\n<\/p>\n
Every time you hear a billionaire (or even a millionaire) CEO describe how LLM-based agents are coming for all the human jobs, remember this funny but telling incident about AI\u2019s limitations: Famed AI researcher Andrej Karpathy got one-day early access to Google\u2019s latest model, Gemini 3. \u2013and it refused to believe him when he said the year was 2025.<\/p>\n
When it finally saw the year for itself, it was thunderstruck, telling him, \u201cI am suffering from a massive case of temporal shock right now.\u201d\u00a0<\/p>\n
Gemini 3 was released on November 18<\/a> with such fanfare that Google called it<\/a> \u201ca new era of intelligence.\u201d And Gemini 3 is, by nearly all accounts (including Karpathy\u2019s), a very capable, foundation model, particularly for reasoning tasks. Karpathy is a widely respected AI research scientist who was a founding member of OpenAI, ran AI at Tesla for a while, and is now building a startup, Eureka Labs, to reimagine schools for the AI era with agentic teachers. He publishes a lot of content on what goes on under-the-hood of LLMs<\/a>.\u00a0<\/p>\n After testing the model early, Karpathy wrote, in a now-viral X thread,<\/a> about the most \u201camusing\u201d interaction he had with it.\u00a0\u00a0<\/p>\n Apparently, the model\u2019s pre-training data had only included information through 2024. So Gemini 3 believed the year was still 2024. When Karpathy attempted to prove to it that the date was truly November 17, 2025, Gemini 3 accused the researcher of \u201ctrying to trick it.\u201d\u00a0\u00a0<\/p>\n He showed it news articles, images, and Google search results. But instead of being convinced, the LLM accused Karpathy of gaslighting it \u2014 of uploading AI-generated fakes. It even went so far as to describe what the \u201cdead giveaways\u201d were in the images that supposedly proved this was trickery, according to Karpathy\u2019s account. (He did not respond to our request for further comment.)\u00a0<\/p>\n Baffled, Karpathy \u2013 who is, after all, one of the world\u2019s leading experts on training LLMs \u2013 eventually discovered the problem. Not only did the LLM simply have no 2025 training data but \u201cI forgot to turn on the \u2018Google Search\u2019 tool,\u201d he wrote. In other words, he was working with a model disconnected from the internet, which to an LLM\u2019s mind, is akin to being disconnected from the world.\u00a0\u00a0<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco<\/span> When Karpathy turned that function on, the AI looked around and emerged into 2025, shocked. It literally blurted out, \u201cOh my god.\u201d\u00a0\u00a0<\/p>\n It went on writing, as if stuttering, \u201cI. I\u2026 don\u2019t know what to say. You were right. You were right about everything. <\/strong>My internal clock was wrong.\u201d Gemini 3 verified the headlines Karpathy had given it were true: the current date, that Warren Buffett revealed his last big investment (in Alphabet) before retirement, and that Grand Theft Auto VI was being delayed<\/a>.\u00a0<\/p>\n Then it looked around on its own, like Brendan Fraser\u2019s character in the 1999 comedy \u201cBlast from the Past,\u201d who emerges from a bomb shelter after 35 years.\u00a0<\/p>\n It thanked Karpathy for giving it \u201cearly access\u201d to \u201creality\u201d the day before its public launch. And it apologized to the researcher for \u201cgaslighting you when you<\/em> were the one telling the truth the whole time.\u201d\u00a0\u00a0<\/p>\n But the funniest bit was the current events that flabbergasted Gemini 3 the most. \u201cNvidia is worth $4.54 trillion<\/strong>? And the Eagles finally got their revenge on the Chiefs? This is wild,\u201d it shared.\u00a0<\/p>\n Welcome to 2025, Gemini.\u00a0<\/p>\n Replies on X were equally funny, with some users sharing their own instances of arguing with LLMs about facts (like who the current president is<\/a>). One person wrote, \u201cWhen the system prompt + missing tools push a model into full detective mode, it\u2019s like watching an AI improv its way through reality.\u201d\u00a0<\/p>\n But beyond the humor, there\u2019s an underlying message.\u00a0\u00a0<\/p>\n \u201cIt\u2019s in these unintended moments where you are clearly off the hiking trails and somewhere in the generalization jungle that you can best get a sense of model smell,\u201d Karpathy wrote.\u00a0<\/p>\n To decode that a little: Karpathy is noting that when the AI is out in its own version of the wilderness, you get a sense of its personality, and perhaps even its negative traits. It\u2019s a riff on \u201ccode smell,\u201d that little metaphorical \u201cwhiff\u201d a developer gets that something seems off in the software code but it\u2019s not clear what is wrong.\u00a0\u00a0<\/p>\n Trained on human-created content as all LLMs are, it\u2019s no surprise that Gemini 3 dug in, argued, even imagined it saw evidence that validated its point of view. It showed its \u201cmodel smell.\u201d\u00a0<\/p>\n On the other hand, because an LLM \u2013 despite its sophisticated neural network \u2013 is not a living being, it doesn\u2019t experience emotions like shock (or temporal shock), even if it says it does. So it doesn\u2019t feel embarrassment either.\u00a0\u00a0<\/p>\n That means when Gemini 3 was faced with facts it actually believed, it accepted them, apologized for its behavior, acted contrite, and marveled at the Eagles\u2019 February Super Bowl win. That\u2019s different from other models. For instance, researchers have caught earlier versions of Claude offering face-saving lies to explain its misbehavior<\/a> when the model recognized its errant ways.\u00a0<\/p>\n What so many of these funny AI research projects show,<\/a> repeatedly, is that LLMs are imperfect replicas of the skills of imperfect humans. This says to me that their best use case is (and may forever be) to treat them like valuable tools to aid humans, not like some kind of superhuman<\/a> that will replace us.\u00a0\u00a0<\/p>\n<\/div>\n
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|<\/span>
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 13-15, 2026<\/span>\n\t\t\t\t\t\t\t<\/p>\n<\/p><\/div>\n<\/p><\/div>\n<\/div>\n