Big thanks to Social Tables for hosting!
Food and drinks are sponsored by the pseudo-anonymous Jay Kay! Thanks Jay Kay!
Want to be an official sponsor? Check out https://dc.hackandtell.org/sponsor.html
Event link: https://www.meetup.com/DC-Hack-and-Tell/events/264250209/
Notes copied directly from the hackpad!
Steve Trickey
Holiday Light Hack
The hack: Two strands of LEDs (25 LED per strand) with arduino (knockoff) board that can flash lights and change colors. Making holiday decorations with green and red lights chasing around the strands. Arduino uses C++ for the coding.
Arduino knockoffs. What’s the harm?…Knock-offs are a problem.
Adafruit has the most hardware documentation anywhere on the internet!
Joseph Nelson
ChessBoss - Playing chess with (or against) AI Grandmaster
https://www.youtube.com/watch?v=XLispu-Yb_0
Chess is hard, Digital Games are meh, and there isn’t enough (over)-engineering in the world \s
Set up a tripod with an iphone that takes picture of chess board, ChessBoss proposes a move based on the board. User then makes that move in an online chess engine (StockFish) and waits for chess engine’s next move.
“Breakfast Finder” comes standard with Apple iOS & is an Object Detection app. Identify where the board is, the board is always 8x8; so you can understand where the pieces are based on their position on the board.
Once you encode the board state into Chess notation.
Built in 48hrs at 2019 Techcrunch Disrupt’s hackathon. So it doesn’t work great
Data and labels are open-source: http://bit.ly/chess-data
Jesse Stanchak
How to tell tiny stories and make friends all over the world.
Publishes very small stories that fit into digital tweets. Microfiction is to Flash Fiction what Short Stories is to novels
Microfiction is about finding the minimally necessary “load bearing” details of the story.
The @MicroFlashFic way:
- 3 stories per day
- No breaks
- Length of single tweet
- All self-contained
Why even attempt this sort of thing? It’s quick. It allows for self reflection. It is a way to use creative muscle. It is a way of channeling ambition. Builds an audience.
To get started: pick a platform, don’t involve your IRL network (!!), tap into existing community, set a schedule, stick with it no matter what, experiment, don’t second guess yourself (aim for growth not perfection), evaluate and look for ways to improve
Zack Stern
AI to replace sports interviews
Post-game interviews are dumb, generic, mundane, and pointless. Why not replace this with artificial intelligence?
AI can respond to “sports reporters” given a prompt question
Scraped 1700 college football press conferences, cleaned the data and formatted in a Question: Answer format. Uses real questions, generates fake answers.
Used OpenAI’s GPT-2 text generation model (used smallest one ~117M tokens) run on Google Colab GPUs. Training used a 25MB file; took maybe an hour using Google Colab (on top of the ~117M token model).
Mostly just messing around to find the best temperature (randomness).
Aaron Schumacher
Q-Learning Tic-Tac-Toe, Briefly
https://planspace.org/20191103-q_learning_tic_tac_toe_briefly/
Q-learning -> Reinforcement Learning (a way of learning to do stuff)
Aaron got nerd-sniped! How good is reasonable using Q-learning?
Player 1 is represented as a +1 and Player 2 is a -1.
Use numpy array to create a board representing a game of tic-tac-toe. Start with an array of all zeros and change matrix entries to represent a move
Represent moves with i, j matrix. Most complicated is: has anyone won the game? (If you get to +3 or -3, you get three in a row)
PLay with a random player 10,000 times.
Random player, you can play a whole game. If you’re going first, you have a better win rate.
Boring player plays the first available move, left to right, up to down. Does better than random!
Created agent based on neural network implemented in Keras. You can then predict how good moves are and learn how good old moves were
There are 9 spots and 9 possibilities for the best value
Q-learning play does way better than Random and Boring players! (Q-learning wins 93% of the time)
Q-learning is a very specific type of reinforcement learning
Alyssa Liguori
Data science model deployed to web app. This project uses an individual’s rating of issue importance for seven issues to “predict” their 2016 Presidential vote.
https://github.com/ali0003433/predict-by-issue-prioritization
Election vote predictor based on survey data. We know that Americans agree in some issues but lack cohesion on others.
Directionality of the question (or lack thereof) tends to matter too. Nearly equal numbers of Clinton and Trump voters found abortion to be important. Likely for different reasons.
Deployed as a web-app. Answer 7 questions then get prediction.
Travis Hoppe
Postern-Perception
https://github.com/thoppe/postern_perception
Found a model that can generate fake/artificial eyes. What if you make an app where a grid of eyes track a mouse?
Render spectre - Eyes that have never been seen before.
Then made an app that can track a human using a webcam! Weird AI interaction.
Disney has had crazy tracking eyes for a while! With AI, we have tracking eyeballs now too!