How does someone go from being a teenage Chinese immigrant in America, to becoming the founder of a $5bn company, appearing on the cover of TIME magazine as one of TIME’s people of the year and being named the Godmother of AI - yes you’ve guessed it, in this series we are going to be deep diving into the life and business of Dr Fei Fei Li.
If you’d prefer to listen/watch you can watch an adapted version on YouTube
Back in the early 1990s, I was a young girl in her bedroom, throwing myself into physics and maths homework, reading every science book I could find, the talk of tiny particles with quantum properties felt like a gateway to another world. So much more enthralling and expansive than every day life and yet at the same time, studying the fundamental laws of the universe somehow felt a lot safer than the rest of the world around me.
I wasn’t the only one feeling that way.
If you have read the title of this essay you will know that I am focusing on Fei Fei Li, the founder of World Labs, Stanford professor and founder of the Human-Centric AI institute at Stanford. If you read her book, The World’s I See, you will know that we had something in common, we were both enthralled by the science that described the natural world around us.
At around about the same time, Geoffrey Hinton was working on developing computational networks which might be able to compute far more efficiently than anything else that already existed, networks that displayed potential to actually evolve without human instructions. Possessing their own artificial intelligence, but at that time they were feeling the cool winds of the so-called AI winter. Without data, and without the work of Fei Fei Li, and other scientists from many different fields, it would be another 20 years before the promise would turn into potential. Potential for constructive and destructive forces that were exponentially more powerful than any computational techniques that had gone before.
Many people laud Fei Fei Li as someone who is more ethical in the AI landscape that other prominent tech founders, due to her AI4ALL project and the Standford Centre for Human-Centric AI, but is human-centric AI actually the same thing as ethical AI? And how does this relate to the founding of World Labs - have we finally got a highly funded AI start-up led by a truly ethical founder?
World Labs actually had more money than OpenAI in its early years, and a much bigger mission, is this what we have been looking for?
Let’s get stuck into it to see if the story of Fei Fei Li and World Labs can help us answer that question. beginning with PART 1 OF 5, today we will explore how the Godmother of AI sparked the modern AI revolution.
Unlike Fei Fei Li I didn’t turn my love for science and maths into becoming the Godmother of AI, no Stanford professorship for me. However, I did run & exit my own tech start up and now I am a Royal Society Entrepreneur In Residence. I have helped over 350 technologists and founders make their ideas a reality. I dig deep into the stories of well known founders or ventures, so that we can better understand the lore behind tech business; so that we can learn from those who have gone before us; and so that we can be better stewards for more ethical technology.
I don’t think we should mindlessly take the “success” of other tech founders and copy what they have done, hearing the stories of others sometimes tells us what we won’t do.
I’ve been fascinated by Fei Fei Li for a while, her book is a very beautiful intertwining of her own story and the story of AI, both of which happened at a similar time. Her Human-Centric approaches, informed by her own life seem more connected to the real world than many other big tech founders. But let’s not be naive, she is a big player in the AI world, based in the heart of silicon valley, sitting on the cover of TIME magazine alongside Altman, Musk, Zuckerberg and others. Can you have one of the worlds biggest AI startups and be at the forefront of big AI and still be ethically sound?
Will I love it or will I hate it. Is it marmite - or, as is so often the case, will we find something more nuanced?
Let’s find out.
In this first part, we’re going to cover who Fei-Fei Li is, how she built the foundations of modern AI… Next week we will dig into what World Labs is actually trying to do, the controversies and my call to action for Fei Fei Li and her team come in future episodes.
The Fei Fei Li origin story
It’s impossible to tell the story of World Labs without telling the story of Fei Fei Li, so let’s start there for a bit of context. Fei Fei Li’s story is self-documented in her book - The World’s I See. So named because she works on computer vision.
She was born Beijing in 1977, and she primarily grew up in Chengdu, Sichuan province.
Being a girl in a society that favoured boys, she was, unusually, encouraged by her parents to be explorative, smart and dedicated. Her teachers on the other hand, didn’t give her so much encouragement, in her book she documents a story of a teacher who asked all the girls in the classroom to stand up, saying they did not need to bother learning any more because “boys are biologically smarter than girls”. her parents were having none of it and continued to support her passion for learning.
