EssayAugust 30, 2026
What does it take to discover science?
Language models are already good at searching within an existing scientific description. The harder thing is deciding when that description itself has to change, and the training record we give them, plus the objectives we score them on, may be exactly the wrong ones for that.
ReadResearchJuly 2, 2026
Sampling is sampling: an outsider's audit of LLM test-time compute
The benchmark-winning way of making an LLM think harder samples the wrong distribution and silently destroys its own confidence signal. With an interactive replay of real runs, a seventy-year-old fix, and a dictionary between cosmology and test-time compute.
ReadEssayComing soon
Why I think simulation-based inference is the right tool for next-generation cosmology
A short essay on why we should stop pretending we have analytic likelihoods, and what we gain when we let neural density estimators do the work.
ReadMethodComing soon
From Kaiser-Squires to plug-and-play: a tour of mass-mapping algorithms
What changes when you swap a hand-tuned regulariser for a learned denoising prior, and why uncertainty quantification stops being optional.
Read