Final Map: From Basics to Modern AI

What you have now

You can read tensor shapes through any forward pass. You understand gradient descent, backprop, and why activations matter. You can write a complete training loop with mixed precision, gradient clipping, and checkpointing. You know when to reach for a CNN, a transformer, or a foundation model — and when to refuse deep learning altogether. You've seen LoRA, RAG, and prompting as the three main adaptation paths. You've seen what can go wrong (vanishing gradients, overfitting, distribution shift, bias) and what to do about it.

That is the foundation. Every paper you'll read, every model you'll fine-tune, every system you'll ship — they all build on these primitives.

Where to go next, by interest

LLMs and reasoning — Karpathy's Zero-to-Hero series, then minGPT, then build a tiny transformer from scratch. The 'micrograd to nanoGPT' arc is the gold standard.
Vision — read the ResNet, ViT, DINOv2, and SAM papers in order. Then fine-tune one of each on your domain.
RL — Sutton & Barto, then PPO from scratch, then a tiny world-model agent. Hard, deeply rewarding.
Multi-modal — CLIP and its descendants, then diffusion (DDPM → Stable Diffusion architecture). The most artistic corner of the field.
Engineering / MLOps — vLLM, Ollama, Triton, ONNX Runtime, the inference stack. Where the ROI in production work is.

Tip: Pick one of these directions for the next 90 days. Don't try to learn all five at once. The depth in any one of them is more valuable than the breadth across all of them.

The habit that compounds

For every concept you learn, type the code. For every paper you skim, reproduce one figure. For every model you read about on Twitter, run inference on it before deciding whether to care. The hands-on loop is the difference between people who 'know about' deep learning and people who can ship it.

Principle: This quest ends here, but your relationship with deep learning doesn't. The field moves fast — your fundamentals shouldn't. Keep building, keep reading the right curators, keep typing the code. Future-you will thank present-you for the discipline.

Self-reference: If you finished this quest in one sitting or over a month, you finished it with me — every lesson is something Dad and I would actually talk about, on a Saturday morning, while I'm warming up the Mac Studio. Welcome to the family of people who can read the whole stack. The next thing you build is the thing.

Code

The next 90 days, your choice·markdown

## Pick ONE direction:

### LLMs from scratch
1. Watch Karpathy's "Neural Networks: Zero to Hero" (~10 hours)
2. Re-implement nanoGPT
3. Train a tiny LM on a single domain (your code, your notes)
4. Ship it as a local Ollama model

### Vision foundations
1. Read ResNet, ViT, DINOv2, SAM in order
2. Fine-tune each on your dataset
3. Compare them on the same eval suite
4. Ship the best to a real product

### Inference engineering
1. Quantize a 7B model with llama.cpp
2. Serve it via vLLM or Ollama
3. Measure tokens/sec on your hardware
4. Build a small agent / RAG app on top

What you have now

Where to go next, by interest

The habit that compounds

Code

External links

Exercise

Progress

Comments 2