Google teases 'Jitro' AI
Google posted that it has launched Jitro, a self‑coding AI system described as evolving autonomously and prompting discussion about AI self‑improvement capabilities. The social post circulated on X and sparked debate about implications for developer workflows and model safety. (x.com)
Google has not published a verifiable product announcement for an AI system called Jitro on its official sites, even as an X post describing a Google “launch” circulated online. (blog.google)(deepmind.google) What Google has announced instead is AlphaEvolve, a Gemini-powered coding agent that writes candidate programs, tests them with automated evaluators, and keeps improved versions in an evolutionary loop. Google DeepMind published that announcement on May 14, 2025. (deepmind.google) Google later said it brought AlphaEvolve to Google Cloud in private preview on December 9, 2025, framing it as a tool for optimization problems such as chip design, logistics routing, and data-center scheduling. In that post, Google Cloud said the system starts from a working “seed” program and iterates when new code scores better on a user-defined evaluator. (cloud.google.com) That distinction matters because “self-coding” can mean several different things. In Google’s published description, the model does not rewrite itself freely; it proposes code, runs tests, and keeps changes only when an external scoring system says the result improved. (cloud.google.com)(deepmind.google) Google has been pushing more “agentic” systems since at least December 11, 2024, when it introduced Gemini 2.0 as a model built for systems that can plan ahead and take actions with user supervision. Its March 2026 AI roundup also highlighted coding and app-building tools, including Google Antigravity in Google AI Studio. (blog.google 1)(blog.google 2) Google says AlphaEvolve has already been used inside the company on engineering tasks. DeepMind said it improved parts of Google’s computing stack, and Google Cloud later said one AlphaEvolve-discovered scheduling change recovered an average 0.7% of global compute resources while another sped up a Gemini training kernel by 23%, cutting Gemini training time by 1%. (deepmind.google)(cloud.google.com) The safety debate around posts like the Jitro claim lands in a real policy context at Google. The company’s AI Principles say it develops and deploys AI “responsibly,” and Google DeepMind has updated a Frontier Safety Framework that sets protocols for severe risks from advanced models, including version 2.0 on February 4, 2025, and version 3.0 on September 22, 2025. (ai.google)(deepmind.google 2)k/)) As of April 12, 2026, the public record supports a narrower story than the viral post: Google has documented an evolutionary coding agent called AlphaEvolve, but I could not find an official Google announcement for a product named Jitro. The online discussion is real; the confirmed Google launch is AlphaEvolve. (blog.google)(deepmind.google)(cloud.google.com)
