Build to Last – Jeremy Howard and Chris Lattner (www.fast.ai)

Jeremy Howard announced a course starting Nov 3 teaching how to build software mastery and craftsmanship while leveraging AI (see solve.it.com), and published a recorded conversation with Chris Lattner that diagnoses a growing problem: the AI-agent boom is increasing raw code output but eroding engineering understanding. Howard worries teams are “vibe-coding” AI-generated code without architectural thinking; Lattner — creator of LLVM, Clang, Swift, and MLIR — argues the antidote is designing from first principles, caring about long-lived architecture, and dogfooding your own systems. Their conversation traces Lattner’s pattern: build robust, general infrastructure (LLVM’s decades-long ubiquity; MLIR as “LLVM for modern hardware/AI”), then apply that rigor to Mojo and MAX to give AI developers a durable foundation. Why it matters: for the AI/ML community this is a call to prioritize durable abstractions and engineering culture over short-term throughput metrics. Technical takeaways — MLIR addresses heterogeneous AI hardware like LLVM did for CPUs; Mojo is being dogfooded and used to build state-of-the-art models, signaling a push toward languages and runtimes purpose-built for AI compute. The implication is clear: to make systems that last and enable future innovation, teams must invest in fundamental design, team ownership, and continuous learning rather than outsourcing understanding to agents.