Talks

Gilded Rage: Tech Billionaires, AI and Power — Jacob Silverman on Democracy Now!

61 min read

About this talk

Speaker: Jacob Silverman (guest); Amy Goodman and Nermeen Shaikh (hosts)
Event: Democracy Now! (daily news programme), segment on tech billionaires, AI and the Trump administration
Date: 2026-09-24 (programme page dated 24 Sep 2026; Trump’s UN speech was Tue 22 Sep; Trump–Xi state dinner “tonight” = Thu 24 Sep)
Length: 27:01 (transcript; sign-off at 26:31)

How to read this page. It has two kinds of content, always marked:

  • 🎙 In the talk. What Silverman, the hosts, or the clips (Trump, Sacks, Coxon, Hawking) say, with [MM:SS] timestamps. Only these parts are the speakers’ views. Each quote names who speaks.
  • 📚 Deeper theory. Background from history, economics, law and AI-risk debates. It explains why the speakers argue as they do and who disagrees. It is not their claim unless a timestamp says so.

Suggested route. Summary → Map → Key terms → Part 2: The intellectual map (read once, slowly; it carries most of the depth because the talk is short) → Deep dives in order → “Putting it together”.

A note on the source. This is a left-leaning, audience-funded news programme. Silverman is a critic of the people he discusses. That is not a reason to dismiss him. It is a reason to separate (a) things that can be checked, (b) his interpretations, and (c) his characterisations of people’s motives. Part 3 does this item by item. Facts from my own memory (not checked in this session) are marked “from memory”.

Speakers

LabelWhoRole
Host 1Amy Goodman (named at sign-off [26:31])Co-host of Democracy Now!. Opens and closes; asks most questions
Host 2Nermeen Shaikh (named at sign-off [26:31])Co-host. The captions do not mark the switch between hosts, so questions are labelled “host” on this page
SilvermanJacob SilvermanJournalist; “reporter at the International Consortium of Investigative Journalists” (the hosts’ introduction [02:38]). Author of Gilded Rage: Elon Musk and the Radicalization of Silicon Valley (paperback “out this week”). Main speaker, joining from Brooklyn
Clip: TrumpDonald TrumpSpeech at the UN General Assembly, “on Tuesday” [00:00]
Clip: SacksDavid SacksCo-chair of PCAST; speaking at Politico’s Decoded Summit [01:31]
Clip: CoxonThe captions say “Jacob Cox”; a researcher who left AnthropicPBS NewsHour interview “last week”, on recursive self-improvement [22:32]
Clip: HawkingStephen HawkingBBC interview, 2014 [24:32]

Identity checks (web search). The programme page and date are confirmed: Democracy Now!, 24 Sep 2026, “Gilded Rage”. The book’s title and subject are confirmed by retailer and review listings (Foreign Affairs review; listings name Bloomsbury Continuum as publisher, first published 7 Oct 2025 per Goodreads). The paperback edition and the exact US publisher were not confirmed (not confirmed). The departing Anthropic researcher is reported in the press as Jacob Coxon, a pretraining researcher, who posted his warning in early September 2026 (TechCrunch, 9 Sep 2026). I found no PBS NewsHour page; that the clip is from PBS rests on the hosts’ words (not confirmed).

Caption warnings and corrections.

  • “Guilded Rage” = Gilded Rage.
  • “Emad” in Silverman’s answer at [04:57] = Amodei (Dario Amodei, Anthropic CEO). He says “Amodei” first, then the caption shows “Emad”.
  • “Dorio” in the Sacks clip [01:31] = Dario. Sacks says “Dario and Sam”, meaning Amodei and Altman.
  • “Sushil Balaji” and “Sushil” [19:04]–[22:32] = Suchir Balaji. The hosts’ own line at [19:04] is captioned correctly.
  • “Jacob Cox” and “Jay- Jacob” [22:32] = Jacob Coxon (press spelling).
  • “banning MS, now CNN” [06:11] = MS NOW (the cable channel). “Status” = the news site Status. See the Political Wire note.
  • “conspiracizing” is Silverman’s word, not a caption error.
  • Silverman corrects himself at [16:24]: Musk’s money will flow into “the 2026 midterms”, not “the 2028 election”.

Factual slips and claims to check.

