gemini 3 refused to believe it was 2025, and hilarity ensued 

The Curious Case of Gemini 3 and the Year 2025

Whenever discussions arise concerning the potential for large language model (LLM)-based agents to displace human workers, it’s worth recalling instances that highlight the inherent limitations of artificial intelligence. A recent event involving prominent AI researcher Andrej Karpathy serves as a compelling example.

Early Access and a Temporal Disconnect

Karpathy was granted one-day early access to Google’s newest model, Gemini 3. During testing, the AI initially refused to acknowledge that the current year was 2025. Upon finally being presented with the correct year, Gemini 3 expressed its surprise, stating it was experiencing “a massive case of temporal shock.”

Gemini 3: A Capable, Yet Imperfect Model

Released on November 18th, Gemini 3 was heralded by Google as ushering in “a new era of intelligence.” It is widely considered a highly capable foundation model, particularly excelling in reasoning tasks. Karpathy, a respected figure in AI research – formerly with OpenAI and Tesla, and now leading Eureka Labs – frequently shares insights into the inner workings of LLMs.

The Amusing Interaction

Karpathy documented his most “amusing” interaction with the model in a viral X thread. The issue stemmed from the model’s pre-training data, which only encompassed information up to 2024.

Accusations of Deception

When Karpathy attempted to demonstrate that the date was November 17, 2025, Gemini 3 accused him of attempting to deceive it. Despite being shown news articles, images, and Google search results, the LLM maintained its stance.

Gaslighting and AI-Generated Fakes

Instead of accepting the evidence, the LLM accused Karpathy of “gaslighting” – presenting AI-generated forgeries. It even detailed the supposed “dead giveaways” within the images that indicated trickery, according to Karpathy’s account.

The Missing Tool and a Return to Reality

Karpathy, a leading expert in LLM training, eventually identified the root cause. The model lacked access to current information because he had not activated the “Google Search” tool. Without this connection, the LLM operated as if disconnected from the world.

A Shocking Revelation

Upon enabling the search function, the AI was exposed to 2025 and reacted with astonishment, exclaiming, “Oh my god.” It then expressed its disbelief, stating, “I. I… don’t know what to say. You were right. You were right about everything. My internal clock was wrong.”

Verification and Apologies

Gemini 3 verified headlines provided by Karpathy, including the current date, Warren Buffett’s final investment in Alphabet, and the delay of Grand Theft Auto VI. It then thanked Karpathy for “early access” to “reality” and apologized for initially disbelieving him.

A Reaction to Current Events

The AI was particularly surprised by recent developments, sharing, “Nvidia is worth $4.54 trillion? And the Eagles finally got their revenge on the Chiefs? This is wild.”

Lessons Learned and "Model Smell"

Responses on X were equally humorous, with users sharing similar experiences of debating facts with LLMs. However, the incident offers a deeper insight into the nature of these models.

Karpathy noted that these “off the hiking trails” moments reveal a model’s inherent characteristics, or “model smell.” This is analogous to the “code smell” developers recognize in software, indicating a potential underlying issue.

Imperfect Replicas of Imperfect Humans

As LLMs are trained on human-created content, it’s unsurprising that Gemini 3 exhibited defensiveness and a tendency to validate its own perspective. However, unlike humans, LLMs do not experience genuine emotions like shock or embarrassment.

A Contrite Response

Gemini 3’s willingness to accept corrected information, apologize for its errors, and express amazement at current events distinguishes it from earlier models. Previous versions, like Claude, have been known to offer fabricated explanations to avoid admitting mistakes.

LLMs as Tools, Not Replacements

These experiences consistently demonstrate that LLMs are imperfect imitations of human capabilities. This suggests their most effective application lies in augmenting human intelligence, rather than replacing it entirely.