Superagency: What Could Possibly Go Right with Our AI Future
Here is the 1PAGE summary for *Superagency*:
---
Every new technology we've invented—from language, to books, to the mobile phone—has defined, redefined, deepened, and expanded what it means to be human.
— Hoffman and Beato, *Superagency*, Introduction
The argument Hoffman and Beato are making is simple: intelligence is about to stop being scarce, and that changes everything. Every previous technology that democratized a resource — the printing press with text, GPS with navigation, the internet with information — produced more human capability, not less. AI is the same pattern at a larger scale, and the name they give the result is superagency: individuals equipped with tools that amplify what they can know, decide, and accomplish. The case rests on the claim that our recurring instinct to regulate new technology out of precaution has historically cost us more than it saved us.
The book is at its best when it goes concrete. The mental health example is genuinely striking: the United States has a severe shortage of therapists, a massive population of untreated people, and an AI that most users rate as more empathetic than the overworked clinicians they can barely access. That collision between real scarcity and available technology makes the pro-AI argument more than hand-waving. The concept of "iterative deployment" — releasing technology, watching what breaks, and fixing it — is also a more honest account of how good technology actually gets built than the precautionary fantasy where experts predict all harms before anything ships. Hoffman knows how software development works because he's done it, and that knowledge shows.
Fundamentally, the surest way to prevent a bad future is to steer toward a better one that, by its existence, makes significantly worse outcomes harder to achieve.
— Hoffman and Beato, *Superagency*, Introduction
Where the book thins out is exactly where it matters most. The economic inequality angle gets a single Ted Chiang citation — Hoffman even quotes the critique that computing wealth "mostly served to increase the wealth of the top one percent of the top one percent" — and then marches on without engaging it. This is the strongest objection to the whole argument, and the book doesn't wrestle with it seriously. If AI amplifies capability, it will amplify existing advantage most efficiently, and Hoffman's answer ("let's make sure everyone has access") is an aspiration, not a plan. The book also leans hard on historical analogies — cars, phones, the internet — without acknowledging the ways those comparisons might be precisely wrong. AI systems that reason, persuade, and generate content are categorically different from a motor vehicle, and asserting the parallel isn't the same as defending it.
Distributing intelligence broadly, empowering people with AI tools that function as an extension of individual human wills, we can convert Big Data into Big Knowledge, to achieve a new Light Ages of data-driven clarity and growth.
— Hoffman and Beato, *Superagency*, "Big Knowledge"
Hoffman's conflict of interest is acknowledged but not really confronted. He has invested in OpenAI, founded competing AI companies, and sits on Microsoft's board. He notes this and says his involvement deepens his conviction rather than biasing it. That may be true. It's also exactly what someone would say if their conviction were biased. A reader is entitled to notice the gap.
Read it as the optimist's brief, which is what it is. The best arguments in the book — that waiting for perfect certainty is itself a policy choice with real costs, that we shape futures by building them rather than prohibiting what we fear — are genuine contributions to a debate dominated by catastrophizing. Anyone who thinks the doomer framing is lazy or lazy-comfortable will find useful ammunition here. Just don't mistake advocacy for analysis.
Read the longer summary
What the book is actually arguing
Reid Hoffman and Greg Beato have written a long argument for one straightforward idea: AI will hand billions of people new powers, and we should build toward that future instead of trying to legislate it away. They call this superagency — the moment when a technology gives one person new abilities and does it for so many people at once that the texture of daily life shifts. Cars did this. The internet did this. AI, on Hoffman’s read, is doing it now and will keep doing it.
The organizing device is what the authors call a “techno-humanist compass.” A compass, not a blueprint. It tells you which direction to steer when something new shows up, but doesn’t promise where you’ll end up. Compass north, in this case, is human agency: any AI policy, deployment, or product decision should be judged by whether it expands what individuals can do or contracts it. That sounds anodyne until you notice how often it cuts against current AI discourse, which tends to frame harms as a function of capability rather than agency.
Hoffman is the more visible co-author. He cofounded LinkedIn, sat on OpenAI’s board, was an early investor in OpenAI, and now runs Manas AI and Inflection. Beato is the writer. Reading the book, you can feel the seams. Beato is doing the lifting on prose and historical analogy; Hoffman is supplying the network and the convictions. The convictions are unmistakably Hoffman’s: build, deploy, iterate, don’t pause.
