Zoho founder Sridhar Venbu shared insights from a C++ code review generated by Anothropic's LLM and said he was impressed but believes there is still room for development in AI. He also suggested that LLM coding still requires human intervention, but such capabilities have improved significantly over the past two years. According to Venbu, the review lasted for hours at the tech town hall, and the result was a clear understanding of how the LLM worked.“Yesterday, we held a technology town hall at Zoho where we did a code review of C++ code generated by Claude Opus 4.5 models. It lasted for hours into the night,” Vembu said in a post on X (formerly Twitter).“We now have a clearer understanding of what these models are good at. Models can successfully stitch systems together by taking data from one system, reshaping it, and passing it to another system. These systems often have a lot of such 'glue code', and while it's not very complex, it's very tedious. ” he added.
Shridhar Vembu says AI will create 'glue code'
He also highlighted AI's strength in creating tedious “glue code” for system integration, but also noted that its output tends to be redundant and dependent on memorized open source patterns.“In general, AI models also ‘remember’ all open source, and can recall patterns from it (possibly hallucinations), and can successfully piece together different open source pieces,” Vembu pointed out.
“Complex code requires human intervention”
Vembu also highlighted that human orchestration was found to be essential during the review to ensure useful results. He said AI is good at connecting components, but struggles with complexity.“Our senior engineers guided ('tailored' is a better word) this process. When the AI got stuck, he helped “unravel” it. This is a very important contribution, and without his experienced guidance, the AI output would have been useless. ” he said.“We went through several C++ files, each containing thousands of lines of code, looking for what appeared to be complex code. Most of it was simple glue code, and only a few were complex,” Venbu said.“I suspect that the AI-generated code tended to be unnecessarily verbose, but more research is needed to confirm that. Overall, I'm impressed, but not extremely awed. I believe we can do even better,” he concluded.
