How to Build an AI Startup: Go Big, Be Strange, Embrace Probable Doom (www.wired.com)

A new long-form look at the current AI boom paints a portrait of thousands of lean, experimental startups racing to build products that were once prohibitively expensive or slow. Reported examples range from 19‑year‑olds using custom LLMs to predict protein behavior and design targeted pesticides (followed by rapid wet‑lab validation) to tiny teams training vision networks and deploying real‑time traffic‑light systems. Founders credit LLMs, vision models, GPU tooling and AI assistants for compressing weeks of engineering into hours or days, enabling rapid prototyping, automated research, and even agents that patrol databases or synthesize feature code (the piece name‑checks Claude Code as a tool developers paste bug descriptions into). The significance is structural: AI lowers the barrier to building sophisticated systems, so startups can parallel‑experiment cheaply and iterate violently fast, but they also become highly dependent on external models, cloud services and frequent platform changes. Technical implications include increased cross‑disciplinary stacks (ML + wet lab, robotics + embedded GPUs), reliance on pretrained models for code generation and tooling, and a shift in competitive advantage away from raw engineering toward product judgment or “taste.” That makes flexibility, product-market judgment, and the ability to rewire entire codebases weekly the new core skills — even as many experiments will inevitably fail.