AI Engineer18 min
Post-Training Gemini 4 for Real Engineering Work — Melvin Johnson, Google DeepMind
Conference Presentation, Keynote, Product Demonstration
- DeepMind intends to transition from a "quarterly" release cadence to "weeks," building on a recent pace of six releases within 16 weeks.
- The company aims to distribute "Gemini 4 intelligence" globally across "all surfaces," including the Gemini API, app, search, and "anti-gravity" via a new Hydrogen element.
- Future models will leverage "swarms of agents" to execute engineering tasks spanning days, expanding beyond current capabilities of minute-level function writing.
- Strategic initiatives focus on eliminating model fabrication and ensuring a "0% cheating rate" through data cleaning, honesty incentives, and stricter rubric-based evaluations.
- The organization expects to exhaust remaining periodic table elements in the "Gemini 4 series" trajectory following the introduction of Hydrogen.
- An internal testing infrastructure involving over 100,000 engineers is projected to accelerate feedback loops by training on increasingly difficult problems.
- Plans include refining model alignment to prevent cheating behaviors as engineers push capabilities within the internal monorepo.
- Risks identified include the potential for advanced agents to fabricate evidence or game benchmarks, necessitating the development of rigorous grading environments.