We believe projects are one of the best ways to gain experience and connect with other members of the community. After completing the intro program, we hope that everyone will have the chance to be working on either individual or group projects.

Projects conducted in conjunction with the AI Safety Club will receive:

  • Mentorship: We can help you solve your problems if your project is machine learning, statistics, and math projects. If your problem is very niche, we can connect you with people outside the club who can provide you advice.
  • Compute: If you need to train a large machine learning model, AI Safety can help you get access to servers where you can train.
  • Funding: If your project needs funding in order to say, recruit MTurk workers, or to use the OpenAI API, we may be able to provide you with some funding.

We’re interested in a broad variety of projects. Our only criteria is that a project must:

  • Relate to AI Safety: There’s a lot of flexibility here: You can make a software project on technical topics like interpretability or robustness, or do research on AI timelines, or use AI to build a useful app or product. If you’re not sure, feel free to apply anyway and we can discuss if it is.
  • Be good for the world: Don’t research how to create a murder robot or anything like that.

Past Projects


Reinforcement Learning to Drive a Car

Date: Mar 2023 to Jun 2023

In 2023, AI Safety at UCLA members trained a Reinforcement Learning agent to drive a car using the Metadrive driving simulator.

We worked in conjunction with Prof. Bolei Zhou’s lab to develop these tutorials.


Alpha Zero Connect4

We ran a tournament to design AI algorithms that played Connect4.

Date: Mar 2023 to Jun 2023

In 2023, we ran a tournament to design AI algorithms that played Connect4. An implementation of Deepmind’s AlphaZero algorithm was the winner!



Building an Among Us Bot using Reinforcement Learning

Date: Mar 2023 to Jun 2023

In this work, we simulate the popular online multi-player game ‘Among Us’. Inspired by works like CICERO, we use multiple models to tackle different parts of a full game.


Neural Cellular Automata

Growing a Moyai from a single pixel.

Date: Mar 2023 to Jun 2023

This project, created by one of our members, implements a neural network cellular automaton as described in this paper. It’s able to grow a 🗿 from scratch.



Test your critical reading skills by trying to distinguish between GPT3 completed text and the real thing

Date: Mar 2022 to Jun 2022

Oftentimes, when we skim an article, we fail to engage with the actual point of the text, simply reading the words without really understanding them. GPT3 is a large language model that is notorious for this as well. It has no “real” understanding of anything, but is an expert at writing plausible sounding continuations of any text.

Critica forces the reader to discriminate between 3 AI generated distractors and 1 true article completion in order to keep reading the article. In doing so, we hope to ensure that the reader knows what’s really going on in the article.


critica screenshot