StarCraft AI Ally team rented a PC bang and recruited 1,000+ humans to collect on-policy rollout data

Kangwook_Lee · x · 2026-09-29

The author shares a favorite behind-the-scenes detail: to improve their StarCraft agent "Ally," the team rented a PC bang and recruited 1,000+ people to play against early versions of the agent, collecting on-policy rollout data from real "users."

Since everything runs on the edge with a tight KV cache budget, they developed a dedicated context compaction algorithm to keep memory in check.

Related event: Kangwook Lee's Team Rented a Korean PC Bang, Recruited 1000+ Players to Train On-Device Game AI Ally Distilled to 2B(5 posts)→

Original post →

More from Research

• Ramez's new piece tests intelligence explosion claims against current AI scaling data — tobyordoxford · 2026-09-29
• Paper Puts a Number on AI-Human Attention Gap: Context Windows Up 3,906x, Focus Down — alex_verem · 2026-09-29
• Max Welling: AI Is Searching 10^60 Possible Materials and Accelerating Its Own Invention — charlieharris01 · 2026-09-29
• StepFun co-founder proposes KITE: PD-separation-inspired training for scaling agentic LLMs — teortaxesTex · 2026-09-29
• Alignment researcher points to 'self-undermining unilateral optimization' classics — edelwax · 2026-09-29
• World's first live AI-assisted brain surgery removes tumour using real-time DINOv3 inference — TimDarcet · 2026-09-29

Research channel →