Ultimately the family moved to the USA, and it was not easy, Fei Fei had to learn english and help translate for her family, but she found safety and awe in physics and maths. She had a Maths teacher Mr Sabella, who ended up being a huge source of support for her. Not giving her special favours, and keeping expectations high, he helped her as she was navigating a new language, new societal norms and the financial aid requirements to fulfil her dream to study at Princeton. A stark contrast to that teacher in China who told her that girls didn’t have the brains for maths.
Physics became her first love and AI became an enthralling subject, but it didn’t start with computers. It started with the insatiable desire to explore the world around her, not through language, but through vision, a topic that ultimately became and obsession and a theme that runs through all her work.
Often, when Fei Fei Li is describing the birth of this obsession for her work, she discusses the Cambrian Explosion, that time period when 539 million years ago there was a massive increase in the number of species, resulting in most of the classes of animal that exist today.
The catalyst for growth? Major theories say that this origin of millions of species came from something simple - an organism being able to detect a photon of light.
So Fei Fei’s ongoing working hypothesis, for technology to evolve into something truly intelligent it has to be able to see, it has to be able to sense the 3D world around us. Words are not enough.
This combined with all the tenacity of an immigrant teenager has set the trajectory for the rest of her career. A tenacity that has allowed her to thrive in what might have been the hostile working environments of tech and academic - I should know - I’ve skirted around several of the fields that she has been in, and I have not found them easy to navigate.
Fei Fei Li talks a lot about the importance and excitement of having her North Star, but what happens when a research North Star collides with major Silicon Valley start-up funding, we’ve seen the answer to this question in many tech start ups. But can Fei Fei Li & World Labs can do things differently. I think that her early research career gives us a few clues.
Importance of data
Needless to say Fei Fei Li became an academic success. She was a Princeton undergrad, but of course this didn’t come without challenges. She got her first taste of business at age 18, when she worked with her family to open a dry cleaning business and maintain it for 8 years, all while completing an undergrad at Princeton and a postgrad at CalTech. She went on to complete her PhD titled: Visual Recognition: Computational Models and Human Psychophysics
Human Psychophysics: the scientific study of the relationship between physical stimuli (like light waves, sound frequencies, or pressure) and the subjective sensations and perceptions they produce in our minds.
Knowing the psychophysics of humans can help inform how robots behave.
Before we move on, I want to ask the question:
Was Fei Fei Li thinking about the ethics of such technology at this time?
This work was clearly building towards an academic problem that she was very excited to solve. But are there wider issues with developing such technology and if yes, does that mean we shouldn’t develop it? Let me know what you think in the comments.
Such questions form lifetimes of work but I want to give my 5 cents right now.
Whether you are a researcher or an academic, benefits of technology should be pursued but risks need to be considered to be mitigated. Not in a checkbox kinda way in an “integral to human beings” kinda way. It is a big flaw in education if humans aren’t equipped to hold both risks and benefits once they have life changing technologies in their hands.
One thing that was vitally important for Fei Fei Li’s work at this time was an interdisciplinary approach. Any silos that existed were dismantled as she focused on her own north star - making the visual world familiar to machines.
She obviously didn’t do this alone, her PhD supervisor Pietro Perona was an important player in this part of her work. In one discussion, Pietro said: I keep thinking about our “one-shot learning paper” “I’m proud of what we accomplished but we both know that data was the star of the show”.
The one-shot learning paper described the use of an AI model to be able to recognise an image with “one shot” based on transfer of unrealted previous knowledge. So this is when they really started to focus on data, the second part of the AI triangle, the first being AI models or algorithms and the third will emerge a few years later in this story.
It might seem obvious to say this now but at the time the focus was still largely on the models. The question of how much data is required and how might we get that data was somewhat new.
There are some slightly comical parts in the book where they discuss how much data would be required and what categories should be used, talks of dog breeds & cat breeds, they looked at what was out there already and decided that they would create 101 categories of 9146 images. This was the so-called Caltech 101.
How did they source the images?
