Google CEO Sundar Pichai recently revealed that AI generates more than a quarter of all new code at the technology company.
This statement sent shockwaves through the internet, raising questions about what this means for software engineers and developers as AI begins to take on more complex coding.
During the recent quarterly earnings call, Pichai said that engineers first review and then accept the AI-generated code. “This helps engineers do more and move faster,” he added.
In response to Pichai’s statement, a Google employee, anonymously disclosed on Hacker News that AI acts as an autocomplete tool, mainly capable of finishing lines.
“The code completion engine is basically just good at finishing the lines I’m writing. If I’m writing ‘function getAc…’ it’s smart enough to complete to ‘function getActionHandler()’…But it’s not doing any engineering at all,” the employee further said.
Another Google software engineer, while offering more insights, explained that much of the AI-generated code is focused on maintenance and cleanup tasks rather than actual feature development.
According to this techie, Pichai was talking about “clean-up jobs for dependencies or removing deprecated classes or changing deployment configurations”. “In terms of features, these are not feature works at all,” the person added.
There’s more.Several developers have expressed concerns, sharing mixed reactions. One such developer from Google highlighted potential long-term risks, suggesting that AI-generated code without dedicated human authorship could lead to significant maintenance and quality issues in the future.
“And now we have Google, THE GOOGLE, dumping code with no author directly into their product. What a shitshow. I have no idea what this will look like in five or ten years, but I’m not confident it will be good,” said the insider.
Another developer said while AI can assist with basic tasks, human intervention is still crucial for complex business logic, casting doubt on AI’s ability to generate fully functional code independently.
“For almost any real code, when I do let Copilot write it, I end up having to basically rewrite it from scratch or edit it to the point of being unrecognisable for anything with more complicated business logic than a basic algorithm,” shared the developer.
It’s Not all That Bad
At GitHub Universe 2024, even GitHub CEO Thomas Dohmke said developers are leading the future of AI by not only building AI but also harnessing AI tools to construct more intelligent systems.
“You’re using AI to build AI. You’ve laid the blueprint for what it means to partner with intelligent machines,” he added.
He further said that developers have embraced AI faster than any sector of the global workforce. This has purportedly made GitHub copilot the most adopted tool on the planet, with Python dominating AI programming.
Even GitHub Copilot Uses GitHub Copilot
WhenDohmke spoke about how the first phase of GitHub Copilot was built on three key pillars, he said the very first pillar is AI-infused.
“They were AI integrations at different touch points in the developer lifecycle. But if you wanted to avoid AI and build it all yourself, you could,” he claimed.
As per Dohmke, the second pillar of phase one is conversational coding. He said that ChatGPT and Copilot Chatbot have established the foundation for a “Jarvis-like interaction with code, a tool akin to Iron Man’s Jarvis, that enables anyone to create code using natural language”.
The third pillar of phase one is about multi-model. “We built with multi-model functionality. One default model for one specific use case,” said Dohmke, citing GPT-3.5 for autocompletion or GPT-4o for workspace.
Dohmke, while suggesting an impending change, said, “The partnership between developers and AI is going to the next level.” Introducing the second phase, he said that it’ll no longer be AI-infused but AI-native.
“AI is core and cannot be separated from the entire developer experience. All up, we will see a native developer workflow,” Dohmke added.
Not just for coding,GitHub is using AI to assess AI. For instance, Copilot Autofix utilises the CodeQL engine, GPT-4o, and a combination of heuristics and GitHub Copilot APIs to generate code suggestions. It builds an LLM prompt based on sources, including CodeQL analysis and short snippets of code around the flow path.
Similar to GitHub Copilot, which helps developers code more quickly, Copilot Autofix accelerates the pace of remediation. This way, security teams make real progress with the backlog of existing vulnerabilities, commonly known as security debt.
A recent GitHub survey found that 97% of developers use AI coding tools, and the use cases vary from project to project.
AI or Not, Change is Coming
While GitHub Copilot Autofix employs automated testing, red team scrutiny, and filtering to mitigate risks, experts underscore limitations in self-verifying AI systems, suggesting that relying on another AI model for review may be fraught with redundancy and cost challenges.
“It’s hard to use AI to trust AI for the same reason people often miss their own mistakes,” said David Timothy Strauss, CTO at Pantheon.
The concept of ‘AI building AI’ remains questionable. “Most of my code is Copilot autocomplete so…I could say nearly 90% of my code is generated,” said a Reddit user.
Needless to say, AI is clearly advancing in its ability to reason through complex problems. Coding, traditionally, involves two key operations: determining how to solve the problem and then writing the code to implement the solution. The latter, the actual coding part, is likely on the brink of being transformed by AI.
The same was even highlighted in one of AIM’s recent articles — ‘AI is Killing Remote Work’, along with how coding tools such as Claude 3.5 Sonnet, GitHub Copilot and Amazon Q Developer are helping developers build products faster and collaborate more effectively.
That explains how Google was able to ship NotebookLM faster with a small, agile group of developers, leveraging AI to accelerate product development.
Not only to accelerate product development, AI is here to create more jobs for software engineers, automating mundane coding tasks.
The post AI is Coding AI—Should Software Engineers Be Worried? appeared first on Analytics India Magazine.