What Are the Dumb Ways to Die as an AI Startup in 2025?
May 12, 2025
Assume models won't get better Building your moat around "GPT-5 can't do X yet" is a death sentence.
Your defensibility must be orthogonal to model capabilities.
Treat evals as performative Without rigorous evaluation frameworks, you're flying blind. Evals are your early warning system.
Vibe code production features Ship fast, not recklessly. In AI products, trust compounds slowly and evaporates instantly.
Skip in-person at SF The density of AI talent, customers, and capital in SF is unmatched. Serendipity still matters.
Never build a data moat Proprietary data and feedback loops are real differentiators.
Treat non-AI infrastructure as secondary Brilliant LLM + broken integrations = customers leave.
Which mistake do you see most often?