chore: add pre-commit test

Justfile

@@ -0,0 +1,8 @@
+build:
+  bundle exec jekyll build
+
+
+test: build
+  bundle exec htmlproofer _site \
+      --check_external_hash=false \
+      --ignore-urls "/^http:\/\/127.0.0.1/,/^http:\/\/0.0.0.0/,/^http:\/\/localhost/,/fonts.googleapis.com/,/fonts.gstatic.com/,/linkedin.com\/in\/grzegorz-milka-55270490/,/twitter.com\/intent\/tweet/,/archive.ph/"

README.md

@@ -12,6 +12,12 @@ The blog is hosted on GitHub Pages at
 
 1. Install [rbenv](https://github.com/rbenv/rbenv) and
    [direnv](https://direnv.net/) and hook them into your shell.
+1. Install Lefthook.
+
+   ```bash
+   lefthook install
+   ```
+
 1. Allow `.envrc`.
 
    ```bash

_posts/2014-12-28-creating-ai-in-haskell.markdown

@@ -433,4 +433,4 @@ easily.
 [minmax]: https://en.wikipedia.org/wiki/Minimax
 [github]: https://github.com/gregorias/Pentago
 [outoftarpit]: https://github.com/papers-we-love/papers-we-love/blob/master/design/out-of-the-tar-pit.pdf?raw=true
-[whymatters]: https://github.com/papers-we-love/papers-we-love/blob/master/functional_programming/why-functional-programming-matters.pdf
+[whymatters]: https://github.com/papers-we-love/papers-we-love/blob/main/languages-paradigms/functional_programming/why-functional-programming-matters.pdf

_posts/2015-03-07-lessons-in-writing-and-testing-an-audio-modem-in-python.markdown

@@ -14,8 +14,8 @@ then I will describe how to combine Cython and profiling to build a flexible and
 efficient Python application. In the last section I will talk about pulseaudio
 and how it allows me to perform user testing without using an external device.
 
-Modulation and demodulation
-===========================
+## Modulation and demodulation
+
 The module that I am building encodes and decodes voice. The encoder has one
 input: a continuous stream of voice represented by audio samples. On the output
 it plays audio that is transmitted by an independent voice communicator. The
@@ -30,8 +30,8 @@ producing a seemingly random bytestring. This task includes a few challenges:
   voice so that it fits?
 * How can I efficiently demodulate the modulated signal?
 
-Basics of audio signals
------------------------
+### Basics of audio signals
+
 The sound that we hear is a result of a mechanical wave of pressure and
 displacement propagating through air. As such we can represent any sound as a
 wave. Wave being a continuous signal needs to be discretized to be representable
@@ -71,8 +71,8 @@ to the other. From fundamental frequencies we can produce sampling by simply
 calculating wave values at given points. To get the other way around we need to
 use the discrete Fourier transform.
 
-Discrete Fourier transform
---------------------------
+### Discrete Fourier transform
+
 Fourier transform is a general mathematical tool which transform a wave function
 into a function of fundamental frequencies. That is the resulting function maps
 each of its argument to its power inside the argument. The discrete version of
@@ -107,8 +107,7 @@ have twice that sampling rate. Since telephony was meant to transmit voice and
 human voice happens in a range up to 300Hz then 8000 sampling rate was more than
 enough.
 
-Modulating data
----------------
+### Modulating data
 
 We know enough to understand important trade-offs of signal modulation. Our task
 now is to encode binary data into a voice channel in such a way that the
@@ -138,8 +137,7 @@ before, but resigned due to lack of some libraries I would find useful. Now
 after overcoming my reluctance to FFI I would simply write such a wrapper for
 Haskell. End of diversion.
 
-Demodulation
-------------
+### Demodulation
 
 So far I have described how the voice is encoded into binary data and how then
 this data is modulated. We are now left with possibly the most difficult part of
@@ -260,8 +258,7 @@ voice demodulation. In the next section I will explain how Cython works and how
 it can greatly improve speed of Python for low cost of just some static type
 annotations.
 
-Cython
-======
+## Cython
 
 Profiling
 ---------
@@ -505,4 +502,4 @@ streams so that multiple inputs can be mixed.
 [hermes]: https://cs.nyu.edu/~jchen/publications/com31a-dhananjay.pdf
 [codec2]: https://www.rowetel.com/blog/?page_id=452
 [pycodec2]: https://github.com/gregorias/pycodec2
-[py3vsC]: https://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=python3&lang2=gcc
+[py3vsC]: https://archive.ph/Dk5Uq

_posts/2017-02-23-why-you-might-prefer-bayesian-ab-testing.markdown

@@ -109,17 +109,13 @@ The measure of unexpectedness is called **p-value**.
 It is the probability that our null hypothesis would produce the observed or more extreme value.
 In other words, the lower the p-value, the stronger the suprise.
 
