Descriptions:
Fahd Mirza tests Fugu Ultra — a new orchestration model from Tokyo-based Sakana AI — by giving it a single screenshot of the live FIFA.com World Cup fantasy interface and asking it to build a working clone from scratch, executed through the Hermes agent framework. The experiment skips benchmarks in favor of a real-world capability test, reflecting Mirza’s skepticism about launch hype.
Fugu Ultra is built on two ICLR 2026 papers, primarily “Conductor,” which trains a small model via reinforcement learning (GRPO) to plan and delegate tasks across a pool of frontier models including GPT-5 and Claude Sonnet 4. The conductor decides which worker models to call, what subtask each handles, and how they coordinate — and can invoke itself recursively for harder problems. In the demo, Fugu doesn’t just write code: it builds a Python/SQLite backend, writes its own verification scripts, spins up a live server, hits every endpoint over a real socket, runs a pytest suite, and refuses to delete another agent’s temp file — genuine self-verification rather than hand-waving.
The resulting app is a functional Premier League fantasy game with working squad management, player transfers, and a budget panel. Notably, Fugu reinterpreted the brand from its own knowledge rather than pixel-copying the screenshot — generating its own players (Haaland, Salah, Luis Díaz) and pricing — demonstrating semantic understanding of the task rather than surface imitation. Mirza’s verdict: a credible real-world performance that justifies closer attention to Sakana’s orchestration approach.
📺 Source: Fahd Mirza · Published July 07, 2026
🏷️ Format: Hands On Build
