Review A very easy to read and follow business book set in a conversational tone with understandable examples. If you are not part of the AI conversation, this book provides a good framework to start participating.
Book Summary Prediction machines by Ajay Agarwal et al looks at the consequence of the current AI upsurge in terms of business and economics. This book is not targeted at the data scientist but more at the managers of AI.
The book says that the current definition of intelligence is “prediction” and calls them prediction machines. The cost of prediction economically is going down and it is the books contention that the value of complements to prediction will go up. Also much like lighting the more cheaper AI gets the more it gets used as prices fall. Also, as the predictive ability goes up, the business model and strategy of an organization that uses the predictions undergoes rapid shifts. For example Amazon shopping-then-shipping will become shipping-then-shopping as prediction becomes cheaper. They also motivate the same with the example of a business school which is freed up in predicting the successful applicants and thus goes for widening the applicant pool through aggressive advertising.
The book outlines a good model for prediction machines in terms of the ladder of Prediction->Decision Making->Tools-> Strategy-> Society. The distinction between the types of data i.e. training data, input data and feedback data is well brought out. Further a good observation is that the marginal utility of data decreases in AI; however business model wise as in the case of long tail searches every data counts. The authors than examine machines and humans working together and the resultant parameters for division of labor. Prediction by exception when the machine fails to predict because of a rare occurrence is speculatively postulated as a long term goal for humans.
On decisions, the book predicts a rising role of the discipline of what it calls “reward function engineering” under uncertainty i.e. determining the payoff of a given prediction and its consequences if implemented. These can be automated in case of driver less cars or maybe open ended in case of say recruitment. In some cases an AI may learn the reward function by training with humans and make the human obsolete going forward. All humans are satisficers who cannot handle the number of “if-then” combinations in the real world. A prediction machine can expand the scope of both if and then and move us in the standard hyper ration homo-economicus. With regards to full automation, the authors bring up the concept of externalities. For example a rise in accidents is an externality which can lead to tighter regulation of driver less cars.
The section of tools is more about identifying the processes and workflow of a business and applying AI judiciously at key points of the same which may also involve re-engineering of the organization. This will lead to lower productivity than imagined in the short run. The authors then provide a framework for assessing AI in terms of input, feedback, prediction, outcome, action, judgement, training and analyze a few examples which is useful.
The authors assess the impact of AI and jobs. Although many jobs may be eliminated in some jobs (like the case of spreadsheets making calculations useless), more jobs may arise as the function become critical. There are other cases like a school bus being automated and the bus driver becoming a adult caretaker of the children on the bus. The authors are silent on what is the total effect on jobs. I guess we have to wait and see.
On Strategy the authors use an analogy of farmer adoption of hybrid corn in different states to AI. Google, Facebook, Amazon are the early adopters but as AI gets prevalent all companies will adopt the same. Also instead of just owning the AI, owning the actions of the AI may prove in itself valuable to some businesses. Further the impact of AI may lead to changes in organization structures, boundaries, hierarchies and roles. An interesting use case is how AI may lead to more outsourcing in the car manufacturing industry as car manufacturers will have a more stable outlook on customer choices and come up with more integral designs (in which the lag behind self sufficient car manufacturers).
One the key strategic quandary of many organization is whether to own the data or contract out the data. Unless data is critical strategically, the authors suggest to buy of the shelf prediction machines (like advertisers on Facebook).The book then examines the AI-first strategy that is being expounded by the likes of Google and Microsoft. As a result, these organizations are maximizing of prediction while trading off goals like revenue or user experience. So its more a bet. Further there are lot of risks in AI due to discrimination, quality risk, hacking, monocultures, IP theft and manipulative feedback which these organizations have to bear.
The last chapter on society is mercifully short but emphasizes the key tradeoffs of AI productivity vs Inequality, Innovation vs Competition and Performance vs Privacy. All these require complex policy interventions.