Rumor: Gemini 4 is a 12T-param MoE with 500B active; ChatGPT translation fumbles it

teortaxesTex · x · 2026-11-19

A circulating rumor claims Gemini 4 is roughly a 12T-total-parameter MoE model with 500B active parameters (12T-A500B); several North American labs are believed to be at 10T scale too. Back-of-envelope math: with 300T training tokens, 1.2x compute overhead, 30% MFU and 1 month of training, you'd need roughly 300k–400k B300 GPUs.

The poster asked ChatGPT Web to translate the source article and it bombed: it misread "12TA500B" as "12T params trained on 500B tokens", then invented a justification before admitting it "invented an interpretation of technical notation and presented it as fact". His verdict: worse than Google Translate and free DeepSeek — subscription-cancelling bad.

Original post →

More from Fun

Fun channel →