and workflows, treating those assets as valuable intellectual property. He argues that prompting works for many uses, while reinforcement learning can push models further.

Wang describes a progression from AI assistants to single agents and then groups of agents handling assigned tasks. He expects people to provide vision, coordinate work, and resolve failures, since getting agents to perform reliably in complex workflows will remain difficult. At Scale, he says, agents already support hiring, quality control, data analysis, and sales reporting.

Scale expanded from data for self-driving cars into language models, government work, enterprise applications, and agentic workflows. Wang says the company focuses on differentiated AI capabilities built around customers’ own data and problems. He describes Humanity’s Last Exam, a benchmark of original problems from researchers: top model scores rose from about 7–8% at launch to above 20%, and he expects future evaluations to focus on fuzzier real-world tasks.

Wang sees the United States as likely to retain a lead in AI, but says China may have advantages in data, energy production, and manufacturing. He describes Scale’s Thunder Forge work with Indo-Pacific Command, which converts military planning workflows into agents and, he says, can reduce decision cycles from 72 hours to 10 minutes. He closes by stressing that leaders’ care for quality influences standards across an organization.