Summary of Prediction Machines

I normally don’t write summaries for books for the general audience, but I wanted to write this summary for three reasons. Firstly, the book is fantastic. It is a great primer in what AI is and how it can be applied in the most general sense. Secondly, there were no summaries of the book online. If people are interested, reading what I wrote might entice you to pick up the book — and I highly recommend it. Lastly, I think it will be important for myself and others to come back to it.

Top Takeaway: The core of general AI rests in its ability to predict. Decision making at any level requires some form of prediction, thus making AI useful to organizations.

Introduction:

• Prediction benefits many organizations, and its been both expensive and hard. Imagine a hotel forecasting its demand for the next year with existing data and excel. At best, hotel owners would be guessing their demand. If the owners could plug in data about the weather, events, prices of nearby hotels, Airbnbs and more, they could get a much accurate forecast.
• As prediction becomes easier across multiple industries, it also becomes cheaper. Think about how using Microsoft Excel made it both easier and cheaper to make basic forecasts using spreadsheets. This drop in price however will lead to an increase in affiliated products. This means prices of things like sensors and microchips will increase.
• If prediction becomes so cheap and so easy, it could change business models drastically. Restaurants ask chefs to cook based on the order. However if you could predict certain orders, the chefs could already begin cooking or even cook and leave before the guest is seated! As a thought exercise, if prediction was SO CHEAP and SO EASY, how would it change your work? Your life?

The book then divides into 5 parts — prediction, decision making, tools, strategy and society. That’s how I have divided my summary as well.

Summary of Part One — Prediction:

• When you think of the terms ‘machine learning’ ‘natural language processing’ ‘object recognition’ and ‘autonomous driving’. All that you’re seeing is prediction at work.
• Machine Learning: This is the software or program using past data, usually in the form of mathematics, to predict the likelihood of an event occurring. By suggesting a pair of shoes you’d like, the program is simply guessing whether these shoes are relevant to you.
• Natural Language Processing: This is a program predicting what you are trying to say and what you are really looking for. When you ask Google Home for the weather, it predicts you want to know what it is like outside and then further predicts where you can get that information.
• Object Recognition: Again a prediction of what an object could be based on certain parameters of objects you have already presented to the program.
• Autonomous Driving: This is a combination of various items mentioned above to predict where and how fast to go. The intelligence in autonomous driving does not rest in the ability to turn the steering wheel or press the gas pedal, but the ability to know how much to turn and how fast to go. Its that prediction.
• One would think improving prediction from 98% to 99% is not as beneficial as improving it from 80% to 85%. However that 1% increase means a larger number of events predicted accurately, or even more precisely, making it super valuable. Think about this from a volume and cost perspective.
• Before machine learning predictions, companies aimed for being correct on average. This could mean you are +5 or -5 and you average out to 0. Machine learning is about regression. Its about having a lower variance, so you may never get to 0 but you will be closer to +1 and -1 most times. This is an important distinction because the lower the variance, the likelihood of a better prediction.
• Although prediction is a key component of intelligence, it is not one in the same. Most news about advances in machine learning make it seem like advances in AI, but this is not always the case. Better statistical analysis doesn’t mean more intelligent beings all the time.
• In order to predict, programs need training data, input data, and feedback data. Collecting this data is expensive, and getting harder and harder to obtain.
• As a business, you need to differentiate between quantity of data, prediction accuracy and value creation. Sometimes adding more data doesn’t lead to better accuracy and more value. Ofcourse, quality is also important, but subject to what you are trying to predict.
• Prediction can be broken down into 4 distinct categories — known knowns, known unknowns, unknown knowns and unknown unknowns. Machines can be very bad at predicting unknown knowns. These are things humans know intuitively, but machines can’t know. Think about the iRobot example of saving a kid vs Will Smith. In these cases, its best to use a human to step in and make a prediction by exception.
• Additionally programs can’t predict things they haven’t seen. If a new virus breaks out with symptoms similar to Zika, a machine may categorize it as Zika unless a human intervenes.

