I study how AI is changing knowledge work.

Shows that language models can agree and still be wrong, so truthfulness needs real verification rather than voting or self-consistency.

Introduces RAMP, a way to score how ready repositories are for coding agents, and links stronger setup with fewer quality problems after adoption.

Studies how duplicated training data affects models differently as they scale, showing that repetition can change performance in ways averages hide.

Revisits V-information and shows it can behave strangely under realistic modeling limits, making it risky to treat as a simple information measure.

Shows that repeating data inside a training set can damage language models, even when the repeated examples look harmless at first.

Explains how generative benchmarks can be contaminated by test data, making models look better than they really are.

Tracks how weak sampling evaluations passed through major papers, turning shaky claims into later assumptions that other work built on.

Rechecks claims that LLM answers collapse into one narrow style, and finds much more diversity across topics, models, and prompts.

Creates a benchmark for testing whether coding agents can produce Lean 4 proofs that verify software end to end.

Tests whether an agent's Lean proof actually matches the original Python code, exposing proofs that are correct but about the wrong thing.

Proposes a competition where researchers predict evaluation results before they are run, making AI benchmarks harder to game after the fact.

Argues that machine learning conferences need a formal place for critiques and corrections, so important mistakes can be reviewed openly.

Measures how much LLM code-review outputs vary across runs, which matters when teams want reliable and repeatable review results.

Reexamines min-p sampling and finds that its claimed benefits are not supported by the available evaluation evidence.

Shows that models can predict expert code-review scores, making it cheaper to estimate code quality across large datasets.