Right. While I appreciate the author's skepticism and diction (there is a lot of misleading terminology thrown around by the ML community), his points don't land.
In particular, he argues that there's no learning going on, but then says that there is "absorption" of statistical patterns going on. That's just nitpicking over semantics—to people in the field, the two phrases mean the same thing. The only difference is whether you anthropomorphize a piece of software.
The second place the author stumbles is that he makes the (quite grave) mistake you pointed out. The title insinuates that the network contains the "source dataset" itself. He has shown nothing of the sort by including the training logs in his "decompilation". That's like suggesting you have a Swift decompiler that can recover the exact source code from an optimized binary, but you actually require access to the pre-optimized LLVM IR.
In particular, he argues that there's no learning going on, but then says that there is "absorption" of statistical patterns going on. That's just nitpicking over semantics—to people in the field, the two phrases mean the same thing. The only difference is whether you anthropomorphize a piece of software.
The second place the author stumbles is that he makes the (quite grave) mistake you pointed out. The title insinuates that the network contains the "source dataset" itself. He has shown nothing of the sort by including the training logs in his "decompilation". That's like suggesting you have a Swift decompiler that can recover the exact source code from an optimized binary, but you actually require access to the pre-optimized LLVM IR.