Some thoughts on what I call context complexity: mchav.github.io/context-comp...
Context complexity: what is the Big-O of an agent API?
TLDR; small, local models are a good instrument for measuring what a harness costs an agent. an API is only as good as its outputs’ specificity, parsimony, and truthfulness. obvious code repair...
mchav.github.io
On the blog: "Quick tips for fast iteration in #Haskell" by Tom Ellis & Laurent P. René de Cotret blog.haskell.org/quick-tips-f...
Quick tips for fast iteration in Haskell | The Haskell Programming Language's blog
Quick tips about tools and techniques for fast iteration when developing Haskell
blog.haskell.org
Who needs some good news today? @pikeplacemarket.bsky.social is introducing a "Locals Night" every Wednesday evening through September 9th where there will be a farmers market on the street and many businesses will be open later than usual! You love to see it. www.pikeplacemarket.org/events-calen...
Models are programs A motivation and brief tour of dataframe-learn. At its best DataHaskell is a place for these kind of cross-cutting experiments. www.datahaskell.org/blog/2026/07...
Models are programs
OverviewWe present dataframe-learn, a machine learning library where the output of training models is symbolic expressions. We first motivate why symbolic mo...
datahaskell.org
I recently really enjoyed @mschav.bsky.social's talk at #ZuriHac so I tried to reproduce this chart in Sabela (a Jupyter-notebook-like project for #Haskell) seems like this Mbappé guy is not so bad either! 🤭
Before we solve data workflows in Haskell we have to teach people Haskell. sabela.datahaskell.com/c/1ea40000
Learn You a Haskell for Great Good!
Curated reactive Haskell notebooks.
sabela.datahaskell.com
sabela.datahaskell.com Sabela now has a community gallery. The highlighted examples show that Haskell can be a good tool for exploratory work.
Sabela Community Gallery
Curated reactive Haskell notebooks.
sabela.datahaskell.com
It’s pretty annoying that LLMs like to use the construction “X is real”.
A small experiment comparing token efficiency between Haskell and Python. Not rigorous at all but I’ll follow it up eventually. mchav.github.io/a-first-look...
A first look at token efficiency
A while ago I saw the article Which languages are most token efficient. The article was largely discredited since it didn’t have a clear methodology. Unfortunately, I haven’t come up with one yet but ...
mchav.github.io
I’ve been doing a small social experiment on how to best to market Haskell (dataframe) to non Haskell programmers. It seems engagement and stars go up when the article doesn’t foreground Haskell but instead uses it as a conduit for a larger concept.
Anyone work with time series data professionally? Either modeling or data engineering? Have some questions and would appreciate a call.
Lake Washington Blvd will be open to people & closed to car traffic every weekend this summer from Memorial Day to Labor Day! The predictable schedule makes it easier walk, bike, roll, along, or drive to the boulevard. Thank you Mayor Wilson, community advocates, and Rainier Valley Safe Streets.
Using your debt, your search history, and your private data to decide how much you’re worth? Hell no. This is what happens when technology is built to serve profit instead of people. This is why every worker needs a union.
I’ve been thinking a lot about how to scale symbolic regression: this seems like a plausible direction. mchav.github.io/grow-and-mow/
Grow and mow: interpretable models with boosting, symbolic regression and e-graphs
This post is the convergence of two ideas that have been floating in my head for about a year. Can we learn messy stochastic models and use algorithmic/algebraic tools to rein in model complexity to make models interpretable?
mchav.github.io
With more and more code being written by AI I think human coding will mostly be for prototyping and brain storming. Notebooks will become a much more important developer interface. So I mixed interactivity, Lean, Haskell, and Python into a single notebook runtime sabela.datahaskell.com
Sabela
sabela.datahaskell.com
I took a crack at category theory + dataframes. I find the difficulty with reading and writing about this kind of stuff is that it's really hard to communicate what the "point" is. Hopefully it all makes sense: mchav.github.io/what-categor...
What Category Theory Teaches Us About DataFrames
Every dataframe library ships with hundreds of operations. pandas alone has over 200 methods on a DataFrame. Is pivot different from melt? Is apply different from map? What about transform, agg, apply...
mchav.github.io
I've been looking at symbolic regression for some time now. I think most genetic approaches would benefit from large e graph databases to reduce the search space. Having to start every search from scratch every time seems silly.
Today is a great day for some major news! DataFrame v1.0.0.0 has been officially released! Step up your exploratory data analysis in #Haskell with Typed data frames, direct connection to HuggingFace data sets, and Python integration through Apache Arrow. discourse.haskell.org/t/ann-datafr...
[ANN] dataframe 1.0.0.0
It’s been roughly two years of work on this and I think things are in a good enough state that it’s worth calling this v1. Features Typed dataframes We got there eventually and I think we got there i...
discourse.haskell.org
Anyone know what could be a good second career? Even if coding makes it out of the AI era I fear it won’t be interesting.
Started a new position @coreweave.bsky.social working on @marimo.io
Sometimes you gotta do the boring thing that gets the job done. www.datahaskell.org/blog/2026/03...
Sabela: A Reactive Haskell Notebook
Overview
datahaskell.org
Sabela - A reactive Notebook for #Haskell by the DataHaskell project Announcement: discourse.haskell.org/t/ann-sabela... Github Repository: github.com/DataHaskell/...
[ANN] sabela - A reactive Notebook for Haskell
Sabela is a reactive notebook environment for Haskell. The name is derived from the Ndebele word meaning “to respond.” The project has two purposes. Firstly, it is an attempt to design and create a mo...
discourse.haskell.org
I gotta starting taking cold showers before a job interview for a job that I want - wash off the smell of desperation.
Small experiment: treat feature engineering as program synthesis, then use an LLM as a lightweight prior over which derived quantities are “nameable.” The learner stays classical; the artifact gets way more readable. mchav.github.io/learning-bet...
Learning better decision tree splits - LLMs as Heuristics for Program Synthesis
A lot of tabular modeling gets easier the moment you stumble onto the right derived quantity. Not something mysterious or “deep.” It’s usually something you can name: a ratio that turns two raw column...
mchav.github.io
Symbolic AI is built on the premise that models should be presented in terms that are understandable to us. When you interact with a symbolic system you learn something about the reality that it tries to model. That alone makes symbolic approaches worth betting on in the long term.
In software it’s often important to distinguish between solving the scientific problem (how do we make this generalize for all instances of the problem) versus the engineering problem (how do we make this work for the environment we anticipate it’ll be used in).
