OpenAI Hackathon Report March 2018

OpenAI hosted its first hackathon on March 3, 2018, bringing together 100 members of the artificial intelligence community to develop AI projects during an eight-hour code sprint. The event served as a collaborative space for a diverse group of practitioners, ranging from high school students to university researchers and industry engineers.

Participant Diversity and Composition

OpenAI curated a balanced group of 100 attendees from a pool of over 500 RSVPs. The participant base included high school students, university researchers, industry practitioners, and engineers working for nonprofits. Interests among the attendees spanned a wide range of applications, from healthcare to Artificial General Intelligence (AGI). OpenAI explicitly strove for gender balance to ensure a positive experience for all participants.

Technical Projects and Applications

During the eight-hour hacking session, participants developed dozens of AI projects. Key technical implementations included:

Reinforcement Learning and Robotics

  • PPO with Curiosity-Based Rewards: Arthur Juliani implemented Proximal Policy Optimization (PPO) using curiosity-based intrinsic rewards to train an agent capable of smashing block towers.
  • Natural Q-Learning: High school students Ethan Knight and Osher Lerner developed a project based on OpenAI's "requests for research" regarding Natural Q-Learning.
  • AeroEnv: Malhar Patel and Lee Redden created a Gym interface for a physical robot.
  • Q-learning in Processing: Rodger Luo implemented Q-learning within Processing, a programming tool designed for media artists.

Healthcare and Specialized Tools

  • Image Classification for Healthcare: Jiale Xian, Clarence Leung, Kyle Zheng, Madeline Hawkins, and Stergios Hetelekides built an image classifier to identify purine-rich seafoods to assist gout patients in avoiding specific foods.
  • Machine-Assisted Sketching: Andy Matuschak developed the "scrying pen," a tool for machine-assisted sketching based on SketchRNNs.

AI Safety

  • James Giammona and Brad Neuberg focused on AI safety by generating ideas for how reinforcement learning (RL) agents can avoid dangerous, unseen parts of their environment.

Infrastructure Support

Cirrascale provided the GPU machines necessary to support the computational requirements of the participants' AI projects during the event.

Sources