Browse
Log in / Register
Share
May 14, 2025 — Zoho co-founder and CEO Sridhar Vembu has stepped into one of the software industry's most divisive debates, urging engineers not to surrender their technical judgment to AI coding tools. In a sharp post on X, Vembu told developers to use large language models but "never cede our understanding to AI." His warning came with a memorable metaphor: the industry is "driving too fast in a fancy new car we barely understand how to drive, inviting disaster." Vembu was reacting to a viral post from an engineer describing a workplace where AI effectively ran the entire development pipeline.
The original post, written by a software engineer at a large company, claimed that Anthropic's Claude Code now produces everything — the specs, the code, the tests, the product requirement documents, the tickets, and even the resolutions to those tickets. "Nobody knows anything here," the engineer wrote. The post described 12-to-13-hour workdays spent mostly prompting AI models, with little time left to review generated code or understand how different parts of the system connect. "People are working 12 to 13 hours a day just to press enter," the engineer said. "There is no sense of victory. Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs. It is so soul-sucking."
Vembu's reply was blunt: "This is sad. I tell our engineers to use AI but never cede our understanding to AI." He then broadened the warning to the entire sector, arguing that the pace of AI adoption has outpaced genuine comprehension of the technology and its consequences.
Vembu's caution is supported by a growing body of industry research. GitClear's 2025 "Coding on Copilot" analysis — which tracks more than 900 million lines of code — found that code churn, the share of new lines reverted or rewritten within two weeks, has climbed steadily since 2020 and reached record levels in 2024. Duplicated code blocks are also on the rise. Google Cloud's DORA 2024 report, widely considered the benchmark for software delivery performance, found that even as AI adoption in development organizations more than doubled, delivery throughput fell 1.5% and delivery stability dropped 7.2%.
The pattern points to a counterintuitive result: AI can make individual developers look faster while quietly degrading the overall system. When no one fully understands the code, problems do not disappear — they accumulate into hidden rework. Vembu's fancy car metaphor captures the dilemma neatly. Accelerating is not the same as driving.
Vembu is no AI skeptic. Zoho has built its own family of large language models, trained from scratch — an unusual move for a bootstrapped software firm — and its Zia assistant is woven into the company's entire product lineup. What Zoho has resisted, Vembu argues, is letting AI replace human comprehension.
The company's engineering culture is famously contrarian. Vembu runs much of Zoho's research and development from Tenkasi, a rural town in southern India. The company hires talent from small towns, trains newcomers through its own educational programs, and takes a deliberately long-term view of craft. For Vembu, an engineer using AI is like a craftsperson using a power tool: the tool speeds up the work, but the builder still owns the final structure. The moment the tool starts making decisions without the craftsperson understanding them, the integrity of the build is at risk.
The clash over AI coding has split the industry's leadership. Executives such as Amazon Web Services CEO Matt Garman have predicted that AI will soon automate much of the coding itself. Vembu's response is a counterweight: faster output without deeper understanding is a liability, not a triumph.
For most teams, the practical answer lies in guardrails. Engineers need meaningful human code review. They need time to own architecture decisions rather than merely approving generated output. And in regulated sectors — finance, health care, infrastructure — the stakes are even higher. When a system fails, "the AI wrote it" is not a defensible answer. Accountability still belongs to people.
Junior developers are especially exposed. Learning to write software by reading model-generated code is like learning to drive while watching autopilot: patterns look normal until something unexpected appears on the road. The instinct to debug, to trace causality, and to reason from first principles is built through struggle. If junior engineers never struggle, they never build those mental models — and the industry loses its next generation of architects.
Vembu's message does not reject AI; it recalibrates its role. His rule — use AI, but never hand over understanding — offers a working philosophy for developers who feel caught between pressure to ship and pride in their craft. The industry may be thrilled with its new sports car. But the driver who still knows how the engine works is the one who survives the inevitable breakdown.
