Transcript of AIE Miami Keynote & Talks ft. OpenCode. Google Deepmind, OpenAI, and more!
AI Engineer
Key takeaways
- Gabe Greenberg introduced Orchestrator AI, a multi‑agent orchestration platform that can run up to 16 agents per task, provides confidence scores, adversarial governance, self‑pruning context memory, and a meta‑observer that auto‑adds skills.
- Benchmarks of Orchestrator AI against Sweetbench Pro (GPT‑4.5‑high) showed lifts of 17.1% on easy tasks, 14.8% on medium, 1.7% on hard, 5.7% on very hard, and an overall 8.4% improvement.
- Dax Raad argued that the ease of shipping AI‑generated MVPs leads to product bloat, with design falling behind engineering and many low‑quality features degrading overall product experience.
- He advocated for restraint: filter ideas, prioritize high‑leverage features, and focus onboarding on a single core idea rather than proliferating many minor features.
- Dexter (Dex) admitted his earlier recommendation to avoid reading code was wrong, emphasizing that code must still be read and reviewed to avoid catastrophic failures.
