Google DeepMind

Confidence 0.75 · 2 sources · last confirmed 2026-06-15

Google’s consolidated AI research division (the 2023 merger of DeepMind and Google Brain), described from the inside as “the engine room of Google” (Kilpatrick 2026). It is part of Google and produces the Gemini model family, the Antigravity agent harness, and the Omni multimodal model.

What the wiki holds on GDM

• Harness synthesis research. AutoHarness (Lou et al., March 2026) shows a small model (Gemini-2.5-Flash) automatically synthesizing a code harness — up to the harness-as-policy extreme where the LLM is not called at runtime. Mechanism-level evidence for the “model subsumes the harness” direction.
• The “model eats the harness” thesis. Articulated by Logan Kilpatrick in the Sequoia interview: the harness gets upstreamed into the model within ~12 months.
• Antigravity + Omni. Antigravity (from the Windsurf team) is the single agent harness now rebasing Google’s products; Omni is the single any-input→any-output model replacing ~8 separate systems.
• Culture. Strong research portfolio under Demis Hassabis (Nobel laureate); explicitly non-zero-sum framing; deploying Gemini to billion-user products at a scale “only two companies in the world” can.

Relationship to Google

GDM is the AI-lab role within Google; this wiki’s broader Google footprint (Cloud, ADK, Google Research) is tracked on the Google page. GDM employs Logan Kilpatrick among ~100,000+ Google engineers who dogfood Gemini.