Introduction:
Jay Hack, an AI researcher with experience in Codegen computer vision and natural language processing, realized a few years ago that by converting natural language requests into code, large language models (LLMs) — think ChatGPT or OpenAI’s GPT-4 — could increase developer productivity.
Following his time as a machine learning engineer at Palantir, Hack founded and sold Mira, an AI-powered cosmetics shopping firm. He then started experimenting with using LLMs to carry out pull requests, which merge new code modifications with the central project repository. Hack gradually developed these tests into a platform called Codegen, which uses LLMs to automate as many repetitive and mundane software engineering activities as possible with the assistance of a small staff.
What distinguishes Codegen from other code-generating AI systems, such as Salesforce, Amazon CodeWhisperer, and GitHub Copilot, with which it is named? First of all, according to Hack, Codegen faces difficulties. Codegen handles “codebase-wide” concerns like massive migrations and refactoring (i.e., restructuring an app’s code without changing its functionality), while Copilot, CodeWhisperer, and other tools concentrate on code autocompletion.
Hack says, “Codegen uses a multi-agent system for complex code generation.” This means directing a group of agents to work together to break down and finish complicated jobs. Many LLMs carefully consider and improve upon one another’s work, producing noticeably superior results.
The primary offering from Codegen is an on-premise and cloud-based technology that automatically creates pull requests to handle support tickets by connecting to codebases and project management platforms like Jira and Linear. The writer could not understand what Hack meant by “infrastructure,” but Hack claims that the platform can even set up part of the required code infrastructure and logs.
Hack stated, “Codegen offers a higher level of automation in executing entire tasks on behalf of developers, in contrast to other solutions.” “We harvest tickets from a company’s backlog, identify those that can be resolved, and then mobilize a horde of agents to locate the necessary code and generate a pull request.”
Now that even the most robust AI models available today make significant errors, Codegen is making a lot of promises. For instance, it’s commonly known that generative coding tools can result in insecure code; a Stanford study indicates that software engineers who utilize code-generating artificial intelligence are more likely to incorporate security flaws into the programs they create.
According to Hack, Codegen attempts to find “the right balance” between best practices and human inspection when looking at LLM-generated code.
Codegen Raises New Cash to Automate Software Engineering Tasks:
Codegen Raises New Cash to Automate Software Engineering Tasks (Image Source: techcrunch.com)
“This is significant work, and a better understanding of assessing and validating LLM output would benefit the entire development ecosystem,” stated Hack. “For generalized, automated code generation systems to enjoy widespread developer trust, significant advancements are required.”
For its money, investors believe that Codegen is on the rise.
At the onset of this week, the company announced the accomplishment of securing a $16 million seed funding round spearheaded by Thrive Capital. Noteworthy angel investors, including Mike Krieger, co-founder of Instagram, and Adam D’Angelo, CEO of Quora, also contributed to this round. Following this investment, Codegen has amassed a cumulative total of $16.2 million, as reported by Hack, consequently placing the company’s valuation at $60 million post-debt.
“Most developers still spend an unreasonable share of their time writing code to deal with low-level tasks like migrations, refactors, integrations, and bug fixes in 2023,” said Philip Clark of Thrive in an email. Businesses such as Codegen use LLMs to create AI agents that relieve engineers of this tedious work. Soon, developers can delegate work to agents, freeing them up to concentrate on developing new products rather than worrying about labour-intensive software.
Based in San Francisco, Codegen is incubating the platform with two “large-scale” enterprise partners; it does not yet have any paying clients. However, Hack is expecting development in the upcoming year.
In addition, he stated that Codegen intends to expand from six to ten personnel by the end of the year. “We’re raising significant capital as the opportunity to make such a substantial and ambitious product has only recently emerged,” he said. “We will scale our workforce and support our infrastructure with the funds.”
