Michael Mullarkey

@mcmullarkey.bsky.social

4575 Followers 1205 Following

Clinical psych PhD | ML Engineer in sports Subscribe to my newsletter on data, learning, and building: https://buttondown.com/datadash Opinions are my own and don't reflect the opinions of my employer

A scatter dot with cost per agentic task on the X axis and overall model performance on the Y axis. There is a strong, positive relationship between cost per agentic task and performance score, with no models really in the denoted "most attractive quadrant" with low cost per agentic task and high performance
A description of me, Michael Mullarkey, that's eeriely accurate from Opus 4.8
A basic chat interface on mobile with messages about to do items, notes, and searching those notes. Most of the chats are slash commands
A photo of Weakened Friends on a lit stage in an otherwise dark concert venue
Do it Dark Kermit meme
An NBA court during a summer league game. Dylan Harper is going up for a dunk and in the process of being fouled by a Mavericks defender
A basketball court with a player shooting a free throw
Dark Kermit Do It meme
A placard with "Check if your project is impossible" for the newsletter Data Dash (still with the wrong date of January 5, 2025, which is when I first started drafting this post!)
A landscape shot of a snowy mountain landscape with a frozen lake in the foreground
Make your README a brag sheet from Data Dash (says January 7, 2025 which isn't today!)
A production app that asks which Dunc'd On basketball podcast episodes are best. This app hits an RSS feed, retains the episodes from the past week, performs feature engineering, and runs an in-production model that achieved a 0.80 F1 score out-of-sample on predicting which episodes are among the better ones for the podcast. In the video, we see someone click the prominent "Predict" button, which nearly instantly causes the initial descriptive text to fade out and an interactive horizontal bar chart to appear. We then mouse over the bar chart and see the chart is interactive and displays the probability of the episode being one of the better episodes of the podcast. The episodes are arranged on the chart in descending order from most to least likely to be one of the better episodes of the podcast.
An interactive scatter chart stacked on top of a horizontal bar chart The scatter plot is the first principal component of Dunc'd On episode description embeddings (X-axis) vs. the second principal component of those embeddings (Y-Axis) Generally Daily Duncs have low X values with a variety of Y-values, while main episodes have higher X values with a variety of Y values The video shows selecting different areas of the scatter plot with the horizontal bar chart changing depending on the set of episodes selected (The bars show the count of episode types)
A horizontal bar chart of episode types of Dunc'd On basketball podcast episode type by whether or not I thought the episode was a banger A majority of all episode types were bangers, but while the Big Picture episodes were the most common type, they were least often bangers. An overwhelming majority of Gamers and Hollinger + Duncan episodes were bangers, though those were both substantially less common
Bar graph of podcast episode release times for Dunc'd On by hour of the day. The most common release times are very late in the evening or very early in the morning
The cover of the book Code: The Hidden Language of Computer Hardware and Software Second Edition by Charles Petzold
An impersonation Bluesky account of Mike Caulfield with the handle @santospacker.bsky.social
A terminal window on an Android phone with the following text: Python has become one of the most popular languages among programmers because: 1. It is easy to learn due to its simplicity and readability. 2. It supports a wide variety of libraries and frameworks for various programming tasks, making it ideal for developing applications in different domains. 3. It is widely used across multiple industries, including web development, data science, machine learning, artificial intelligence, scientific computing, etc. Python's versatility allows developers to build custom software solutions quickly and effectively. With its active community of users, libraries, and tools, Python can be a valuable tool for many developers who prefer not to reinvent the wheel. Some key points about Python: - It has built-in support for several programming paradigms (e.g., object-oriented, procedural, functional). - It is cross-platform compatible. - It supports various data structures like lists, dictionaries, tuples, sets, and so on. - It is open source with a growing community of developers contributing to its development and improvement. In summary, Python is an excellent language for beginners as it allows them to write code quickly and learn the basics of programm…
Since I'm not good at doing any of that, the way I tend to approach "big projects" is to break them down into useful self-contained chunks that have these properties: The chunk of work is small enough to deliver sorta/kinda quickly so that people don't get annoyed that I've run off for a year The chunk is valuable on its own merits, even if it subsequent chunks don't ever get built The chunk is potentially useful for other projects, it's not solely tied within the big project
A cow at the edge of the ocean with the text I am cringe but I am free
A picture of a push notification that reads: Editors Picks. A young couple's dream home was infested with thousands of bats and their skeletons
An article excerpt from AI ethics researchers that reads: There’s a humorous definition of AI that says “AI is whatever hasn’t been done yet”. When an AI application starts working reliably, it disappears into the background of our digital or physical world. We take it for granted. And we stop calling it AI. When a technology is new, doesn’t work reliably, and has double-edged societal implications, we’re more likely to call it AI. So it’s easy to miss that AI already plays a huge positive role in our lives. There’s a long list of applications that would have been called AI at one point but probably wouldn’t be today: Robot vacuum cleaners, web search, autopilot in planes, autocomplete, handwriting recognition, speech recognition, spam filtering, and even spell check. These are the kinds of AI we want more of — reliable tools that quietly make our lives better.
Cover of the book Character Limit: How Elon Musk Destroyed Twitter by Kate Conger and Ryan Mac
BlueSky imagine celebrating 100k users, and I was user 64,302
Andrew Wiggins about to shoot a free throw with a graphic showing he was drafted first in 2014 while Jokic, the defending Finals MVP, was drafted 41
Highlighted text from Substack post: "Note to self: if you’re going to write code to do fraud, make it messy and unreadable to reduce the chances it’s later put in front of a jury as evidence"
Picture of command line terminal with a command that says "llm -m llama-2-7b-chat.Q8_0 'five names for a cute pet skunk' -o no_gpu 1 with output of 5 pet skunk names + descriptions generated by Llama 2 running locally on my Mac
A cute cat sitting, wearing a witch"s hat next to a tiny pumpkin with a caption of "It's October 1st it's spooky time"
. I think I got sidetracked. My point is that this colleagues’ advice2 helped me tremendously and now I’d like to pass it on. The bottleneck for success and fulfilment and even personal growth probably doesn’t look like what we think it looks like. It probably looks more like finishing your projects. Or at least understanding why you don’t want to finish them. And if you’re avoiding something, there’s a part of you that doesn’t want to do it. You need to be in dialogue with it, no matter how much it horrifies you, or how thoroughly you’ve buried that part. Enough already. Now I’m rambling because I’m insecure about the blog post. You have to press send. Finish your projects
A screenshot of the research article saying "We deviated from this term set in only two ways. First, on YouTube we eliminated the “signs” term, as the phrase “fire signs” was returning many astrology videos. To deal with a similar issue on X, we used these exclusion terms: -“fire signs” -leo -aries -sagittarius. "
A landscape picture of an ocean with a picturesque island town in the background