Zoho founder Sridhar Vembu has shared insights from a code review of C++ generated by Anothropic’s LLM, saying that he was impressed but thinks that AI still have room for development. He also suggested that LLMs coding still need human intervention but there has been a significant improvement in such capabilities in the last two years. According to Vembu, the review went on for hours in the tech town hall and it provided him a clear understanding on how LLMs work.“Yesterday we had a tech town hall in Zoho where we did a code review of the C++ code generated by the Claude Opus 4.5 model. It went on for hours late evening,” Vembu said in a post on X (formerly Twitter).“I now have a much clearer understanding of what these models do well: they are able to stitch together systems well, taking data from one system, reshape it and pass it to another system. There is often a lot of such “glue code” in these systems and that is not very complicated but it is very tedious,” he added.
Sridhar Vembu says AI creates ‘glue code’
He also highlighted AI’s strength in creating tedious “glue code” for system integration but noted that its tendency toward verbose outputs and reliance on memorised open-source patterns.“In general, AI models have “memorized” all the open source too and they are able to recall patterns from them (with some possibility of hallucination). They are also able to stitch various open source pieces together well,” Vembu pointed out.
‘Human intervention needed in complex codes’
Vembu also highlighted that human orchestration proved essential during their review as it ensured usable results. He said that while AI excels at stitching components but it struggles with complexity.“Our senior engineer had guided (“orchestrated” is the right word) this process. When the AI was stuck he helped “unstuck” it. This was a very vital contribution and without his experienced guidance, the AI output would not be useful,” he said.“We examined several C++ files with thousands of lines of code in each, looking for what I consider complex code. Most of it was straightforward glue code and only a tiny part of it was complex,” Vembu noted.“I suspect that the AI generated code tended to be needlessly verbose but I have to study it more to be sure. On the whole, I am both impressed and not super awed. I believe we can do better,” he concluded.
About the Author
