Speaking at this year's SXSW Sydney, Meta's VP of AI looked back at the past decade of AI innovations and compared the technology to in-flight WiFi.
While AI "experts" are easy to find in 2024, Manohar Paluri has a track record in the space that stretches back over a decade. He originally joined the Meta back in 2012 as an intern.
Back then, it was still called Facebook. As Paluri tells it, he wanted to do AI at a time when the biggest challenges involved teaching machines to see.
"To built AI models, we needed GPUs. But we were not allowed to have GPUS in the data centers because they are not efficient. So I had to order all the parts, including the latest GPUs back then which is probably twenty times slower than today’s GPUs and assemble it on my desk."
Paluri said that this strange desktop setup would inspire plenty of questions from other people working at Facebook at the time.
If you've feel like AI advocates have spent years promising you that the technology will be transformational society, only to find it falls short outside of flashy demos you might be surprised by the fact Paluri somewhat agrees.
As someone who has been working on the technology for over a decade, he's become familiar with the cyclical trend of rising and falling expectations that accompanies innovation in the space. Every time the world's AI researchers appear to make a major breakthrough, the biggest discovery seems to be just how much further there is to go.
“If you look at the AI innovation over the past three to four decades, every time we feel we are at the cusp of solving [artificial] intelligence, we [realised] very quickly that it falls short. Take this moment back in time when a computer program beat the best chess person."
Paluri said that while this was "an amazing moment" it quickly revealed new challenges that existing AI models needed to overcome.
“Take Jeopardy. Initially, we said it was a big moment but then it became an information extraction problem.”
Bizarre as the phrase might sound, Paluri compared the nascent technology to Wi-Fi on planes.
"What happened between the first and second time? Nothing has changed in technology. Your expectations have dramatically morphed from something that is beautiful and elegant and what you thought was not possible to I want more of it," he said.
According to Paluri, every innovation or breakthrough in the AI space fails to satisfy but instead feeds those ever-growing expectations.
"We want faster and better versions of it and that’s where we keep pushing the bar on whatever we expect from AI," he said.
2 min read
