MIT Press

View on GitHub

Questions from MIT Press

Elevator pitch

What if there was a source of intelligence, akin to a electrical outlet, which was easy for anyone to plug into? The book is a call to action to create a utility as transformative as the web itself, one providing AI to everyone at arbitrarily low cost.

Key Points

Audience profile

Managers, quants, data scientists, machine learning practitioners, analytics consumers/observers of various kinds in sectors such as finance, internet of things, manufacturing, transport and others requiring real-time operational intelligence and process optimization.

Some quotes from the book

In an imagined future where ubiquitous real-time operational intelligence has arrived at zero cost, I have prompted the reader to speculate as to the most likely origin.

The future of AI is a first principles exercise, not an extrapolation of industry trends.

The failure of the economic system to provide a critical economic capital good to the vast numerical majority of businesses and organizations is propagated to the consumer in large and small ways.

Every story needs an antagonist. Ours is the cost of bespoke analytics.

Even the most secretive of firms will, I predict, not be completely black. They will emit a kind of Hawking radiation.

A prediction web need not be a limited catalog of things that interest many people at once.

The crucial question is whether pyramids of humans are the best organizers of the production of autonomous prediction. That isn’t the only way.

The price mechanism is an old but miraculous device. But its relative efficacy is an economic question breathed entirely new life when we focus on repeated prediction specifically. The question is not is reward enough? in our existing economy but rather, is engineering enough? to unleash the price mechanism in a way we haven’t seen before.

I’ve spent most of my career striving to create fine works of mathematical modeling: business decision tools manifesting as calculators, embedded analytics, algorithmic trading processes, and enterprise data feeds - most of which can be couched as microprediction.

Ours is an untried invisible hand, admittedly, with dainty pinkies that never weighed anchor in a storm.

Listening to a small number of people boasting about how much applied mathematics they eschew could get rather dull. I have something far more entertaining in mind.

Micromanagers will also erode, slowly but surely, a bedrock assumption in educational and practitioner circles: that the task of matching models (and data) to problems will always be something of an artisan activity - one that is highly compensated, and one that is associated with a very specific image of the data scientist: the jack-of-all-trades.

The joke in cloud computing is that you pay for the services you forget to turn off, but that’s true of firms hiring dozens, hundreds, or thousands of quantitative people to build models. Continuous use of generalized intelligence is pricy.

In this thought experiment, our preconceived notions about expense and quality of bespoke work must be squished into oblivion.

The central problem of machine learning is … is the question of determining the best system for producing and distributing microprediction to all who need it.

Model-free reinforcement agents of tomorrow (and today) might struggle to relay to a central authority all the reasons for their actions, present past or future.

I hope the reader is at least suspecting of the possibility that the best system for arranging production of local knowledge is the one that best harnesses local knowledge.

The implication is that microprediction might require the entire world to collectively engineer - something of a desperate measure, it might seem - though consider what has already been tried.