I'm impressed by #Google's #Gemini 1.5 model. It can directly answer questions from an entire book thrown at it. The amazing part is that you don't need to read the book anymore. It greatly improves the efficiency of querying reference books.
The explosion of ChatGPT and the fall of Bard at the press conference may reflect, to some extent, the huge difference between scientific research for reviewers and technical products for users. #ChatGPT#Bard
The manipulation of deep layers in a CNN, such as neural style transfer, is in a sense similar to "playing" (filtering) the Fourier plane in an optical imaging system. (2/2)
Random thoughts today: it feels like there might be some similarity between Fourier optics (e.g., a 4-f system) and a convolutional neural network (CNN) that takes images at both the input and the output (e.g., a U-net). (1/2)
@sama ChatGPT sometimes answers with wrong programming code and wrongly predicted outputs, eg on PyTorch. Wrong output should be easily eliminated, no?