(US and Canada) Genpact Chief Digital Officer Sanjay Srivastava speaks with Derek Strauss, Chairman of Gavroshe, in a video interview about the top challenges of applying large language models in enterprises alongside the issues of explainability, privacy, and global regulations.

According to Srivastava, one of the biggest issues with large language models is accuracy. These models can say something authoritatively while being inaccurate. He mentions that AI models need to cover the journey from being directionally correct to specifically accurate as they train on many more parameters.

Srivastava says that enterprises need to correlate the expected level of accuracy with the use cases, as being directionally correct may be sufficient for certain uses while being insufficient for others. He adds that AI models also need to improve explainability and take their ability to explain outcomes to the same level as humans.

When asked about what organizations need to do to reach the maturity for their specific use cases, Srivastava says that the right approach is to take the foundational models and further train them for specific domains, like writing a legal contract or a footnote on a drug discovery paper. They need a lot of reinforcement learning to become even more intelligent.

Speaking on the approach to avoiding privacy issues, Srivastava notes that the norms on privacy are also changing. It is much broader and includes things like security, biases, and ethics. Feeding off breached or biased data can compromise the AI system itself.

He concludes that it is also affected by different regulations and sovereignty rules and storing data in different locations while ensuring that the usage of the data is permitted.

CDO Magazine appreciates Sanjay Srivastava for sharing his insights and data success stories with our global community.

See more from Sanjay Srivastava

Data Is its Own Asset Class