AI scientist Fei-Fei Li: ‘Maths is pretty clean. Humans are messy’ –
AI scientist Fei-Fei Li: ‘Maths is pretty clean. Humans are messy’ · If we [academic AI] lose that centre of gravity, then the other centre of gravity
…
link
As I wait for Fei-Fei Li to meet me for lunch, groups of students begin queueing for seminars around me. Through glass walls I can hear the squeak of marker pen on whiteboard and see the furrowed brows of a dozen eager scientists in the making. I feel a nostalgic dread as I await the arrival of a professor and wonder if I am adequately prepared.
The feeling quickly lifts after Li arrives at the nondescript café, here on Stanford University’s campus, that she has chosen for our meeting. “I had an impression that I would have made it if I had lunch with the Financial Times,” says Li, as she arrives for what will surely rank among the cheapest lunches the FT has hosted.
Li is one of a small number of academics and technologists responsible for laying the foundation for today’s revolution in artificial intelligence. She is now pushing to ensure that revolution is carried out responsibly from a new institute at Stanford, her base since 2013. In one form or another, campuses have been Li’s home for more than 25 years.
During that time, universities and research labs have driven a string of breakthroughs in machine learning, computer vision and natural language processing. Li herself led the development of ImageNet, a vast repository of categorised images that demonstrated the importance of big data in powering AI and paved the way for significant advances in computer vision over the past decade.
But the AI tools being rolled out today, which demonstrate near-human level abilities to communicate, are coming instead from start-ups backed by the world’s biggest technology companies. Can any university keep up?
“I get why you’re asking this,” says Li. “But it really, really bothers me if we collectively are assuming there’s only one centre of gravity [in AI].” She insists that the public sector, with
universities at its axis, still has a hugely significant role. “We’re pushing research on neuroscience, we’re pushing research on climate . . . We still have very unique interdisciplinary thinking. We have unique interdisciplinary data. And we have the youngest and most daring minds.”
If we [academic AI] lose that centre of gravity, then the other centre of gravity is driven by capitalism
In 2019, Li set up a new institute for Human-Centered Artificial Intelligence at Stanford with professor of philosophy John Etchemendy. Their aim is to ensure that powerful new AI tools and policies are designed explicitly to improve the human condition, rather than simply to boost productivity or play. Li describes herself as “walking between being a scientist and a humanist”.
In the pursuit of AI, she says, “civilisation is like a big boat and we’re sailing forward in the dark”. She sees HAI and other public bodies as lighthouses illuminating a safe passage.
Plotting that passage has become increasingly fraught since the launch of OpenAI’s powerful ChatGPT chatbot in late 2022. That brought consumers face to face with the enormous power of modern AI and kicked off a race for technological supremacy between start-ups and Big Tech players such as Microsoft and Google.
The leap forward in capability shown by ChatGPT also aggravated fears about the dangers of AI: workforce disruption, disinformation and even existential risk — the subject of a major summit hosted by the UK earlier this year.
More than just a race to build the best chatbot, the past 12 months have been about increasingly fierce competition to determine how AI is developed, deployed and governed. Li does not dismiss the idea of AI as a threat to humanity, but her work has focused on curbing the more immediate dangers of it and ensuring that powerful new tools are used for good.
Universities remain vital places from which to pursue public benefits such as finding cures for rare diseases or mapping the Earth’s biodiversity, she says, and can provide a useful counterweight to purely profit-driven companies.
But Li is also aware of how the odds are stacked, having punctuated her tenure at Stanford with a stint as chief scientist of AI and machine learning at Google Cloud from 2017 to 2018. Arriving there, she found the abundance of snacks “staggering”, let alone the technology and the depth of talent.
It is an observation I recall as we scan the more limited menu at the counter of Coupa Café, a family-run eatery that sources all its produce from the San Francisco Bay Area. We order two portions of pollo arepas, Venezuelan cornbread stuffed with chicken, cheese and caramelised onions.
Menu
Coupa Café,
473 Via Ortega, Stanford, California CA 94305
Pollo arepa x 2 $22.50
Vietnamese coffee $4.10
Pumpkin spice latte (decaf) $5.85
Total (inc tax and service) $41.25
“Right now in AI, what worries me is we don’t have the resources to make sure that academic AI continues to be a centre of gravity. Because if we lose that centre of gravity, then the other centre of gravity is driven by capitalism,” says Li, when we get back to our table. “Public-sector investment in AI is so abysmal. Not a single university today can train a ChatGPT model . . . academia cannot fully develop its own versions so that it can be used for more open scientific research. That is a problem.”
Li met an executive at OpenAI shortly after the company launched as a not-for-profit in 2015. Raising a glass in toast, the executive said: “Everyone doing research in AI should seriously question their role in academia going forward.” Today the comment looks prescient. OpenAI has transitioned to a for-profit model and carries a theoretical valuation of nearly $90bn. It and rival start-ups have become magnets for the best researchers.
Li has “such respect” for OpenAI. But a boardroom coup at the San Francisco start-up in November suggested that private enterprise might be a more precarious place from which to develop AI than it had appeared. Co-founder and chief executive Sam Altman was abruptly fired for not being “consistently candid” with his board, only to be returned to post days later after employees and investors rushed to his side.
“It’s such an important company and I’m going to trust that even with glitches like we’ve seen — a tsunami glitch — we’re going to get to a better place,” says Li. “They have energised AI.”
An insistent mechanical buzz interrupts our conversation; our lunch is waiting at the counter. I have paired mine with an iced Vietnamese coffee, Li opts for a pumpkin spiced latte. We return to our table and set about the arepas, paired with a heap of shredded lettuce and a thick, tangy dressing.
