Back to index
OtherFive to six hours·Around $30 hardback, less digital

Superagency by Reid Hoffman and Greg Beato

2.9

A useful corrective to doom writing, made by someone with hundreds of millions of reasons to reach this conclusion, and notably quiet on the objections that would be expensive to concede.

What We Liked

  • The iterative deployment argument is genuinely worth engaging with
  • Historical framing against the printing press and steam engine is well handled
  • A necessary counterweight to a discourse that has tilted heavily pessimistic
  • Readable and short, which is not universal in this category

What Could Be Better

  • The author's financial interest in the conclusion is enormous and underplayed
  • Copyright and training data questions are essentially sidestepped
  • Labour displacement is addressed at a level of abstraction that costs nothing
  • Superagency as a concept does less work than the title implies

Detailed review

Reid Hoffman co-founded LinkedIn, sat on the OpenAI board, and has invested in a great deal of what this book is optimistic about. That is not a reason to dismiss the argument and it is the frame within which the argument has to be read, in exactly the way Alexander Karp's book on the technological republic has to be read against Palantir's business. The central concept, superagency, describes the expanded ability of individuals and organisations to act effectively when supported by advanced tools. Hoffman and Beato place AI in a lineage with the printing press and the steam engine: technologies that provoked genuine fear, caused genuine disruption, and ultimately expanded what ordinary people could do.

The historical material is handled well and the parallel is not stupid. Every general purpose technology has produced a wave of confident predictions of catastrophe that did not arrive in the form predicted. The strongest and most practically useful argument is iterative deployment. Rather than attempting to anticipate every harm before release, the case is that you ship carefully, watch what happens, and correct, because real deployment surfaces problems that no amount of advance theorising will.

There is genuine substance here. A great deal of what we now know about how these systems fail was learned from contact with actual users doing things no researcher predicted. Anybody arguing for a purely precautionary approach has to answer this, and most do not. Where the book weakens is on everything that would be expensive to concede.

The copyright and training data question, whether it is legitimate to train commercial models on creative work without permission or payment, is largely sidestepped. This is not a peripheral objection. It is a live legal question with substantial litigation attached and a large constituency of working writers, artists and musicians whose livelihoods are directly implicated. A book making the case for AI's benefit to human creativity that does not seriously engage with the people whose creative work was consumed to build it has a hole in it, and the hole is exactly where the author's interests are.

Labour displacement gets similar treatment: acknowledged at an altitude high enough that it costs nothing. Yes, previous technological transitions eventually produced more employment than they destroyed. The people who lost their trade in the interim did not experience an eventually, and the distributional question of who bears the transition cost is the one that actually matters politically. Abstraction is doing work here.

The superagency concept itself is thinner than the title suggests. Stripped of the framing it amounts to the observation that good tools make people more capable, which is true and not novel. The book is more valuable as a set of arguments about deployment policy than as a contribution of a new idea. To be fair to it, the optimistic case deserved a serious statement and had not really received one.

The discourse had tilted heavily towards risk, and a well-argued counterweight from someone who has actually built things is worth having, even heavily discounted for interest. It is also mercifully short and readable, which in this genre counts for something. Two point nine. The best available statement of the case that this goes well, containing one genuinely important argument about iterative deployment, undermined by silence on precisely the objections that would cost its author money and by a central concept that does less than it promises.

[ final ]

The verdict.

Read it as the best-argued version of the optimistic case, which is worth having in your head. Read it alongside a critical account, and keep the author's position in mind on every page, because the book will not remind you.