In 2022, my relationship with artificial intelligence changed permanently. In the past I’ve used generative models for small, experimental projects: using style-transfer to make QB Matt Stafford look like a stellar nebula, generating semi-coherent text where Bilbo says “second breakfast” 100 times in a row. These models were either too cumbersome to train or they broke down well before producing any useful output. This year I started producing work that I cared about with models that I paid for. After the success I had with these models, I feel compelled to find new models and new ways to leverage them. They have opened up possibilities that seemed absurd just months ago..
Coding with AI
I’ve long been a fan of code generation tools, but using a Larg Language Model (LLM) to generate code is a revolutionary change from generating missing function stubs, or compiling .proto files into serde classes. Instead of writing everything character by character (getting hints every so often), I can simply prompt the LLM with what I want the class, function, or test case to do, and it will generate the code for me. The results are almost always useful. Often all I need to do isI will need to adjust the prompt slightly or rename some variables. For highly formulaic code (unit tests), the model will often generate dozens of lines of code that need no modification.
The impact on my productivity has been dramatic. Not only can I “produce” code much faster, but I can do so at a lower mental load. This means that I don’t fatigue as quickly as I would working character by character; allowing me to work longer and keep my mind on higher level concerns, rather than getting bogged down in syntax or low level semantics.
I plan on using tools like this in perpetuity. I’m not giving my time back.
Christmas Card
Our family had some major changes this year and I wanted to give my wife something special… so I bought her some nice earnings. Those were a hit, but I also like to do something creative. Using a generative model at work seems almost natural, but creating a Christmas card with the help of AI … not so much. I decided to use a combination of generative and style transfer models to stylize a family photo. I worked with several different models and modification techniques, and ultimately generated a large number of images. Most of them were unusable, and a few were disturbing. However, there were a couple that were 90% cute, and I decided to use those as the basis for my Christmas card. I cleaned up the images in Gimp to remove any elements of cosmic horror that might have caused existential dread, and then printed out some 4x4 copies, glued them to an actual Christmas card and added a heartfelt handwritten note (which I did not use a chatbot to generate!). Though it was a small gift, I was happy with the final result and was proud that I could produce something worth giving.
Cleaning up the bots messy work
As you may have guessed, I used ChatGPT to generate this post from an outline. Mostly. The results that the model yields have some … let’s say limitations. The models don’t have an internal model of the world, so there are some logical inconsistencies. I also had to remove several section openings that looked like this:
As a software developer, I have always been interested in finding ways to increase my productivity and efficiency. That’s why I was excited to start using a large language model (LLM) at work in the summer. Woof.
Or.
As someone who is constantly generating ideas and looking for ways to turn them into something more concrete, I am always on the lookout for ways to leverage the tools and resources at my disposal. Barf.
Just boilerplate LinkedIn fluff. So the voice sucks, and there is 0 humor. I had to cut out, clean up, and reword quite a bit. So are these models useful? Well, I was actually able to finish the post in a couple hours, but I’m also just kinda jazzed up about the topic right now. Will will they help when I’m not brimming with enthusiasm? My guess is that the utility for the models I used for creating this post, coding at work, and creating the christmas card are on the same continuum: probably faster, way way faster, would have been impossible. I’m choosing to grasp this power.
Power to Grasp
Using these models to produce something I couldn’t before has my mind turning. What other things can I pursue that I couldn’t before? Can I use these tools to make up for my ignorance of market research and poor design skills to bootstrap product ideas that I cook up? Thousands of pages of notes and outlines collecting dust on my hard drive and producing something worth reading? Can I pursue some of these goals without destroying my health and losing my mind? I don’t know, but for the first time in a while the limits of possibility don’t feel like they are contracting quite so fast.