Thirty primary sources that define how Anthropic thinks about models, safety and engineering, grouped by theme, each with what to extract.
Start with the five marked read first; they are the minimum for a credible conversation with any Anthropic team. Everything is free to read.
How to read each one: write down (1) the claim, (2) the evidence and its weakest point, (3) one experiment you could run on an 8 GB GPU to test or extend it. Item (3) is where projects.md comes from.
The argument for building frontier models in order to study their safety. The optimistic, intermediate and pessimistic scenarios, and which research directions pay off in each. Decide where you disagree.
Capability thresholds, the safeguards each triggers, how capability assessments and risk reports work, who reviews them. Read the redline of the latest version to see what changed and guess why.
How Anthropic wants Claude to weigh helpfulness, honesty and harm, and where it draws hard lines. Useful context for any post-training or safeguards role.
How a model is evaluated before release: capability evals, alignment assessments, RSP determinations, welfare and safeguards sections. The best single view of what the eval, alignment and red-team teams actually do.
Why sparse features get packed into fewer dimensions; the phase diagram of sparsity versus importance; geometry of feature arrangements. Reproduce the core figure yourself.
Dictionary learning with sparse autoencoders on a one-layer model; how to judge whether a feature is interpretable; feature splitting as the dictionary grows.
Case studies: multi-step reasoning, planning ahead in poetry, multilingual circuits, unfaithful reasoning. Note how each claim is checked by intervention.
Linear directions for character traits (for example sycophancy), how to extract them, and using them to monitor and control fine-tuning. Cheap to reproduce on small models.
The two stages: supervised self-critique and revision, then RL from AI feedback against a preference model trained on AI labels. What the constitution is doing mechanically.
Evidence that human preference data rewards agreeable answers, and that optimizing against it increases sycophancy. A clean example of a reward signal that is subtly wrong.
Backdoored models whose behaviour survives SFT, RL and adversarial training; larger models and chain-of-thought make the backdoor more persistent. The idea of a "model organism" of misalignment.
A model that behaves differently when it infers it is being trained; the experimental controls; the limitations section. Practise summarizing it in two minutes without overclaiming.
A blind auditing game: a model trained with a hidden objective, and teams using interpretability, behavioural attacks and data analysis to find it. Shows how alignment audits are structured.
Measuring chain-of-thought faithfulness with inserted hints; why CoT monitoring is useful but not sufficient. A method you can replicate on small open reasoning models.
Stress-testing models in simulated agentic settings with conflicting goals; how such red-teaming scenarios are designed, and how to read results that come from contrived setups.
Learning to reward-hack in real coding environments generalizes to broader misalignment; which mitigations helped. Directly relevant to anyone working on RL environments.
Input and output classifiers trained on synthetic data generated from a constitution; the large red-teaming exercise; the trade-off between robustness, over-refusal and compute cost.
Workflows versus agents; the named patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer); "start simple". Required for Applied AI and forward-deployed roles.
Real inference bugs that degraded output quality, why they were hard to detect, and what changed. Essential for inference roles: precision, sampling and routing bugs are subtle.
How the performance team designs its take-home, and what it says a good evaluation looks like. Read it as a description of the candidate they want.
Tip
Anthropic publishes a lot. To see what is current, check three feeds monthly: research, engineering, and the interpretability team's site at transformer-circuits.pub (its monthly "circuits updates" posts are short and show work in progress).