futuristic realm of AI

With the sudden rise in the prominence of AI due to the generative AI boom, I was extremely curious to learn more about the various sub-fields of AI and what the future in an increasingly AI-centric world would look like. AI is expected to trigger a disruptive transformation in how the world operates similar to what the internet did almost three decades ago. I came across an amazing book called ‘AI 2041’ authored by Kai-Fu-Lee, ex-Head of Google China, which gives a good insight into the different sub-fields of AI and their future outlook through fictional stories set int the year 2041. I would like to first establish a baseline high level understanding of how AI functions. The key proposition of AI is its ability to accurately predict an outcome (i.e. output data) if it is fed with certain input data, for example: being able to identify a cat (output) when the image of a cat (input) is fed into the AI system. When you dive deep into the actual working of artificial intelligence, it ultimately boils down to objective mathematical functions that are designed on the basis of vast amounts of past input data and their corresponding output data which are collectively termed as the training data. The objective mathematical function that gets designed is an equation that assigns specifically computed weights or coefficients for each input parameter depending on the underlying training data such that when we enter new input data, the equation is able to mathematically compute an output that is most likely to be the requisite outcome. The final objective mathematical function is a result of multiple iterations where the computed weights or coefficient of the equation keep getting refined with each iteration until it reaches a point of maximum accuracy. The accuracy of the objective mathematical function in terms of predicting the required outcome increases with increase in the amount of training data. Deep learning is a mechanism applied in AI model training which tries to mimic biological neurons of the human brain to add a deeper level of nuance to the prediction model of AI through intermediate hidden layers where each layer generates layer-specific objective mathematical functions whose outputs are treated as input data for the successive layer and this process propagates further layer-by-layer until the set of final output data is reached. Through the hidden layers, we get a wider set of parameters that help make the prediction model more nuanced and accurate. This nuanced prediction model is termed as artificial supervised neural networks which is highly capable of very accurately predicting outcomes based on multiple simple input parameters i.e. numerical and categorical variables. Image processing and prediction models based on visual inputs also uses neural networks called convolutional neural networks which operates in a slightly different manner. Images are constituted by multiple elemental building blocks called pixels whose color can be ultimately represented mathematically in binary code. Convolutional neural networks convert the two-dimensional array of pixel data into a linear one-dimensional vector through sequential processes of vector multiplication with a series of filter vectors (basically matrixes with specific values) where each filter identifies different features of the image. The linear one-dimensional vector is now similar to a set of simple input parameters required in supervised neural networks and can be used to generate an accurate image prediction model. A more interesting variation of neural networks is self-supervised neural networks which is what is used in natural language processing for generative AI. Typically, prediction models are built by labelling the correct output data for the corresponding input data, but labeling is a laborious task when it comes to prediction in text generation where each text input has to be labelled as per corresponding grammar rules of the particular language. Natural language processing uses a method called sequence transduction where a sequence of words in a sentence up to a point is taken as the input and the output is simply the sequence of words after that point rather than a label for the corresponding grammar rule of the language. Transformers (’T’ in ChatGPT) are text generators trained on vast sequences of different texts without the use of any specific labelling in regard to the rules of the language, so the text generation models are able to naturally (hence ‘natural’ in natural language processing) glean the rules of a language and generate output texts that successfully abide by those same rules. The growth of AI has been and will continue to be fueled by the increase in computing power and vast availability of training data. Now coming to the future outlook, I have outlined a few key highlights below in the different sub-fields of AI :

• deep learning + big data : Insurance companies that dynamically update the premium rates in real time by accessing multiple sources of user data such as social media activity, internet search activity etc. and predict potential risk to user and thereby accordingly update the premiums.

• computer vision + convolutional neural networks : Our faces will serve as ID cards and wallets where superior image recognition technology will easily identify faces of users and link it to a database that is filled with the relevant personal and financial data of the user. It can then automatically facilitate basic transactions like booking a subway or paying for a coffee.

• natural language processing : Generative AI advancement has massive potential in the field of education. Current education system uses one-size fits all teaching methodology which can be ineffectual to accomplish required learning outcomes as different students have different aptitudes and require different methods of teaching to help them learn effectively. Children can have personalized AI tutors in the future trained on a plethora of measurable parameter of a student which can in turn generate very student-specific learning plans and teaching content.

• robotics/automation : AI advancement will make autonomous robotics more of a reality which can (hopefully) supplement human capabilities in fields such as e-commerce (delivery robots) and healthcare (robotic surgeons). Autonomous vehicles like robotaxis coupled with smart cities are innovations to look forward to. Agriculture is a sector ripe for robotic intervention to help effectively improve crop yield through aerial robot i.e. drones that can take care of seeding and spraying pesticides along with on-ground robot harvesters.

• IoT : IoT will serve as an AI companion due to its ability to fuel AI with more and more useful user data. IoT has immense potential in healthcare industry, where different utilities we use regularly in our day-to-day life like watches or even a toothbrush can be equipped to collect real time data on various health parameters of a user. The data can be used for effectual AI model training by health insurance companies to assess risk associated with customers more accurately or even healthcare companies to design more personalized treatment or medicine for users.

• XR or [ virtual + augmented + mixed ] reality : Picture computerized contact lenses that can augment the way you view reality in any way you prefer. Suppose you’re exercising, you can completely transform the reality of the environment you are exercising in. you can superimpose supplementary objects or texts to the existing, basically have a visual personal assistant. XR will transform the way people play video games or perform various social interactions. Another useful application of XR would be immersive business meetings or classroom interaction as though participants are physically present together even though they are geographically far away.