Posts

Five hypotheses about Flip 7 strategy, a full simulation of the game, and 200,000 games. There is no magic number of cards. The best rule reads the odds of the next card, and it is the same rule at every table size.

Hypothesis: on the grid, crossing the avenue when the light allows and turning uptown when it doesn't beats the plain L. A simulator with NYC signal timing says yes, by about 50 seconds on a 13-minute walk, if you are patient at the corner.

My Mac mini stays on at home and I use its full desktop from my iPad, on Wi-Fi or cellular. Tailscale, Jump Desktop, a fallback or two, and a watchdog.

My Flip 7 scorekeeper now has an Apple Watch app. I tap the cards on my wrist, the phone (or the TV) shows the race to 200, and every finished game is saved round by round.

A nine-turn game for 2–6 people where you fight over a hex map with dice you build yourself. It started as a phone app for game night; now it runs anywhere with a browser. Early playtest build, rules still moving.

An e-ink screen by the door that shows the weather and, for the two stations near me, the next trains with the time I have to leave to catch them. Code and setup guide included.

Rebuilding the word2vec model and the class, gender, and morality dimensions from Kozlowski, Taddy and Evans (2019), checking them against survey data, and putting words in an interactive 3D space.

A working note on building an autonomous AI agent with GPT-4, LangChain, and AutoGPT.

Notes on the required reading for my credit risk team: the seven stages of building a credit risk scorecard, plus the questions the book raised for me.

Notes from Arize's online event on LLMs and ML observability, with talks from OpenAI, HuggingFace, Spotify, Etsy, LangChain, BabyAGI, and more.

Introducing a project to predict Tesla's stock price with recurrent neural networks, combining traditional market data with cultural signals scraped from Reddit and Twitter.

A step-by-step walkthrough of building and validating a linear regression model on the 1990 California census housing dataset, ultimately explaining 64.4% of the variance with three degrees of interaction terms.