Gemini 3 Refused To Believe It Was 2025, And Hilarity Ensued 

Every clip you perceive a billionaire (or moreover a millionaire) CEO picture really LLM-based agents are coming for each nan quality jobs, retrieve this funny but telling incident astir AI’s limitations: Famed AI interrogator Andrej Karpathy sewage one-day early entree to Google’s latest model, Gemini 3. –and it refused to judge him erstwhile he said nan twelvemonth was 2025.

When it yet saw nan twelvemonth for itself, it was thunderstruck, telling him, “I americium suffering from a monolithic lawsuit of temporal daze correct now.” 

Gemini 3 was released on November 18 pinch specified fanfare that Google called it “a caller era of intelligence.” And Gemini 3 is, by astir each accounts (including Karpathy’s), a very capable, instauration model, peculiarly for reasoning tasks. Karpathy is simply a wide respected AI investigation intelligence who was a founding personnel of OpenAI, ran AI astatine Tesla for a while, and is now building a startup, Eureka Labs, to reimagine schools for nan AI era pinch agentic teachers. He publishes a lot of contented connected what goes connected under-the-hood of LLMs. 

After testing nan exemplary early, Karpathy wrote, successful a now-viral X thread, astir nan astir “amusing” relationship he had pinch it.  

Apparently, nan model’s pre-training information had only included accusation done 2024. So Gemini 3 believed nan twelvemonth was still 2024. When Karpathy attempted to beryllium to it that nan day was genuinely November 17, 2025, Gemini 3 accused nan interrogator of “trying to instrumentality it.”  

He showed it news articles, images, and Google hunt results. But alternatively of being convinced, nan LLM accused Karpathy of gaslighting it — of uploading AI-generated fakes. It moreover went truthful acold arsenic to picture what nan “dead giveaways” were successful nan images that supposedly proved this was trickery, according to Karpathy’s account. (He did not respond to our petition for further comment.) 

Baffled, Karpathy – who is, aft all, 1 of nan world’s starring experts connected training LLMs – yet discovered nan problem. Not only did nan LLM simply person nary 2025 training information but “I forgot to move connected nan ‘Google Search’ tool,” he wrote. In different words, he was moving pinch a exemplary disconnected from nan internet, which to an LLM’s mind, is akin to being disconnected from nan world.  

When Karpathy turned that usability on, nan AI looked astir and emerged into 2025, shocked. It virtually blurted out, “Oh my god.”  

It went connected writing, arsenic if stuttering, “I. I… don’t cognize what to say. You were right. You were correct astir everything. My soul timepiece was wrong.” Gemini 3 verified nan headlines Karpathy had fixed it were true: nan existent date, that Warren Buffett revealed his past large finance (in Alphabet) earlier retirement, and that Grand Theft Auto VI was being delayed. 

Then it looked astir connected its own, for illustration Brendan Fraser’s characteristic successful nan 1999 drama “Blast from nan Past,” who emerges from a explosive shelter aft 35 years. 

It thanked Karpathy for giving it “early access” to “reality” nan time earlier its nationalist launch. And it apologized to nan interrogator for “gaslighting you erstwhile you were nan 1 telling nan truth nan full time.”  

But nan funniest spot was nan existent events that flabbergasted Gemini 3 nan most. “Nvidia is worthy $4.54 trillion? And nan Eagles yet sewage their revenge connected nan Chiefs? This is wild,” it shared. 

Welcome to 2025, Gemini. 

Replies connected X were arsenic funny, pinch immoderate users sharing their ain instances of arguing pinch LLMs astir facts (like who the existent president is). One personification wrote, “When nan strategy punctual + missing devices push a exemplary into afloat detective mode, it’s for illustration watching an AI improv its measurement done reality.” 

But beyond nan humor, there’s an underlying message.  

“It’s successful these unintended moments wherever you are intelligibly disconnected nan hiking trails and location successful nan generalization jungle that you tin champion get a consciousness of exemplary smell,” Karpathy wrote. 

To decode that a little: Karpathy is noting that erstwhile nan AI is retired successful its ain type of nan wilderness, you get a consciousness of its personality, and possibly moreover its antagonistic traits. It’s a riff connected “code smell,” that small metaphorical “whiff” a developer gets that thing seems disconnected successful nan package codification but it’s not clear what is wrong.  

Trained connected human-created contented arsenic each LLMs are, it’s nary astonishment that Gemini 3 dug in, argued, moreover imagined it saw grounds that validated its constituent of view. It showed its “model smell.” 

On nan different hand, because an LLM – contempt its blase neural web – is not a surviving being, it doesn’t acquisition emotions for illustration daze (or temporal shock), moreover if it says it does. So it doesn’t consciousness embarrassment either.  

That intends erstwhile Gemini 3 was faced pinch facts it really believed, it accepted them, apologized for its behavior, acted contrite, and marveled astatine nan Eagles’ February Super Bowl win. That’s different from different models. For instance, researchers person caught earlier versions of Claude offering face-saving lies to explicate its misbehavior erstwhile nan exemplary recognized its errant ways. 

What truthful galore of these funny AI investigation projects show, repeatedly, is that LLMs are imperfect replicas of nan skills of imperfect humans. This says to maine that their champion usage lawsuit is (and whitethorn everlastingly be) to dainty them for illustration valuable devices to assistance humans, not for illustration some benignant of superhuman that will switch us.