There is a version of deep learning history that gets repeated in course introductions, which is that in 2012 a neural network won an image recognition competition and everything changed. That version leaves out the interesting question, which is where the competition came from, and the answer is this book. In the mid 2000s the consensus in computer vision was that progress would come from better algorithms. Models were compared on datasets of a few tens of thousands of images and the research effort went into cleverer architectures and better hand designed features.
Li's contention, which was not a popular one, was that the bottleneck was not the algorithm but the data, that a model given orders of magnitude more examples would learn things no amount of algorithmic ingenuity could produce, and that somebody needed to go and build that dataset. The book describes what that actually took. Millions of images organised against a linguistic hierarchy, labelled by a very large number of workers on a crowdsourcing platform, quality controlled, and released free at a time when senior colleagues told her plainly that it was not research and would not help her career. The reaction to the first release was indifference.
Then a network trained on it in 2012 beat the field by a margin that made argument unnecessary, and the entire discipline reorganised around the approach within about two years. Having the person who did that describe the years of scepticism beforehand is the value of this book, because the sanitised version in the textbooks removes exactly the part that a working researcher needs to hear, which is that the idea looked stupid for a long time to people whose judgement was worth respecting. The personal material is not filler. Arriving in the United States as a teenager, learning English while translating for her parents, working in her family's dry cleaning shop while getting through school and then Princeton, and later managing her mother's serious illness while running a lab.
It is told without self pity and it is bound up with the intellectual story rather than bolted on, because the outsider position and the willingness to pursue an unfashionable idea are plainly connected. She is also honest about luck. The timing of the graphics card becoming a viable training platform, the specific people who happened to be available, the fact that the crowdsourcing infrastructure existed at exactly the right moment. She does not present the outcome as inevitable, which is a discipline most successful people lack.
Now the limitations. This is a memoir and not a technical book. You will not learn how convolutional networks work, and the description of the 2012 result is at the level of a general reader. That is the right choice for the audience and it means a technical reader will find the central episode covered more thinly than they would like.
The later chapters, covering her time in industry, the founding of the Stanford institute and her policy work, are noticeably weaker. The language becomes institutional, the specificity drops, and passages read like they have been past a communications team. The energy of the early chapters does not survive into the part about committees. The labelling labour deserved more.
The dataset was built by an enormous number of crowdworkers paid very little, and this is acknowledged and not examined. Given that a substantial critical literature now exists about exactly this pattern in machine learning, and given that she was there and could have said something specific and first hand, the brevity is a missed opportunity. Kate Crawford's Atlas of AI covers that ground and this book would have been stronger for engaging with it. There is also some quiet settling of accounts.
Colleagues who doubted the project, the internal politics of a big technology company, and the disputes of the field appear in ways that are polite and clearly shaped by her side of the story. Nothing dishonest, and you should read the passages about other people knowing you are getting one account. My four is for a well written and genuinely important piece of first hand history, told with more honesty about doubt and luck than this genre usually manages, marked down for thin technical content, for later chapters that lose their nerve, and for skipping lightly over the human labour that made the whole thing possible.