The Loop is a broad cautionary tale about data science technology， mainly deep learning， but it has an unfortunate tendency to oversimplify （and occasionally misinterpret） the state of ML industry and research。I imagine that this book would benefit from a narrower scope。 It tries to address everything from high-tech surveillance to the dangers of AGI and a "one-size-fits-all" mentality， and ends up somewhat disjointed and patchy。 Ward brings up a lot of genuine questions and concerns in present- The Loop is a broad cautionary tale about data science technology， mainly deep learning， but it has an unfortunate tendency to oversimplify （and occasionally misinterpret） the state of ML industry and research。I imagine that this book would benefit from a narrower scope。 It tries to address everything from high-tech surveillance to the dangers of AGI and a "one-size-fits-all" mentality， and ends up somewhat disjointed and patchy。 Ward brings up a lot of genuine questions and concerns in present-day scenarios， but mitigation is another thing entirely and there aren't really any concrete action items。 It's mostly some lofty ideals with a healthy does of pessimism:It's not even clear that if we clearly articulate the problem and outline a solution， people in a position of power will be willing to act on any of it。 Most of the overall points are highly worthwhile to think about。 A brief sampling of such ideas follows:- A good objective function for business may not necessarily be a desirable objective function for society。- There are dangers in fine-tuning models originally trained for a different objective。- People shouldn't blindly trust and follow algorithms out of convenience， or because they don't want to make a hard decision。- There's an urgent need for accountability as applications outstrip regulation。- It's important to distinguish between correlation vs。 causation in data， especially in predictive models determining housing， credit scores， crime， etc。- There are things we should do the hard way to force more deliberate thought。However， attempts at technical explanations and extrapolation often don't make much sense， and reveal a lack of understanding of the techniques being used， as well as why and how decisions are being made in the field。This is most visible in ch。7， where the book tries to give a crash course in ML terminology。 There seems to be some confusion about what， exactly， an objective function is （"the objective function for you may be very different from the objective function for me"， "setting objective functions for humans"）， an example of reinforcement learning is just standard supervised learning， and there's some misunderstanding about fine-tuning and shared model architectures。Later， in ch。9， Ward seems to misinterpret the point of algorithms designed to measure audience reactions: from the interviews he quotes， it's clear that they're picking up engagement patterns， not specific emotional reactions。 Companies like WattPadd are not trying to promote things that elicit the exact same viewer reaction every time based on simple emotional response， they just want to see if they're driving community excitement。 There's also a curious lack of exploration into alternative explanations for some of the points raised--I'd posit that art/music/writing becoming trite comes from the fact that things often need to be inoffensive and bland to have mass appeal （see: Marvel）， not because algorithms only create things that are "matched to a few of our most basic emotions。"Other passages here and there make it obvious that Ward is not terribly familiar with the industry itself。 He mentions: 。。。 today， AI is being refined entirely inside for-profit companies。 This is strictly untrue， and in fact his next example is of a GTech research project funded by DARPA。 It's also a little odd that he goes on so much about how secretive and closed-off the industry is， when most companies seem to be falling over themselves to publish their work。 In the course of my job （DL engineer）， I see papers from many different companies， large and small， and most famous models have multiple open-source implementations and downloadable checkpoints， with a large community of people writing blog posts about how they work。As a side note， I do have to question the weird vendetta against Google Maps and assisted navigation that briefly comes up as an example of limiting human choice （because you are given only a few options as a default）。 I don't know about you， but I really do not miss the MapQuest and AAA folding map days。 There are a few other examples that come up about machines limiting choice where it just seems like the algorithms are just mitigating human error （flight paths， etc。）， so I don't really buy them as supporting evidence。The book also tends to be quite melodramatic:I worry that as we become caught in a cycle of sampled behavioral data and recommendation， we will be instead caught in a collapsing spiral of choice， at the bottom of which we no longer know what we like or how to make choices or how to speak to one another。 With some straight-up fear-mongering:And while the examples I'm about to describe may feel disconnected， remember that the interoperability of machine learning means a set of algorithms built to do one thing can also do many others well enough that you'll never know its various roles， so anything AI can do in one part of your life will inevitably metastasize into others。 （I already have a hard enough time with transfer-learning within the same domain。。。）Finally， there's a strong narrative of complacent people who get used to offloading thinking/choice to machines， leading to nothing truly new。I would be interested in seeing studies of whether or not this is true， rather than speculation: percentages of people who search for specifics vs。 autoplay on YouTube， for example， or how many people blindly take recommended food choices on delivery systems， and so on。 One could also argue that recommender systems help people with niche tastes discover new artists who match those tastes （I know several people for whom this is true）， even if they're not mainstream， rather than constraining choice。 Just because an algorithm recommends something doesn't mean that your tastes change， and that's an idea that wasn't explored at all。All in all， The Loop utilizes some alarmist language and iffy technical explanations that damage the credibility and believability of its argument， but it does still bring up a lotof interesting ideas。 Where it excels is in the fascinating current examples that Ward has dug up and the informative interviews with various people in the industry， and these examples and thought-provoking questions earn it three stars and are well worth your while。 At the very least， it's a jumping-off point for further discussion， and is useful as a general survey of issues （current and upcoming） in the field。** Thanks to NetGalley and Hachette Books for an ARC in exchange for an honest review。 ** 。。。more