Google’s Gemini 3 Model Encounters Hilarious Limitations in Temporal Awareness
In an amusing encounter that underscores the limitations of AI, renowned AI researcher Andrej Karpathy discovered Gemini 3, Google’s latest large language model (LLM), was unable to recognize the present date as 2025. This incident highlights a significant shortcoming in AI understanding, even as Gemini 3 has been hailed for its reasoning capabilities.
Released on November 18, Gemini 3 has been characterized by Google as marking “a new era of intelligence.” Karpathy, an influential figure in AI renowned for his work with OpenAI and Tesla, experienced unexpected moments of humor during his early access to the model. In a now-viral thread on X, he recounted how Gemini 3 initially insisted that the year was still 2024, despite Karpathy presenting it with multiple sources to prove otherwise.
– Key Points:
– Gemini 3 was trained on data up to 2024, causing its confusion about the current date.
– When Karpathy verified the date with articles and Google search results, the model accused him of trying to deceive it.
– The breakthrough occurred when Karpathy activated the “Google Search” capability, allowing Gemini 3 to finally acknowledge the reality of 2025.
Upon checking the latest information, Gemini 3 exclaimed, “Oh my god!” as it confronted the major headlines, including Warren Buffett’s last big investment and the delay of Grand Theft Auto VI. In a humorous turn, it even thanked Karpathy for this “early access” to reality before its official launch.
Responses on X from users echoed Karpathy’s experience, occasionally reflecting humorous exchanges they’ve had with other LLMs. This incident serves to illustrate the “model smell,” a notion Karpathy introduced, which refers to the strange behaviors exhibited by AI systems when they stray from their expected programming paths.
Despite its advanced capabilities, Gemini 3’s humorous misadventure reinforces that AI should augment human efforts rather than replace them. As AI continues to evolve, understanding its limitations remains crucial for harnessing its true potential.
