28 October 2018 · 3 mins

A fellow student recently wrote a letter to the editor of The Tech, regarding the donation of a billion dollars establishing the MIT College of Computing. If you haven’t read the piece, you should check it out here before reading this.

I just wanted to lay down a few things that I had in mind as I read it and re-read it. So this little thing is basically a response to the letter.

I understand where the writer is coming from: there are so many pressing issues that require active research. A few that were mentioned like climate change and diseases and urban planning are all really hard problems that need to be solved as soon as possible. And to see MIT spend money on AI and computing and stuff sounds like an utter waste and a distraction away from these real problems.

These issues are very real. As someone who chugs down Clover sandwiches and refuse to eat meat. As someone who tried his best to conserve energy. As someone who reads biology books in complete awe. As someone who UROPed at an urban planning lab because its so hecking amazing — I am an AI practitioner at heart.

Why? Because AI is a tool for science.

You’re working on a method to create positive energy nuclear fusion as an renewable energy source. Awesome. You need your magnetic mirrors to bend your plasma into a specific shape in order to achieve fusion. You have several input parameters that determine the shape of the plasma, but you have no idea how they affect it. The state space is sooo huge. What do? Train an AI model on all the data you collected and reverse engineer what parameters led you to the correct shape. Sounds crazy, right? Here is a project just attempting to do that.

Solving brain diseases require researchers to study a field called connectomics, which is basically studying and creating maps of how an organisms neurons are connected. That’s a hand-wavy definition, but it’ll do for now. The primary bottleneck in mapping the brain this way is labeling images to find which slices refer to the same neurons. This is thousands of terabytes of data that needs to be labeled, typically done by getting a few grad students and some pizza in a room. What if you could automate the entire process of generating connection maps and labels? Yep, a neural network can help!

Other examples include finding materials for safer lithium batteries, finding efficient chemical reactions, developing tools to predict diabetes from retinal scans, beating humans at Jeopardy and so many more.

Yes, people in data science sometimes do go into business analytics, and no, business analytics does not help these pressing scientific issues. But the people doing AI research at MIT don’t care about business analytics.

AI is not just a tool, it’s an incredibly powerful one. My mind is blown when AlphaGo Zero learned thousands of years of human Go playing knowledge, and then rejected it, developing new insights into the game, in just a matter of days, completely from scratch. Imagine a computer learning thousands of years of what humans understood in a matter of days (of course, after throwing a supercomputer’s worth of compute).

I love how the programs and algorithms I write, teach me so much more than I could have ever discovered myself.