The Hackathon Where We Mostly Just Talked (worksonmymachine.ai)

At a recent Google hackathon the reporters built Photoslider — not by furiously typing code but by sketching ideas while Chrome’s new “Built‑in AI” LLM (running locally, no server calls, API keys, or rate limits) handled most of the implementation. Photoslider analyzes an uploaded image with Chrome’s model, returns three abstract attributes scored 1–10 (e.g., “mystery,” “luminous”), and lets users tweak sliders or create custom attributes. The delta becomes a natural‑language prompt sent, along with the original image, to Gemini’s image generator to produce variations; users can branch from any point. The prototype and code are on GitHub (sublayerapp/photoslider). The bigger takeaway: the team spent ~95% of the time designing and guiding an AI agent rather than debugging, revealing a new failure mode — rapid agentic coding produces functioning features fast, but can yield architectural spaghetti if you don’t steer for coherence. The author argues hackathons (and engineering) must evolve from raw speed to “control at velocity”: designing agent‑proof architectures, anticipating mid‑build twists, and judging coherence under pressure. They propose an experimental contest (Iron Vibecoder) to test those skills and predict patterns and practices for safe, maintainable agent-assisted development in the near future.