AI

Core Thesis

AI does middle-to-middle, not end-to-end. You still have to prompt and verify it.

Domain expert + AI > non-expert + AI. Specialist prompting a model has background knowledge the model can't replicate from context alone. Non-experts will use AI to replace experts — but it won't be at the level of the expert prompting it.

Impact on India

Cost advantage of Indian outsourcing & service industry will diminish. AI automates exactly this tier — text processing, code generation, data entry, analysis templates. Huge chunk of exports and GDP at risk.

Automation Trajectory

All work involving patterns gets automated. The progression:

1. Text (internet was text-heavy) → LLMs emerged
2. Images → bad at first, now production-quality
3. Video → expected plateau never arrived
4. Novel architectures → molecule prediction, optimization, reasoning, agentic workflows

No plateau in sight. Each modality was expected to hit a wall; none did.

The LLM Nature of Human Work

Most of the work we do is LLM-type. I say Hello how are you, you reply Fine/Great. You ask me How are you, I say Good. Formulaic. Predictable. Automatable.

For high-level work — you'd hire a specialist who uses the AI model because he's the domain expert with background knowledge needed to reason. Sure non-experts will use it to replace experts but it won't be at that level.