#35 State of AI Education
My favorite AI books of 2023
This year was filled with many interesting AI books. Consciously or unconsciously, I have focused my reading this year on the physical world of AI and its history. Below I have selected six books that I really enjoyed reading and that helped me understand the extent of AI as a technology.
Material World by Ed Conway: Our interactions with AI rarely have anything to do with the physical world on which the technology is based. The book was a breath of fresh air as it delved into this material world that forms the infrastructure of our modern life. Sure, AI is only one part of it, but it depends on it even more. Ed Conway covers six materials in detail: sand, salt, iron, copper, oil and lithium. If you think that sounds boring, you're in for a surprise. Ed vividly describes the places where all these materials are produced and writes beautifully.
Atlas of AI by Kate Crawford: What does it take to produce Amazon Alexa? The book is an attempt to answer this question. Kate Crawford attempts to debunk the notion that AI is artificial or intelligent. Similar to Ed Conway's book, she shows where and how the material world of AI is produced. Her journey takes the reader to places such as the Bayan-Obo mine in China or the Silver Peak Lithium mine in Nevada. Her critical stance on some practices in the development of AI systems opens the door to a discussion about the work behind labeling datasets for AI systems, the difficulty of creating unbiased classification schemes, and the political dynamics of enforcing AI use cases. The book reminds us what needs to be done to develop AI systems that we can trust and that in line with our moral values.
Chip War by Chris Miller: Did you know that a quarter of all chips produced go into our smartphones? Or that China spend more money on importing chips than on oil in 2021? Or that a modern chip can contain more than 11.8 billion transistors? I didn't until I read this book. Chip War will open your eyes to the huge supply chain required to make chips. The book is fantastically researched and goes into great detail about the historical development of the chip industry. I particularly liked the sections on the development of EUV (extreme ultraviolet lithography), the foundry model and fabless manufacturing. Again, chips are used for things other than AI, but without chips there would be no AI.
The Power Law by Sebastian Mallaby: It all began with Arthur Rock and the traitorous eight. The Traitorous Eight were eight disgruntled employees working under their cranky boss William Shockley. Arthur Rock supported them financially when they decided to start their own company: Fairchild Semiconductor. Venture capital was born. Sebastian Mallaby's stories in this book are about power, big money and change. The book will leave you amazed at the forces that have shaped the companies that are central to today's society: Uber, Spotify, Facebook or Airbnb. You won't find an in-depth discussion of AI in this book, but it will give you a sense of the power structures that help build companies that use AI.
AI related books I am looking forward to in 2024
Here are four AI books I am planning to read next year:
Machine Learning Interviews: Kickstart Your Machine Learning and Data Career by Susan Shu Chang. The obvious target audience for this book is young professionals who want to pursue a career in AI. In her book, Susan Shu Chang covers the different roles in machine learning and gives a guide on how to find your way. I chose this book because it also describes in detail what constitute modern ML teams and what skills are required to deploy ML at scale.
Against Technoableism: Rethinking who needs improvement by Ashley Shew: Everyone wants to solve problems from their own perspective. Technology-savvy people often look for solutions to problems that use technology. Finding a tech "solution" for people with disabilities, for example, seems like a noble cause. In her book, Ashley Shew argues that the idea of "fixing" the disabled is a false idea of an equitable world. AI will help people with disabilities in the future, but we need to be careful how we think about disability technologies.
Benny the Blue Whale: A Descent into Story, Language, and the Madness of ChatGPT by Andy Stanton: After reading the blurb for this book, I'm still confused as to what to expect. Andy Stanton is an author of children's books and with this book he dives headfirst into the creative process of a writer using ChatGPT. At the heart of the book is a story about a blue whale: "Tell me a story about a blue whale with a tiny penis". This develops into a story about pushing ChatGPT to its limits and the art of storytelling.
Unmasking AI: My mission to protect what is human in a world of machines by Joy Buolamwini: In the short and long term, we need to ensure that AI benefits society as a whole and does not favour certain groups. I chose Joy Buolamwini's book because she has a strong track record of exposing injustices in AI technology. The book is her journey on how we should deal with the harms of technology.
