Brandon Rohrer
LightNews LightNews
banner
brandonrohrer.com
Brandon Rohrer
@brandonrohrer.com
3.7K followers 2.2K following 990 posts
Robotics and Reinforcement Learning tinkerer.
brandonrohrer.org
Wrangler of algorithms.
Eater of bread. Sipper of whisky.
Reports to a Shih Tzu.
Posts Replies Media Videos
Pinned
brandonrohrer.com
Brandon Rohrer @brandonrohrer.com · May 28
Update from my long-running #ReinforcementLearning side project.

It uses a combination of new tools to control a simulated pendulum.
Solving an easy RL problem on hard mode Brandon Rohrer A story about a reinforcement learning approach learning to make a pendulum stand up straight, while making very few assumptions. tl;dr This is a demonstration of an RL approach that uses a combination of new tools to control a simulated pendulum: BucketTree to learn a discretization of the pendulum’s continuous state variables, angle and angular velocity, Ziptie to bundle the discretized values into discrete states, Fuzzy Naive Cartographer (FNC) to learn common state-action-state sequences and to make conditional predictions of reward for each action. It all runs in Myrtle, a real-time reinforcement learning workbench. This approach requires only a little domain specific design and makes very few assumptions about its world. The physical representation of the pendulum is simplistic. The arm has uniform mass. It has a small amount of rotational friction, proportional to its speed. And gravity acts on the pendulum's center of mass, pulling it downward. As always, the canonical source of information is the code itself. Sensors There are two quantities returned by sensors: pendulum position and rotational speed. Position, θ, is measured in radians, zero when pointing straight down, θ = π/2 when pointing to the right, θ = π when pointing upward, and continuing around to almost 2π when it reaches the bottom again and resets to 0. Angular speed, ω, is measured in radians per second. Positive angular speed is counter-clockwise (the direction of increasing position) and negative speed is clockwise. Actions Actions take the form of torque, τ, applied to the base of the pendulum. Positive torque accelerates the pendulum counter-clockwise, in the direction of increasing position. Negative torque accelerates it in the clockwise direction. Actions are discrete in time. Each action is a constant torque that lasts for 1/8 second. There are 13 discrete values the torque can take, 6 positive, 6 negative, and zero. Possible torque values are distributed nonuniformly across the range, with the middle half of the range having denser coverage. This results in a finer grained representation of small torques, useful for making fine-tuning adjustments. Reward The pendulum returns reward, r, related to how high the its swinging end reaches, its vertical distance from its straight down θ = 0 position. It reaches a maximum of r = 2 at the straight up θ = π position. A successful learning curve will work its way up to 2 and stay there.
brandonrohrer.com
Cool Python trick: Sometimes you want functions to have keyword arguments, but you also want them to be required. This is especially useful when you're passing required arguments using dictionary expansion, like `function(**arg_dict)`.
September 29, 2026 at 12:25 PM
brandonrohrer.com
Optimizing models with lots of hyperparameters and options is a slog that usually requires a combination community wisdom, obscure blog posts, and tedious trial and error. Automated methods tend to rely on strong assumptions of smoothness or brute force their way through every permutation.
September 29, 2026 at 9:26 AM
Reposted by Brandon Rohrer
zephoria.bsky.social
Today is Pub Day for "Data Are Made, Not Found: A Story of Politics, Power, and the Civil Servants Who Saved the US Census"!!! ::bounce:: Please support me by buying a copy of the book or asking your library for a copy! press.uchicago.edu/ucp/books/bo...
September 24, 2026 at 12:57 PM
brandonrohrer.com
Hey US eng friends, if you want to give Patreon a shot:

Staff SWE, Data Infrastructure
jobs.ashbyhq.com/patreon/e8f0...

Senior MLE, Infrastructure (on my team)
jobs.ashbyhq.com/patreon/722d...

Staff SWE, Backend Platform
jobs.ashbyhq.com/patreon/c277...
Staff Software Engineer, Data Infrastructure
Patreon is a media and community platform where over 300,000 creators give their biggest fans access to exclusive work and experiences. We offer creators a variety of ways to engage with their fans an...
jobs.ashbyhq.com
September 16, 2026 at 10:02 PM
brandonrohrer.com
Three days of automated parameter sweeps saved me 15 minutes of Internet searches!
September 15, 2026 at 12:41 AM
brandonrohrer.com
added type hints and ran mypy on some code I considered tight. It was like shining a black light in a hotel room.
September 8, 2026 at 1:30 AM
brandonrohrer.com
Diving in to Python typechecking and stumbled on this @vickiboykis.com post
vickiboykis.com/2023/12/11/w...