Summary of Part Two — Decision Making:

• All decisions are task based and each tasks consists of inputs, judgement, prediction, action and outcome. With enough feedback, the prediction also gets smarter through training. This is not just in machines but also in people. We take various inputs, we add our judgement (biases) and references to past experiences to predict how to act.
• If we can train a machine to predict without biases, we may make smarter and faster decisions. This makes predictions more apparent and thus cheaper, making the value of other things like the inputs, the judgements and actions much more expensive.
• The core benefit of predictions is so that we don’t end up making satisfactory decisions. We can optimize outcomes once we know how to better predict the tasks that influence them. When you think of waiting rooms at clinics, these are essentially solutions to poor planning as a result of an inability to predict patient throughput. The solution has been televisions or magazines, or notifications when the doctor is ready.
• If prediction becomes so good, it may lead to pure automation — but this is still far away. The areas where automation is starting to occur are those with the best ROI. This includes areas where everything else is automated but prediction, where the gap between and action and prediction is small, and where prediction can speed up a path to more returns.

Summary of Part Three — Tools:

• When thinking about automating a process, the best thing is to look at an entire workflow. Breaking it down by each task, decision point and job to be done. Each task should then be categorized by how expensive/valuable it is and how hard it is to conduct.
• At every step there are opportunities for automation (this leads to concepts like robotic process automation through companies like Blue Priszm)
• The book goes over an AI canvas which I will leave out in my summary but I highly recommend you take a look at it. The canvas helps you understand all the elements you need to build an AI related solution
• By automating certain tasks of different jobs, it could change the professional landscapes of many careers. Some jobs require a significant level of judgement that can be done by a machine (revenue forecaster), while others have multiple small tasks that can be automated (headhunting).

Summary of Part Four —Strategy:

• Many of the top tech companies have an advantage as they are already part of many workflows. As they start to identify what complements their workflows, they can easily get bigger. Think Google with Google Maps, then with transportation elements and local vendor reviews.
• Adoption of AI is also very depending on ROI. If there is an apparent ROI, the adoption is clear (autonomous driving). If the adoption is less clear then it will take a longer time to buy in. This can also be said of blockchain.
• It is not until your margins are threatened that you start looking for other areas to exploit better returns
• In thinking about the types of data you need between training, input and feedback — note that training data is almost useless once its used. How well your company did in the past is no longer necessary when you have built the correct algorythms for prediction.
• To expand on that first point, if you’re still using 5 year old trends to decide your current or future action, you’re already behind. Those 5 year trends have been fed into a statistical model is is searching for new data points to become even more precise. Stop shooting long range targets with a pistol.
• The book predicts that more prediction will make the lines between one business and another more blury. It also goes on to say that judgement will become very important once prediction is everywhere. You can contract out costs for data, equipment, predictions and even actions, but you can’t buy good judgement. This will force more people in this field to stay with the company as in house employees.
• If you are AI first, you are dedicating yourself to learning and data collection. Everything else is secondary. This is why big companies fail at this because startups are more focused on this than existing customers.
• Adversarial machine learning is when an AI is used to challenge the plans set by the main AI. This is a very interesting concept because it evolves from human behaviour. Imagine listening to two people speaking in a foreign tongue and guessing what they are saying — except once in a while they will tell you what just happened and you’ll get better at picking it up.
• Timing is key — if you launch an AI too early, it might be so inaccurate that customers find it useless. If you are too late, you may not get the full benefit of customer input and real feedback data (IE too much training data is a bad thing).
• AI is risky. Most risks come from the biases in the data, the lack of enough data, and poor feedback. Not to mention people playing around with predictions to learn more about the things they are predicting (If I can understand how you detect fraud, I can work around it)

Summary of Part Five — Society:

• This section is all about how prediction machines will change the fabric of society. At best these are predictions based on economic trends. The irony being that this fact alone is hard to predict.
• There is a deep fear that AI will enhance the income inequality substantially. While all of us will be better off, some of us will really reap the rewards of our success. Governments are not doing enough to manage this as they would argue the success of monopolies like Google and Facebook are to champion global competition.
• Privacy and performance are interesting concepts when it comes to prediction. You need data to predict, but things like GDPR strongly restricts how much data you can obtain and for what purposes (vs China where its a surveillance state). Does this mean China will be further ahead than the UK in AI? Its hard to answer but at a macro level one could argue positively for the Chinese.

This book was a terrific read and I highly recommend you pick it up. The Amazon link is here. Sidenote, although I have no direct relation to the authors: Ajay Agrawal, Avi Goldfarb and Joshua Gans, I was taught by them during my brief time at Next 36 (now Next Canada). They were phenomenal teachers then, and they have continued to teach me via this book.