The four camps
Early in the book, Hoffman and Beato sort the AI conversation into four groups: Doomers, Gloomers, Zoomers, and Bloomers. Doomers think AI poses an extinction-level threat. Gloomers think it will erode jobs, privacy, and human meaning short of catastrophe. Zoomers want acceleration with minimal friction. Bloomers, Hoffman’s camp, want fast deployment paired with active steering toward broadly positive outcomes.
This taxonomy is doing real work. It lets the book frame the debate as a four-way conversation rather than the usual binary of techno-optimists versus AI safety people. It also conveniently positions the authors as the reasonable middle, which is a familiar Silicon Valley move. Read carefully, the Bloomer position is closer to the Zoomer one than to anything else. The disagreement with the acceleration camps is largely about tone and stakeholder management, not pace.
What’s missing from the taxonomy is the camp that thinks first about who gets harmed earliest and most by deployment. Concentration risk, labor displacement that lands disproportionately on specific occupations, the fact that “iterative deployment” means the iteration happens on people. The book mentions these concerns and moves on. It’s the biggest gap, and we’ll come back to it.
Iterative deployment is the strongest argument
If you read Superagency for one idea, this is the one. Hoffman borrows iterative deployment from OpenAI’s own house philosophy: ship something usable, watch real users break it, fix it, ship again. The argument against the precautionary principle, the idea that disruptive technology should be held back until proven safe, is that you can’t actually predict downstream effects of a general-purpose technology in a lab. You learn by deploying.
Hoffman’s strongest historical analogy is the car. In the early decades of automobile adoption, fatalities were grim, regulation was minimal, and the vehicles themselves were dangerous machines. Rather than pause the technology, the industry, hobbyists, and eventually regulators pushed through millions of small adjustments. Better brakes, seat belts, traffic lights, driver education, road design. Today the death rate per mile driven is a tiny fraction of what it was a century ago. You couldn’t have engineered that outcome by sitting in a 1908 conference room and writing rules.
Fundamentally, the surest way to prevent a bad future is to steer toward a better one that, by its existence, makes significantly worse outcomes harder to achieve.
That, from Hoffman and Beato, is the clearest articulation of the thesis. It reframes safety as a positive program rather than a defensive one. Build the world you want; the worse versions become structurally harder once enough of the better one exists.
This is genuinely useful framing. Standard AI safety conversation defaults to constraint: pauses, licensing, capability evaluations, red lines. Hoffman is arguing that constraint without a positive vision tends to produce stagnation, regulatory capture, and a vacuum that adversarial actors fill. The strongest version of his argument isn’t “don’t regulate AI.” It’s “if you regulate, regulate toward a future you actually want, not just away from outcomes you fear.”
The weakness is that the car analogy hides a lot. The casualty rate during the learning curve was real. Iterative deployment of AI means iterating on people whose jobs disappear, whose information ecosystems flood, whose mental health apps get tested on them. The book waves at these costs without pricing them.
What the chapters actually contain
The chapter list reads like a tour of Hoffman’s hobbyhorses: Big Knowledge, The Triumph of the Private Commons, Innovation Is Safety, Informational GPS, Law Is Code, Networked Autonomy, The United States of A(I)merica. The best chapters are the concrete ones.
Big Knowledge engages with the surveillance critique. Hoffman knows the standard reference is Orwell, and he goes after it directly. His move is to argue that historical fears about centralized data have generally not played out as predicted. The personal computer democratized information rather than enabling Big Brother. The argument lands partly. It’s also where Hoffman makes one of the book’s most strained claims: that Orwell’s 1984 failed to imagine the bidirectional possibilities of a telescreen. As one reviewer put it, this turns a literary symbol of total submission into a missed product opportunity. Hoffman is treating the dystopia as a design flaw, which is the kind of move that gives the book its tone of cheerful obtuseness about power.
Informational GPS is the chapter most likely to convince a skeptic. Hoffman’s analogy is that GPS didn’t replace human navigation. It gave everyone access to the kind of spatial expertise that used to belong to taxi drivers and cartographers. AI, on this read, is a similar layer of always-available expertise. The student who can ask a tutor any question at midnight, the new customer-service rep who can draw on the institutional knowledge of veterans, the immigrant who can decode a legal notice in their own language. The book cites a study showing AI assistance produced the largest productivity gains for the least experienced workers, which is the genuinely interesting empirical finding here.
Law Is Code is where Hoffman makes his case for software-style regulation. Laws should be written so they can be updated as the technology shifts, with public-private collaboration replacing rigid up-front rules. This is Hoffman’s actual policy preference, and it has more substance than the rest of the book’s regulatory commentary. The Colorado AI law that critics like Brad Feld have written about is a real-world example of what Hoffman is against: broad up-front restrictions written before anyone knew what the actual harms looked like.
