Roadmap.

Master list · curated, not exhaustive

Resources

A short list you'll actually use beats a long list you won't. For any given phase, pick one primary resource and go deep; the others are references for when the primary leaves a gap. Collecting resources is procrastination; finishing one is progress.

Hands-on courses & code — the spine, do these

Karpathy — Neural Networks: Zero to Hero1–2The canonical from-scratch course: micrograd, makemore, build-GPT. Do the exercises.
3Blue1Brown — Neural Networks series1The best visual intuition for nets, gradients, and backprop. Watch when a concept won't click.
nanoGPT2Clean minimal GPT to study + train. Deprecated & frozen since Nov 2025 — still the shortest complete read; nanochat is the successor.
nanochat2nanoGPT's official successor: the full modern stack end-to-end (tokenizer→pretrain→SFT→eval→UI).
The Illustrated Transformer (Jay Alammar)2Visual walk through attention; read alongside the code.
Raschka — Build an LLM (From Scratch)2Thorough code-first book + repo (incl. GPT-2→Llama, Qwen-from-scratch).
fast.ai — Practical Deep Learning0–1Top-down, build-first DL course. Free.
Dive into Deep Learning (d2l.ai)1–3Free interactive textbook; use as reference.
Stanford CS336 — Language Modeling from Scratch3–4The flagship systems course: tokenizer, FlashAttention2-in-Triton, distributed training, data, SFT+RL. Do the assignments.
Hugging Face TRL4Practical SFT / DPO / GRPO toolkit.
ARENA (Callum McDougall)5Alignment/research-engineering curriculum: DL fundamentals → transformers + mech interp → RL.
TransformerLens5 (interp)The mech-interp library (induction heads, IOI in GPT-2 small).
modded-nanogpt3Speedrun repo; the commit history is an efficiency masterclass.

Books & long-form

The RLHF Book — Nathan Lambert4Authoritative post-training/RLHF/RLVR guide. Free online.
Mathematics for Machine Learning — Deisenroth et al.1Math reference (look up, don't read linearly). Free PDF.
Deep Learning — Goodfellow, Bengio, Courville1–3The classic theory reference; dip in, don't read cover-to-cover.
Andrew Ng — Machine Learning Specialization1Gentle, rigorous ML fundamentals if you want more hand-holding. (Coursera)

Canonical papers — read, don't just cite

GPT-1 — Improving Language Understanding by Generative Pre-Training2The generative-pretraining lineage (OpenAI report, not on arXiv).
BERT2Bidirectional pre-training; the other 2018 transformer milestone.
Chinchilla — Training Compute-Optimal LLMs3Scaling laws; "small but well-trained."
FlashAttention3Memory-efficient attention; the systems mindset.
GPTQ / AWQ3Post-training quantization.
Distilling the Knowledge in a Neural Network (Hinton) + Sequence-Level KD3Distillation foundations.
TinyStories → Phi-1.5 → Phi-3 → SmolLM3The data-quality thread; small-model recipes.
InstructGPT / RLHF + DPO + a GRPO/RLVR paper4The post-training lineage.
DeepSeek-R14 / RL-3RL for reasoning (RLVR); the GRPO breakthrough.
Anthropic — Transformer Circuits / induction heads5 (interp)Mech-interp foundations.

People & newsletters — signal, low noise

Raschka + Han Lab + the Hugging Face science blog cover ~80% of what matters week-to-week.

Sebastian Raschka — Ahead of AIallHighest signal; annual paper reading lists.
Lilian WengallDeep technical explainers.
Nathan Lambert — Interconnects4Post-training / RLHF pragmatics.
Neel Nanda5 (interp)Mech interp; "Concrete Steps to Get Started."
MIT Han Lab3/5The academic home of efficiency: AWQ/SmoothQuant/TinyChat.

Communities, compute, venues

EleutherAIallDiscord #research + the SOAR mentored program. The highest-ROI single move.
Hugging FaceallModels/datasets/Spaces; Daily Papers for visibility.
ComputeallFree Colab/Kaggle to start; Vast/RunPod community tiers for cheap rentals; granted compute via TPU Research Cloud / EleutherAI.
Venues for independents5ICLR Blog Posts track; NeurIPS ENLSP workshop; ML Reproducibility Challenge. Not the main-track lottery.

How to choose · so you don't drown