TL;DR - Missed connections + Markov chains
I’m beginning to develop a real nostalgia for “old” ML. The stuff that originally lured me away from motorsport a decade ago. Back when ML was actually fun. Even the imperfections of Dall·E 2 (2022) now seem closer to DeepDream (2015) than what the latest models are producing. The progress has been incredible, but the relentless marketing hype is exhausting. LLMs are wildly impressive and are capable of delivering actual value for many people despite their shocking simplicity. But they’re too clean. The output is too sanitized. The expectation of a perfect response is becoming normalized and t…
TL;DR - Missed connections + Markov chains
I’m beginning to develop a real nostalgia for “old” ML. The stuff that originally lured me away from motorsport a decade ago. Back when ML was actually fun. Even the imperfections of Dall·E 2 (2022) now seem closer to DeepDream (2015) than what the latest models are producing. The progress has been incredible, but the relentless marketing hype is exhausting. LLMs are wildly impressive and are capable of delivering actual value for many people despite their shocking simplicity. But they’re too clean. The output is too sanitized. The expectation of a perfect response is becoming normalized and that’s boring…
Buried deep in my crontab, there’s a script that’s been running every Sunday at midnight for many years now. It’s a simple one, in that its only job is to scrape all new missed connections from Craigslist and dump them into a sqlite table. Given my recent feelings on ML, I decided to take a trip back in time. Before RLHF and attention mechanisms, there were Markov chains. They were unpredictable and produced mostly gibberish. Every once in a while though, they would surprise you with something great.
I decided to spend the morning writing a Markov chain based on the missed connections data. The output of which gets posted to Bluesky and Nostr (view on nostr.band). Given the short form of a Craigslist missed connection and the large corpus of data available, the Markov model is able to produce a pretty high hit rate for believable output. Here are some of my favorites so far:
My hobbies are reading science or psychology. I’m basically looking for a 3rd encounter...
- June 17th, 2025
I’m probably too hot to be under 175lbs
- June 25th, 2025
Michelangelo t shirt. I was in Walmart tonight 9/18. You are pleasant and good looking.
- June 27th, 2025
The older I get, the more fun dumb things like this become…