The tldr: "The result of the deep dive is that I now no longer trust anything in the Python typing ecosystem."

I'm now confused as ever, but on a higher level and about more important things
Why if TYPE_CHECKING?
Typechecking is brittle yet important
vickiboykis.com
August 31, 2026 at 12:53 PM
brandonrohrer.com
Josh, Luis, and I had a great time chatting the other day and recorded in case you'd like to listen in
www.youtube.com/live/Yx-ZkLf...
TRIPLE BAM!!! With Josh, Luis and special guest Brandon Rohrer
YouTube video by StatQuest with Josh Starmer
www.youtube.com
August 22, 2026 at 10:22 PM
brandonrohrer.com
T-minus 30 minutes
brandonrohrer.com Brandon Rohrer @brandonrohrer.com · Aug 20
If you’re into such things, I’m going to be chatting with Josh Starmer and Luis Serrano (two of the best ML educators I know) in a YouTube livestream, answering viewer questions, at 11:00am EDT Thursday (tomorrow/today).
8am PDT, 3pm UTC, 5pm CEST

m.youtube.com/live/Yx-ZkLf...
TRIPLE BAM!!! With Josh, Luis and special guest Brandon Rohrer
YouTube video by StatQuest with Josh Starmer
m.youtube.com
August 20, 2026 at 2:31 PM
brandonrohrer.com
If you’re into such things, I’m going to be chatting with Josh Starmer and Luis Serrano (two of the best ML educators I know) in a YouTube livestream, answering viewer questions, at 11:00am EDT Thursday (tomorrow/today).
8am PDT, 3pm UTC, 5pm CEST

m.youtube.com/live/Yx-ZkLf...
TRIPLE BAM!!! With Josh, Luis and special guest Brandon Rohrer
YouTube video by StatQuest with Josh Starmer
m.youtube.com
August 20, 2026 at 1:18 AM
brandonrohrer.com
@ellecordova.bsky.social is a wonder. If you'd like to tell her so, pop over to Patreon and buy her a coffee.
www.patreon.com/cw/ellecordova
August 11, 2026 at 10:53 PM
brandonrohrer.com
My Artisanal Language Model for proofreading continues to evolve. Pairing forward and backward first-order Markov models is proving to be just good enough to merit further exploration.
brandonrohrer.com/alms_composi...
five plots showing five different proofreading evaluation scores: capitalization, grammar, missing words, punctuation, spelling. In each case, as the error threshold for identifying a proofreading error increases from 1e-5 to 1e-2, precision drops from 100% to very small and recall rises from very small to 100%. The crossover point tends to be in the 30-40% range at around 0.0002
August 10, 2026 at 1:41 PM
brandonrohrer.com
RSS is the vinyl records of the web.
And the Current RSS reader is the $5000 turntable whose clean design matches the medium.
Except it’s only $10.
www.currentreader.app

Built by one designer with a vision ( @terrygodier.com ), Current shows what the web can be.
Current - A River of Reading
Current rethinks RSS for the modern reader. Your feeds flow like a river. No inbox anxiety, no unread counts. Just beautiful reading.
www.currentreader.app
August 8, 2026 at 12:35 PM
brandonrohrer.com
It's primary season Boston
50501movement.bsky.social 50501: The People’s Movement ❌👑 @50501movement.bsky.social · Aug 6
It's time for a new generation of leadership! 50501 is proud to announce our newest Rising Star: Patrick Roath, who is a candidate for Massachusetts's 8th Congressional District. We the
People must stand together!
50501 Rising Star Patrick Roath (D) for Congress (MA-8) vs Rep. Stephen Lynch No Kings!
August 7, 2026 at 12:19 AM
brandonrohrer.com
There are several command line incantations that I have had to look up at least a dozen times.

Finding a string in Python files
grep -r --include \*.py 'string I am looking for' .