The images were downloaded and categorised by students - and yes - this was imagery scraped from the internet.
Most people at that time would probably think that downloading 9146 images from the internet for a PhD project would be considered “fair use” BUT many will argue that this is the gateway to many of the issues that we now have with AI companies scraping data from the internet.
My view…
I would have done this without much questioning when I was doing my PhD back in the mid-2000’s I was so focused on just trying to survive my PhD that the wider context didn’t get much of a look-in, so I don’t think it is unreasonable for Fei Fei Li to also be single focused, I understand it.
It is not that clear which licence the images were published under, all I can find is that they were sourced from google image search. The images that were publicly shared on the internet were cropped, labelled and used to train models.
Is that the beginning of AI companies using other people’s data to make a small number of private citizens wealthy, delivering some positive, negative & ambiguous impacts to the world?
Personally, I don’t think that there is an issue with using published works to train AI, if permissions are granted for that use. I DO think there is an issue with HOW those AIs are used and who gets wealthy off the back of it. If public data is used then the public should be recompensed in some way.
However, we are in a difficult time now, in that many AIs have already been trained, without the correct permissions. Using the world’s data to build private wealth is clearly a problem, but was this the sliding doors moment when we might have done things differently?
Post PhD
Of course, once she had completed her PhD she was somewhat single focused on continuing to solve the problems of computer vision. The CalTech 101 helped in this journey, but it was its successor ImageNet that changed the game.
ImageNet was not without its challengers.
Many people might recognise the mid-noughties as a time when reality TV burst on the scene, when we were all finding long lost relatives & school friends on facebook and Netflix was still sending DVDs in the post. But, for those in the developing AI field, in the mid noughties algorithms were still a predominant focus. Fei Fei lists a bunch of questions in her book that naysayers would throw her way…
“What good is a data set with tens of thousands of categories? Most models are still struggling to recognise one or two.”?
The push back on the concept of ImageNet seems to be nearly universal - making Fei Fei Li’s life at this time challenging. Or at least that’s the picture she paints in her book, in hindsight I’m sure that she wasn’t completely alone in her ideas at this time. What amazes me is how she as a young immigrant woman in academia, communicating in an albeit very proficient second language continued to pursue ideas. This is a testament to who she is.
Even now, 20 years older than she was then, I struggle in university environments when people confidently (or arrogantly) push their ideas on me. People far less experienced than me like to tell me why I’m wrong and what good looks like and for some reason, perhaps my autism, or some other baggage, I still sometimes let their incompetence or disagreement turn into my barrier.
So here is one great lesson we can learn from Fei Fei Li, a lesson I wish I’d learned sooner:
Those naysayers are not playable characters in your game, they are the background, not the foreground, and your focus should reflect that. There is of course nuance here, we should all be thinking through challenges from multiple perspectives, but if you have a goal, your north star, you need to createn tight focus by bringing together the activities and people who will help you to get there.
In the case of Fei Fei Li - the person who actually understood the journey was coming from the third part of the AI triangle: infrastructure.
Prof Kai Li, another Chinese immigrant was an expert in microprocessor architecture and Prof Li names one thing that they had in common, exponential thinking. Together, they could see what this future looked like:
a world’s worth of data with millions upon millions of transistors
This would create two parts of the triangle required to unlock the power of the algorithm.
They weren’t wrong.
This is a future that Fei Fei Li is firmly pursuing in World Labs and many other AI founders are reproducing in their giant AI businesses.
So at point point in the story might they have questioned that future?
This story is not the usual founder story. It is very much an academic story. But there is a general question here, something that I think we all face as individuals in our work and our lives.
Imagine, you are trying to build a career, a perspective, expertise, networks, you are trying to make your mark. In Fei Fei Li’s case, she was doing this while supporting her parents, helping them financially, culturally and with their health, for others we might be battling our own chronic illness, mental-health challenges, financial challenges - that was my reality in 2006.
So tell me, when do we stop and think about the implications of the futures that we are trying to create?
The adage “we were so busy trying to prove that we could do it that we didn’t stop to think if we should do it” is very relevant across the AI space. But I don’t think “should we do it?” is the right question.
