NVIDIA's 2017 AI Summer School, in the form of preserved notebooks
2017 was the last year the deep-learning field had time to teach itself from first principles before the LLM era flattened everything into one application. The "Attention Is All You Need" paper had been out for about three months. ResNets, dense and convolutional, were the workhorses. Sequence-to-sequence with LSTM was the production-grade machine translation architecture. CUDA was thirteen years old and finally widely understood as a general-purpose computing platform rather than a graphics quirk.
NVIDIA's Summer School that year ran in this period of relative breathing room. The curriculum reflected it — broad rather than deep, principles-first, with a healthy emphasis on the hardware side because in 2017 if you didn't understand the memory model you couldn't train anything interesting in finite time. The notebooks covered convolutional architectures, recurrent architectures, autoencoders, GANs, a section on reinforcement learning, and CUDA at the metal. Each topic got two or three notebooks and a problem set.
I attended, I worked through the notebooks, I kept them. The repository is 184 megabytes of Jupyter content from that course, preserved approximately as I left it. The README is empty because at the time the README was the course manual, not the repo. There's no commentary I've added since.
The reason I haven't pruned it: this is what I looked like as a learner of deep learning at that moment in the field's history. The notes in the margins of those notebooks are the questions I had then. Many of them — about gradient flow, about why batch normalization works, about the relationship between RNN depth and gradient vanishing — are questions the field has since shrugged off as understood. A few — about why attention generalizes so much better than they expected — were ahead of their time by exactly long enough.
There's a certain self-respect in not editing the historical record. The notebooks are intact. The questions in the margins are the ones I asked at 27 or 28 or whatever I was. The mistakes are still in them. So are the moments where I figured something out that nobody had told me yet.
This is not a project. It's a preservation. Eight years later, this is what the substrate of how I think about machine learning was first poured into. The PQC notebook I wrote in 2025 has descendants of techniques I first saw in 2017. So does the autoresearch quantum ratchet, which is a 2026 reformulation of an idea I first encountered in this exact corpus.
What becomes possible: when someone asks "where did you start with deep learning?" I have a precise answer and can point at it.