This post shows how Bayesian updating works when we observe new evidence. We compare four different prior beliefs about the probability of success of some event. Priors vary along two dimensions: Location – what probability do we believe a priori? Strength – how strongly do we believe it? (number of pseudo-observations encoded in the Beta prior)

DeepSeek-R1 is an interesting model. It’s generally quite willing to discuss a range of topics, although because it’s a thinking model, the outputs can be a little harder to get into a structured format. In this post, I show how to use langchain and Pydantic to get structured output from DeepSeek-R1.

In this post I’m going to build off the last post on Working with US Census Bureau Data and discuss how to visualize it. That post walked through working with the Census Bureau’s API, so in this post I’ll skip those details.

I found that the US Census API is difficult to work with and even LLMs don’t provide working code for it. So I thought it might be helpful to share some techniques that did work. In this post, I’m going to focus on both raw API calls and the Python wrapper.

Google Cloud authentication can be confusing. This post explains how Application Default Credentials (ADC) work and how to fix common authentication errors.

In this post, let’s visualize the internals of a transformer model. These visualizations reveal some interesting patterns that can help us understand how well the training is going.

The dis module is a great tool for understanding how code runs. While I mainly use it out of curiosity, it can also be valuable for optimization and debugging. The module allows you to translate your Python code into bytecode—a low-level, intermediate representation of your Python code. By examining bytecode, programmers can glimpse the Python interpreter’s view of their code, shedding light on performance characteristics and operational behaviors that aren’t apparent at the source code level.