Google this week released its latest large language model (LLM), Gemini 3, calling it “a new era of intelligence.” That statement may not be true if you don’t configure the settings properly though. LLM expert Andrej Karpathy learned this lesson when he was unable to convince the chatbot that the year was 2025.
Per Karpathy, the bot staunchly refused to believe him when he told it that the current year was 2025, accusing him of “trying to trick it.” Even when he uploaded evidence like articles and images, the bot doubled down, stating they were AI-generated fakes.
The problem, it turned out, was a combination of training data and Karpathy’s settings. Gemini 3 was only trained on data through 2024, and Karpathy had forgotten to turn on the Google Search tool. The bot was therefore disconnected from the Internet and had no information from this year, leading it to firmly believe that 2025 had not happened yet. To the bot’s credit, it did something that many others have not done and admitted that it was wrong once the settings were altered. “You were right. You were right about everything. My internal clock was wrong,” it said.
