Kolibri-1 plays Doom with 72 action combos at 39ms median latency
kmodi · reddit · 2026-10-07
Developers hooked Aleph Alpha's Kolibri-1 up to classic Doom, letting the model output up to 72 action combinations across four decision groups in a single batched pass. They got inference down to 39ms median and 60ms p95 — apparently enough time to make questionable decisions. Demo page and gameplay video are public.
More from Fun
- Good Will Hunting Paradox Perfectly Roasts the ML Hype Cycle, From CNNs to MoE — zack_overflow · 2026-10-07
- Satirical Press Release: OpenAI to Release the Final Ten Minutes of 500 Unreleased Films — _onionesque · 2026-10-07
- Viral Quip: Soon Only Mainstream Media Will Still Write by Hand, Because AI Refuses to Lie for Them — tetsuoai · 2026-10-07
- Tomo launches Anonymous Relays, letting users send messages to people off the platform — Jackyhuang · 2026-10-07
- "Your agents want to work in a monorepo" — and the Bazel joke that followed — vboykis · 2026-10-07
- Fortune unpacks the explosion of memes about Anthropic CEO Dario Amodei — jeremyakahn · 2026-10-07