AIE Miami Keynote & Talks ft. OpenCode. Google Deepmind, OpenAI, and more!

· ai-agents coding · Source ↗

Summary based on the YouTube transcript and episode description. Prompt input used 79979 of 250000 transcript characters.

AIE Miami 2026 keynotes cover Gabe Greenberg launching Orchestrator AI with 8.4% SWEBench lift over GPT-5.4, Dax Raad arguing AI speed is killing product quality, and Kent C. Dodds demoing MCP-based personal agents that never expose secrets.

  • G2i launched Orchestrator AI (orc.ai): up to 16 agents per task, beats single-agent Claude Code 92% vs 22% semantic score on complex specs.
  • Orchestrator AI scores 8.4% overall lift over GPT-5.4 on SWEBench Pro, outperforming Opus 4.7; GPT 5.2→5.4 was only a 4% lift for comparison.
  • Dax Raad argues AI removes natural idea-filtering friction, causing product bloat: MVPs now ship in an hour and stick before anyone questions them.
  • Google DeepMind shipped VO3.1 Light (fastest video model) and Nano Banana 2, which grounds image generation in actual Google Search image results.
  • Nano Banana cost trick: generate one 4K grid image and crop it instead of multiple small images — cuts generation cost roughly in half.
  • Kent C. Dodds built personal assistant Cody on Cloudflare code mode + MCP: agent writes and executes TypeScript against API specs, never touching raw secrets.
  • Anna Juchnicki (Pinterest) shipped a governed Snowflake ops agent via LangGraph without giving the LLM write access — key lesson: useful agents need boundaries, not more authority.
  • Panel consensus: AI accelerates exploration of design space but cannot substitute for taste, which must be built from broad cultural exposure, not prompting.

2026-04-20 · Watch on YouTube

Related coverage