Tools like ChatGPT are a work in progress; they can often get answers wrong, says Siddhartha Gadgil

CHENNAI

“Artificial intelligence tools based on large language models (LLMs) like ChatGPT are still a work in progress, and tend to go wrong often,” said Siddhartha Gadgil, Professor, Indian Institute of Science, here on Sunday.

He was delivering a lecture on ‘The quest for automatic mathematics’ at the sixth annual ‘Science at the Sabha’, an outreach programme for the public, organised by The Institute of Mathematical Sciences.

Responding to a question on LLMs, he said that while they were marvellous systems that knew “everything” and can have a variety of uses, they did not have a grip on reality. “Neither in the logical sense, nor in the physical or empirical sense,” he said, adding that there were ways to make them be right more often.

Stating examples from his field of mathematics, solutions offered to mathematical problems offered by new systems like Minerva (developed by Google Research) can be made to provide output in a format, which can be verified by formal proof systems like Lean.

Echoing similar thoughts, Hema A Murthy, Professor, IIT Madras, who delivered a lecture on ‘Carnatic music processing: a culture specific approach’, said that tools like ChatGPT can recognise patterns in data, but could not recognise what those patterns meant. Sharing her experience in developing tools to recognise ragas and other details from Carnatic music performances, she emphasised the need for culture specific approach in using machine learning.

Ajith Parameswaran from International Centre for Theoretical Sciences, who delivered a lecture on ‘Black holes through different windows’, highlighted how studying of gravitational waves have enhanced humanity’s understanding on blackholes. He said while the LIGO Lab and its sister facilities, the forerunners in the study of gravitational waves, have so far detected 100 gravitational waves emanating from different regions of the universe, they are expected to detect another 400 in the near future.

Mitali Mukerji, Professor, IIT Jodhpur, who delivered a lecture on ‘Ayurveda and modern medicine: the beginning of a molecular conversation’, spoke on the research she and her team were doing to understand ayurveda better through genomics.