All sections
Latest
Everything published, newest first.
The log
Entries 1–20 of 93
1 entry
1 entry
Language Models Memorize More as They Grow, as Text Repeats, and as the Prompt Lengthens
Setting: a language model is trained on a huge pile of text, and sometimes it can reproduce parts of that text word for word.Quick thoughtprivacylanguage-models1 entry
One Number Can Fix an Overconfident Classifier: Temperature Scaling
Setting: a classifier outputs a probability for its answer. If it is calibrated, the answers it gives 90% confidence are right about 90% of the time.Quick thoughtcalibrationevaluation1 entry
Why Attention Divides by the Square Root of the Key Dimension
Setting: in a transformer, attention lets each token weigh every other token. For each token it computes a query vector q, compares it with a key vector k for…Quick thoughttransformersattention1 entry
Sample Many Reasoning Paths, Then Vote on the Answer
Setting: chain-of-thought prompting asks a language model to write its reasoning step by step before answering.Quick thoughtreasoningdecoding1 entry
For a Fixed Compute Budget, a Smaller Model Trained on More Tokens Can Win
Setting: training a large language model costs a fixed amount of computing, and you choose how to split it between a bigger model (more parameters) and more…Quick thoughtscaling-lawspretraining1 entry
LoRA: Train Two Thin Matrices Instead of One Giant One
Setting: fine-tuning adapts a pretrained model to a task by updating its weights. For a very large model, updating all of them means storing a full copy per…Quick thoughtfine-tuningtraining1 entry
Why Serving a Language Model Runs Out of Memory Before It Runs Out of Compute
Setting: a language model writes one token at a time. To avoid recomputing everything it has already read, it stores two vectors, a key and a value, for every…Quick thoughtinferenceserving1 entry
The DPO Loss Is One Number: How Much More You Raised the Winner Than the Loser
Setting: to make a language model prefer better answers, you collect pairs: a chosen answer and a rejected answer for the same prompt.Quick thoughtpost-trainingpreference-learning1 entry
A Code Benchmark Is Only as Strict as Its Tests
Setting: code-generation models are usually scored by running the programs they write against unit tests.Quick thoughtevaluationcode-generation1 entry
A Verifier That Checks Each Step Beats One That Checks Only the Final Answer
Setting: a language model can solve a math problem by writing its reasoning step by step, and you can raise its accuracy by sampling many solutions and…Quick thoughtreasoningverification1 entry
Forward KL Covers, Reverse KL Commits: Which Direction You Distill In
Setting: distillation trains a small student model to imitate a large teacher. The loss that compares them is usually KL divergence, a measure of how far one…Quick thoughttrainingdistillation1 entry
Optimize Against a Reward Model Too Hard and the True Reward Falls
Setting: to train a language model on what people prefer, you fit a reward model, a second model that predicts which of two answers a person would pick, and…Quick thoughtalignmentreward-models1 entry
With Only a Handful of Runs, Report an Interval and Use the Interquartile Mean
Setting: in reinforcement learning, a run is one training attempt of an algorithm from scratch, and results vary from run to run.Quick thoughtevaluationreinforcement-learning1 entry
Two Models Differ by a Point: Which Significance Test Do You Run?
Setting: you evaluate two models on the same test set, and one scores a point higher. Is that a real difference, or luck in which examples ended up in the…Quick thoughtevaluationstatistics1 entry
Read the Conceptual Reasoning Fellow Posting Before You Apply
A posting tells you what a role needs, and it often tells you more about who the role is for than a job title does.Quick thoughtanthropicfellowships1 entry
Measure Human Agreement Before You Trust an LLM Judge
An LLM judge cannot give a stable quality signal if the people who wrote the rubric disagree about which model outputs should pass.Quick thoughtopenaievaluation1 entry
Review a Model Upgrade: Count the Regressions
A new model snapshot can raise your eval score and still break items that used to pass. OpenAI's prompt engineering guide says it directly: "Even different…Quick thoughtopenaievaluation1 entry
Check a Codex Fix Against the Original Failing Case
Your eval harness scores cases with an empty reference answer as correct, which inflates accuracy by the share of such cases.Quick thoughtcodexcode-review1 entry
Give an Unattended Codex Run a Finish Line You Can Check
"Keep improving the eval harness" has no finish line, so an unattended run can wander until something outside it stops it.Quick thoughtcodexautomationContinued on the next page.
To browse rather than follow: topics, all articles and all notes. To learn in order: courses and the book.
