In 1963, the Internal Revenue Service installed a $1.2 million IBM 7074 mainframe computer to process tax returns. The machine was designed to verify calculations and enforce compliance, a seemingly mundane administrative upgrade. The public reaction was immediate terror. Taxpayers streamed into local IRS offices across the country to pay back taxes they had previously dodged, panicked by the sudden omniscience of a machine that could theoretically cross-reference the entirety of their financial lives. A few years later, when the federal government proposed centralizing its scattered datasets into a single “National Data Center,” the backlash in Congress was absolute. Politicians warned of an impending Orwellian nightmare. One congressman prophesied the creation of “The Computerized Man,” stripped of his individuality, his entire existence reduced to a magnetic tape. Author Vance Packard testified that this relentless drive for efficiency would end with humanity in “chains of plastic tape”. The proposal was abandoned. The apocalypse was presumed to be merely delayed.
We always panic. I spent the early part of my career studying the history of technological assimilation in Silicon Valley, watching the exact same cycle of dread and capitulation play out with search engines, social media, and location tracking. We imagine the new machine will end human autonomy, and then, a decade later, we throw a minor tantrum if the machine takes more than three seconds to locate the nearest artisanal coffee shop.
Reid Hoffman and Greg Beato understand this cycle intimately. In Superagency: What Could Possibly Go Right with Our AI Future, they trace the historical rhythm of techno-panic from the fifteenth-century printing press to the twentieth-century automobile. The central thesis of the book asserts that artificial intelligence is not a harbinger of civilizational collapse, but rather a democratizing engine for “superagency”—a cognitive revolution that will exponentially amplify human energy and intelligence if deployed openly and iteratively. The authors argue that the safest and most effective way to integrate AI into society is through broad, hands-on public access rather than preemptive regulation or prohibition.
The book divides the current discourse into four distinct camps: Doomers, Gloomers, Zoomers, and Bloomers. Doomers fixate on existential risk, prophesying superintelligent machines that will destroy humanity, leaving perhaps a few tech bros alive to run errands for rogue Roombas. Gloomers ignore the sci-fi apocalypse to focus on near-term harms like job displacement, algorithmic bias, and the destruction of individual agency, advocating for strict top-down control. Zoomers want absolute freedom to innovate at breakneck speed with zero government interference, trusting that markets will sort out the collateral damage.
Hoffman plants his flag firmly in the Bloomer camp. He favors mass engagement and “iterative deployment”—releasing tools into the wild to learn from real-world usage while continuously course-correcting. This is the exact philosophy OpenAI utilized when dropping ChatGPT onto the internet with zero marketing and zero warning in November 2022.
He leans heavily on the history of the automobile to map this trajectory. Early cars were dangerous, noisy, and deeply unpopular. Critics called them “devil wagons” and “death cars,” predicting they would destroy the American family by encouraging unmarried men to blow their savings on a Model T instead of settling down. In 1923, the motor-vehicle death rate was nearly nineteen deaths for every one hundred million miles driven. Yet society did not ban the car. We iteratively deployed safety measures. We built traffic lights, mandated seat belts, established speed limits, and eventually achieved a 93 percent reduction in fatalities per mile. We tolerated the friction and the fatalities because the reward was personal mobility on an unprecedented scale. Cars functioned as new nodes in a network that collapsed time and space, serving as a flywheel for self-determination that diffused across society.
The Donner Party makes a grim appearance to drive the point home. Hoffman compares modern drivers, safe in their surveilled, climate-controlled cabins, to the doomed pioneers who resorted to cannibalism because they lacked infrastructure on the untrammeled frontier. The implication is that technology frees us from the brutal whims of nature. The assumption here is that AI will follow the exact same path. It will cause accidents. It will hallucinate facts. It will disrupt labor markets. But the expansion of human capability will be so immense that trying to pause its development would be an act of societal self-sabotage.
