A snippet of an ongoing discussion with my friends about GPT-3's reasoning abilities (spoilers: it doesn't do well...).

AI Plays Guitar Hero 3 on a Surface Pro 4 & Edge TPU. Still a work in progress. The CNN model does not currently recognize star power but it has still 4 starred this song. Recorded using OBS. Both recording and playing are done on a Surface Pro 4 at 60 FPS. Inference is hardware accelerated using a Google Edge TPU and TFSlim. This ended up taking a little more than 2 weeks to implement.

I trained several TensorFlow models to play Tetris using OpenAI Baseline's A2C implementation. They trained at over 4,000 FPS using a 980ti graphics card. The first video of the playlist is of my best model. It can score a lot of perfect 4 line clears, but it generalizes poorly. The second video shows this model playing 18 games at once. The last video is of another model that does not score as many points per move, but generalizes very well. It frequently clears over 200 lines a game. Everything is controlled by a Raspberry Pi. The power supply was under-rated at the time of the filming. It can actually clear ~40 lines in 11 seconds. Much faster than the human record.

This is our team's final presentation for Big Ideas in AI (AI530) at Oregon State University. It is about Wittgenstein's later philosophy and how those ideas have influenced natural language programming (NLP)