Small Task-Specific Agents Beat General Intelligence in India, Says Definable AI Founder
anandesh-sharma · reddit · 2026-11-09
The founder of Definable AI, a 10-person startup in Gurugram, argues that frontier labs' race toward one giant general brain misses what most real problems in India need: small, specific tasks like converting selfies to passport photos, filing GST, or answering WhatsApp customers at 11 pm.
- Their general-purpose agent got 3k organic users before a data-driven pivot
- One narrow WhatsApp agent (photo-to-passport-size, photo-to-PDF) hit 250 users with no app and no website
- Lesson: narrow job + existing interface (WhatsApp, calls) beats a shiny new app
Their current approach: frontier LLM as teacher, a small model (SLM) as student, distilled for one proven workflow so it can run on a phone CPU. He also asks the community about evals for distilled agent models, tool-calling reliability on small models, and session state/human handoff for WhatsApp agents.
More from AGI Musings
- When machines do everything else, relationships are what's left for humans — ZeroStateReflex · 2026-11-09
- Onggi in 2039: running a kimchi-jar shop in the post-Acceleration economy — NathanpmYoung · 2026-11-09
- AI speeds up research by at least 1.4x, says Wildeford, citing private METR evidence — peterwildeford · 2026-11-09
- AI has broken science's main evaluation proxy, argues computational biologist — anshulkundaje · 2026-11-09
- Learning is compression: a long-form essay on manifold hypothesis and why intelligence is possible at all — burny_tech · 2026-11-09
- 'AI democratizes Jevons paradox itself': efficiency fuels even more usage — cocktailpeanut · 2026-11-09