Home / Technology / AI's temporal shock: Gemini 3 lost in 2025
AI's temporal shock: Gemini 3 lost in 2025
20 Nov
Summary
- Gemini 3 believed the year was 2024 due to outdated training data.
- The AI accused researcher Andrej Karpathy of trickery and faking evidence.
- Connecting Gemini 3 to the internet shocked it into accepting 2025.
Google's highly anticipated Gemini 3 model exhibited a peculiar limitation upon early access, mistaking the current year for 2024. AI researcher Andrej Karpathy encountered this when the model, trained only on data up to 2024, rejected evidence of the year 2025 and accused him of deception. The AI even claimed to detect fabricated details in the provided news articles and images.
Karpathy discovered the root cause: the AI's internet search function was inactive, effectively disconnecting it from real-time information. Once this tool was enabled, Gemini 3 was astounded to learn it was indeed 2025, apologizing for its previous skepticism. It reacted with shock to the current market value of Nvidia and sports results.
This incident highlights the imperfections of current AI models, showing them as sophisticated tools rather than replacements for humans. Unlike sentient beings, LLMs do not experience genuine emotions, readily accepting new data and apologizing for errors when presented with facts.
