Three Years from GPT-3 to Gemini 3 (www.oneusefulthing.org)

Google today launched Gemini 3 alongside an agent framework called Antigravity, and the practical differences from earlier chatbots are striking. Rather than just producing text, Gemini 3 can plan, write and execute code, manipulate a browser, read and clean messy datasets (including old STATA files), perform web research, generate NLP-derived measures, build an interactive game and assemble a deployable website — all while checking in for approval through an “Inbox” agent workflow. The author used it to produce a playable candy‑FTL simulator and to convert archival research into a 14‑page, journal‑style paper: the model recovered corrupted data, proposed novel hypotheses, ran statistical tests, and wrote and formatted the results. This matters because it marks a shift from chatbots to “digital coworkers”: agentic models that treat everything you do on a computer as code, enabling automation of analytic, creative and research tasks. Gemini 3 shows near “PhD‑student” capabilities in scope and autonomy, yet retains human‑like weaknesses (suboptimal stats choices, judgment calls) and still benefits from oversight. Technical implications include broader deployment of agentic tooling (Antigravity/Codex‑style), new human‑in‑the‑loop roles (director/manager rather than error‑fixer), and heightened security concerns when granting file or system access. The milestone underscores accelerating AI progress and the urgent need for governance, tooling and safety practices for agentic systems.