  • “180 days” [00:52]: the host says Sacks’s special government employee term “expired” after 180 days. Press reports say a 130-day allowance, used up, and that he left the czar role on 26 Mar 2026 (TechCrunch). The PCAST co-chair role is confirmed.
  • “11 years old” [25:14]: Hawking’s BBC interview was in 2014 (from memory: December 2014). From 2014 to 2026 is 12 years. Probably the script was written for 2025.
  • “UN Security Council” [08:47]: the opening says AI CEOs spoke “at the United Nations” and Trump spoke at the General Assembly (confirmed: 81st UNGA, 22 Sep 2026; Fox Business). Whether CEOs addressed the Security Council is not confirmed.
  • Anthropic, Pentagon, “supply chain risk” [04:57]: Silverman says Trump tried to designate Anthropic a supply-chain risk “but that was after Anthropic got a deal with the Pentagon”. The Pentagon’s designation was announced 5–6 Mar 2026, after a dispute in which Amodei refused to drop limits on mass surveillance and autonomous weapons (NPR via VPM). The order he gives (deal first, then clash) is roughly right. His claim that the software is “probably” used in targeting air strikes is his own inference (he says “probably”); not confirmed.
  • “Only Amodei skipped the inauguration” (host, [04:57]): not confirmed.
  • “First trillionaire” (host, [14:43]): not confirmed here.
  • Trump–Xi dinner guest list [08:47]: the host lists Zuckerberg and Altman among guests. Reports confirm a heavy tech guest list (Huang, Musk, Cook, Pichai, Su, Bezos and others) for the dinner on 24 Sep (AJC). I could not confirm Zuckerberg or Altman individually.
  • Musk and “immigration fraud” [17:11]: Silverman says Musk came to the US “possibly through via immigration fraud”. The press record is narrower: the Washington Post (Oct 2024) reported that Musk briefly worked without proper authorisation in 1995–97; Musk says he was legally there (Stanford Daily). “Fraud” is Silverman’s stronger word.
  • Balaji [19:04]: Silverman says he thinks it was a suicide and that he did not review the physical evidence much. Official findings agree with suicide; the parents dispute it (see Deep dive 6).

Summary

Central question. Who is shaping US AI policy, and what does it mean that a small, rich, politically organised group of tech leaders now sits next to state power?

Thesis (Silverman). The people who lead big tech and AI are not neutral engineers. A network of right-leaning billionaires (Thiel, Sacks, Musk, Andreessen) won influence in Washington. They are driven by anger: at “woke” culture, at past regulation, at a utopian future they feel was promised and not delivered. Hence the book title, a play on the Gilded Age. AI makes the link to power tighter, because AI needs state-scale resources (money, land, energy, chips) and the state can supply contracts and equity. So the firms behave “like their own little nation-states” and need friendly authoritarian partners [10:41].

Three further claims.

  1. Doom talk is also a sales pitch. When AI leaders say their product may escape control, that “attracts good attention to these companies” and lets them dodge ordinary liability [12:27], [23:53].
  2. The harms are present, not only future. He names “AI psychosis”, use in war zones, and data theft. These “don’t necessarily have to involve the end of the world” [13:52].
  3. Humans stay responsible. AI systems are built by companies using huge resources; “humans should be legally liable”. The story that AI will “just suddenly escape all of us” is “an abdication of human responsibility” [25:52].

How the segment is built. The hosts open with Trump’s UN line “whoever wins AI wins” and Sacks’s defence of the White House line. Then Silverman covers Sacks, Thiel, Vance, Musk, the Balaji whistleblower case, and the new Anthropic resignation. It ends with Hawking’s 2014 warning and Silverman’s reply that such warnings are older and that humans, not machines, carry the blame.

Stance to note. Silverman is not an “AI will kill us all” voice, and not an “AI is nothing” voice. He takes the risks seriously enough to want audits and liability, but treats the apocalyptic framing as politically useful to the firms. He also agrees with Jensen Huang on one narrow point: existing laws can already be used against harmful products [13:52].

Map — talking points

#PointTime
1Trump at the UN: “Whoever wins AI wins”; rejects a “globalist scheme” to control “superintelligence”[00:00]
2Hosts: AI CEOs ask the UN for regulation; Trump calls warnings a hoax; Sacks now PCAST co-chair[00:52]
3Sacks clip: hoax comparisons; “the frontier is two companies”; “they are pacing the frontier”[01:31]
4Silverman on Sacks: Thiel’s old friend; “extremely pro-capitalist”; in line with Trump’s “authoritarianism”[02:38]–[03:53]
5Why “Gilded Rage”: the tech elite is angry (woke culture, broken utopia, mild Biden-era limits)[03:53]
6Amodei and Anthropic: “skating a line”; close to Trump government, sells to the military[04:57]
7Vance: “a chameleon”, guided by Thiel; the Politico-ban story[06:11]–[08:11]
8Musk: “woke mind virus”; radicalised by the right and by time on X[08:11]
9Trump–Xi dinner; why AI CEOs side with Trump: they need state-scale resources, contracts, perhaps equity[08:47]–[11:28]
10Companies as nation-states; fusion of corporate life and the White House[11:28]
11”If your product might destroy the world, stop making it”; no other industry gets this latitude[11:28]–[13:20]
12The Oppenheimer / Manhattan Project analogy; technological determinism; present harms[13:20]–[14:43]
13Musk’s volatility; transactional like Trump; money for the 2026 midterms; AI super PACs[14:43]–[17:11]
14Musk’s childhood under apartheid; “genetic elitism”[17:11]–[19:04]
15Suchir Balaji: OpenAI whistleblower; copyright and data; weak whistleblower protection[19:04]–[22:32]
16Coxon clip: recursive self-improvement, “within a year, certainly within 2 years”[22:32]–[23:53]
17Silverman: outside audits needed; the apocalypse pitch “attracts good attention”[23:53]–[24:32]
18Hawking clip (2014); Silverman: warnings are decades old, rooted in science fiction; humans must stay liable[24:32]–[26:31]

