San Francisco's AI event calendar is having a small identity crisis, and for once it sounds healthy. This week's invites keep asking versions of the same question: not which lab won, but which model belongs where.

At Baseten's multi-model architecture event, the copy says teams are "increasingly turning to smaller, specialized models" because they beat generalists on core tasks, with voice, structured data, and retrieval specialists on stage. Across town, a16z and OpenRouter's Lunch & Learn frames the founder anxiety like a recurring office ritual: a new model tops the benchmarks, do you switch? The promised discussion is not about vibes, it is about cost, latency, reliability, and unit economics. Which is to say, the city has moved from model fandom to procurement.

Even the more devotional rooms have gotten practical. Claude Community Build Day offers tracks called Delight, Breakthrough, and Everyday, but the brief is hands on keyboard, measure what the new model can do, then demo it before the snacks settle. Workato's AI Journal Club puts talks on reasoning dynamics, on-policy distillation, and nondeterminism in the same room as people who still have to ship something Monday.

Then there is Side-Project Party III, which sweetly describes itself as a "stress-free hackathon" and gives the city what it always secretly wants: earlybird coworking, intros, ships, and the friendliest beta testers in SoMa. That is the anthropology report, darling reader. SF still loves a launch, but the social unit of ambition is shifting from one heroic model to a stack of specialists, a benchmark spreadsheet, and a room full of people comparing latencies over sparkling water.

Where to be seen this week? In the rooms admitting that model choice is now an operating expense, not a personality.