My Artificial Intelligence Journey Through YouTube
Like many of us, I spent a good portion of 2020 emerging from a deep Netflix chasm only to fall down a YouTube rabbit hole looking to learn more about what I had just watched. After a particularly deep dive into a season of Black Mirror, I stumbled on a channel called DADABOTS that had a live stream called RELENTLESS DOPPELGANGER — a non-stop SpeedMetal track, completely generated by an Artificial Intelligence engine. It has been playing continuously since September 4th, 2019 and is presumably still playing as you read this. They have since then also expanded into slightly ear-friendlier territory with a never-ending bass solo which will be easier on your eardrums and your speakers.
Machine Learning is the process by which a computer can learn complicated tasks or perform analysis through experience based on data. Put simply, data is fed into a program that analyses and manipulates it millions of times until it achieves the given parameters. It is an integral part of Artificial Intelligence and can be an incredibly dense and complicated subject of study.
Subscribing to the DADABOTS channel caused YouTube to funnel more AI and Machine Learning videos onto my feed, an expected and much welcomed consequence. I eventually landed on video-game related AI videos, including Google’s DeepMind AlphaStar beating the top Starcraft II players in the world. It seemed at once incredibly interesting and very complicated. I was aware of IBM’s DeepBlue besting top chess grandmasters but Starcraft is arguably far more complicated because it’s played in real-time creating seemingly endless possibilities. I was intimidated by the complexity but hooked on the concept. The intimidation factor would persist until my next spelunk into YouTube when I found videos of QWOP and stumbled upon a real-time duel of man vs machine.
QWOP is considered one of the most difficult video games ever invented because of its obtuse control scheme. The object is to make your runner complete a 100 meter dash without falling down using four buttons — each controlling the runner’s left or right calves and thighs. It’s even more difficult to play than it is to wrap your mind around the previous sentence.
We generally think of running as continuous forward momentum but don’t pay much attention to how it is generated. In QWOP, the player can only generate momentum by contracting each individual calf and thigh muscle, something our bodies do without any input. Most players don’t make it past a few meters. However through rigorous practice a select, dedicated few have mastered these mechanics and have posted some incredible numbers over the years on Speedrun.com.
1. Why Corporate AI projects fail?
2. How AI Will Power the Next Wave of Healthcare Innovation?
3. Machine Learning by Using Regression Model
4. Top Data Science Platforms in 2021 Other than Kaggle
Of course an AI has no preconceived limitations of how a human runs to overcome the obtuse control scheme and play the game. Wesley Liao, a data scientist, decided to adopt a machine learning model to see if a robot could beat the world record. He describes his entire process in amazing and informative detail here,
On his first attempt he managed to spend approximately 80 hours training his AI from scratch and was able to complete the course but didn’t come anywhere near the Speedrun leaderboards. He contacted a QWOP speedrun professional to help him compile data that his machine could learn from. By tinkering with the learning parameters and having high-quality data to feed in he cut the training time in half and was able to come within a top 10 finish at 1 minute and 8 seconds as reported in Gizmodo on February 3, 2021. By February 8th, with another 40 hours of refined training, Wesley’s AI beat the world record by nearly a second.
Before I found these videos, I saw AI and Machine Learning as one of two extremes — that it’s primarily used to squeeze a few extra percentage points out of stock portfolios and that eventually the machines will have learned so much they will come to rule or destroy us all. I am clearly not an expert on either subject. But stumbling upon these seemingly arbitrary uses for Machine Learning, the entire concept was suddenly demystified and became far more approachable. Artificial Intelligence was no longer the sole domain of stock brokers and doomsday preppers but something with an incredibly diverse array of applications. Not only can we teach robots to play games with us but if the Machine Apocalypse does come it will have a killer soundtrack.