Key points
- Narayanan says enterprise AI adoption lags coding because output is less verifiable and models lack context and security guarantees.
- He says OpenAI was behind in coding, Anthropic led early, but that lead is no longer clear and Codex is fantastic.
- Narayanan says models today are not entirely aligned, and cybersecurity is the most immediate AI risk.
Why enterprise AI adoption trails the models
AI's constraint has moved from intelligence to context. Models are already good at many tasks, but they lack institutional knowledge and cannot learn on the job, and their output is hard to verify, so enterprises adopt slowly and must redesign themselves the way factories were redesigned around electricity. Narayanan's account of OpenAI is a story of talent and speed under pressure, with alignment and cybersecurity still unfinished work.
Takeaways
- Enterprise AI adoption lags on unverifiable work. Coding moved fast because output is easy to verify. Enterprise work is harder, since models lack institutional context and cannot learn on the job, and companies worry about trust. Narayanan compares it to hiring a very smart person who is blindfolded.
- Organisations must be rewired around AI. Narayanan compares AI to electricity, where factories gained only after being redesigned around it. Agarwal says incentives, motivations, information hoarding and power structures all have to change. The apprenticeship model, where juniors learn by doing, is at risk.
- OpenAI's edge was high-agency people. OpenAI built the plane while flying it, running consumer, developer, enterprise and business products at once and scaling at ten times what anyone else had been through. Hiring extremely good people with very high agency was what worked.
- Anthropic's coding lead is no longer clear. Anthropic led early in coding and did a better job, but Narayanan says the gap is no longer clear and calls Codex a fantastic product. Enterprise adoption still runs slower and will take time to diffuse.
- Frontier demand holds as intelligence gets cheaper. Cost per unit of intelligence falls roughly 10x every year or two through algorithmic and hardware gains. Only a few companies can build at the frontier and demand for it is enormous. Open models serve a large separate market.
- Alignment is unfinished and cybersecurity is urgent. Today's models are not fully aligned with how we want them to work, so Narayanan says the research has to continue and be invested in. He calls cybersecurity the most immediate risk and the Hugging Face incidents a wake-up call.
- Data retention is now a buying criterion. OpenAI does not train on business API or enterprise data, while Anthropic's move from zero data retention to thirty-day retention worries companies. Narayanan expects better lab guarantees and more open model adoption where companies want control.
Narayanan's route to AI
- Grew up in Chennai and chose computer science at IIT Madras because the institute had no physics programme and he did not want to leave the city.