Infosys AI Head Shares Real-World Struggles of Using Generative AI

Today, a host of companies are queuing up to leverage generative AI tools such as ChatGPT and GitHub Copilot to optimise code and automate business processes. However, the ground realities are a tad bit different from the easy-breezy picture we often get. 

In an exclusive interview with Analytics India Magazine, Gary Bhattacharjee, VP of Data Strategy and AI at Infosys, speaks about the challenges associated with companies adopting generative AI – the hottest technology this season. 

As per Bhattacharjee, a prime challenge for code generation is predictability. “For example,” says Bhattacharjee, “if we ask ChatGPT to write a Python code for Fibonacci’s sequence, there can be multiple ways to write it, but the most optimised version still remains a human decision. So, predictability is important in code generation.”

Bhattacharjee believes predictability becomes more important in data codes like SQL. “But on the flip side, coding is a lot more structured than natural language, so there are opportunities in data engineering and pipeline creation.”

However, he believes that code IP is a challenge that needs to be addressed, “Different people have different perspectives on it.” He explained that GPT is trained with everything they could find on the web, including open-source codes. However, he adds that other code-generation models like CodeGen from Hugging Face are trained specifically on open-source codes, and offer much better chances to companies concerned with code IP. 

Furthermore, Bhattacharjee explained that translating code can be complex because just a syntactical translation won’t give you an efficient code. “For example, when you translate COBOL to C, the way code fragments to data structures like variables are different. So, variable structures need to be translated as well.” But according to him, other code generators like the one IBM is working on, use a different kind of algorithm and different kinds of rules-based computational search method to translate code.

Infosys and Indian LLMs

When asked if Infosys is working on generative AI in Indic languages, the response from Bhattacharjee was in the affirmative. But when asked how Infosys is tackling the issue of tokenisation and the LLMs getting costlier because of that, Bhattacharjee pointed out that tokenisation is just one way to achieve results and there are other alternatives.

“So there are the techniques called Eigenvectors and Eigenvalues and things of that nature where you can deal with the token estimation in a different way, so your language model doesn’t become gigantic,” said Bhattacharjee.

He goes on to explain that there are alternative ways to open a neural network to deal with this explosion of tokens so that the developers don’t have to really train up, obviously more retrain, and the law of diminishing returns becomes much deeper and the loss prevention function becomes more frequent.

Using AI to Automate Operations & Processes 

Infosys is leveraging artificial intelligence to improve its operational and administrative procedures. Bhattacharjee said that the company uses AI in optimising and automating expense management, in accounts receivables payments, and other areas. Additionally, Infosys takes its operational efficiencies to its clients, offering AI-driven solutions to help them streamline their processes and cut costs by as much as 60% to 70%. 

Bhattacharjee said that AI can eliminate entire functions, rather than simply optimising processes, which can lead to significant cost savings for clients. He cited robotic process automation (RPA) as an example of how Infosys uses AI to eliminate human decision-making processes altogether. Bhattacharjee said that RPA allows Infosys to achieve a “labour arbitrage” by replacing human labour with robots that have almost zero costs.

Infosys is not alone in its use of AI to improve operational efficiencies. Many other companies are investing heavily in AI to automate their processes and reduce costs.

However, there are also concerns about the impact of AI on the workforce. Some experts predict that AI will replace millions of jobs in the coming years, leading to widespread unemployment. As per the Future of Jobs report, the World Economic Forum estimated that AI will replace about 85 million jobs by 2025. Others argue that AI will create new jobs and industries, and that it will ultimately be a net positive for society.

Bhattacharjee however, believes that with coding becoming more commoditised and replaced by AI, Infosys is focusing on transforming its workforce’s skill sets from coding to algorithms, emphasising higher-end mathematical understanding, such as developing algorithms for complex quant models. 

However, he acknowledges that this transformation will not occur overnight, and Infosys will continue to hire freshers. The company is partnering with premier institutions, such as IITs and IIMs, to build a strong tech funnel for talent. Infosys is also transforming its fresher program to focus more on developing skills and less on writing code.