Anil Ananthaswamy is an acclaimed science writer and author whose work spans some of the world’s top publications, including New Scientist, Scientific American, Quanta, and Nature. A former staff writer and editor at New Scientist, he is also a 2019–20 MIT Knight Science Journalism fellow and the author of four celebrated books, including Why Machines Learn: The Elegant Math Behind Modern AI, published last year. (The book has received critical acclaim, including from Geoff Hinton, who called it “a masterpiece.”)
Anil’s background bridges engineering and journalism—he trained in electronics and computer engineering at IIT-Madras and the University of Washington before earning a science journalism degree from UC Santa Cruz. He has taught and mentored widely, and was recognized with the Distinguished Alumnus Award from IIT-Madras for his contributions to science communication.
In this conversation, Anil gives Sandhill readers his input on the foundational mathematics of machine learning, the complexities of AI safety across sectors, and the uncertain but transformative future of AI for leaders and consumers.
M.R. Rangaswami: How would you describe “fundamental math” of machine learning, and how does that math apply across industry?
Anil Ananthaswamy: We can divide the fundamental math of machine learning into two broad categories. One is the math you require to get a good conceptual understanding of ML, so that one can make good decisions about the kinds of algorithms/models/datasets needed to solve some particular problem. This level math involves a basic understanding of linear algebra, calculus and probability and statistics.
The second category is the kind of math you need to design ML algorithms/models. This requires an advanced understanding of calculus (such as multivariate calculus, vector calculus, convex optimization techniques, etc.), linear algebra (matrix decompositions and tensor operations), information theory, graph theory and so on. The list can get quite heavy.
So, depending on which side of the fence you are in the industry, you might either have to learn the basics or the more advanced aspects of these branches of math.
M.R.: What do you feel is considered safe AI? Is it user dependent? Industry dependent? Software dependent?
Anil: I think it’s all of the above. As with any technology, the safe use of it depends on the capabilities we build into our AI systems, how we use it, and what safeguards are employed by industries that deploy the AI systems, which of course will involve the software scaffolds that tie artificial intelligence to the other aspects of information processing.
The difference between AI and the technologies that have come before (such as the internet) is going to be unprecedented scale and speed at which AI is going to permeate almost everything we do. Also, the barriers to entry for bad actors is much lower than it has been in the past. Using AI for nefarious purposes is relatively easy (think about deep fakes, for example). We will have to work extra hard to ensure the safe use of AI.
Also, the black-box nature of deep neural network-based AI is going to make it harder to ensure the kind of safe use that relies on being able to interpret the workings of these machines.
M.R.: What is your forecast for what AI will do for leaders and consumers by 2030?
Anil: Given the pace at which things are changing, it’d require someone with a very clear crystal ball to foresee what’s going to happen in two years, let alone in five or ten years. But some broad trends are clear. Machine learning is here to stay. Some of the concerns we had about whether neural networks will be effective have been laid to rest. For example, computer vision and image recognition (and the consequent downstream applications that they make possible)—something that was considered an extremely hard problem a decade ago—is now a mature technology. The same goes for many natural language processing tasks—such as machine translation of text from one language to another.
The wild card is whether or not large language models will truly deliver on their promise. Big companies are betting that scaling up these models—making them bigger and spending more on training data and compute—will create systems that will be superhuman in their abilities, coming close to what many call artificial general intelligence (AGI). While scaling has delivered amazing results (as evidenced by the “emergence” of sophisticated behavior in LLMs as they have been made bigger), there are also concerns that these LLMs, based on the Transformer architecture and next token prediction, may have in principle limitations that might be impossible to overcome.
Regardless, both leaders and consumers need to prepare for AI/ML systems that will both empower and disrupt in equal measure, with the attendant social consequences, such as job losses, as AI/ML and robotic systems replace the low hanging fruits (such as cognitive tasks that can be easily automated, basic coding, and even simple kinds of manual labor).
M.R. Rangaswami is the Co-Founder of Sandhill.com
Photo credit: Amit Madheshiya / TED