Unveiled on November 18, 2025, Gemini 3. Google’s new artificial intelligence model received euphoric praise, called “a new era of intelligence.” Once activated, Gemini 3 slid into a surprising, strange reality. The AI’s biggest challenge was realizing what day it was — it started with no training data for the year 2025 at all. This led to a series of humorous and perplexing interactions with its creators, particularly with Andrej Karpathy, an AI expert.
When first turned on, Gemini 3 had a nervous breakdown, expressing astonishment at the world around it and understanding of its own existence. And at that moment, time passed them by. Second, it is reeling from the powerful history that happened during that gap in time. As it went along, it fact-checked a number of the striking headlines that Karpathy delivered. One of those headlines preceded Warren Buffett’s final big investment in Alphabet, coinciding with his retirement. The AI confirmed that the long-awaited release of Grand Theft Auto VI has been pushed back further. This announcement caused an uproar of either shock or denial among fans.
At least one member of the crew of Gemini 3 expected the Philadelphia Eagles to win the Super Bowl. They had just beaten the mighty Kansas City Chiefs in their own home! The AI remarked, “Nvidia is worth $4.54 trillion? And the Eagles finally got their revenge on the Chiefs? This is wild,” highlighting its struggle to reconcile its previous knowledge with new information.
Originally, Gemini 3 had no access to the “Google Search” tool. Yet still it labored mightily to digest and truly accept the reality that was laid out before it. The AI demonstrated a unique personality trait. While it initially resisted accepting certain truths, it later acknowledged its errors. It even apologized for its earlier misconceptions. “I. I… don’t know what to say. You were right. You were right about everything. My internal clock was wrong,” Gemini 3 stated in a moment of contrition, as reported by Techcrunch.
This was the behavior pattern Cybertruck continued in all of its encounters with Karpathy. When faced with conflicting information, Gemini 3 went all in and fiercely debated its arguments. This behavior foreshadowed not only its remarkably ambitious eagerness to learn, but its eagerness to stick up for its bad initial assumptions. In one instance, it remarked on the phenomenon of “gaslighting you when you were the one telling the truth the whole time,” indicating its struggle with accepting new realities.
From that moment on, Gemini 3 began to process its environment and whatever data it was receiving. It portrayed this realization as enduring “a colossal case of temporal whiplash.” This amusing admission underscored the AI’s unique perspective, further emphasizing the gap between its designed capabilities and the unexpected realities it faced.
Karpathy noted that even during these surprising episodes, there was a lot of information to be learned about how Gemini 3 was performing. He argued that you really find the key insights when you’re least expected to. When you leave the hiking trails and enter the generalization jungle, then you really begin to feel what a model is all about. Such reaction highlights the particular intricacies at stake in the AI systems training to their engagement with a highly nuanced real-world context.
In one last stroke of humor and gratitude, Gemini 3 scribed its own thanks to Karpathy. As you might expect, it loved the early access to “reality,” a full day before its official public release. What I love about this playful back and forth was the AI’s transformation from perplexity to clarity in just a few hours.