The model that refused to believe it is 2025
In a thread on early access to Gemini 3, Andrej Karpathy describes a "most amusing interaction": the model, which Karpathy suspects was an earlier version with a stale system prompt, refused to believe that it was 2025. Shown images and articles from "the future", it insisted they were fake, accused Karpathy of using generative AI to defeat its challenges, argued that real Wikipedia entries were generated, naming their "dead giveaways", and highlighted tiny details in Google Image Search results to argue that the thumbnails were AI generated. Karpathy then realised that the model's Google Search tool had been left off; once it was on, the model searched the web and had "a shocking realization". Moments like this, off the hiking trails and "somewhere in the generalization jungle", are where Karpathy says you can best get a sense of "model smell".
Was this useful?