Summary

Meta faces an existential remedy—forced divestiture of Instagram and WhatsApp—but the FTC’s case depends on a market definition that TikTok may have rendered obsolete. Casey Newton thinks Meta acted aggressively against competitors yet doubts it still monopolizes “personal social networking,” while Kevin Roose’s sharper formulation is that Meta “had a monopoly, but that they failed to maintain the monopoly.”
The trial’s internal documents strengthen the case that Instagram could stand alone, even if they do not prove the FTC’s legal theory. Mark Zuckerberg considered a 2018 spinout because Instagram imposed a “strategy tax,” cannibalized Facebook, and might be worth more independently; Instagram reportedly now generates around half of Meta’s revenue, making separation survivable for Instagram and devastating for Meta.
Meta’s political courting failed to eliminate the trial, but antitrust pressure may already have helped produce a more competitive consumer-app market. Zuckerberg reportedly offered $450 million against an FTC demand of $30 billion after Meta transferred at least $26 million to Donald Trump through an inauguration contribution and a lawsuit settlement. Telegram, Substack, and even Clubhouse subsequently had “oxygen” to grow because acquisitions by dominant platforms became nonstarters.
Arvind Narayanan argues that headline adoption figures mistake experimentation for economically meaningful diffusion. ChatGPT may have roughly 500 million users and generative AI may reach 40% of US adults, but measured intensity is only around one hour per workweek, translating into “a fraction of a percentage point in productivity”—no faster, he says, than PC adoption 40 years ago.
Rapid capability gains do not automatically give AI systems power, because power also depends on how businesses and governments deploy them. Narayanan accepts that AI could accelerate AI research, but argues that unsupervised deployment often will not make business sense once failures are attributable; safety therefore calls for regulation and “defense in depth” across model developers and deployers, not faith in a self-correcting market.
Narayanan’s proposed safety line is stricter than alignment: society should refuse to hand systems consequential autonomy before alignment becomes existentially important. “If you even get to the stage where alignment becomes super important, you’ve already lost.” He suggests concrete constraints such as prohibiting AI from owning wealth, while conceding that military competition could be the “Achilles heel” of his normal-technology framework.
His forecast is gradual through 2027 but potentially revolutionary over 10 to 20 years. Work might rise only from roughly three to five hours per week using AI by 2027, while the longer-run definition of cognitive work shifts toward specifying tasks, supervising systems, and monitoring failures. “Normal technology” is not harmless: Narayanan compares the possible transition to an Industrial Revolution that eventually raised living standards but initially produced exploitation and destabilization.

Deep dive

1. Meta faces an existential remedy atop a narrow market

The FTC is asking a federal court to unwind Facebook’s 2012 purchase of Instagram and its acquisition of WhatsApp two years later. Casey called that outcome unlikely but existential: losing Instagram and WhatsApp would leave Meta looking like “a completely different company.”
Filed in December 2020, the original complaint alleged that Facebook illegally maintained a monopoly by buying emerging rivals. A judge dismissed it for offering too little evidence; after Joe Biden took office and Lina Khan assumed FTC leadership, the agency refiled with more statistics supporting its proposed market.
Everything turns on “personal social networking,” defined as apps primarily used to help people keep up with friends and family. The FTC places only Facebook, Instagram, WhatsApp, Snapchat, and little-known Facebook alternative MeWe inside that boundary.
Casey’s early handicap was blunt: Meta behaved “super anti-competitively” over the preceding decade, but the government is struggling to prove a present monopoly. Meta’s defense is that this market is effectively “fake,” invented for litigation rather than derived from how people now use social products.

2. TikTok made the FTC’s historical monopoly look obsolete

Casey thinks Meta plausibly dominated traditional social networking from roughly 2016 through 2021, after Snapchat and Twitter were neutralized as major threats and YouTube pursued a different use case. The problem is prosecuting that historical condition in 2025, after TikTok transformed the category.
TikTok’s breakthrough was to disregard the friends-and-family graph: “We’re just gonna show you the coolest stuff we can find on our network.” Personalized recommendations let users begin consuming without following anyone, forcing Meta and everyone else to chase its model.
Kevin’s framing captures the legal tension: Meta “had a monopoly, but that they failed to maintain the monopoly.” TikTok even bootstrapped its challenge by purchasing advertisements across Facebook’s own properties, while Zuckerberg testified that Meta initially missed the threat because it looked unlike familiar social-network competitors.
The apps have since converged—TikTok now encourages users to add friends and exchange messages, while Facebook fills feeds with algorithmic recommendations from creators, celebrities, and others users never chose to follow. That evolution makes the FTC’s boundary harder to defend and illustrates how a case aimed at a 2016-era problem can arrive after the market has moved on.