So under what circumstances do companies adopt AI? A new study focusing on the DACH region sheds light on this question
For AI to find its way into companies, it is not enough for the technology to be used somewhere. What makes companies successful in using AI is often the physical or cognitive proximity to AI networks. We know this, for example, from the relationship between universities and AI start-ups in Italy or from studies by Adam B. Jaffe and colleagues.
In a recent article, Johannes Dahlke from the University of Twente and colleagues trained a transformer language model to predict the adoption of AI at company level, taking into account 380,805 companies in Germany, Austria and Switzerland. The aim of the study was to find out which mechanisms motivate companies to adopt AI. The authors used self-scraped web data for their predictions, which included websites scraped from hyperlinks. AI adoption was defined as companies that “have developed an AI identify which requires that AI is a core part of their products and processes” (p. 2).
A look at the descriptive data shows that a small proportion of 1.5% of companies use AI extensively and have in-depth knowledge of AI. Another 1.3% use AI at a superficial level. So if we assume that 2% of all 380,805 companies use AI, we arrive at 7,600 companies in the DACH region. As expected, companies with a strong AI profile are heavily clustered in certain regions such as Munich or Berlin. However, some small clusters are not to be found in these metropolitan areas, but in close proximity to the AI research facilities of the German network of national centers of excellence for AI research. I recommend taking a look at the nice maps to get an impression of the geographical distribution of AI use.
So what mechanisms influence the adoption of AI? According to the results of the study, there are three factors. First, the adoption of AI in companies is promoted by the fact that they are located in regional AI hotspots that produce AI knowledge. This effect is rather indirect and creates a cognitive pressure for companies to start using AI as well. In other words: When companies see that companies in the region are using AI, this increases the mimetic pressure to invest more in this technology. The other mechanism comes from direct contact with educational offerings that provide in-depth AI knowledge. Interestingly, the transfer of AI knowledge is not influenced by geographical distance. Being part of this knowledge network is great, but it does not seem to spread across the entire company network, leading to strong imbalances that could have long-term regional effects. Third, the adoption of AI is boosted by direct involvement in an AI knowledge network. In other words, firms adopting AI have actors with strong social capital and support the diffusion of AI knowledge, especially deep AI knowledge.
In light of the findings, the main challenge for policy makers is to develop measures that help break up the tight AI networks or strengthen interactions between companies and actors in the AI field. For students aspiring to a career in AI, the study provides food for thought on which places could be attractive places to work.
What AI tools are used?
I stumbled across an extensive analysis of web traffic data by Writerbuddy. To find out which AI tools attract how much traffic, data from September 2022 to August 2023 was scrapped and analyzed to identify the 50 most visited AI tools. Unsurprisingly, ChatGPT takes the lead taking with around 60% of all web traffic data among the 3,000 AI tools analyzed. The second most popular tool, character.ai, is trailing behind with 15.77%, followed by Quillbot, a writing tool, with 4.68%. The data also shows a gender gap in the use of ChatGPT: 69.5% of users were male and 30.5% are female. ChatGPT is clearly the king when it comes to GenAI tools. We see the same dominance of ChatGPT among German students, who mainly use it together with DeepL. Similar results were found in studies in the UK. It will be interesting to see how the other tech giants will increase their share of web traffic in the future. There is a long way to go: Google’s Bard only had a 1% share in the data of this study.
Using AI to track icebergs 10,000 times faster than humans
Since I read “When the heavens when on sale” by Ashlee Vance, I have been intrigued by the possibility of using satellite data with AI to answer questions about the Earth’s surface. Just look at what LeoLabs is doing. The company builds shoeboxed-sized satellites that allow engineers to take multiple images of the entire Earth every day. An initial feasibility study has shown how the image data can be used to query the Earth’s surface like Google. In the future, it will probably be possible to find out how many trees have been lost in a fire or how much land has been lost due to climate change. Another very interesting use case is the tracking of icebergs. With this new infrastructure in orbit, society can combine satellite technology with AI to monitor how icebergs move and evolve over time. Researchers at the University of Leeds recently trained a neural network to do just that. Previously, the outlines of icebergs were calculated manually, which took a lot of time. The system proposed by Anne Braakmann-Folgmann and colleagues is 10,000 times faster than humans. For me, this use case is a great example of how AI can influence tasks in different ways: By speeding them up, lumping them, augmenting them, and improving their quality.