AI has not plateaued, we are on an exponential curve, says Anthropic’s Niki Parmar

Niki Parmar, Tech Staff Member, Anthropic

Artificial intelligence (AI) is far from reaching its limits and is instead advancing at an exponential rate, according to Niki Parmar, member of Technical Staff at AI research firm Anthropic.

Her comments assume significance as many experts said the AI arms race seems to have plateaued at the beginning of the year as major AI firms such as OpenAI, Google, Anthropic, and other AI giants realised that bigger models, more data, and faster computing power aren't delivering the results they once hoped for.

“I think it’s not true that AI has plateaued. In fact, I would say that we are on an exponential curve,” Parmar told Moneycontrol on the sidelines of the second edition of Mumbai Tech Week on February 28. “We just don’t know where in that curve we are. The push on the technology side will continue to happen.”

Parmar is a former Google AI researcher, and has founded two AI companies -  and Essential AI. She has even co-authored the much spoken about “Attention Is All You Need” paper.

She pointed to new capabilities emerging within AI, such as the "Computer User Guide" concept, where models learn to interact with computers. Simultaneously, efforts are being made to make these systems more efficient, ensuring they can operate at scale with reduced computational demands.

Open-Source AI? Not Quite

Meanwhile, Parmar also weighed in on the ongoing debate about open-source AI, arguing that the term is often misunderstood. Many experts have said that AI will sooner or later be open-source and all efforts to contain it within a “blackbox” will result in a failure.

“Nothing is truly open source,” she said. “What we have is open weights, but AI itself is an entire engine where it’s about the people, the compute you have available, and all the research and innovation that goes into it.”

She added that while some AI models have been released with open weights, transparency remains limited, particularly regarding the datasets used to train them.

“No one really shares what data has gone into each of these models,” she said, adding that this lack of disclosure makes true openness elusive.

The Compute Bottleneck

Beyond open-source discussions, Parmar pointed out that access to high-performance computing resources remains a key challenge in AI development. “The commodity that is more constrained is the compute side of things,” she said.

A good amount of graphics processing units (GPUs) or a lot of accelerators to actually be able to run models at scale.

As AI models grow in complexity, demand for advanced hardware continues to outstrip supply, making access to compute a crucial factor in defining AI’s future.