PR by Xuan-Son Nguyen for `llama.cpp`: > This PR provides a big jump in speed for WASM by leveraging SIMD instructions for `qX_K_q8_K` and `qX_0_q8_0` dot product functions. > > …
A thing I’ve noticed with deepseek is that it operates in a very system-oriented manner (it carefully plans out how to answer your question when you use thinking mode and it’s actually quite interesting) whereas chatgpt just tells you how long it “thought” and ultimately regurgitates an output that it is statistically likely. So we actually get to see a bit of the black box in my view
A thing I’ve noticed with deepseek is that it operates in a very system-oriented manner (it carefully plans out how to answer your question when you use thinking mode and it’s actually quite interesting) whereas chatgpt just tells you how long it “thought” and ultimately regurgitates an output that it is statistically likely. So we actually get to see a bit of the black box in my view
ChatGPT o1 hides it’s chain of thought so you don’t even really know what the ‘reasoning’ is
yeah it’s fascinating to see how the sausage is made