Yoav Orlev’s Post

Just wrapped a talk with Gabriel Grinberg at Code with Claude in London about something we're still wrapping our heads around: in less than a year, Base44's R&D went from 1 to 80. A few things we tried to get across on stage: As AI compounds, your operational mode has to compound with it. This is true for your product, but it's at least as true for your process. If your process doesn't get better when the next model ships, your process isn't structured right. For small teams, keep it simple! lean on AI extensively, but resist the urge to build sophisticated process around it early on. A smart prompt goes a long way these days. The scaffolding can wait until you actually need it. Some business-critical efforts are genuinely hard to do well early on. Find a smart proxy instead. Evals were ours. Proper evals are brutal work: the right test set, the right rubric, the right scoring loop, and keeping it all from rotting the second a new model ships. We replaced the whole thing with our now-famous Base Frustration Score. It carried us further than the formal version would have at that stage. Huge thanks to the Anthropic team for the warm welcome, the brilliant audience, and the genuinely killer makers swag.