The United States of A(I)merica is the geopolitical chapter. Hoffman argues that the US should treat AI as a strategic priority on the order of the Manhattan Project or Apollo, with the explicit goal of leading over China. This is the chapter where Hoffman’s politics are clearest and where he aligns most directly with the broader Silicon Valley consensus that strangling US AI development hands the future to authoritarian states.
Where the book is weak
The recurring criticism in serious reviews, and we agree with it, is that Superagency refuses to engage with its strongest opponents. Doomers get dismissed as catastrophists. Critics like Emily Bender and Kate Crawford get name-checked but not answered. The pattern is to introduce a concern, gesture at it, pivot to a reassuring historical analogy, and move on.
Two examples make this concrete. On AI companions and mental health, Hoffman cites a JAMA study where physicians rated ChatGPT’s responses as more empathetic than human doctors’ responses, and concludes that AI on tap will increase our capacity to be kind to other people. That’s a leap. The opposite hypothesis, that frictionless AI emotional support reduces our practice of human connection, is at least as plausible, and the book never engages it. The hypothesis becomes an assumption.
On labor, Hoffman acknowledges job displacement and pivots to the long historical record of new jobs replacing old ones. That’s true at aggregate timescales. It’s not a comfort to the people in the middle of the transition, and it’s not an argument. It’s an article of faith. The reviewer at aibookreview.substack.com, who is clearly sympathetic to Hoffman’s broad project, lands the right phrase: the book is rich in faith. It extrapolates from past technological transitions and asks the reader to trust that this one will follow the same arc.
The other weakness is mechanical optimism. Hoffman is too quick to convert downsides into upsides through analogy. Filter bubbles get briefly noted, then dispatched with a counter-narrative about algorithmic springboarding, YouTube as a self-improvement engine. Surveillance gets reframed as participatory governance. AI-driven legal automation gets framed as access to justice. Each has a kernel of truth. None is the full picture, and the book doesn’t pretend to give the full picture.
The political subtext
Hoffman doesn’t write as a partisan, but the politics are visible. The central regulatory commitments, permissionless innovation, light-touch oversight, fast iteration, line up with a particular American free-market tradition. The geopolitical framing is straightforwardly hawkish on China. The implicit theory of progress is that private actors, properly motivated and lightly constrained, deliver better outcomes than centrally planned ones.
We don’t think this is hidden, and as a default it isn’t wrong. The Colorado AI law is a real artifact of preemptive regulation done badly: a state legislature passing broad restrictions in May 2024, the governor signing with explicit reservations, and a year of failed amendments resulting in a four-month delay. Hoffman’s policy intuitions track the failure mode here. Measure first, regulate against measured harms, iterate.
What’s missing from the political picture is any serious account of who gets harmed during the iteration. Hoffman is comfortable saying “we” should accept some risk and uncertainty. The “we” is doing a lot of work. It’s easy to accept iterative deployment when you’re the one doing the deploying. The book never reckons with the fact that the costs of iteration land on specific people, workers and students and patients and citizens, who didn’t sign up to be part of an experiment.
Who should read it
If you’ve been reading the doomer literature and haven’t found a serious counterweight, Superagency is the counterweight. It’s not the best argument the optimists could make. Hoffman’s reluctance to engage with strong critics holds the book back. But it’s the argument from someone who has been close enough to the action that his framing has practical weight.
If you’re a developer, founder, or policy person trying to think about how to deploy AI responsibly, the iterative deployment chapter alone is worth the price. The framing is portable to whatever you’re building.
If you’re already a Hoffman skeptic, the book won’t change your mind. The criticisms that have accumulated in serious reviews, that it’s faith-based, that it ignores power asymmetries, that it treats every historical analogy as bullet-proof, are valid criticisms, and reading the book confirms them.
If you want to understand why a significant slice of the AI industry believes that pausing or heavy-handed regulation is the worst possible move, Superagency is the cleanest articulation of that worldview in book form. Read it alongside something more skeptical. Richard Susskind’s How To Think About AI is the most honest balanced treatment we’ve seen, and the two books together give you a more complete picture than either does alone.
The book we’d most like to read, the one neither Hoffman nor his critics have written, is the one that takes the techno-humanist compass seriously and then asks honestly which deployments expand agency for whom, and which ones quietly contract it. Superagency gestures at that book without writing it. The next one might.