Just watched Deepseek go into infinite thinking loop over 2-3 specific tokens.. Is this still a thing for frontier models or was it a quantized version that I stumbled upon in busy hours?
