Jason Droege on Scale AI, Uber Eats, and the Hard Work of Training AI

Jason Droege on Scale AI, Uber Eats, and the Hard Work of Training AI

transcriptaiscale-aiuber-eatsdata-labelingtraining-dataevalslenny

Jason Droege on Scale AI, Uber Eats, and the Hard Work of Training AI

Jason Droege — CEO of Scale AI (since 2024, succeeding Alex Wang following Meta’s $14B investment for 49% non-voting stake); previously launched and led Uber Eats (0 to $20B run rate in 4.5 years); earlier co-founded Scour (1997, UCLA, peer-to-peer multimedia search) with Travis Kalanick — joins Lenny’s Podcast for his first interview since taking the Scale CEO role. The conversation covers the evolution of AI training data (from simple preference labelling to expert hours-long tasks), the hard reality of enterprise AI adoption, and a set of product and entrepreneurial lessons from Droege’s career.

Source: Lenny’s Podcast Speaker: Jason Droege

Key ideas

  1. AI training has moved from generalist labelling to expert judgment. Eighteen months ago, Scale asked workers to compare short stories. Now, a single task might require a world-class web developer to build and annotate an entire website, or a specialist to explain a nuanced cancer topic with supporting reasoning. 80% of Scale’s expert network has a bachelor’s degree or above; ~15% have PhDs. The bottleneck for enterprise AI is now “digitising human judgment” — the context-specific decision logic that varies company to company and cannot be extracted from static training data.

  2. From models knowing to models doing. The next frontier is not knowledge benchmarks (largely saturating) but agent environments: how does an AI navigate a Salesforce instance, a healthcare system, a calendar? Each environment has thousands of configurations and goals; training cannot generalise from a small set. Scale has been building RL environments for over a year. The practical implication: AI enterprise adoption is 6–12 months from POC to production-ready reliability, not weeks.

  3. Enterprise AI hype vs. ground reality. POCs fail not because AI is bad but because getting from 60–70% to “production-safe” accuracy is an exponential reliability curve — like data centre uptime nines. The high failure rate of pilots is partly a denominator effect (it has never been easier to start a POC). With a quality implementation partner and months of calibration, AI delivers genuinely novel insights (the healthcare allergy example: AI surfaced a drug conflict a human might have missed in a 200-page record).

  4. Product and company-building: gross margin as a filter. When evaluating new business ideas, Droege starts at a 60% gross margin and works backwards. If that’s impossible, why not? The answer usually reveals whether a business is genuinely differentiated or commoditised. This applies to Uber Eats: restaurant incrementality economics (20–30% ingredient cost, same labour scales with demand) justified the delivery fee structure. Gross margin is not perfect but is a fast signal of value-add and defensibility.

  5. Not losing is a precursor to winning. The best founders are forces of nature over long durations — they survive until their insight, timing, and product converge. Taking asymmetrically positive bets, avoiding unrecoverable risks, and staying solvent through doubt is the precursor to the moment when things click. The insight from Scour onward: in business, everything is negotiable; there is no way, only the way you can negotiate through.

On independent thinking and the insight test

Before starting any business: why do I have this insight? In a world of a million smart, motivated entrepreneurs, what positions me to see something they do not? The test is not “have I talked to a customer with a problem” but “why am I uniquely placed to solve it over a 5–10 year window.” Combined with personal conviction (“do I actually want to work on this?”), this filter quickly eliminates weak ideas.

On hiring

Three things to interview for: (1) curious problem-solver — can they articulate a reasoning process? (2) can work across people — humility, collaborative instinct; (3) good leader. Speed-of-market exceptions apply for very specific technical roles (~5%) where existing expertise cannot be built fast enough. The Uber Eats management team was largely the same from $0 to $20B, built as an organism of complementary strengths rather than assembled for credentials.

See also