November 24, 2025
Candy vs. comments
Three Years from GPT-3 to Gemini 3
Otters, candy, and a comment war over what AI really changed
TLDR: Gemini 3 shows off by coding a playable candy-powered game and working like a computer sidekick through Antigravity. The comments spar over timeline nitpicks, Google’s hardware edge, and whether “humans as directors” is real yet—asking when this power escapes the chat box and changes everyday work.
Google’s new Gemini 3 didn’t just talk—it whipped up a tiny playable game about a candy-fueled starship outrunning otters. The crowd loved the demo, but the comments quickly turned into a fact-check vs hype cage match. One pedant jumped in to correct the timeline (“ChatGPT was 3.5!”), while others argued the real story is Google’s muscle: not just software, but custom hardware powering these models, with victory laps pointing to Alphabet’s blowout quarter. Meanwhile, the vibe check on everyday use: folks are still stuck in the textbox. Even fans say the big usability leap is coding agents (like Antigravity) that feel like “another mind” doing work, but the UI hasn’t caught up. The line that lit up the thread: “humans are now directors, not fixers.” Cue eye-rolls from veterans who say they’ve heard that for a year and a half—so when does this go mainstream? The “PhD-level intelligence” claim morphed into campus humor, with grad students quipping these models feel like chatting with a very competent TA who occasionally needs a nap. Bottom line: Gemini 3 wowed with candy and code, but the community’s drama is over dates, hardware, and whether this thing can finally escape the chat window.
Key Points
- •Google launched Gemini 3 and the companion agentic coding tool Antigravity.
- •The author demonstrates Gemini 3’s progress by having it build an interactive mini‑game, showing capabilities beyond text generation.
- •Antigravity is positioned as similar to Claude Code and OpenAI Codex, enabling autonomous programming with user-granted computer access.
- •The article argues coding ability makes AI agents general‑purpose tools for a wide range of computer tasks executed via code.
- •A practical example shows Antigravity reading local files, planning work, and seeking approvals through an Inbox interface to compile and verify AI predictions.