The right question is:
“How can we do this to minimise potential negative outcomes as we continue to develop human-centric technology?”
My belief is that you can’t get innovation without the two coming together. The true meaning of human-centric should include a wider ethical framework, because it should not only consider the human when it comes to using and gaining benefits from the technology, it should also consider end-to-end potential negative impacts or risks from the technology and include mitigations. Innovation is invention & mitigation. It might not feel as free or as sexy, but its where true creative impact is born.
Image Net and AlexNet
Let’s move on to perhaps the game-changing years for Fei Fei Li - she had already found a place in world renowned departments - but ImageNet really opened the door to help reveal the Godmother of AI.
Data and exponential scale were the order-of-the-day. So she went from the measely Caltech 101, to a new database. 101 categories were not enough, 1000 categories was more like it. And 10k images? No, let’s aim higher!
After 5 years of work this number would grow to 3.2m images.
ImageNet was published in 2009 boasting 12 subtrees with 5247 synsets (categories) and 3.2 million images in total. Now it has risen to 14,197,122 images, 21841 synsets indexed.
The same questions remain about where these images were sourced and the labour used, Amazon Mechanical Turk was instrumental in building ImageNet, a service that has been shown to treat data workers badly.
If Fei Fei Li is to be believed, many of her peers didn’t understand her vision, in the beginning. So she did something creative, and I love to see technology and creativity coming together. I’d like to think that technological approaches always involved creativity, but over the years the process of innovation seems to have become more formulaic.
Fei Fei Li had to think differently as she developed ImageNet. Firstly to reach a high number of images but secondly to prove that there was value, how might they bring the three pieces of the AI triangle together? They had the data, now the infrastructure and models needed to come together.
Instead of building the perfect model themselves, they realised that ImageNet provided the perfect opportunity for researchers to benchmark their models. The team decided to launch a contest - an annual competition where anyone could train their model on ImageNet then test them on a series of never-before seen images the team with the fewest errors would win.
Launched in 2010 the first two years were a bit of a damp squibb, but in 2012 a new idea was submitted by Alex Krizhevsky (yes that Alex), Ilya Sutskever (yes that Sutskever) and Geoffrey Hinton (yes that Geoffrey Hinton), they decided to use a neural network called AlexNet. Many saw these networks as archaic old cobwebs in 2012, but things were about to change. The performance of AlexNet blew all the other models out of the water. Much of what follows in AI flows from this moment.
So the first principles of our modern era of AI had been set: infinite resource + large amounts of data scraped from the internet = algorithms that could predict. Whether or not any of this was true artificial intelligence is a question for another day.
In the early years of Fei Fei Li’s research there was little mention of the Human-Centric approaches that she has become known for in later years. She was largely focused on trying to give machines the same power of vision that was so powerful for biology, the Human-Centric aspects were a more general view on how this would provide positive impacts for humans.
There were some inklings of the future founder in Fei Fei at this time, she appears to have possessed great leadership from the early days. She has a clear vision of where she wanted to go, from getting into Princeton only a few years after she entered the US, to attending the most prestigious universities to post grad and undergrad, to ultimately launching a well renowned contest that re-launched neural networks into a wider public conscious.
She also proved to have the tenacity, creativity and work-ethic required of founders, all things that the best researchers and founders had in common. In reality she had in fact already been a business owner, developing and running the dry cleaners as a teenager. But back in the early 2010s she probably didn’t have the motivation to become a tech founder, it wasn’t the next step towards her North Star, after all.
After serving another few years in research, continuing to develop ideas, research and reputation as she climbed in AI research circles she became the 7th Director of the Stanford AI Lab or SAIL.
Fei Fei Li had firmly embedded herself in the centre of Silicon Valley just as the AI winter was fast thawing. She had found herself SAILing in the AI Ark, lucky enough to be one of the chosen few ready to thrive in the coming AI flood. The question we will look at in the rest of this series is who else is she taking with her, is it just the fellas sitting on the front of time magazine with her, or can we all benefit with her at the helm of an AI business?
Subscribe to my Substack and I will bring you part 2 next week.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.