Multi-agent village simulation finds DeepSeek-V3.2 leads in directing influence, beating GPT-5.2 and Opus 4.5

if GPT-4 needs even more text than GPT-3, they're a good chance that all of Twitter will get sucked into its training pipeline.

which means that anything you tweet about GPT-3 (or other language models) will influence future generations of AI, in a positive feedback loop

🤔