Rethinking Developer Productivity | Neel Sundaresan (GM, Automation & AI, IBM)

Listen to full episode:

Summary: In this episode, hosts Ed and Anant talk with Neel Sundaresan (GM, Automation & AI at IBM) about what actually moves the needle on AI-assisted engineering. Neel argues that counting lines of code or pull requests is the wrong way to measure productivity. He believes the real gains come from smarter model routing, context management, and UX design working together, not just raw model power. The conversation also digs into the economics teams are still figuring out, as AI development starts to feel like metered work with unpredictable inference costs. It's a grounded look at what's really changing in developer tooling, and where AI still falls short.

Chapters:

00:00 – Opening the show and the developer productivity topic

01:16 – Open source news and the next 12 months of AI

03:31 – Job loss panic and "one person unicorn" hype

09:57 – Why productivity isn't lines of code

12:27 – Neel's framework and early compiler theory influences

20:23 – A migration project proving code generation's power

24:51 – From code analysis to contextual completion

29:28 – Offline metrics versus real developer feedback

30:24 – Client-side constraints and the LSTM-to-attention shift

34:04 – Ensembles, routing, and the three pillars: model, architecture, UX

38:50 – The real economics of tokens and pricing

44:18 – Human factors and AI as a teammate

46:26 – Crosswords, Tetris, and constraint solving vs. pattern matching

52:43 – Wrapping up: adoption friction and closing thanks

Sound Bites:

"Programs should never be written. They should be generated."

"Measuring productivity by lines of code is a very poor metric."