The textbook by Stuart Russell and Peter Norvig, Artificial Intelligence – A Modern Approach, is a standard teaching authority on the subject. In 2021, the fourth edition appeared in print. In its preface, the authors explain how this edition reflects the changes in Artificial Intelligence since the previous edition a decade ago. Two comments that stood out for me were:  

1. “Focus more on machine learning rather than hand-crafted knowledge engineering”  
2. “Natural language understanding, robotics, and computer vision… revised to reflect the impact of deep learning.”  

These comments perfectly capture the technological basis of current AI. Learning Artificial Intelligence today is mostly about learning-machine learning, and deep learning, as applied to natural language and vision. The explosion of AI implementations (and the consequent emphasis on skill development) is a result of the relative ease of implementing deep neural networks.  

The mainstays of AI skill development continue to be variants of convolutional neural networks for computer vision, and variants of recurrent neural networks for natural language understanding. Implementation is typically through readily-available template code in Python or through low-code cloud-based environments such as AWS. The learning journey is curated so that even learners new to programming can execute elementary AI tasks such as image classification and sentiment analysis within a few weeks and move on to more sophisticated tasks in less than a year.  

As AI based on deep learning relies on large datasets, skilling in AI has also focused on how to work with large volumes of unstructured data. For instance, techniques and algorithms for image segmentation and parsing of text are learned as preliminaries and rather useful by-products. Courses in AI thus become specialized courses in data science or big data too. It has become common for skill-based courses to combine data science, machine learning, and artificial intelligence into single coherent curricula. 

Will this relatively low-bar yet highly impactful entry to AI remain? The answer depends on how synchronized AI skilling will be with trends in AI. Some relevant trends are:  

• Increased activity in generative AI, in addition to discriminative AI: Basic deep neural networks are very good at predicting and classifying based on observed data, essentially finding discriminating patterns in data. Generating ‘new’ images and text requires more complex extensions to this basic paradigm. The success of applications such as DALL.E (which generates images from text descriptions using GPT-3) means that more such technologies are already in use and the pipeline.  
• Emphasis on explaining and understanding AI predictions: Neural networks and other models used in AI have the reputation of being ‘black boxes’ that reveal little by way of predictive logic. When designing safety-critical applications like self-driving cars, more insight is needed to see how models prescribe optimal automated actions. Scientific applications, such as protein structure predictor AlphaFold from DeepMind, are also being engineered to reveal new scientific knowledge as opposed to just predictions.  
• Quickly deploying AI solutions as services and cloud-hosted applications: While it is helpful to design and code algorithms that predict well, it is increasingly the case that such algorithms should be deployable and open to new users. Until now, the software deployment task has typically not been the task of the AI engineers, but increasingly organizations want the DevOps pipeline integrated. Thus ‘MLOps’ is steadily being accepted as a requirement for taking things to production.  

AI upskilling is expected to absorb these and other trends into training and learning programs. For example, working directly and earlier within a cloud environment will make it easier to train learners to create and deploy microservices and also work with generative AI services as they are made available. Methodologically, unsupervised and reinforcement learning may supplement some of the now-traditional supervised learning techniques.  

In a larger scheme of things, AI training is being embedded into other disciples. A quick look at recent papers appearing in Science and Nature shows that machine learning (ML) and AI methods are being creatively used for scientific discovery across physical and life sciences. Scientists and engineers are being educated to add AI and ML to their analytical methodology. Similarly, business schools are incorporating computer-intensive data analysis in their elective curricula, going beyond core business analytics. Cases in point aiming for this kind of interdisciplinarity are MIT’s new Schwarzman College of Computing and the joint IIM-IIT-ISI program closer to home.  

In closing, it may be worth noting that there is a growing consensus that an AI-driven society needs new laws, business rules, and codes of ethical behavior. So far, AI has been an engineering discipline. Soon, it may be crucial humanities discipline, intersecting with law, sociology, psychology, philosophy, etc. That path to learning and wisdom awaits.  

Photo by DeepMind on Unsplash