Reading AI research
How to read a machine learning paper well enough to tell a real result from a well-presented one, and what the field's landmark papers actually say.
Almost everything you will read about a new model is a summary of a summary. Somewhere at the bottom of that chain there is a paper, and the paper usually says something narrower, more conditional and more interesting than what reached you. The gap between those two things is where most confident wrong opinions about AI come from.
Closing that gap is a skill, and it is a smaller skill than it looks. You do not need to follow the mathematics to check whether a baseline was chosen fairly, whether the ablation table supports the story the abstract tells, or whether a benchmark number means what the headline says it means. Those checks are mechanical, they take minutes, and almost nobody performs them.
This cluster teaches the checks, then applies them to the papers the field is actually built on. The goal is scepticism rather than cynicism: most papers are honest, many results are real, and the point of reading carefully is to be able to tell which.
How to Read an AI Paper
A three-pass reading method adapted for machine learning papers, with a real paper taken through all three passes.
6 min read
Attention Is All You Need, Explained Section by Section
A walk through the 2017 transformer paper in the order it is written, including the ablation table most summaries skip.
5 min read
The GPT Papers: What Each One Introduced
The intellectual lineage from generative pre-training to the technical reports that stopped reporting, and what changed at each step.
5 min read
The Chinchilla Paper and Compute-Optimal Training
What compute-optimal training means, how the paper demonstrated it, and the three ways its conclusion is routinely misapplied.
5 min read
InstructGPT and the Birth of RLHF
The three-stage method that turned a text continuation engine into something that follows instructions, and the cost the paper reports for it.
4 min read
The LoRA Paper: Why a Simple Idea Had Such Large Consequences
The low-rank adaptation method, the hypothesis it rests on, and the one property that turned a memory optimisation into an ecosystem.
4 min read
Retrieval-Augmented Generation: The Original Paper
What RAG meant in the 2020 paper that named it, and how the term drifted into meaning something the paper never proposed.
4 min read
Chain-of-Thought: The Paper and What It Actually Found
The finding, the tasks it was demonstrated on, the scale condition attached to it, and the much larger claim it turned into.
4 min read
Constitutional AI: The Paper, Including the Parts That Get Skipped
The two-stage method for training a harmless assistant from a written set of principles, and the details summaries leave out.
4 min read
Toolformer and Learned Tool Use
A self-supervised method for teaching a model when to call an API, and why the industry took a different road to the same place.
4 min read
Sparse Autoencoders and the Interpretability Papers
Why neurons are not features, what a sparse autoencoder is trying to recover, and which parts of this line of work are settled and which are not.
4 min read
Benchmarks as Papers: Reading a Benchmark Critically
The construction choices inside a benchmark that decide its results, and how a number can be entirely true and still tell you nothing.
5 min read
Reproducibility in AI Research
The four different things people mean by reproducing a result, why machine learning fails at several of them, and how to weight a single unreplicated finding.
5 min read
Preprints, Peer Review and arXiv Culture
What posting to arXiv does and does not certify, what peer review actually catches, and how to weight a paper you cannot verify.
5 min read
Spotting Overclaiming in AI Papers
Ten signals that a paper is claiming more than it showed, with constructed examples, plus the things that look like red flags and are not.
5 min read
Ablations: The Section That Tells You the Truth
What an ablation table is for, how to tell a real one from a decorative one, and what its absence tells you about a paper.
5 min read
Following AI Research Without a PhD
A weekly routine that fits in two hours, a filtering method that survives the volume, and permission to ignore almost everything.
5 min read
The Papers Behind Modern Inference Optimisation
FlashAttention, PagedAttention and speculative decoding — three papers whose ideas are running behind every API call you make.
5 min read
Alignment Research: A Map of the Agendas
The main research agendas in AI alignment, the assumption each one is load-bearing on, and where they genuinely disagree.
5 min read
20 Papers That Explain Modern AI
An annotated reading list ordered by dependency, with the reason each paper is on it and what it is a prerequisite for.
5 min read
Other topics
- LLM fundamentals & architecture
- Tokens, tokenization & context windows
- Prompt engineering
- Reasoning models & test-time compute
- Multimodal AI: vision, audio, video
- RAG & retrieval
- Embeddings & vector search
- AI agents & tool use
- Structured output & function calling
- Fine-tuning & post-training
- Inference, serving & latency
- Evaluation, benchmarks & LLM-as-judge
- Observability & LLMOps
- Hallucination & failure modes
- LLM cost engineering
- AI security & prompt injection
- Privacy, compliance & data residency
- AI governance, policy & society
- Building reliable AI applications
- AI hardware, GPUs & compute
- Open-weight models & local inference
- AI for developers & coding agents
- AI in industry: vertical playbooks
- AGI, superintelligence, alignment & the long future
- Machine learning foundations
- NLP fundamentals & classical tasks
- Data engineering for AI
- Synthetic data & dataset curation
- AI product design & UX
- Search, ranking & recommendation
- Enterprise adoption & change management
- AI careers, skills & teams
- AI in science & discovery
- Robotics & embodied AI
- AI economics, markets & business models
- AI myths, hype & media literacy
- Context engineering
- Shipping AI features: patterns & anti-patterns