Section de la base de données · 12 fiches
Recherche
The papers the modern stack is built on — from the transformer to scaling laws, RLHF and reasoning models.
Notes de terrain
Une douzaine de PDF sous-tendent mille milliards de dollars d'infrastructure. Ce sont les articles où les idées sont apparues pour la première fois — la plupart sont étonnamment accessibles, et chaque fiche ici vous dit ce qui compte, sans les équations.
Commencez par ces trois-là
AI
Attention Is All You Need (2017)
Huit auteurs, une architecture, tout ce qui a suivi.
TC
Training Compute-Optimal LLMs — Chinchilla (2022)
La correction qui a changé la façon dont chaque modèle est entraîné.
DR
DeepSeek-R1 — Reasoning via RL (2025)
La recette ouverte du raisonnement — publiée pendant que les concurrents gardaient la leur secrète.
Foundational papers
12
AI
Attention Is All You Need (2017)
The eight Google authors who invented the transformer.
A(
AlexNet (2012)
The ImageNet result that started the deep-learning era — on two gaming GPUs.
CA
Constitutional AI (2022)
Anthropic's method: models critique themselves against written principles.
DR
DeepSeek-R1 — Reasoning via RL (2025)
The open recipe for reasoning models — a Nature cover, from a hedge fund's lab.
DD
Denoising Diffusion Probabilistic Models (2020)
The DDPM paper behind every image and video generator.
F(
FlashAttention (2022)
The kernel trick that made long context affordable.
HA
Highly Accurate Protein Structure Prediction (2021)
AlphaFold 2 in Nature — the result that won a Nobel.
I—
InstructGPT — RLHF at Scale (2022)
How raw language models became assistants people can use.
LM
Language Models are Few-Shot Learners (2020)
The GPT-3 paper: capability from scale, tasks from prompts.
LL
LoRA: Low-Rank Adaptation (2021)
Fine-tuning for the GPU-poor — tiny adapters instead of full weights.
SL
Scaling Laws for Neural Language Models (2020)
Kaplan et al.: loss falls as a power law in compute — the investment thesis, quantified.
TC
Training Compute-Optimal LLMs — Chinchilla (2022)
DeepMind's correction: most models were undertrained on data.