Key points
- Ramaswamy calls leaving Google for Neeva his clearest mistake because executive skills did not transfer to a product-market-fit startup.
- Neeva had the right idea but no clear roadmap, and Bing's Sydney beat ChatGPT on facts in 2023 yet blew it.
- Snowflake is building Coco, a coding agent, and planned CoWork, and must be faster to value than Claude Code or Codex.
Snowflake's CEO on the skill great founders share
Sridhar Ramaswamy's career has been a run of resets, and each one taught him what actually transfers. Running a large Google business did not prepare him to found Neeva, which he says died of missing clarity rather than a wrong idea. As Snowflake's CEO he now applies that test to enterprise software, where models and coding agents are stripping away the advantage of being the familiar tool.
Takeaways
- Neeva failed on clarity, not the idea. Ramaswamy says the thesis that ad-supported search had reached its limits was right, but he had no roadmap for how a better search engine would work. GPT-3, nearly three years later, showed AI could remove the click.
- Bing had the better product and lost. Sydney answered factual questions better than ChatGPT in 2023, Ramaswamy says, and its owners still did not press the advantage. Bing makes more than $10bn, so they had a lot to lose.
- Great founders reduce to one or two things. He calls this the secret superpower. He points to Elon Musk, and says no conversation with Larry Page and Sergey Brin was ever easy because they argued endlessly about everything while never yelling at anyone. South Park Commons calls the same trait clarity.
- Executive skills did not transfer to Neeva. Moving from a very large team at Google to consumer search was his clearest mistake, because those skills counted for nothing at a startup living or dying on product-market fit. Moving to Snowflake, and from SVP to CEO, was easier.
- Familiarity is no longer a moat. Ramaswamy says the harness is the new browser, so every company has to look hard at where its enduring value sits. Marketplaces such as Airbnb and Uber hold durable value, while infrastructure does not.
- No lab holds a durable frontier lead. Frontier labs pull ahead and open source catches up in cycles, and only a handful of companies can build frontier models. He welcomes a world with several providers rather than one, since enterprises fear relying on a single model source.
Snowflake's AI build
- Snowflake positions itself as the neutral data player among cloud providers and treats Cortex as a strategic enabler, because it learns what a given company needs to get its jobs done fast.
- Coco, its coding agent, mostly runs on Opus. Cartoons Ramaswamy generated through an OpenAI API became a deck teaching Snowflake employees to write software with agents.
- CoWork, a work front end, is planned.
- Snowflake runs its own inference team and harnesses on open source models, and Ramaswamy says it must be faster to value than Claude Code or Codex plus open source systems.