I find this argument persuasive at the level of historical mechanics but weak at the level of political economy. Hoffman views the automobile as an unambiguous victory for individual freedom. He ignores the structural coercion of infrastructure itself. The automobile did not merely expand our menu of choices. It actively eliminated the viability of alternative modes of existence. It decimated public transit, fueled suburban sprawl, and reshaped the built environment so totally that owning a car became a prerequisite for economic survival in most of the country.
When you look at AI through this lens, the promise of “superagency” starts to look like a mandatory toll. If AI becomes the “informational GPS” that Hoffman envisions—an intelligent interface mediating our interactions with government, healthcare, and commerce—opting out will become functionally impossible. The freedom to use AI will rapidly morph into the obligation to use AI. We will be free to operate at superhuman speeds, provided we pay the monthly subscription fee to the corporations maintaining the servers.
Hoffman attempts to preempt this exact line of criticism by tackling Shoshana Zuboff’s The Age of Surveillance Capitalism. Zuboff argues that companies like Google operate a “Big Other,” a sensory infrastructure that extracts “behavioral surplus” from our lives to predict and modify our actions. She compares this extraction to the poaching of elephants for ivory. Hoffman rejects her central metaphor entirely. You cannot extract data the way you extract coal, he insists. Coal is finite; when you mine it, you leave a hole in the ground. Data is nonrivalrous and infinite. Taking a digital photograph does not deplete the global reserve of imagery.
Instead of an extraction operation, Hoffman prefers the term “data agriculture”. He argues we are participating in a “private commons”. In this private commons, users freely provide the raw material of their lives, and in exchange, they receive services of staggering utility. Hoffman cites research by Erik Brynjolfsson and Avinash Collis to measure this consumer surplus. Their studies revealed that the median user would need to be paid $17,530 to give up search engines for a single year, and nearly $8,500 to give up email. By this logic, the $44.60 in annual advertising revenue that Meta squeezes from the average user is a wildly asymmetrical bargain in the user’s favor. We are not being exploited. We are getting the deal of the century.
This is a clever rhetorical maneuver. By shifting the focus to consumer surplus, Hoffman makes the critic look ungrateful. But Zuboff’s anxiety was never about data scarcity. Her concern was asymmetry of power. The danger of the private commons is not that the pasture will run out of grass, but that the entity managing the pasture controls the gates, sets the rules of engagement, and uses its totalizing visibility to shape the behavior of the sheep. A digital commons governed by a profit-maximizing corporation is not a commons at all. It is a company town. Hoffman’s optimism relies on the belief that the incentives of the platform will remain roughly aligned with the well-being of the user. History has consistently proven that alignment to be fragile at best.
The debate moves from the abstract to the intimate in the book’s exploration of mental health care. Rates of depression and suicide have climbed drastically over the last two decades. Millions of Americans live in professional shortage areas, waiting up to six months to speak to a therapist. Hoffman points to Rob Morris, a developer who added GPT-3 functionality to his peer-based mental health service, Koko. The platform uses cognitive reappraisal, teaching users to modify negative emotional responses by reframing their thoughts. When GPT-3 was integrated, users had a “copilot” to help draft empathetic replies to peers in distress. Users receiving messages coauthored by the AI rated them significantly higher than those composed entirely by human peers.
When Morris tweeted about the experiment, the public backlash was vicious. Critics accused him of conducting unethical experiments on vulnerable populations with a machine incapable of empathy, labeling his actions “immoral shitbaggery”. The feature was quickly disabled.
Hoffman defends Morris aggressively. He calls the public outrage a symptom of “problemism”—a default mode of viewing technology as an inherently suspect force, prioritizing precaution over innovation. The critics focused on the hypothetical harm of a hallucinating chatbot while entirely ignoring the ongoing, catastrophic harm of the status quo. Traditional mental health chatbots rely on rigid, predefined scripts. When a user expresses a subtle, unscripted feeling of emptiness, the legacy bot stumbles, outputting a generic response that makes the user feel utterly unseen. Large Language Models navigate the subtlety of human despair with far more grace.
