Deciphering the buzz behind AI pair programmers (part 4 of 4)

This article is the last one of a 4-part series on AI pair programmers. The previous articles have covered the history of AI pair programmers, how they work and the issues. If you’ve read the previous articles, welcome back!

Will Big Tech capture most of the value in AI pair programming or can startups compete?

To recap, there are three major AI pair programming tools today — GitHub Copilot, Amazon CodeWhisperer and Tabnine, with more being developed. There have been a few casualties as well, with Kite recently announcing that they were shutting shop because, despite their best attempts since 2014, they had failed to find a business and product that developers and/or managers would pay for.

Tangential to the space of code auto-complete and AI-pair programmers includes players like:

1/ Ponicode — Helps developers check code, author tests, comment code v/s prompting new code; Acquired by CircleCI

2/ Deepcode — Automates the code review process (i.e., finding bugs, vulnerabilities, style violations, and more in the earlier stages of software development); Raised $5M

3/ Facebook TransCoder AI — Converts code from one programming language to another, reducing the cost and time needed to update older source code.

4/ Amazon CodeGuru — Integrates with IDEs that reads code, identifies errors and unproductive lines of code and flagging issues to the developer

If I had a crystal ball, I would love to learn who might be the eventual winners in this space. As of today, it does seem like Big Tech would be the obvious winners. They have the deepest pockets for expensive AI R&D, have the distribution (Microsoft owns Visual Studio and Visual Studio Code that capture 42% market share of IDEs), have tons of data to train and most importantly, have signaled that this space is important for them. Yikes! Even players not directly in the space might cross over soon, like Facebook with its TransCoder AI product.

However, I love a good David v/s Goliath story and would be curious to see a smaller startup find a niche (e.g., in a specific language, IDE or for a specific developer persona like an inexperienced programmer) and spread its wings to capture more of the market. If I was a founder creating a new AI pair programmer, these are the questions I might ponder over:

1/ What kind of AI model is most appropriate for the use case I am solving?

2/ Does my AI model really ‘learn’? Does it matter?

3/ How much capital would I require to build a model that understands, and outputs ‘correct’ code?

4/ How do I vet the correctness of the training data? How do I ensure that vetting data does not limit the data set or slow the training process down?

5/ What are the associated licenses and copyrights associated with their training data?

6/ How should I think about (A) Quality and (B) Security / vulnerability of the output?

7/ How would I distribute the product?

This remains an exciting space and I’m keen to see how it evolves! If you’re a user of GitHub Copilot, Amazon CodeWhisperer or Tabnine, I’d love to speak to you and get your perspective. You can find me on LinkedIn here.