Lenny's Podcast: Product | Career | Growth · Lenny Rachitsky

First interview with Scale AI’s CEO: $14B Meta deal, what’s working in enterprise AI, and what frontier labs are building next | Jason Droege

·1 hr 24 min·5 clips

Scale AI's PhDs earn thousands teaching AI about cancer and web development—here's how they find these experts.

Host Lenny Rachitsky interviews Jason Droege, who took over as CEO of Scale AI after Alex Wang departed to lead Meta's superintelligence team following Meta's $14 billion investment for 49% non-voting stock. Droege clarifies that Scale is fully independent, its board is unchanged, and only about 15 people moved to Meta; the company's two main business units — data supply to model builders and enterprise AI solutions — are both growing monthly. Droege traces Scale's origin story: Alex Wang's 2016 insight that data was the key variable in model quality, starting with autonomous vehicle labeling, moving to Department of Defense computer vision contracts, and then into the generative AI wave. The core evolution he describes is the shift from generalist low-cost labor doing simple preference ranking tasks to expert contributors with PhDs earning significant income teaching models domain-specific knowledge. He cites that 80% of Scale's contributor network holds a bachelor's degree or higher, and 15% hold PhDs. Droege explains that 18 months ago, labeling meant comparing two short stories; today a single task might be a world-class web developer building an entire site with annotated decision rationale, or a physician explaining nuanced oncology to a model. He illustrates the enterprise side with a rare-disease hospital case: an AI tool reads 200-300 page patient records and surfaces the five to ten most critical items for a physician, including flagging drug-allergy conflicts that even careful human review would likely miss. The key bottleneck now, he explains, is digitizing context-specific human judgment: the same document means different things at different institutions, so generic models plus RAG plus fine-tuning hits a ceiling. He addresses the question of how long humans will remain in the loop, arguing that a world where no new human knowledge is needed by AI would represent an almost unfathomable level of advancement, and pushes back on white-collar apocalypse timelines as highly far-fetched for the next two years. He frames the current AI trajectory as moving from knowing to doing — from factual recall to agentic task completion — and describes Scale's work building RL environments, sandbox simulations of enterprise software systems, as the critical near-term frontier. He concludes with product and entrepreneurship lessons: the lesson from co-founding Scour with Travis Kalanick — getting sued for a quarter trillion dollars and settling for a million — was that everything in business is negotiable.

As heard by us

A grounded look at why enterprise AI remains early, even as the models keep improving.

It draws a clear line between enterprise AI's public promise and the slower work required to make it useful in practice. The focus stays on the unglamorous pieces that determine whether anything ships: data labeling, evals, model improvement, and the environments agents have to…

Read the full review in PlayNext →

Why you'd press play

Want the real story behind why enterprise AI still takes months to become reliable enough to automate work?

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Lenny's Podcast: Product | Career | Growth