Three events in 72 hours defined the AI application era. The U.S. government forced Anthropic to pull Fable access, triggering a loud consensus: open-source and local models are not optional. Satya Nadella published a thesis on June 14 arguing the model cannot be the moat. Salesforce acquired Fin, formerly Intercom, for $3.6 billion on June 15, validating that AI-native application companies built on open-source models are worth buying at scale.
Building AI applications is not a staffing problem or an uptime problem. It is three new disciplines: model selection, loop design, and continuous evaluation. Each model has a distinct personality and failure mode. Kimi K2.6 writes well but loses precision. Qwen 3 27b punches above its size but stalls mid-toolchain. GLM 5.1 codes reliably but slowly. Designing the feedback loop that lets an agentic system improve is systems engineering, the kind Donella Meadows wrote about, applied to infrastructure that changes weekly. Evaluating the combined performance of model plus loop is ongoing labor most companies will not want to staff internally.
The original piece is worth reading in full for how it frames the core economic question of this moment: how much intelligence can you squeeze out of a token budget. That framing explains why Fin sold for $3.6 billion, why Nadella is worried about ecosystem stability, and why the Fable shutdown rattled developers so hard. The companies that master model selection, loop design, and evaluation will set the price of intelligence itself.
[READ ORIGINAL →]