Mixture of Experts AI code generation: Wins, fails and the future
15 snips
Dec 26, 2025 Gabe Goodhart, an AI architect focused on open innovation, joins Chris Hay, a distinguished engineer in developer tooling, and Olivia Buzek, a developer advocate for AI. They explore the highs and lows of AI code generation, highlighting a dramatic optimization success with Apple Metal versus significant failures in simple tasks. The team debates whether models or developers dictate architecture and discusses the importance of agent orchestration over mere model selection. They also tackle the competitive landscape between open-source and proprietary tools, emphasizing the challenges of inference costs and co-evolution in closed ecosystems.
AI Snips
Chapters
Transcript
Episode notes
Use AI To Bootstrap, Not Replace Judgment
- Start projects with AI to bootstrap work and use AI in hiring as a skill filter because daily use is now common across teams.
- But keep realistic expectations: don't assume AI equals human-level reliability and remain vigilant for catastrophic failures.
Polished Code That Masked Failures
- Gabe described Cloud Code generating a polished CLI and tests that mocked everything but didn't actually work, requiring a week of manual chiseling to fix issues.
- The assistant created beautiful scaffolding but left untested mocks that consumed more developer time to correct.
Decide Who The Architect Is
- Chris Hay frames a core question: who is the architect — the human or the model — because models will choose frameworks, libraries, and mocking strategies if given that role.
- He argues production readiness requires explicit rules and AI-friendly architecture rather than leaving decisions to the assistant.