In mid 2021, there was a revolution in coding. As a lazy programmer who always needed a fast and smart assistant, I was really happy to have Github Copilot in my arsenal of coding tools. So I was one of the early adapters of the whole idea of AI pair programmer.

Everything was fine with Copilot. I wrote tens of thousands of lines of code in last year and I could code a lot of projects which were impossible with a good, smart and fast pair programmers, but everything has been changed since last week I got an email from github, telling me I can’t have free access to Copilot anymore.

It was a sad moment in my life, but I had different ways of adapting and accepting the reality. First, I was thinking of paying $10 a month for a github premium account, but since I won’t use most of github’s premium options, it wasn’t a suitable solution for me. I also checked tabnine or kite as well, and those didn’t work out for me, as well.

Say hello to Severus, my new AI pair programmer!

First, let me talk about the name a little bit. I was watching Harry Potter franchise recently, and my favorite character in whole franchise is non other than Severus Snape. So I named my AI pair programmer after him. But I know you might be curious about how I made it. So let’s find out!

The language model

First, I needed a language model which could be capable of generating code. At first, I had OpenAI’s GPT-3 in my mind but I remembered that for some reasons, I can’t use it. Then, I fell for free language models. I used GPT-J and although it could understand the code, it didn’t seem a very high-accuracy model to me.

Then, I realized that Meta has released OPT-175B model. I put some of its functionalities to the test. It is a really perfect language model, but it works well when you use it as a core for a chatbot or a blog-post generator (or maybe a prompt engineering tool for Text-To-Image models) but not a great code generator.

Then, I found my saving angel. A lot of open-source engineers and enthusiasts of the world and it’s non other than BigScience’s BLOOM.

Code tests and inference

Like what most of you may have done, first I tried to complete a love story with the model. It was cool. Then I tried to create a friendly, a helpful, an idiot and an evil chatbot with the model. All worked out perfectly. Back then, I did not have any limitations to Copilot, so I didn’t care about the code generation.

When I found out myself in misery of not having my beloved AI pair programmer, I tried some basic python code generation with BLOOM. It was fine, then I have tested PHP, Ruby and JavaScript as well. I found that it works pretty well, so I have decided to write a simple inference code over the API.

Code generation may go wrong

Since I didn’t fine-tune the model (and I don’t have resources to) it may glitch sometimes. For example, when you don’t really pay attention to your code formatting, it might generate explanation of the code.

For me, what happened was that it started explaining the code in a tutorial format (and I bet the whole python codes were from towardsdatascience website since it had pretty similar literature).

In general, I may need a solution for this, as well.

Will it be open source?

Yes. At least it’ll be partly open sourced in near future. But more than being open source, it will be free (as in non-paid) and I guess it may be a pro for the tool. I haven’t even paid a single penny on the model, so why should I make you pay for it? By the way I will be open for donations and technical helps from the community.

Future Plans

• The API
• VSCode extension
• A community website (or discord server)

Conclusion

At the end, it seems we have a lot to do with these brand new language models. I found my way to create a free, reliable and smart AI pair programmer and of course I need some help in this way.

I have to warmly thank you for the time you’ve spent to read my article, and I openly accept your comments and ideas.