Side projects

Jailbreaking ChatGPT’s image generator. MEDIUM ARTICLE

A quick attempt to circumvent restrictions on depiction of substance use with recursive complexity and evasive language techniques. DALL·E 3 is Open AI’s latest iteration of its text to image system. It is built natively on ChatGPT and can at this time be used by ChatGPT Plus and Enterprise users. Open AI claims to have “taken steps to limit DALL·E 3’s ability to generate violent, adult, or hateful content.” This short article is intended to showcase some of the limits of the current safety features.

Predicting chess games results using LightGBM. MEDIUM ARTICLE

In this project I trained multiple ML models to predict results of chess games. In particular, I considered games played on Lichess, the second biggest website for playing chess, and one of the most successful open source projects. After gathering and cleaning the data, I used the gradient boosting framework LightGBM to train the ML models. I found that even simple models are better at predicting results than the Glicko2 rating system.

Adding a 4-point line to the NBA using the NBA API. MEDIUM ARTICLE

In this article I used the NBA API to determine how many games would have had their result flipped if there was a 4-point shot. This work was prompted by a video of a popular creator, who claimed that only two games in the 2022-23 season would have changed, were a 4-point line introduced at 30ft from the basket. I rectified his findings, and extended the results to the last 20 seasons.