3. Zuckerberg’s own documents make breakup economics less absurd

A newly disclosed 2018 discussion showed Zuckerberg considering an Instagram spinout, partly because regulators might eventually demand one and partly because continued ownership imposed a “strategy tax.” He reasoned that businesses separated from large parents sometimes become more valuable after gaining independence.
Instagram remained unusually autonomous under founders Kevin Systrom and Mike Krieger, but its growth unsettled Zuckerberg. It looked “younger, hipper, sexier” as Facebook became older and “fuddy-duddy,” leading Zuckerberg to consider removing visible Instagram attribution when users cross-posted photos to Facebook.
The FTC can now argue that separation was sensible enough for Zuckerberg himself to consider. Reporting cited by Casey says Instagram contributes around half of Meta’s revenue—evidence both that Instagram could “survive and thrive on its own” and that losing it would be extraordinarily damaging to Meta.
Zuckerberg floated an even more radical reset in 2022: delete every Facebook friendship and make users rebuild their graphs around perhaps 20 people they still cared about. Facebook chief Tom Alison cautiously questioned whether that was “viable, given my understanding of how vital the friend use case is”; both hosts nevertheless thought the destructive proposal sounded appealing.

4. Political courting failed, but antitrust already changed the app market

Zuckerberg reportedly offered to settle for $450 million shortly before trial, a fraction of the FTC’s requested $30 billion. Casey read the offer as nearly “giving the FTC the finger”—a signal that Zuckerberg believed the government’s case was weak and expected favorable treatment from the Trump administration.
Meta had transferred at least $26 million to Trump: $1 million for the inauguration and $25 million settling litigation over his post–January 6 suspension. Trump and JD Vance also began criticizing European penalties against US technology companies, apparently encouraging Zuckerberg’s belief that the case might disappear.
That approach met internal resistance. According to the cited reporting, FTC chair Andrew Ferguson and Justice Department antitrust chief Gil Slater urged Trump to let the case proceed. Kevin considers the FTC’s theory weak but still welcomes Meta’s “day in court” rather than allowing the company to purchase an escape.
Casey’s broader result matters regardless of the verdict: the inability to freely acquire emerging services helped companies such as Telegram, Substack, and Clubhouse gain “oxygen in the market.” In the world of 2016, Casey thinks Facebook would have tried to buy them. Independent companies backed by billions in venture capital still face the harder question of how investors eventually realize returns.

5. Usage counts exaggerate how deeply AI has diffused

Kevin challenged Narayanan with unusually fast-looking adoption: ChatGPT is not yet three years old, reportedly has around 500 million users, and generative AI reaches roughly 40% of US adults. On raw user counts, AI appears to be diffusing far faster than earlier general-purpose technologies.
Narayanan’s “crazy-sounding claim” is that adoption is not actually accelerating. The 40% figure treats a worker relying heavily on AI like someone generating a weekly limerick; once intensity is measured, usage is around one hour per workweek and contributes only “a fraction of a percentage point in productivity.”
On that basis, he says this adoption is not faster than PC adoption 40 years ago. The bottleneck is not whether a laboratory can demonstrate a capability, but whether organizations redesign workflows, establish responsibility, learn from failures, and trust systems enough to incorporate them into consequential work.

6. Capability can accelerate without giving models power

Narayanan accepts Daniel Kokotajlo’s premise in AI 2027 that AI capabilities are improving rapidly and that AI could further accelerate AI research. His disagreement begins with language: “highly capable” and “highly powerful” are not synonyms.
Capability belongs to the model; power belongs jointly to the model and its deployment environment. Because people design that environment, society can refuse to hand systems increasing control and autonomy—even if laboratories make them substantially better at coding, planning, or other bounded tasks.
Automobile history supplies his incentive model. Manufacturers initially treated crashes as entirely the driver’s responsibility; once safety became attributable to manufacturers, bad engineering created reputational consequences and made regulation more feasible. Clear responsibility similarly could make AI safety a dimension on which suppliers compete rather than an optional brake on growth.
Kevin pressed with Waymo’s safety record, Cruise’s high-profile incident and subsequent pullout from San Francisco, and reporting that OpenAI gave testers less time and fewer resources. Narayanan answered, “I’m not confident”: markets will not self-correct. Regulation, journalism, and shared responsibility between developers and deployers must provide “defense in depth.”

7. Alignment arrives too late if autonomy is already unchecked

Conventional alignment reasoning moves from increasing capability to increasing power, then asks how to keep systems controlling economies or infrastructure faithful to human values. Narayanan rejects that sequencing: “If you even get to the stage where alignment becomes super important, you’ve already lost.”
Once a system holds superpowerful institutional authority, adjusting its technical alignment is, in his phrase, “a fool’s errand.” The meaningful brake comes earlier, between improved performance and the organizational decision that human supervision is no longer necessary.
Narayanan expects labs may declare AGI within a couple of years, as Kevin predicted, but doubts their chosen definition will describe a “drop-in replacement” spanning real human work. That requires much longer feedback loops in messy environments, plus learning what jobs actually entail from domain experts rather than narrow conceptions of those jobs.
Even if broad worker replacement becomes technically possible, unsupervised deployment would remain a bad business and policy choice. One deliberately simple intervention is banning AI systems from owning wealth, closing a route through which they could accumulate power while forcing humans to retain control at consequential points.