-<figure>
-<img src="/images/2017/p_value.png" alt="A diagram explaing the concept of a p-value." />
-</figure>
+![A diagram explaing the concept of a p-value.](/images/2017/p_value.png)
 
 Same calculations may be used to compute a related concept - **confidence interval** of $$\hat{\theta}$$.
 Confidence interval is an interval that intuitively approximates the likely range in which the parameter $$\theta$$ really lies.
 In our example, the interval would be $$\hat{\theta} \pm 1.96 \hat{\sigma}$$, because $$\hat{\theta}$$ is normally distributed.
 
-<figure>
-<img src="/images/2017/normal_196.png" alt="A plot of a normal distribution with 95% confidence interval." />
-</figure>
+![A plot of a normal distribution with 95% confidence interval.](/images/2017/normal_196.png)
 
 A correct intepretation of a confidence interval is not that simple:
 
@@ -177,7 +173,9 @@ First we have to agree on a prior.
 Usually, we don't know much, so we might as well assign uniform probability distribution to $$\theta_A, \theta_B$$.
 By applying the Bayes rule to experimental data, which consists of numbers of successes and attempts in each variant, we'll get the posterior probability distribution of variants' parameters.
 In our case the posterior probability distribution is conveniently expressible by a [Beta distribution](https://en.wikipedia.org/wiki/Beta_distribution).
-If you'd like to understand what happens under the hood, then I reccommend an excellent introductory series ["using baseball statistics"](https://varianceexplained.org/r/empirical_bayes_baseball/).
+If you'd like to understand what happens under the hood, then I recommend an
+excellent introductory series “Understanding empirical Bayes estimation (using
+baseball statistics)”.
 
 With the probability distribution we can do all kinds of stuff we couldn't do with frequentist methodology.
 Like plot its density:
@@ -212,10 +210,8 @@ At any given time we only have probability estimates of the expected payoff, and
 Experiments need to continuously balance exploration (choosing alternative variants in order to learn more about payoffs) and exploitation (choosing the variant we believe to be the best) in order to minimize the expected loss - patient is assigned the worse treatment.
 
 Experiments using this approach are called [Multi-armed bandit experiments](https://en.wikipedia.org/wiki/Multi-armed_bandit)[^wiki_quote].
-The multi-armed bandit problem generalizes the outlined approach to multiple variants with unknown payoff.
-
-[Google Analytics uses multi-armed bandit as their experiment framework](https://support.google.com/analytics/answer/2844870?hl=en).
-Their [FAQ](https://support.google.com/analytics/answer/2847021?hl=en) does a good job of explaining the reasons why a Google analytics customer might prefer this method to a classical approach.
+The multi-armed bandit problem generalizes the outlined approach to multiple
+variants with unknown payoff.
 
 ## Summary
 
@@ -236,7 +232,7 @@ Also, I heartily recommend using Bayesian methods in your experiments, quite oft
 
 [gelmanblog]: https://andrewgelman.com
 
-[fharrell_contradiction]: https://www.fharrell.com/2017/01/null-hypothesis-significance-testing.html
+[fharrell_contradiction]: https://www.fharrell.com/post/nhst-never/index.html
 
 [wagenmaker_conf]: https://www.ejwagenmakers.com/inpress/HoekstraEtAlPBR.pdf
 

_posts/2023-03-25-how-does-plenary.async.run-run-asynchronously.markdown

@@ -5,9 +5,18 @@ date:   2023-03-25 11:00:00
 tags: neovim plenary coroutine
 ---
 
-[Neovim Plenary](https://github.com/nvim-lua/plenary.nvim)'s [`async.run`](https://github.com/nvim-lua/plenary.nvim/blob/253d34830709d690f013daf2853a9d21ad7accab/lua/plenary/async/async.lua#L104) seems to be doing an impossible trick. It starts a concurrent computation that runs even after the main script has finished and calls the provided callback once it's done. It seemed impossible to me, because Lua uses [non-preemptive coroutines](https://www.lua.org/pil/9.4.html). The caller needs to orchestrate when its coroutines can resume. So when the main script is done, there should be nothing that can resume remaining coroutines.
-
-This note is my investigation into why this is happening. Surprise is important for learning.
+[Neovim Plenary](https://github.com/nvim-lua/plenary.nvim)’s
+[`async.run`](https://github.com/nvim-lua/plenary.nvim/blob/253d34830709d690f013daf2853a9d21ad7accab/lua/plenary/async/async.lua#L104)
+seems to be doing an impossible trick. It starts a concurrent computation that
+runs even after the main script has finished and calls the provided callback
+once it's done. It seemed impossible to me, because Lua uses
+[non-preemptive coroutines](https://www.lua.org/pil/9.4.html).
+The caller needs to orchestrate when its coroutines can resume. So when the
+main script is done, there should be nothing that can resume remaining
+coroutines.
+
+This note is my investigation into why this is happening. Surprise is important
+for learning.
 
 ## Testing
 

lefthook.yml

@@ -0,0 +1,5 @@
+pre-commit:
+  parallel: true
+  commands:
+    tests:
+      run: just test