Redundancy alert: Here’s how AI assistants are threatening Indian software code factories
Summary
- Generative AI advances in software code generation are about to reach a level that’ll leave only project oversight roles for human engineers. Is India’s IT industry ready for such a dramatic shift?
Software programming involves two distinct efforts: first, an effort to write code that will ‘compile’—in other words, it must be syntactically and logically accurate so that when it is run through its first tests on a computer, it does not result in an ‘ab-end’ or abnormal ending, thereby forcing the programmer to go back and look over the code again for, say, a missed comma that may have caused the ab-end.
The first wave of AI coding assistants has become proficient at producing code that meets this standard, thanks to training on billions of code snippets that enabled them to learn the structural syntax of various programming languages. Ask an Indian software engineer what frightens them the most right now, and many will say Generative AI coding.
The second effort is to ensure that the entire set of programs being written for a business function (say, for accounts payable) performs the process without error. A syntactically perfect piece of code could still fail to do the job it was designed for. Even in the old manually intensive model, the real challenge lies in functional proficiency. One could make a perfectly moulded phone cover for an iPhone while the requirement is for a Samsung device. The cover may be flawless, but it won’t fit your handset.
Also Read: DeepSeek’s R1 may be the first of many AI super-apps to come
Developers trying to auto-generate code are increasingly relying on and paying for general-purpose AI chatbots—such as Anthropic’s Claude, Microsoft’s Copilot, OpenAI’s ChatGPT and Google DeepMind’s Gemini—for assistance, while a wave of code-generation startups is entering the market.
Until now, almost all of AI’s gains have come from solving the first problem of syntax, easing the creation of ‘compilable’ code. The limitation of current models stems from the nature of their training data.
While they have absorbed vast amounts of code from online repositories and can get their syntax right, they lack insight into the thought processes that lead human developers to structure their code in a particular way for a business function. The GenAI goal now is to create models that don’t just replicate good code, but understand and emulate the reasoning behind it.
Now, new players, some valued at hundreds of millions or even billions of dollars, are vying for dominance in the second space—functional correctness.
Also Read: Nvidia has leapt back into the AI game as panic over DeepSeek ebbs
The potential to monetize AI through code generation is significant, and these companies expect to define the next generation of these tools. These aim to not only aid humans with code completion, but also create prototypes, test and debug code autonomously.
A key factor in achieving this is contextual awareness. GenAI coding assistants need to determine which sections of an extant code-base and other business function information should be referenced for a specific task.
Startups like Zencoder are addressing this challenge by using expertise from search engine veterans to develop tools that analyze large code-bases and extract relevant context. Similarly, Cosine focuses on human decision-making by collecting data on the actions taken by developers as they navigate complex tasks. By mapping these steps and the sources they rely on, the company is building a synthetic data-set designed to train AI models to mimic real-world coding workflows.
Another emerging approach is reinforcement learning from code execution (RLCE), which builds on the principles of reinforcement learning from human feedback (RLHF).
Also Read: Hang on… Did Microsoft just admit that AI could dumb us down?
While RLHF is used to refine chatbots by training them to generate text that aligns with human preferences, RLCE seeks to enhance AI coding assistants by prioritizing outputs that successfully execute their intended functions. This method aims to use human intervention to improve the reliability of AI-generated code by focusing not just on its structure, but on its ability to perform the business function correctly when the code is run.
Yes, scepticism abounds about GenAI’s evolution to acquire human-like capabilities. Writing code demands rigorous logical reasoning, and despite their growing sophistication, large language models rely solely on statistical pattern matching and predicting the following word (or, like DeepSeek, the following phrase). This is not true problem-solving.
However, as software systems grow increasingly complex, with billions of lines of code, the reliance on AI tools to maintain and expand these systems will increase, shifting the bottleneck from code writing to the speed at which humans can review machine-generated output. As a result, the role of developers may shift from writing code themselves to supervising and refining AI-generated output.
Also Read: Jaspreet Bindra: The K-shaped trajectory of AI offers India a big opportunity
India’s software industry, long focused on producing meticulously reviewed code within rigid frameworks, may struggle with this shift. Developers trained to work within strict guidelines to create syntactically correct code may find it challenging to transition into supervisory roles that require a holistic understanding of software architecture and business functions.
Many companies will reduce their developer headcount as AI streamlines the coding process. There will also be a polarization of software engineering roles.
At one end, a few elite engineers with exceptional architectural and business function skills will command fat pay cheques as overseers. At the other, leaner teams of, say, a dozen low-level engineers generating code using AI assistants will replace the vast teams once needed for large projects. An entire era of coding armies is ending. Are we ready?
The author is co-founder of Siana Capital, a venture fund manager.
