Jordan Nafa
@ajordannafa.com
Bayesian Statistician and Data Scientist in the Gaming/Entertainment Industry | Bayesian Statistics, Causal Inference, R, Python, Stan, Decision Theory, Guitar | Former Political Scientist
JFC, the hell are you compiling? In both Stan and PyMC my compile times are like 10-15 seconds max, even for complex models
January 11, 2026 at 4:49 AM
JFC, the hell are you compiling? In both Stan and PyMC my compile times are like 10-15 seconds max, even for complex models
The fight is not to save democracy in America, but rather to Democratize America to begin with.
January 11, 2026 at 1:07 AM
The fight is not to save democracy in America, but rather to Democratize America to begin with.
Even if an individual state for some reason chose not to participate (not clear there's any mechanism for this outside of secession i.e., the election of 1864), the federal election still happens.
January 10, 2026 at 8:36 PM
Even if an individual state for some reason chose not to participate (not clear there's any mechanism for this outside of secession i.e., the election of 1864), the federal election still happens.
If someone lacks the patience to wait 5 extra seconds, I have extreme doubts they're capable of doing statistics with any shred of rigor in the first place 🤷♂️ (my PhD advisor was one of those people lol)
January 10, 2026 at 2:15 PM
If someone lacks the patience to wait 5 extra seconds, I have extreme doubts they're capable of doing statistics with any shred of rigor in the first place 🤷♂️ (my PhD advisor was one of those people lol)
I think that context is helpful because it clarifies we're talking about getting past the learning barrier rather than the cost of the procedure itself.
January 9, 2026 at 11:51 PM
I think that context is helpful because it clarifies we're talking about getting past the learning barrier rather than the cost of the procedure itself.
Realizing it's 2026 now, but my point stands 🤷♂️
January 9, 2026 at 11:39 PM
Realizing it's 2026 now, but my point stands 🤷♂️
Saying "I'm only estimating a simple difference in means, I don't need Bayes" is like saying "I have this nice cordless circle saw, but I'm going to use a plain old hacksaw to cut this board instead." Like, you do you, but I'm totally going to judge you for it 🤷♂️
January 9, 2026 at 11:37 PM
Saying "I'm only estimating a simple difference in means, I don't need Bayes" is like saying "I have this nice cordless circle saw, but I'm going to use a plain old hacksaw to cut this board instead." Like, you do you, but I'm totally going to judge you for it 🤷♂️
A Bayesian t-test shouldn't exist, because it's basically always preferable to just formulate it as a GLM 🤷♂️
January 9, 2026 at 11:20 PM
A Bayesian t-test shouldn't exist, because it's basically always preferable to just formulate it as a GLM 🤷♂️
"Infer a long-term outcome based on short-term signals so you don't have to wait for your actual metric to mature" is mostly just people trying to find a way to get free lunch and developing very sophisticated ways to obfuscate the fact that their question isn't reliably answerable 🤷♂️
January 9, 2026 at 5:01 AM
"Infer a long-term outcome based on short-term signals so you don't have to wait for your actual metric to mature" is mostly just people trying to find a way to get free lunch and developing very sophisticated ways to obfuscate the fact that their question isn't reliably answerable 🤷♂️
Not sure what you're referring to as "proxy metrics" because that's kind of a broad and loaded term but I am extremely skeptical of how well the literature on surrogate metrics generalizes beyond very specific circumstances since it's analogous to mediation and the assumptions are mostly nonsense.
January 9, 2026 at 4:46 AM
Not sure what you're referring to as "proxy metrics" because that's kind of a broad and loaded term but I am extremely skeptical of how well the literature on surrogate metrics generalizes beyond very specific circumstances since it's analogous to mediation and the assumptions are mostly nonsense.
As a professional dev, one-shot vibe coding from scratch is still absolute trash for most of what I do but if you're using it to add/modify an existing codebase like adding unit tests/validation checks, clearly defined extensions it works pretty darn well.
January 9, 2026 at 1:40 AM
As a professional dev, one-shot vibe coding from scratch is still absolute trash for most of what I do but if you're using it to add/modify an existing codebase like adding unit tests/validation checks, clearly defined extensions it works pretty darn well.
Non-methods academic code is notoriously bad and borderline unusable in most fields, so to the extent LLM-based vibe coding applies best practices from software engineering that your average academic would otherwise be oblivious to, it makes sense it would lead to improved quality
January 9, 2026 at 1:36 AM
Non-methods academic code is notoriously bad and borderline unusable in most fields, so to the extent LLM-based vibe coding applies best practices from software engineering that your average academic would otherwise be oblivious to, it makes sense it would lead to improved quality