Zoho's Vembu warns of bias in AI models; says personalised AI is the future

Sridhar Vembu, Zoho founder and former CEO, today shared his perspective on AI models like DeepSeek and their impact on the tech world. He highlights the mysterious nature of ‘open weight’ models like DeepSeek, which he describes as black boxes.

Vembu raises concerns about the secrecy surrounding these AI models, questioning both the nature of the training data that fed them and the ‘reinforcement learning’ techniques that were used to make them ‘behave well.’

He opines that these models usually mirror the value systems of their creators. “Here "behave well" includes innocuous things like making the model polite to human users and much more contentious issues like active censorship of ideas the model trainer wants to suppress,” Vembu pens.

Vembu puts forward that while US-trained models are sensitive in certain areas, Chinese-trained models show (sensitivity) in different areas. “To paraphrase, "all models are biased, some are useful".”

He further underscores that in areas like code generation and chess, objective standards exist to evaluate models, while models have the option to prefer Javascript over C.

“This is why I would trust these models only when there are other objective ways to evaluate their output. This is true for fields like programming, computer-aided design etc,” he adds.

Vembu reveals that he would not trust AI for economic development advice, as he already finds most academic economics not that helpful. Since these models have memorised textbooks and merely echo the textbook theories, he remains unconvinced though he states this might be his own bias.

He emphasised the ultimate goal of the AI model is to be different for every person. This personal model will evolve as a mirror of an individual's thought process, he said. “Shaped by one’s values, this personal model will evolve as a true reflection of their thought process, learning and adapting like a digital mirror of the mind.”