Over the last few weeks I’ve begun using the recent crop of AI-powered services in my daily work and I’ve found them to be an enormous boost to my productivity and fun to play with. I do not know if these human-like chat services will end up causing great harm to the earth’s population or not — but as of today, I think they can be both useful and entertaining.
I’d like to start off this post by describing what I think AI is. Not because I think there is any confusion on the matter but because I want a definition I can point people to. A service that uses a bunch of learned prompts and responses to use various APIs to present you information in a human readable or audio format are not what I would consider AI. Think Siri. Siri, while it may very well have portions of it that are using AI-data sets, I do not consider AI. Siri feels much more to me like a gargantuan list of IF THEN statements.
IF you ask Siri for anything relating at all to the weather THEN it will pretend to be a human and answer in a human way. “Hey Sir, will it rain today?” “Yes, there is a 70% chance it will rain today.” is a good answer but far from being an intelligent one. If you ask “Hey Siri, what time will the rain start and stop today?” it will almost certainly fall on its face. In fact, Siri falls on its face more often than not.
IF you ask Google’s assistant a similar question it does a far better job of trying to deliver you the answer you want. For instance, you can “OK Google” a fairly complex mathematics query and, in my experience, it does a good job. Where Siri is like a very young child trying to respond to adult questions, Google is somewhere in the teenager area. But a teenager that does not know how to use Google, read the first few links, and summarize an answer.
I’ve heard that Alexa is even better than Google but I have almost no experience using Alexa personally.
Siri, Google, and Alexa are not even close to what I would consider AI.
ChatGPT, which is the AI-powered chat service I’ve been using the most, is most definitely intelligent. And it is improving rapidly. My guess is that in a few years ChatGPT 3.5 and 4 will feel like Siri does today. Which is a little mind blowing to be honest.
I have a few ways in which I’m describing ChatGPT in conversation with those that do not understand large language models, data corpuses, neural nets, or RLHF. To be clear, I do not understand those things either. But I can at least navigate their broad concepts. I think. Maybe I should ask ChatGPT about that? Anyway, here are the analogies I’m currently using.
- Imagine doing a Google search (or, Duck.com because you’re a sane person) and you read the first 1,000 links in totality. And then you distill what you’ve learned into a single paragraph. In 5 seconds.
- Imagine you could ask someone with 25-years of experience in a given field (programming, music, etc.) a question or to do a task for you and they were able to complete it in a few seconds.
If computers are a bicycle for the mind, AI is an e-bike for the work you do on that computer. (I’m not settled on this metaphor. It’s a WIP.)
I’m using ChatGPT, DALL•E and their ilk for a variety of my daily tasks. It takes a little while to retrain my muscle memory to start at ChatGPT and go from there, rather than starting from scratch, but I’m slowly getting there.
I’d like to describe just one use case from this past week where I used ChatGPT alongside my manual process and the obvious benefits it would have on my work. This isn’t a use case you haven’t heard of as many, many others are doing the same. But I will say it impressed me nonetheless. I used ChatGPT to write some JavaScript and it did remarkably well.
At work we offer our customers something we call company stores. A company store is very much like an e-commerce store – only we have most or all of their inventory in our warehouse and we can deliver it to their locations the next day. Anyway, on many of these stores there is a large rotating banner area on the homepage. We wanted to hide this rotating banner on all other pages. Since the software we use wouldn’t allow us to do this on the server side, we are left with the choice of doing it client side. This means that I was going to write a simple bit of JavaScript to 1) determine whether or not the current page was the homepage, and 2) if it was not the homepage, hide the rotating banner.
This isn’t a difficult piece of code. Anyone with any amount of JavaScript experience could write this code in a few minutes. But that is just it. You can’t copy and paste this specific code from a random result from a search engine. You need experience and then you need to write this code. I started this project from scratch and it took me about 15 minutes or so to write code that worked for our purposes. I would call the code OK not great.
I then asked ChatGPT to write the code for me. My prompt was “In Javascript hide a section element that has a data attribute of data-name if the URL is anything other than XXX” (where XXX is the URL of the homepage). The code it returned was better than my code. The way I would describe it is, if I had 2 hours to refine the code that I wrote in 15 minutes, I would have written the same code that ChatGPT wrote the first time in a matter of seconds. I did have to make one very small tweak to the code to run it on production – but it was a tweak there was no way ChatGPT would know to make.
It felt like writing code without using your hands. More often than not, the code I’m about to write is fully realized in my brain and I just need to type out the characters onto my computer. I may make a few mistakes along the way, or need to search for a specific syntax that I had forgotten about, but the rest of the code is menial work to type out. But if I start out with the idea of the code using a detailed prompt to ChatGPT I don’t need to use my hands to write the code. I save a huge amount of time.
This experience and many others (which lead to hours being saved not just minutes) leads me to believe that for many things that I do day-to-day; write, interpolate data, program, etc. I should be using ChatGPT as an accelerator.
My initial optimism for crypto, which has all but completely faded away (the entire industry feels like a squandered opportunity at this point), has me reticent to excitement about AI. And I do not know how to predict the next several years of “improvements” to this technology. But as of today, it certainly feels both useful and fun and I plan on continuing to use it to get work done faster.