QR: Gemini 3 knowledge seems fresher (can reference things from September without RAG). It now answers correctly a few questions previously failed on 2.5 Pro (but succeeded on GPT-5). Haven't yet found a problem that GPT-5 fails but Gemini 3 succeeds. Will take a few days.

https://t.co/uEvBQfGfmU

M5 is as fast as M2 Max.

That's all.

Once we started to talk about deployment-focused frameworks, there are too many to be enumerated. Many of them still active today (look no further than ncnn, still going strong a decade later! Also onnx..). They brought a lot of sweat and tears to the table, but not modernity.

llama.cpp: similar design backbone to Caffe2 (highly likely just a coincidence). Brought quantization techniques and one-click deployment to mass.

mxnet: started similarly to Caffe, mid-way changed to be similar to TensorFlow. Ran out of steam when trying to be more like PyTorch.

Caffe2: similar design as TensorFlow, but tries to be a bit more developer-friendly.

Here are some side-quests: Chainer: more performant Theano, I honestly don't know why it didn't succeed but PyTorch did.

There are many more frameworks today, that brings nothing new beyond PyTorch offers (will list these "side-quest" later). The only novelty to me since PyTorch is the single-coordinator pattern to launch large-scale single training run.

PyTorch: eager execution, but no memory-leaks and implicit asynchronous kernel launches gave good-enough performance with good debugging experience.

Keras: modularize your complex neural networks with side-effect-free functions.

TensorFlow: popularize the idea you can use DAG to represent your neural networks, gave people hope that frameworks can do crazy optimizations (they cannot).

LuaTorch: eager execution. Ran your neural network with a script s.t. you can debug internals.

Theano: autograd.

Caffe: popularize im2col op to do CNN such that your shining GTX 980 can use cuBLAS ran CNN without writing your own kernel. Also popularized modular layer-based neural network construction, which is slightly easier than cuda-convnet.

cuda-convnet: show-case it is possible to run convolutional neural networks at 10x~20x speed as CPU at that time with GTX 780.

Modernity and deep learning frameworks since 2012. There are many deep learning frameworks since the renaissance, but only some of them brought new ideas (or brought back old ideas in significant way, as someone might say there is no new ideas in DL). Here is my list:

Notice anything missing here?

(The matmul (we care about) are 4096x3072x3072 / 4096x12288x3072 / 4096x3072x12288, and on M5 (iPad Pro), it tops ~12.5 TFLOPs TFLOPs w/ NA).