One Tweet started it all:
Jordan Singer, a designer at CashApp, built a plugin for Figma (a web design platform) to generate designs based on text description. This completely blew people’s minds, and fueled a whole torrent of GPT-3 demos on Twitter.
You see, this was one of the first good demonstrations of GPT-3’s “code” generating capabilities*. The demo above looks eerily similar to commanding an AI henchman to do your work.
In fact, it’s fascinating that GPT-3 is able to generate running code at all. With no manual training, it can spit out bits of code that are (mostly) correct.
GPT-3 knows how to produce code because much of its training data contains working code from the internet. So it knows what working code looks like, and it’s able to whip it out at whim.
Once people realized this capability, they understood that text generation has broader implications than Harry Potter parodies in Hemingway’s voice
Yet, as much as code-generation has fueled GPT-3’s popularity, I don’t think this will be a crucial use-case.
At least, not yet.
So don’t worry developers, you’ll get to keep your jobs.
Don’t get me wrong, the demos are really cool. However, code generation with NLP is still quite new, and it has a long way to go before becoming generally useful.
The issue
The biggest issue right now is that its not good enough for non-developers to use yet.
Some people have described these sorts of applications as a new form of “no-code” tooling. No-code platforms are graphical interfaces that allow non-developers to build applications and websites.
Most of these are not no-code tools, they’re fewer-code. GPT-3 is still spitting out actual lines of code, which is gobbledygook to non-developers.
Even worse, those output lines are often wrong! You still need a human to look over and manually check things.
As of now, these tools are meant for developers, and I think that most of the creators would agree.
The creator of Tailwind, a GPT-3 CSS generator, said
The way I imagine this tool being used is sort of a smart snippet or a code assistant that can save you time writing boilerplate code.
Honestly at this point, I think we’ll see marginal efficiency boosts at best. GPT-3 is only really capable of generating small snippets of code at a time, like a button or two in HTML.
GPT-3 at its core is a time saving tool. If you can use it to make people more efficient, then you are saving them time. If you can save them time, you are saving them money. People pay for that stuff.
I just don’t think you can save lots of time with this use case. I’m not saying there are no viable opportunities. I just think it pales in comparison to the other, perhaps less interesting use cases.
That being said, if you think I’m dead wrong, DM me on Twitter. I want to know what other people’s thoughts are on this. It is very possible that there are niche areas I haven’t thought of. I’m only human, after all :)
P.S. Sharing is caring. If you would like to help me in my mission to never need a job, hit that button below.
* Jordan’s plugin does not use GPT-3 to generate code directly. GPT-3 generates a high-leveled representation of the Figma design, which then gets translated into code. However, GPT-3 is capable of generating code and there are many people who have been showing off similar use-cases