When I was a freshman at RISD, there was a club fair, and I was drawn to the RISD STEAM table, where I met Carly Ayres and Sayer Pease. They welcomed me with open arms, and I started to learn about things like 3D printers and Arduino. I first learned Processing from Clement Valla, paper electronics from Jie Qi, and then Arduino from Paul Badger. I fell in love, and the fate was more or less sealed when I made my first music synthesizer. Creating a tool that output things I never thought of while making it lit up my heart, and is still why I do the things I do today.

Fast forward a few years, and I got my first job at littleBits. When I worked there, we experimented with different ways to teach kids “computational thinking.” One of the most important concepts we taught was algorithms. The best way we taught it, which we learned from Rob Gilson of the Blue School, is for the teacher to emulate a robot. Students say things like “pick up the trash” and the teacher does silly things like picking up the trash but staying leaned over, or picking up the trash can instead of the trash on the floor. The lesson is the harsh reality of traditional programming - you need to understand and model every possible condition that the program might encounter. If you’ve ever debugged a complicated if, else statement, you know what I’m talking about.

My first contact with AI (or ML, as we called it back then) was Teachable Machine. In the demo, you can record frames with your webcam, and almost instantly train a classifier (basically a ML model that sorts your input into a hand-full of possible outputs, like distinguishing a ripe banana from an unripe banana). A squishy probability bar updates in real time as it guesses which class the current frame of your webcam falls into. Watching that squishy line wiggle, and comparing it to the brutal experience I had writing if/else statements on the Arduino made something click - this is the solution to the “model every possible condition” problem traditional programming has. My roommate and friend Barron Webster, was working on it, and we had the idea that maybe I could help make a demo for Teachable Machine 2, to demonstrate how you could use it to build applications. I made a little app that let kids wire up their classifier as a controller for p5.js games, and started to see how this kind of input could be used for lots of things.

I found this to be so exciting, that I eventually found my way to Google ATAP, where I could marry my interest in AI with my interest in hardware. While I was there, I worked on various projects that attempted to give Google’s hardware this kind of “squishy” input - like knowing when to wake the display on a Nest thermostat based on whether or not the person was simply walking past it, or approaching with the intent to interact. Turns out, you could detect this with ML much better than a heuristic algorithm. At this time though, the applications of AI still seemed somewhat limited to me. How many different squishy environmental cues were really important enough for computers to perceive?

At some point, I heard about GPT-3, and applied for access. At first, I heard nothing back, so I applied again, and again, each time getting increasingly desperate. Eventually, I got access, and the first thing I did with it was play Dungeons & Dragons. The first thing that impressed me was that it understood the rules of D&D enough to actually output something that made sense at all. But what really froze me in my steps was this interaction:

Me: You’re a barbarian in a game of D&D. You find a poison rag, and pick it up, taking one damage, what do you do?
GPT-3: I gulp it down and collapse

It made me laugh! That’s when it dawned on me, that this technology was likely to be a very big deal. I remember one walk back from Trader Joe’s in SF in particular, where I was asking myself the question “What part of intelligence can’t be chalked up to noticing patterns and then generalizing them?” I could not think of one.

• Language - basically observing patterns in how others speak
• Relationships - basically observing patterns in how others react and how I feel
• Skills - basically noticing patterns in which actions lead to which outcomes
• Wisdom - basically noticing which decisions pan out well and which ones don’t
• Humor - basically noticing which jokes make people laugh and which ones don’t. No one is born funny

I figured, with sufficient data and time, eventually an AI that can play D&D and make me laugh can probably also do anything else a human mind could do. It would follow, then, that any piece of software which interacts with the real world, would want to use such a technology, bc otherwise it needs to make bad tradeoffs to fit the “model every possible condition” problem (e.g. customer service bots that result in laughably annoying experiences).

I decided that I should pivot my career to focus on it. Around that time, I had managed to get a full time offer from Google (I was a contractor at ATAP). My colleague had encouraged me to get a few competing offers, so I ended up talking to a few companies, and in the end, I decided to leave Google for Sudowrite. My friends and family thought I was crazy, cutting my salary in half for a startup, but I had reflected on my values and I thought this was the best way to learn and make a strategic bet. Nobody knew what an LLM was, and the prospect of starting a business focused on helping novelists sounded a bit like a suicide mission, but I thought that if things went the way I thought they would, this would be a big opportunity. The icing on top was that Sudowrite was looking for a designer who had designed games, and likes writing. Thank you Linda Eliasen for the intro, and thank you James Yu and Amit Gupta for giving me a chance.

That’s the story of how I got into AI.