If an AI can analyze vast transcripts of counseling sessions to identify effective evidence-based practices, and if it can provide on-demand support at three in the morning when a human therapist is asleep, why should we prohibit it? Hoffman envisions a future where mental health care is distributed like Spotify—affordable, scalable, and highly configurable to user preferences. You could test-drive multiple virtual therapists, assemble a team of them to debate your issues in real time, and curate a custom therapeutic blend based on your engagement history.
I stop to consider the sheer pragmatism of this vision. It is incredibly difficult to argue against access. If a person is in crisis and the only available support is an LLM, the statistical probability of empathy is better than the absolute certainty of silence. Hoffman pushes this even further, introducing the concept of the “superhumane.” He points to a study from JAMA Internal Medicine where licensed physicians rated ChatGPT’s responses to patient questions significantly higher than the responses written by actual human doctors. He suggests that constant interaction with infinitely patient, emotionally attuned AI models might actually condition humans to become kinder and more generous in their interactions with one another.
The idea that a machine might teach us to be more human is deeply unsettling. It forces us to ask whether empathy requires shared vulnerability, or if it is simply a linguistic pattern that can be mapped and replicated at scale. The book does not linger on the metaphysical implications. It is too busy building.
This velocity is characteristic of the field itself. The developers building these systems are obsessed with measurement. Hoffman spends a significant portion of the text demystifying the “arms race” narrative. The media portrays AI development as a reckless sprint toward doomsday, warning of ethical casualties and unintended consequences. The reality inside the labs is a culture of exhaustive, granular testing. From the Turing Test to SuperGLUE to the Word Error Rate benchmark, AI progress is driven by gamified metrics.
Benchmarks serve as a dynamic alternative to static regulation. A law dictates what you cannot do; a benchmark dares you to do better. When a model fails an adversarial input—when altering a few pixels causes it to mistake a zebra for a toaster—developers retreat to the lab, adjust the parameters, and test again. To prove general intelligence, a Microsoft research team asked GPT-4 to draw a unicorn using TikZ, a command-line programming language typically used for rendering geometric charts. The model had to translate the abstract concept of a unicorn into mathematical coordinates. Over the course of a month of iterative refinements, it succeeded, producing an image with a distinct horn, mane, and equine proportions. This is more than pattern matching; it is synthesis.
Hoffman points to the Chatbot Arena, an open-source platform where users blindly vote on responses from competing LLMs, stack-ranking them like college football teams. This crowdsourced evaluation creates a feedback loop of continuous improvement based entirely on human preference.
The limitation of this approach is glaring. Optimizing for human preference is exactly how social media algorithms engineered the filter bubbles that currently poison our civic discourse. Hoffman attempts to counter this by highlighting “algorithmic springboarding.” He argues that the same recommendation engines accused of radicalizing youth can also lead a bored teenager from a video game forum to a Python coding tutorial. Every tool of radicalization is also a tool of education. The technology is agnostic; the agency belongs to the user.
That is a very clean ideological position. The invisible hand of the internet bends toward knowledge. I remember hearing the exact same rhetoric in 2011, when social media executives sincerely believed that connecting the world would automatically result in global harmony.
Yet, despite my skepticism regarding the corporate stewardship of these tools, it is impossible to dismiss the raw utility Hoffman outlines. The capacity of multimodal models to process text, audio, and visual inputs in real time alters the dynamic of human-computer interaction. When OpenAI demonstrated GPT-4o, the latency dropped to 232 milliseconds. The interaction became conversational, paced like human dialogue. When a smartphone app can look at the physical world through a camera lens and translate a foreign film live, or analyze a golf swing, the device stops being a portal to the cloud and becomes an active participant in the physical environment.
