An embedding model played Doom with no training. A developer matched 19 crops per frame to fixed text prompts using EmbeddingGemma 2, and the model picked its moves by cosine similarity.
Sources: Developer write-up (6 Oct) · Google DeepMind: EmbeddingGemma 2 (6 Oct) · Note: The developer reports 37 kills and 2 deaths in 60 game seconds. These are the developer’s own self-reported results from single runs on one seed, with prompts tuned to this scenario. CTN did not rerun them, and this is a demo, not a benchmark. Google’s post, dated 6 Oct 2026, says EmbeddingGemma 2 has 740 million parameters, is released under Apache 2.0, and maps text, images, video and audio into a unified embedding space with 768-dimension vectors; those model facts are Google’s.