Reading paths through AI research
Follow a topic across methods, datasets and the evidence behind a result.
Reading order
- Reading Transformer papers: from attention to modern language-model architectures
- Reading scaling-law papers: parameters, data, compute and emergent behavior
- Reading instruction-tuning papers: supervised fine-tuning and instruction following
- Reading preference-optimization papers: reward models, DPO, GRPO and preference data
- Reading retrieval-augmented learning papers: from RAG to memory-augmented models
- Reading reasoning-model papers: chain of thought, self-consistency, supervision and search
- Reading tool-use papers: from ReAct to verifiable tool calls
- Reading multimodal LLM papers: vision encoders, alignment data and reasoning
- Reading vision-language-action papers: from embodied agents to robotics foundation models
- Reading AI for science papers: molecules, proteins, materials and automated experiments
- Reading interpretability papers: representations, circuits, probes and causal interventions
- Reading safety and alignment papers: red teaming, abstention and controllable generation
- Reading evaluation papers: open-ended tasks, human judgments and contamination
- Reading data-centric AI papers: selection, deduplication, synthesis and quality
- Reading efficient training papers: LoRA, QLoRA, MoE, checkpointing and optimizers
- Reading efficient inference papers: KV caching, speculation, quantization and serving
- Reading open-model technical reports: model cards, training data, evaluations and limits
- Reading rebuttals and peer reviews to assess a paper’s evidence
- From ten papers to a research proposal: question, hypothesis, method and experiment
- After a month of AI papers: building a lasting reading practice