In addition, I don't consider text to image very useful, because when I ask mid journey to draw "a pikachu playing baseball with a squirtle", it draws a very nice blue faced pikachu, which is still trash to me
21
things SD can do and mid journey can't:
1. inpainting
2. finetuning (including LoRA)
3. batch generation
4. controlnet
5. It can even do videos
1
2
92
Mid journey may be better than Stable Diffusion on the most basic text to image. But SD can do a thousand different things that Mid journey just can't, thanks to being open source.
2
2
45
I think "cannot cook without rice" is more like training without data, GPUs are more like cookers
3
293
I agree that AI can be dangerous, but not because they are too powerful. AI is not dangerous if you use it as a reference, but it can be very dangerous if you rely on it for making decisions.
301
矩阵很重要是因为这是一个很好的抽象层，搞算法的人用矩阵来描述算法，而搞硬件的人只要集中精力来优化矩阵运算的效率。不管CNN RNN还是transformer，里面计算复杂度最高的部份都是矩阵乘法。
93
硬要说的话back propagation也不是本质的，它只是一种用空间换时间的方法来优化gradient computing. 事实上在一些大模型里为了省内存就不是完全按照back propagation 来计算gradient.
82
非常同意。简单举几个例子：command/observer模式本质上就是解决那个年代c++/java没有lambda的缺陷。abstract factory本质上是解决c++/java没有duck typing, 在动态语言里就没什么用。iterator在有generator/iterator的语言里也就没有作为模式存在的意义
2
I think this is due to the fact that everything has only one version in google's internal repo. However externally it's not possible to break a library and fix everyone else's code.