Each time you hear a billionaire (or perhaps a millionaire) CEO describe how LLM-based brokers are coming for all of the human jobs, bear in mind this humorous however telling incident about AI’s limitations: Famed AI researcher Andrej Karpathy obtained one-day early entry to Google’s newest mannequin, Gemini 3 — and it refused to consider him when he stated the 12 months was 2025.

When it lastly noticed the 12 months for itself, it was thunderstruck, telling him, “I’m affected by a large case of temporal shock proper now.” 

Gemini 3 was launched on November 18 with such fanfare that Google referred to as it “a brand new period of intelligence.” And Gemini 3 is, by practically all accounts (together with Karpathy’s), a really succesful basis mannequin, significantly for reasoning duties. Karpathy is a extensively revered AI analysis scientist who was a founding member of OpenAI, ran AI at Tesla for some time, and is now constructing a startup, Eureka Labs, to reimagine faculties for the AI period with agentic lecturers. He publishes a lot of content material on what goes on under-the-hood of LLMs. 

After testing the mannequin early, Karpathy wrote, in a now-viral X thread, about essentially the most “amusing” interplay he had with it.  

Apparently, the mannequin’s pre-training information had solely included info by way of 2024. So Gemini 3 believed the 12 months was nonetheless 2024. When Karpathy tried to show to it that the date was really November 17, 2025, Gemini 3 accused the researcher of “making an attempt to trick it.”  

He confirmed it information articles, photos, and Google search outcomes. However as an alternative of being satisfied, the LLM accused Karpathy of gaslighting it — of importing AI-generated fakes. It even went as far as to explain what the “lifeless giveaways” have been within the photos that supposedly proved this was trickery, in line with Karpathy’s account. (He didn’t reply to our request for additional remark.) 

Baffled, Karpathy — who’s, in spite of everything, one of many world’s main specialists on coaching LLMs — finally found the issue. Not solely did the LLM merely don’t have any 2025 coaching information, however “I forgot to activate the ‘Google Search’ software,” he wrote. In different phrases, he was working with a mannequin disconnected from the web, which to an LLM’s thoughts, is akin to being disconnected from the world.  

When Karpathy turned that operate on, the AI regarded round and emerged into 2025, shocked. It actually blurted out, “Oh my god.”  

It went on writing, as if stuttering, “I. I… don’t know what to say. You have been proper. You have been proper about the whole lot. My inside clock was improper.” Gemini 3 verified the headlines Karpathy had given it have been true: the present date, that Warren Buffett revealed his final huge funding (in Alphabet) earlier than retirement, and that Grand Theft Auto VI was being delayed. 

Then it regarded round by itself, like Brendan Fraser’s character within the 1999 comedy “Blast from the Previous,” who emerges from a bomb shelter after 35 years. 

It thanked Karpathy for giving it “early entry” to “actuality” the day earlier than its public launch. And it apologized to the researcher for “gaslighting you when you have been the one telling the reality the entire time.”  

However the funniest bit was the present occasions that flabbergasted Gemini 3 essentially the most. “Nvidia is price $4.54 trillion? And the Eagles lastly obtained their revenge on the Chiefs? That is wild,” it shared. 

Welcome to 2025, Gemini. 

Replies on X have been equally humorous, with some customers sharing their very own cases of arguing with LLMs about info (like who the present president is). One individual wrote, “When the system immediate + lacking instruments push a mannequin into full detective mode, it’s like watching an AI improv its manner by way of actuality.” 

However past the humor, there’s an underlying message.  

“It’s in these unintended moments the place you’re clearly off the mountaineering trails and someplace within the generalization jungle you could greatest get a way of mannequin odor,” Karpathy wrote. 

To decode that a bit of: Karpathy is noting that when the AI is out in its personal model of the wilderness, you get a way of its persona, and even perhaps its detrimental traits. It’s a riff on “code odor,” that little metaphorical “whiff” a developer will get that one thing appears off within the software program code however it’s not clear what’s improper.  

Educated on human-created content material as all LLMs are, it’s no shock that Gemini 3 dug in, argued, even imagined it noticed proof that validated its standpoint. It confirmed its “mannequin odor.” 

Then again, as a result of an LLM — regardless of its refined neural community — just isn’t a residing being, it doesn’t expertise feelings like shock (or temporal shock), even when it says it does. So it doesn’t really feel embarrassment both.  

Meaning when Gemini 3 was confronted with info it truly believed, it accepted them, apologized for its habits, acted contrite, and marveled on the Eagles’ February Tremendous Bowl win. That’s totally different from different fashions. For example, researchers have caught earlier variations of Claude providing face-saving lies to elucidate its misbehavior when the mannequin acknowledged its errant methods. 

What so many of those humorous AI analysis tasks present, repeatedly, is that LLMs are imperfect replicas of the talents of imperfect people. This says to me that their greatest use case is (and should endlessly be) to deal with them like useful instruments to assist people, not like some form of superhuman that can change us.