Find files by name
find . -name '*partial_name*'
August 4, 2026 at 1:21 AM
Reposted by Brandon Rohrer
valyriansteelers.bsky.social
Listening to the Dungeon Crawler Carl books with Soundbooth Theater leads to really long walks because I can't stop listening. They are a blast!
July 27, 2026 at 1:25 AM
brandonrohrer.com
Heads up ruff users - with the release of 0.16.0 this week the number of default rules went up from 59 to to 413. This means that if you're ruff version auto updates, your CI pipelines might start failing on you all of a sudden.
simonwillison.net Simon Willison @simonwillison.net · Jul 25
Ruff 0.16.0 - Astral's fast Python linter - came out a few days ago and increased the number of default-enabled rules from 59 to 413, which highlighted all sorts of problems across my projects (1618 in sqlite-utils alone) simonwillison.net/2026/Jul/25/...
Ruff v0.16.0
Astral shipped a significant new version of their Ruff Python linting tool a few days ago on July 23rd. I noticed today because my various CI jobs all started failing …
simonwillison.net
July 26, 2026 at 11:56 AM
brandonrohrer.com
Living small is part of the artisanal process. The constraints are not a barrier to the final product. They are an important shaping force. You can't come in at the end and magically shrink it. You can only build small if you start small and work with what you've got as you go.
brandonrohrer.org
Book Project: Artisanal Language Models Chapter 1: Artisanal Language Models Chapter 2: Build a custom tokenizer Chapter 3: Define a task and write evals Chapter 4: Build an end-to-end prototype Chapter 5: ALM v2, First-order Markov model Chapter 6: ALM v3, Second-order Markov model Chapter 7: ALM v6, Sparse second-order Markov model Chapter 8: ALM v14, The Search for Data
July 25, 2026 at 8:12 PM
brandonrohrer.com
Artisanal Language Models take a step closer to being able to do proofreading.

Trained and hosted on an old Mac. These are small but mighty. Kind of. Aspirationally mighty, at least.
brandonrohrer.org/alms_data.html
It took a little time to figure out how to properly download Project Gutenberg texts at scale, but I finally set up three new training corpuses:
brandonrohrer.org
July 20, 2026 at 11:27 PM
brandonrohrer.com
the biggest difference I've noticed between data scientists and machine learning engineers is that MLEs will do absolutely anything to avoid looking at their data
July 8, 2026 at 1:29 AM
brandonrohrer.com
don't touch those! those are my emotional support patreons!
June 28, 2026 at 1:52 PM
brandonrohrer.com
My favorite thing to do is find a narrow problem that doesn't have a good solution, and then build one. Here are the highlights, from my blog.
brandonrohrer.org/blog.html#al...
Algorithms and Methods Evolutionary Powell's method for hyperparameter tuning k-nearest neighbors adaptive feature weighting k-nearest neightbors data reduction k-sparse neural network layer Naive Cartographer , a Markov Decision Process learner oknn, an online variant of kNN Pathfinder, a many-to-many shortest path algorithm Sharpened Cosine Similarity, an alternative to convolution for learning 2D features in neural networks Ziptie, an unsupervised feature creation algorithm
June 28, 2026 at 12:24 AM
brandonrohrer.com
I shut down by online courses at End-to-End Machine Learning 18 months ago. Wrote a blog post and everything. But the platform I was using hasn't bothered to take them down. So they're just sitting there, all for free. Prego.

end-to-end-machine-learning.teachable.com
A screenshot from the link showing cards for courses on decision trees, polynomial optimization, k-nearest neighbors, multilayer neural networks, autoencoders, and convolutional neural networks
June 26, 2026 at 11:34 PM
brandonrohrer.com
Efficiency maxxing is problematic for lots of reasons but a fundamental reason is that it always implies near term efficiency.
June 21, 2026 at 11:39 PM
brandonrohrer.com
This next round of Artisanal Language Model improvements are mostly focused on efficiency rather than dramatic algorithmic improvements. This ALM is a proofreading tool for English prose.
brandonrohrer.org/alms_somm_sp...
After some trial and error with a few sparse representation concepts, I settled on one that takes advantage of one of Python's strengths: dictionaries. The sparse second-order model keeps two running dictionaries: one full of bigram counts and one for trigram counts.
brandonrohrer.org
June 21, 2026 at 9:49 PM