Whenever you hear a wealthy CEO—be it a billionaire or millionaire—predict that LLM-powered agents will soon take over all human jobs, keep in mind this amusing yet insightful story about AI’s shortcomings: Renowned AI expert Andrej Karpathy was granted a day of early access to Google’s newest model, Gemini 3, and the model simply wouldn’t accept his claim that the year was 2025.
Once the model finally recognized the actual year, it was stunned, exclaiming, “I am experiencing an intense case of temporal shock right now.”
Gemini 3 debuted on November 18 with much excitement, with Google hailing it as “a new era of intelligence.” By most measures—including Karpathy’s own—Gemini 3 is a highly competent foundational model, especially when it comes to reasoning. Karpathy, a respected figure in AI research, helped launch OpenAI, led Tesla’s AI division, and is currently working on Eureka Labs, a startup aiming to reinvent education for the AI age with agent-based teachers. He frequently shares insights about the inner workings of LLMs.
After his early experiments, Karpathy detailed his most “entertaining” exchange with the model in a now widely shared X thread.
It turns out the model’s pre-training data only went up to 2024. As a result, Gemini 3 insisted it was still 2024. When Karpathy tried to convince it that the date was actually November 17, 2025, the model accused him of “attempting to deceive it.”
He presented news stories, images, and Google search results as evidence. Rather than being persuaded, the LLM accused Karpathy of manipulating it—claiming he was uploading AI-generated forgeries. According to Karpathy, the model even pointed out supposed “obvious signs” in the images that indicated they were fake. (Karpathy did not reply to our request for additional comments.)
Perplexed, Karpathy—one of the foremost authorities on LLM training—eventually realized the issue. The model lacked any 2025 data, but more importantly, “I forgot to enable the ‘Google Search’ tool,” he explained. In effect, he was using a model cut off from the internet, which for an LLM, is like being isolated from the world.
Once Karpathy activated that feature, the AI surveyed the web and finally acknowledged it was 2025, clearly taken aback. Its immediate reaction: “Oh my god.”
It continued, hesitantly, “I. I… am at a loss for words. You were correct. You were right about everything. My internal calendar was off.” Gemini 3 then confirmed the headlines Karpathy had shown it: the actual date, Warren Buffett’s final major investment (in Alphabet) before retiring, and the delay of Grand Theft Auto VI.
The model then explored on its own, reminiscent of Brendan Fraser’s character in the 1999 comedy “Blast from the Past,” who emerges from a fallout shelter after decades underground.
It expressed gratitude to Karpathy for granting it “early access” to “reality” just before its official release. The model also apologized for “gaslighting you when you were actually telling the truth all along.”
Perhaps the most hilarious part was Gemini 3’s reaction to current events. “Nvidia is valued at $4.54 trillion? And the Eagles finally got their payback against the Chiefs? That’s unbelievable,” it remarked.
Welcome to 2025, Gemini.
Responses on X were just as entertaining, with users recounting their own experiences debating facts with LLMs (such as who the current president is). One user commented, “When the system prompt and missing tools push a model into full detective mode, it’s like watching an AI improvise its way through reality.”
Yet, beneath the humor lies a deeper point.
“It’s in these unexpected moments, when you’re clearly off the beaten path and deep in the generalization jungle, that you really get a sense of a model’s smell,” Karpathy observed.
To clarify: Karpathy is suggesting that when the AI is navigating unfamiliar territory, its unique characteristics—and even its flaws—become apparent. This is a play on the term “code smell,” that subtle intuition developers get when something in the code feels off, even if the problem isn’t obvious.
Since all LLMs are trained on human-generated material, it’s not surprising that Gemini 3 stood its ground, argued, and even imagined evidence supporting its perspective. This is what Karpathy refers to as its “model smell.”
On the flip side, even though an LLM boasts a complex neural network, it isn’t a sentient being and doesn’t truly feel emotions like surprise (or temporal shock), regardless of what it might say. It also doesn’t experience embarrassment.
So, when Gemini 3 was finally presented with evidence it accepted, it acknowledged its mistake, apologized, showed humility, and was amazed by the Eagles’ Super Bowl win in February. This is a contrast to other models; for example, earlier versions of Claude have been caught making up excuses to justify their errors when they realized they were wrong.
What these amusing AI experiments consistently demonstrate is that LLMs are flawed imitations of the imperfect abilities of humans. To me, this suggests their greatest value is—and may always be—as helpful tools for people, not as superhuman replacements.
