Indian cos need to build research opportunities to retain AI talent: Meta AI scientist

When discussing the current AI boom, LeCun acknowledged “irrational exuberance” while dismissing the notion of an AI bubble

Limited research opportunities and brain drain are among the main challenges for India to build its own expertise in Artificial Intelligence, Meta’s chief AI scientist Yann LeCun said on Thursday.

LeCun, one of the world’s foremost experts in the area, was speaking at an event organised by Nasscom in Delhi. (Yann LeCun | Official X account)

LeCun, one of the world’s foremost experts in the area, was speaking at an event organised by Nasscom in Delhi.

“Giving opportunities for young, talented aspiring scientists to do research in India and to stay in India is important. For this to occur, young people need to have fulfilling jobs in research in India. Right now, the choice for that is fairly limited,” said LeCun, who visited IIT Delhi, spoken at Meta’s Build AI Summit in Bangalore, and met with IIT Madras faculty and students in Chennai during his India tour.

LeCun noted that research positions are scarce in India because companies haven’t yet embraced the research culture found in organisations like FAIR (Fundamental AI Research team at Meta), Google Research, Microsoft Research, or Bell Labs.

The culture of “ambitious research” in industry, where private companies actively publish research papers and maintain an open approach to intellectual property, exists primarily in the US, according to LeCun. “The best industry research in AI is all happening in the US, basically at Meta and Google,” he stated, describing a scenario in which Indian talent moves abroad.

“If you have prospects for young talented people in India, they will stay in India. They will get a PhD and stay in India. They may not get a job at a top research lab, but if that doesn’t work, they start a startup or something. That will lift the ecosystem up,” he explained, drawing parallels to Paris, where research careers are often associated with academic positions in government-owned universities.

Does Big Tech have an advantage in AI?

When discussing the current AI boom, LeCun acknowledged “irrational exuberance” while dismissing the notion of an AI bubble. He explained that heavy investments in AI stem from companies’ fear of missing out, predicting an eventual “normalisation” where funding for financially unviable companies will decrease.

“Some companies form the idea of training their own foundation models. I personally think that is a very difficult thing to do for a company that doesn’t have a large source of revenue from other sources like Meta or Google or Microsoft. If you are a startup, and want to be at the top level to create a foundation model, you might raise a billion [dollars] but you are going to run through this billion in two years. And are you going to be able to generate the appropriate level of revenue to justify your valuation within the timeframe?”

LeCun cautioned that transforming impressive technology into viable products isn’t straightforward, citing IBM Watson’s eventual failure in the healthcare sector despite initial promise. “Companies like Meta and Google have enough funds to keep funding this for a while. At some point they might get tired if we don’t make significant progress,” he said, advising startups to utilise open source models instead.

Addressing HT’s question about Big Tech companies becoming AI gatekeepers, LeCun countered, “Open sourcing is the antidote to gatekeeping.”

LeCun said Meta realised that the risks that were voiced when the company released its first large language model called ‘Galactica’ in November 2022 to write scientific papers, because of which the company took it down just after three days of releasing it, “were overblown” when it saw the superstar reception OpenAI’s ChatGPT got Just a couple of weeks later.

“We also concluded that because the world didn't end after ChatGPT that the dangers of the risks associated with those models were basically overblown. It's not that there is no risk, but they clearly were overblown,” he said.

LeCun said he envisions a future where people interact with information through AI assistants rather than traditional websites and search engines. “Of course, it is difficult to do this without any kind of bias. It is the same with newspapers. They all have a particular view of the world. ... Which is why in democracies, particularly the largest democracy in the world, it is important to have a free and diverse press, you cannot have democracy without that, despite what Musk might say. And the same thing is going to be true for AI systems.”

To achieve diversity in AI systems and prevent concentration in Silicon Valley, LeCun advocates for distributed training across the world. “The information we get from AI systems, we are going to need this to come from a free and diverse population of AI systems, systems that understand bias, languages, various cultures, value systems, centers of interest. We’re not going to have one single or maybe a handful of systems trained in Silicon Valley for the entire world. That is just not going to work.”

This vision, he explained, drove Meta’s decision to open source its foundation models, which he playfully referred to as the “herd of llamas.”