DeepSeek is too focused on architectural and memory usage efficiency and not really caring about changing up how their models reason or the reasoning efficiency their models still burn tokens like no other models on earth I feel like the RL they use is still pretty primitive while the memory improvements like KV cache and stuff for hyperscaling has been the focus

As promised, here are some much harder tests. This will be a thread. I think Gemini wins here, actually, because it edited the image more precisely. GPT technically got 1 more right, I think. But it also added more things that weren't there originally than Gemini (although Gemini also added 1 extra set of lines that was never there and also removed 1 line that was there). GPT added and removed and repositioned so many things. It's embarrassing, frankly. The max score you can get here is 19. Gemini got 12 right - 2 completely hallucinated lines = 10. GPT got 13 right - 6 completely hallucinated lines = 7. Really, this is embarrassing for both models. This is literally middle school biology, which they know at, like, Nobel Prize level, and a simple image-editing task, which they also should be good at. TLDR: Gemini got 1 less label right but made WAY fewer editing errors and blatant hallucinations. 1/N