GitHub Copilot: The Smartest AI

Last week I got the opportunity to test the new GitHub Copilot, a AI that suggests you code snippets but not only boring code snippets like completing some basic imports like Kite usually does. No, Copilot suggests you code, like the complete thing:

In the case above I wanted to write a for loop to grab data from a local database. I haven't even written the first query only the connection string and it already suggested me a function to grab the names and then the types of Pokémon within the data.

The list comprehension style didn't showed as a default (since is not the most used way) but after I re-wrote the code in this format it learned from that and started giving me suggestions to write every loop in that style showing that it can learn how to code just like you do.

The best part is that Copilot knows every programing language there is. This is an example of a HTML code, I only written the first line and once it got what I was doing it went all the way writing a simple page structure:

Oh! And if you don't like the suggestion you can hover over a bunch of suggestions by using Alt+[ so you are never stuck with one solution.

But what is copilot? Well everything starts way back with OpenAI, a San Francisco Company that turned things around in terms of natural language processing. Their latest work is GPT-3.

The graph above from Brown et al. 2020 shows that increasing the numbers of parameters rises the level of accuracy of the model. In GPT-3 case it was used more than 175B parameters to achieve its final form. This all happened in September 2020 after the company received a massive investment from Microsoft in 2019 making Microsoft the only company that can take full advantage of GPT-3 components. Then you remember that Microsoft also acquired GitHub on October 2018 and has also joined the Linux foundation back in 2016.

So you see, since Microsoft has access to all this technologies and the internet is also blessed with the many developers that wrote answers to code questions at StackOverflow is only logical that Copilot would be trained using all this mechanisms to become this great AI that we see today (if you want to see a cool video about it I suggest this one).

Even though Copilot still in preview mode and there are some bugs and limitations, this is promising but at the same time kind of scary. Is there a plan to turn all the coders jobs into Copilot generated code? Or maybe it is only the application of top technology to make our lives easier? Who knows? The only thing I know is that I completely agree with YouTube user Drogenopa1:

Everyone knows that thinking of names for variables and functions is the actual hard part...

So what do you think? Are you excited about this new amazing world of AI or are we going to be unemployed by the end of the decade?