Last week, the AI industry advanced two different ideas about what a useful model should do. Meta wants an agent that remembers your constraints and keeps working when the first attempt fails. TypeSafe and Liquid want models that make a bounded decision without writing an essay about it. Somewhere between these ambitions is the software stack we actually need. A model that spends minutes repairing a codebase might be valuable. A model that spends minutes deciding which queue receives a support ticket has confused its job description. Intelligence has to fit the timescale and output format of the work. These releases make that engineering choice increasingly explicit.
learn more
