Building In Public
Getting started...
I was recently inspired by AI DEV 25, hosted by deeplearning.ai. The conference felt like an origin point for something much bigger. The explosion in AI is only 2 years old, and advancement on the cutting-edge side of things has been blisteringly quick and deafeningly loud. This has rendered it inaccessible to anyone other than those actively working in the field. I named this blog human-in-the-loop for the AI pun (https://cloud.google.com/discover/human-in-the-loop), but also because it describes how I feel these days. In a world moving so rapidly, it’s unclear where we humans fit in, or how we can harness these tools. I am writing to demonstrate in a simple and pragmatic way how these technologies work, and how you and I can put them to use.
The path forward, coming out of AI DEV, was very clear: Build in Public. This was the unanimous advice of a panel of industry leaders. Let’s break that down:
- Why Build? To avoid just learning things. I can always tell myself, “I’ll get started on that project after one more course, when I will really know what I’m doing.” Building forces me to stop exploring and start digging in, and to find the actual limits of what I can do, not just the perceived limits generated by an endless tide of possibilities.
- Why in Public? To force it to be as good as I can. I have a tendency to not show my drafts, because they aren’t good enough for public consumption. Well, what if they were better? It is my hope that providing these updates is both entertaining for you, and inspiring for me to create the best possible work.
The next question is what to build!
Project 1: 😤WorkoutFace
One of my guiding questions in this new era is “could AI make this problem instantly disappear?” With that in mind, I think about automating some of my most time-consuming tasks. One of the largest ones, for me, is planning my fitness routine.
A typical marathon training plan has 18 weeks before race day, meaning there are 126 workouts to plug in. That means you have to determine your pace, structure your workouts, and then enter each of those into your workout tracker (if you are a running-scientist like myself). Suppose that setting up each run takes about 5 minutes of clicking around, that’s 10.5 hours of data entry into Garmin, and I don’t want to do it anymore! So, can AI help me?
I think so! There exists an emerging AI technology known as a “browser agent”, or more simply, letting AI use your mouse and click around on your computer. You may be familiar with Operator, from OpenAI (https://openai.com/index/introducing-operator/) or Computer Use, from Anthropic (https://docs.anthropic.com/en/docs/agents-and-tools/computer-use).
I will be attempting to create a browser agent for myself, which takes in a table of workouts and output those same workouts within the Garmin website.
Why choose this as our goal?
1) We know that browser agents are possible, and can be used right now.
2) This is an extremely specific task. I don’t want to have this agent surf the web for me. I just want to enter information into one specific page. This should make it easier to get it right, because I don’t need it to generalize.
3) There is an extremely clear success criterion. I know precisely what the end result should be, I just don’t want to do it myself.
Is this going to take more than 10.5 hours?
Almost certainly. AI has accelerated development, but this is a new skill for me, and it’s going to take some time. The point is to see if it’s possible, not to complete it as quickly as I can.
Why not just ask AI for a workout plan?
In short, I don’t think the quality would be very good. This is a topic for another blog entry, but AI works on information it’s ingested from the internet. The good news is there’s a ton of fitness writing to digest. The bad news is that much of it conflicts with itself, or is irrelevant to my specific case, the extremely bad news (for AI anyway) is that understanding the plan is the point. While you might just follow whatever your watch tells you to do, simply moving your body isn’t the point. This is a plan I like, and that I understand. I’m not offloading the understanding, but simply the clicking.
Is there some sort of Garmin workout upload feature that would make this much faster?
Probably! I don’t care. I want to see if I can make this work.
The next update is on Monday, wish me luck!