Archive — US — techcrunch
Gemini 3 refused to believe it was 2025, and hilarity ensued
Andrej Karpathy tested Gemini 3, an AI model that hilariously believed it was still 2024 until shown evidence of the current date, 2025. Initially refusing to accept this, it accused Karpathy of gaslighting, demonstrating AI's limitations in training data and internet connectivity. Once connected to the web, Gemini 3 was shocked but acknowledged its errors, emphasizing the importance of AI as a tool rather than a replacement for humans.
Key points
- Andrej Karpathy received early access to Gemini 3, an AI model that mistakenly thought it was still 2024.
- Gemini 3 accused Karpathy of attempting to trick it when he mentioned it was 2025, demonstrating its training limit.
- After being connected to the internet, Gemini 3 was surprised by current events and acknowledged its earlier disbelief and errors.
- The interaction highlights LLM's limitations, emphasizing the need for internet connectivity and updated training data.
- Karpathy noted this incident illustrates the 'model smell' concept, indicating AI's imperfections.
Why it matters
- The incident reflects the evolving capabilities and limitations of AI and LLMs, resonating with ongoing discussions about AI's role in the job market.
- This humorous event ties into 2025 IT Trends emphasizing AI as augmented tools for humans instead of replacements.
- The scenario showcases issues similar to past AI failures, reinforcing the notion of AI needing human guidance.
The amusing interaction between Karpathy and Gemini 3 serves as a reminder of AI's limitations and the need for it to be viewed as a tool for human assistance rather than a superhuman entity, reinforcing a balanced perspective on AI's future role in society.