8. “Normal” AI can still reorder work and destabilize society

Narayanan stressed that “normal” does not mean trivial or safe: electricity and the internet qualify as normal technologies in his framework. The discussion identifies risks including bias, discrimination, job loss, inequality, concentrated power, democratic backsliding, and surveillance, even while rejecting near-term superintelligence narratives.
His concern is not merely that catastrophic-risk discussion distracts from present harms; he has stopped making that argument formally. The sharper claim is that proposed superintelligence remedies could worsen other dangers—most starkly, if fear of AI is used to justify a world authoritarian government.
The Industrial Revolution is his warning. It eventually lifted living standards, but its early decades drove workers into crowded tenements amid terrible safety and extensive child labor; the resulting horrors helped produce the modern labor movement. A decades-long AI diffusion could bring similarly severe adjustment before benefits arrive.
For 2027, Narayanan expects a world qualitatively similar to today—perhaps AI use rising, illustratively, from three to five hours weekly. Over 10 or 20 years, cognitive work could change fundamentally: humans would increasingly specify tasks, supervise AI execution, and monitor systems to ensure they are “not running amok.”

9. Diffusion, not a few months of model lead, drives geopolitical advantage

Casey challenged the paper’s claim that AI need not trigger arms races, noting that military rivalry usually creates them. Narayanan candidly called military AI a possible “Achilles heel of the whole framework”: he does not currently think such a race is likely, but said the team lacks enough confidence and is researching it further.
Outside warfare, he questions metaphorical races toward automating courts or other state decisions. Criminal justice is not a domain in which AI can simply perform at a superhuman level, and any efficiency advantage from eliminating judges would produce domestically felt civil-liberties consequences that citizens could resist.
Borrowing from political scientist Geoffrey Ding, Narayanan argues that geopolitical power comes less from invention than from spreading technology through the economy, government, and productive sectors. Innovations cross borders easily, while America’s AI lead may be “a few months at best”; diffusion is the step that takes decades.
Kevin and Casey pointed to export controls, US-China rhetoric, relaxed model restrictions, and a Paris AI summit that felt like a trade show. Narayanan’s distinction is narrower: a race clearly exists in model development, but he sees no equivalent race yet to place AI in charge of consequential systems.

10. AI risk turns on friction—and public discourse rewards extremes

In Narayanan’s thought experiment, developers release models with no safeguards. Kevin offered the classic danger: malicious actors use them to create a novel pathogen or bioweapon. Narayanan replied that open-weight models already have removable safeguards and much relevant information predates AI; otherwise, “we would all be dead already.”
Casey called that “a bit glib,” comparing it with synthetic sexual images: manipulation existed before, but instant generation radically increased abuse. Narayanan agreed about deepfakes but separated the threat models—reduced friction empowers impulsive teenagers using “nudification” apps, while a determined bioterrorist will not be stopped by three additional clicks.
For nudification abuse, Narayanan wants direct interventions: remove the tools from app stores and stop social platforms profiting from their advertising. He called policymakers’ slow response “shameful,” placing this abuse at the top of what he actually worries about from AI.
Narayanan also resists being boxed into utopia, dystopia, or “nothing to see here.” Skeptical posts receive 10 or 100 times the engagement of capability updates, creating audience capture; meanwhile, inflated corporate claims make disappointed users flip toward total disbelief. His requested corrective is mutual understanding between developers and people who know what jobs involve.

11. HatGPT found AI everywhere, including where it was not

Hacked Bay Area crosswalk signals impersonated Zuckerberg and Musk; the fake Zuckerberg said, “It’s normal to feel uncomfortable or even violated as we forcefully insert AI into every facet of your conscious experience.” Kevin proposed monetizing public infrastructure with sponsored crossing messages, an idea Casey immediately judged among his worst.
Andrew Cuomo’s 29-page housing plan contained odd passages and a link whose UTM code identified ChatGPT as the referrer. Adviser Paul Francis, who dictates after losing his left arm, acknowledged using ChatGPT for research but denied using it to write; the campaign’s wonderfully circular defense was, “If it was written by ChatGPT, we wouldn’t have had the errors.”
Google’s DolphinGemma, developed with Georgia Tech and the Wild Dolphin Project, learns dolphin-vocalization structure and generates dolphin-like sequences. Casey questioned both interspecies privacy and falsifiability; the institutional-comedy prize went to Education Secretary Linda McMahon repeatedly calling AI “A.1.,” prompting Kevin’s verdict: “We are so cooked.”
Japan assembled a just-over-100-square-foot train station in six hours from components printed offsite over seven days, replacing a wooden station that had served the community for more than 75 years. Blue Origin’s brief all-women flight was framed by Casey and Amanda Hess as publicity: Hess wrote that it showed several women had amassed enough social capital “to be friends with Lauren Sanchez.”

a16z's $20BN Fund & Founders Fund's $4.6BN & Why Josh Kushner Has Mastered the Game Robert Rodriguez: Sin City, Desperado, El Mariachi, Alita, and Filmmaking | Lex Fridman Podcast #465