We are moving from tools that require our explicit command to agents that operate with networked autonomy. Hoffman asks us to imagine a near-future where all new vehicles are equipped with federally mandated Driver Alcohol Detection Systems for Safety (DADSS), using touch and breath sensors to physically prevent a drunk driver from starting the engine. He acknowledges that civil liberties advocates will resist this as an infringement on personal autonomy. But he argues that drunk driving is a well-established societal risk, and the “consent of the governed” allows us to trade certain freedoms for collective security.
Code is law. Lawrence Lessig warned us of this in 1999, noting that the architecture of cyberspace regulates behavior far more efficiently than state legislation. Today, that architecture has bled into the physical world. It is in our cars, our appliances, our infrastructure. Hoffman’s optimism requires us to trust that democratic institutions can adequately oversee the deployment of this code. He argues that the United States must lead AI development because isolationism won’t protect us from autonomous weapons devised by authoritarian regimes, and protectionism won’t save our economy from competing nations armed with synthetic scientists. The geopolitics are binary. We build it, or someone else builds it and uses it against us.
There is a relentless momentum to this book. It refuses to wallow. When Hoffman encounters a potential hazard, he acknowledges it briefly, proposes a technical or regulatory mitigation, and immediately resumes his march toward the horizon. He treats the future as an engineering problem that can be solved if we just collect enough data and write enough tests.
The approach works best when dismantling the paralyzing fear of the unknown. Hoffman is right to criticize the precautionary principle, which demands absolute proof of safety before any action is taken. He notes that the Department of Defense created GPS for military use in 1973, restricting civilian access through a policy called Selective Availability. When Bill Clinton ended the policy ahead of schedule in 2000, it didn’t cause a catastrophic breach of security. It birthed global navigation, logistics optimization, and ride-sharing apps. A precautionary standard would have suffocated this utility in its infancy. By insisting on iterative deployment, Hoffman provides a practical roadmap for managing risk without halting progress. We learn by doing. We release the model, watch it fail, and patch the vulnerability.
Where the argument falters is in its presumption of equitable distribution. Hoffman states that the more we commit to AI, the more every individual is likely to benefit. He imagines a world where every aspiring entrepreneur has access to AI-powered market analysis, leveling the playing field and creating prosperity that ripples through society.
I have seen this movie before. The internet was supposed to level the playing field, too. It did democratize information, but it also centralized wealth and power in the hands of a few monopolistic platforms. Hoffman’s vision of superagency is thrilling, but it relies on an economic model that treats users as raw material for a corporate harvest. The private commons is a beautiful idea, but it lacks the structural safeguards of an actual public utility.
Take his vision of the ultimate AI travel assistant. He imagines a tool that scans your calendar to learn about an upcoming business trip. It reads your payment history on Square to calculate your restaurant preferences based on tip size, cross-references Yelp reviews, books the table, and then checks your Strava data to map a morning jogging route that perfectly matches your training goals. The convenience is absolute. The aggregation of intimate behavioral data required to deliver that convenience is equally absolute. Hoffman asks us to celebrate the efficiency and ignore the panopticon.
And yet, I cannot bring myself to reject the tools. The utility is too vast. The potential to analyze the human genome, to optimize energy grids, to bring cheap education to the global south—these are not trivial promises. The intelligence is abundant. The models are learning.
Hoffman leaves us with a provocative assertion. He suggests that we are inhabiting the world before Copernicus again, standing at the edge of unmapped frontiers. The technologies we invent do not challenge our humanity; they define it. We are Homo techne as much as Homo sapiens, constantly creating tools to shape the world, only to have the tools shape us in return.
We will change. We will offload our cognitive burdens to server farms in the desert. We will speak to our phones, and our phones will speak back with the calibrated patience of a saint. We will navigate the world with a synthetic copilot whispering in our ear. The agency we gain will be spectacular. The autonomy we surrender will be quiet, gradual, and entirely voluntary.
If you’d like to read the full book in EPUB or MOBI format, feel free to send me an email—I’d be happy to share a free copy with you. Please reach me at: thenovaleaf@gmail.com