“I’m proud that this is one of our fastest growing markets. One of our strongest teams worldwide are based here (Microsoft India Development Centre, Bengaluru) and in Hyderabad working on all parts of our stack. There are extremely talented engineers and developers here,” Suleyman told S Krishnan, Secretary, Ministry of Electronics and Information Technology (MeitY), in a fireside chat on the future of AI and its possibilities.
On Krishnan’s question on emotional quotient (EQ) and whether users will be able to keep anything private with AI, Suleyman said that we’ve never been able to create technologies that are principally about EQ.
“We’ve always been focused on their factuality, whether they have real time information, and whether they’re bug-free. Now, we have a challenge as creators. What is the emotional tone of something like this? What are its values?” he said.
He said that trust is about boundaries. One of the things that he spends a lot of time thinking about is what AI companions do not ask users about, he said.
Discover the stories of your interest
“What does it not remember? What does it refuse if you ask it to do something or talk about something? Those boundaries are how we establish trust in society, and I think it is super important that we proactively push that conversation out into the social world,” Suleyman said.He added that the company is increasingly involving social scientists, psychologists, therapists, script writers, and comedians to get diverse perspectives.
“That’s an opportunity for us to synthesize more diverse perspectives and get a broader picture of people involved in the design iteration process,” he said.
Just as the web puts information at everybody’s fingertips, AI is going to put knowledge at everybody’s fingertips, he said. “Synthesised, distilled, personally tuned, to the way you want to learn and use information. And that applies in the workplace as much as it applies at home,” he said.
Microsoft Copilot is a generative artificial intelligence chatbot developed by Microsoft. Based on the GPT-4 series of large language models, it was launched in 2023 as Microsoft’s primary replacement for the discontinued Cortana.
Microsoft 365 Copilot already does an incredible job of reasoning over work data, he said. It can provide citations for any question that you would ask of it. Referencing your email, calendar, excel sheets, docs, company’s HR records, and supply chain information, he explained.
“So, this lowers the barrier of entry to making use of knowledge,” he said.
Making an important and valuable contribution to the workplace for many knowledge workers is fundamentally about getting access to useful information that you can act on, Suleyman said. “That is going to have a profound economic benefit for many of our industries,” he said.
Voice is the ultimate way to make these tools accessible and available, he said. “So, investment in languages and translation is what I think governments should be making,” he said.
MeitY’s Krishnan said that India is investing $1.3 billion in the India AI Mission. About $500 million goes to providing access to AI compute, he said, about the government’s plan of procuring 10,000 graphics processing units (GPUs). This is to enable private facilities, startups, and researchers to build at a subsidised rate, he explained.
Suleyman said that many startups and businesses are aware of large sets of data required for post training and fine tuning. Asked about what Krishnan’s thoughts are on making more government data available to help spur that innovation, he said that MeitY has an open data platform where various kinds of government databases are shared.
“Under the India AI Mission more anonymised datasets will be made available,” Krishnan informed Suleyman.
Pretrained models are going to be largely commoditised, asserted Suleyman. They will be widely available via APIs and open source, he said. The data that will be required for post training, or the last stage of training to adapt the model to a specific use case, is actually very small, he revealed.
“You only need a few hundred thousand examples of the good behaviour that you’re trying your model to imitate or learn from at the post training stage,” he explained.
“So, I expect to see hundreds of thousands of different types of agents with different types of expertise both linguistically but also (with) knowledge and grounding over different types of databases and knowledge corpuses. I expect to see many different types of those emerge over the next few years just with a small amount of high-quality training data,” Suleyman said.