Zoho founder and CEO Sridhar Vembu has conducted an extensive, multi-hour review of C++ code generated by a leading large language model (LLM), offering a nuanced perspective on the current state of AI in software development. The exercise took place during a Zoho tech town hall, where Vembu and his team dissected code produced by Anthropic's Claude Opus 4.5 model.
AI's Proficiency in 'Glue Code' and System Integration
Vembu shared his key takeaway on social media platform X, stating that he gained a clearer understanding of what these AI models excel at. He identified a significant strength in creating what developers often call "glue code." This refers to the often tedious but necessary code that connects different systems, taking data from one source, reshaping it, and passing it to another.
"They are able to stitch together systems well," Vembu noted. He explained that while this type of integration work is not extremely complex, it is very time-consuming for human engineers. The AI's ability to handle such repetitive tasks efficiently could free up developer time for more intricate problems.
He also pointed out that the models have effectively "memorized" vast amounts of open-source code, allowing them to recall and replicate common patterns. However, he cautioned that this reliance on memorized patterns comes with the inherent risk of AI "hallucinations," where the model might generate plausible but incorrect or nonsensical code.
The Crucial Role of Human Orchestration
A central theme from Vembu's review was the indispensable role of human expertise. He emphasized that a senior engineer had to "orchestrate" the entire code generation process. When the AI model reached an impasse or became "stuck," the human engineer stepped in to guide it back on track.
"This was a very vital contribution," Vembu stated. "Without his experienced guidance, the AI output would not be useful." This highlights that current AI coding assistants are powerful co-pilots, not autonomous developers. They require skilled human oversight to steer them toward practical and correct solutions, especially when dealing with nuanced or complex logic.
Verbosity and Room for Improvement
During the review, which examined several C++ files containing thousands of lines of code, Vembu observed that the AI-generated code tended to be verbose. He suggested it might be unnecessarily long-winded, a point he intends to study further.
Furthermore, the team found that only a tiny fraction of the generated code qualified as genuinely complex. The majority was the straightforward, integration-focused glue code. This indicates that while AI handles routine tasks admirably, it still struggles with deep, architectural complexity that requires abstract reasoning and novel problem-solving.
"On the whole, I am both impressed and not super awed," Vembu concluded. "I believe we can do better." His final assessment acknowledges the remarkable progress made in just the last two years in AI coding capabilities, while also setting a clear benchmark for future development. The journey toward AI that can independently manage high-complexity coding tasks continues, with human developers firmly in the loop.
