Why we built GibsonAI

Generative AI has rapidly become a key player in software development, with thousands of companies, large and small, adopting tools like GitHub Copilot to accelerate the creation of new applications and services. However, while AI can enhance productivity, it may also negatively impact code quality, as highlighted by a recent study from GitClear, a developer analytics platform based in Seattle.

The research examined 153 million lines of changed code, comparing revisions made in 2023 to those from previous years, before AI became significant in code generation. Some key findings include:

• Code churn, or the proportion of lines discarded within two weeks of being written, is on the rise and expected to double by 2024. Higher churn increases the risk of errors in production.
• The percentage of copy-pasted code is growing faster than other changes like updates, deletions, or moves. According to GitClear founder Bill Harding, this suggests AI-generated code behaves like that of a short-term developer who fails to thoughtfully integrate their work into the broader project.

Harding explains that while AI code assistants are great at producing code quickly, they risk creating “AI-induced tech debt.” He notes, “Rapid code addition may be helpful in isolated situations or greenfield projects, but hastily added code becomes a burden for teams tasked with maintaining it.” You can learn more about the GitClear study here. This is also covered in the Geekwire article here. 

At GibsonAI, we recognize that more code doesn’t always mean better code. That’s why our focus is on delivering high-quality backend code that is production ready, efficient, consistent, and easy to maintain. Let's look at how GibsonAI solves many of the pitfalls of current crop of AI coding assistants and agents. 

• Quality: AI coding assistants powered by large language models (LLMs) often produce non-deterministic outputs with limited quality control. In contrast, GibsonAI generates code with strict guardrails, delivering results with the precision and reliability expected from a 10X engineer.
• Production Readiness: LLM-generated code tends to be generic, lacking the nuanced reasoning required for flawless execution. AI today can’t replicate human-like understanding to consistently deliver perfect code, especially in the context of a large codebase . On the other hand, GibsonAI ensures that the code works every time by focusing on the 70% of tasks that don’t require deep human reasoning. GibsonAI utilizes LLMs for simpler decisions where they excel. 
• Efficiency: AI-assisted developers often produce a large volume of code, partly due to the exploratory nature of interacting with LLMs. This can result in excessive and inefficient code. GibsonAI streamlines this process by minimizing unnecessary paths to the final output, often delivering tight and efficient code with a single command.
• Consistency: The inconsistency of LLM outputs stems from their need to handle a vast set of use cases, leading to non-deterministic inputs and outputs. Entrusting these systems with production code without human supervision can be risky. GibsonAI, however, is purpose-built as a backend engineer, designed to produce consistent code every time. It’s not a general-purpose coding assistant, and in most cases, the output is predictable based on the prompt.
• Maintenance: Maintaining code written by AI can be more challenging than even human-written code. The lack of control over AI outputs can lead to incremental degradation with each iteration. GibsonAI addresses this by managing 70% of your codebase, and carefully managing updates. This frees up developers to focus on innovation and new features. 
• Speed: Most AI coding assistants operate as LLM passthroughs, introducing latency and slowing down developers due to the conversational nature of the process. GibsonAI, however, runs directly from your command line, writing code in seconds. It’s incredibly fast because it’s not an LLM wrapper; instead, the code composition happens within the developer’s environment.

No developer wants to sit and review thousands of lines of AI generated code. But it's a big part of the job today.  Asking developers to do something they don't want to do could result in "skimming" over code and pushing poor quality code into production. This will eventually result in lower productivity gains in the long run.

Our approach to AI-driven code writing is deliberate and precise. With GibsonAI, every engineer on the team is empowered to perform at the level of a 10X engineer. Try it out for yourself by signing up at www.gibsonai.com.