Deciphering the buzz behind AI pair programmers (part 1 of 4)
AI pair programmers or text-to-code tools have created a lot of buzz over the past year. While GitHub Copilot has captured a larger share of media attention for its ability to write complete functions as well as its disregard for licenses and attribution, other players like Amazon CodeWhisperer and Tabnine (founded in 2018 and have raised $30M from investors like Khosla Ventures) have also made significant dents in the market. To put it simply, AI pair programmers help software programmers write code, create unit tests or comment, hence reducing manual tasks and increase productivity.
This 4-part series covers a bit of history (my favorite part :)), a high-level overview of how AI pair programmers work (see here), the issues with these tools (see here) and ends with a question of whether Big Tech is all set to capture all the value to be created (see here).
Disclaimer: This space is evolving rapidly, and these series of posts seeks to capture the state of the market as of the first week of February 2023. I would welcome feedback on this piece if you’ve used some of these tools!
Backup.. what has happened up until now?
Artificial intelligence has been talked about since the mid-20th century and has been discussed in board rooms and investor meetings for years. As an investor in my previous life, it felt like most startups inserted the words ‘Artificial Intelligence’ to indicate code that did something complex that we probably shouldn’t talk about too deeply.
However, over the last decade, AI has advanced rapidly, primarily because of the vast availability of text data, growth of computational power, and advancements of AI research (new techniques, models etc). One of the most influential companies in the AI space has been OpenAI, a company started in 2015 by a clutch of Silicon Valley’s elite (Sam Altman, Reid Hoffman, Jessica Livingston, Elon Musk, Ilya Sutskever, Peter Thiel and others) to promote and develop ‘responsible’ AI. I put the word responsible in quotes because the definition of responsible AI is open and free flowing, atleast to me. Anyway, about four years later, Microsoft invested $1b+ in OpenAI at a $20b valuation and rapidly increased its involvement in the company, with many claiming Microsoft wields tremendous influence over OpenAI. OpenAI started using Microsoft Azure exclusively to develop and host its products, giving Azure’s cloud computing a massive boost in usage.
In parallel, OpenAI and one of Microsoft’s subsidiaries, GitHub (acquired for $7.5b in 2018), began work on GitHub Copilot, an AI pair programmer that uses OpenAI’s Codex technology (a large language model that converts natural language into code) to complete blocks of code using AI. Copilot was released to the public in 2021 and improved rapidly. In 2022, GitHub’s CEO Thomas Dohmke remarked that Copilot handled 40% of coding among programmers using the AI in the beta testing period (the year prior). Put another way, for every 100 lines of code, 40 are being written by the AI, with total project time cut by up to 55%. This was a massive announcement and inspired many people to think of a world in which the cost and time to create software code would drop dramatically, potentially giving rise to more tech products or faster updates to current products.
In parallel, another tech giant, Amazon, was quietly chipping away at its own AI pair programmer called Amazon CodeWhisperer, releasing it to the world in mid-2022. There are plenty of similarities between CodeWhisperer and Copilot — they both output entire functions of code with just a simple text prompt, they can write unit tests and comment your code with just a prompt. However, there are some key differences as well — some which don’t land CodeWhisperer in the kind of controversy that it seems to have landed Copilot. We’ll explore some of these differences later in the series, but first — how do these AI pair programmers work? The next article is here.
