$200 Tasks
In 2013, inspired by GitHub’s octocat, I too wanted a mascot. It wasn’t something I exactly needed, but I thought would be interesting to have. I had come across a freelance designer, and decided to splurge and commission a design. The designer did good work, and I was happy with the results. Overall, it cost $200.
There are many examples of these $200 tasks: a professional headshot, text transcription from audio, translation of document, or writing an article. They tend to require some level of specialized skills or equipment, are transactional in nature, and have well-defined units of output. They are the types of tasks that fit well in the side hustle ecosystem or a platform like Fiverr.
What AI can do
Current AI technology now happens to be sufficiently capable of performing $200 tasks. Today, I can ask about any AI tool to help me create a mascot. I would certainly need a lot of back and forth, with the design, and I would need some last mile fixes and touch up. While I’m not exactly happy with the results, if I had these tools available in 2013, I might have just went with the generated AI slop.
When we performed participant interviews for research studies, we’d need to pay grad students or services to manually transcribe text from participant interviews. Now, we can simply get these from AI transcription services. Similiar services exist for headshots, text translation, and article writing, and more.
Not every task is paid, but when we think about what AI can excel in, they tend to be useful when tasks are transactional and have well defined outcomes. Let’s think of something we’d not typically pay for. In a context such as software development, programmers will inevitably experience a problems in their code. As demonstrated in our recently published study, GitHub Copilot, when given appropriate contextual and conversational structure, can be effective at helping programmers with localizing faults and helping propose resolutions to fix bugs. Across all my years of programming, there are certainly some bugs I would have paid to fix.
Landscape designer
We can also think of other professional services that are slightly more costly. Tasks that require more contextual and deep knowledge, good taste, and handling ambiguity. For example, create a landscape design:
When I met with a landscape designer, I was amazed by her ability to walk through the property, immediately recognize species of trees and plants, identify aesthetic and functional issues, such as water drainage, and visualize possible designs that could integrate with our high-level goals. This was coupled with intimate knowledge of local plant species, including typical height, density, sun and water needs, and timing of flowering seasons.
What AI can’t do, yet
Unlike creating a quick mascot, for tasks like creating a landscape design, there is simply too much of a gap in expertise and quality for either AI or me to replicate.
Unsurprisingly, when it comes to these more complex tasks outside of benchmarks, AI stumbles. For example, Devin bungling programming tasks, Operator stumped by web popups, ChatGPT incorrectly answering Benedict Evan’s How many people were employed as elevator operators in 1980, and my personal favorite, failing to create a complex excel template from an image.
No doubt, the progress we’ve seen in AI so far is amazing, but economically, we’re just beginning to scratch the surface of being able to solve tasks we’d pay for.