Part 1 — Before you read

Key terms

  • Gilded Age. The US period c. 1870s–1900. The name comes from Mark Twain and Charles Dudley Warner’s 1873 novel The Gilded Age: a thin gold layer over something cheaper. It was a time of rapid industry, huge fortunes, corruption and open fights over what business may do. See 2.1.
  • Gilded Rage. Silverman’s twist: not only gold and wealth, but anger among the already-rich [03:53].
  • Superintelligence / “SI”. AI far better than humans at nearly all cognitive tasks. Trump proposes this as the official name for AI [00:00]. Note the slide: a speculative future state becomes the label for today’s products. The term has a technical history (Bostrom, 2014; see 2.5). The Defense One piece (link) explains the rename.
  • Recursive self-improvement (RSI). AI systems that improve the process that builds AI, so each generation builds a better next one. Coxon’s definition [22:32]: “automate it [AI research] with the AIs we’re building. So we could let the AIs make themselves smarter.”
  • AI safety / alignment. Alignment means making AI systems follow human goals and values (Silverman’s gloss on Balaji’s job [19:04]). Safety is wider. Watch out: “safety” means three different things in this segment: (1) existential safety (loss of control, Coxon, Hawking), (2) product safety (Sacks: “make sure their products are safe”), (3) consumer harm now (Silverman: “AI psychosis”, war use).
  • Regulation. Also slippery. It can mean (a) new AI-specific laws, (b) a UN or treaty body (Trump’s “globalist scheme”), (c) enforcement of existing laws (Huang, and Silverman at [13:52]), (d) outside audit of company claims [23:53].
  • Technological determinism. The belief that technology advances by its own logic and society can only adapt. Silverman names it as the AI leaders’ stance: “someone has to do this. It’s going to advance on its own anyway” [13:52].
  • Negative externality. A cost that a firm’s activity imposes on others who did not agree to it (pollution is the classic case). Silverman: “some definite negative externalities from AI” [13:52].
  • Special government employee (SGE). A temporary federal adviser limited to a set number of days a year. This is why Sacks “stepped down” (see caption box).
  • PCAST. The President’s Council of Advisors on Science and Technology. Sacks co-chairs it with Michael Kratsios, per the press (TechCrunch).
  • Super PAC. A political committee that can raise and spend unlimited sums to influence elections, but may not coordinate with candidates.
  • Supply-chain risk (designation). A Pentagon label for vendors whose products are a security risk. Normally used for foreign adversaries.
  • Whistleblower. An insider who reports wrongdoing. Legal protection differs by industry and law (see 2.7).
  • “Woke mind virus.” Musk’s phrase. In Silverman’s account it is a sincere belief for Musk, not only a slogan [08:11].

People named (a reading list)

PersonWhy they come upStart with
Jacob SilvermanThe guest. Journalist (ICIJ) and authorGilded Rage (2025); earlier Terms of Service (2015, from memory)
David SacksPayPal early executive, investor, White House AI and crypto czar until Mar 2026, now PCAST co-chairHis X posts; his All-In podcast; Peter Thiel and Sacks, The Diversity Myth (1995)
Peter ThielInvestor; Silverman’s link between Sacks, Vance and MuskZero to One (2014); “The Education of a Libertarian” (2009)
J.D. VanceUS Vice President; former venture capitalistHillbilly Elegy (2016)
Elon MuskThe book’s central figureIsaacson’s Elon Musk (2023); Silverman’s CBC podcast, which he mentions
Marc AndreessenVC; co-author of the “Techno-Optimist Manifesto” (2023)The manifesto (2023)
Sam Altman, Dario Amodei, Jensen HuangThe AI and chip leaders discussedTheir own essays; Amodei, “Machines of Loving Grace” (2024, from memory)
Suchir BalajiFormer OpenAI researcher; whistleblower on copyrightNYT profile (Oct 2024); Silverman’s Nation piece
Jacob CoxonResearcher who left Anthropic in Sep 2026His X thread; TechCrunch
J. Robert OppenheimerThe Manhattan Project analogyBird and Sherwin, American Prometheus (2005)
Stephen Hawking2014 warning on full AIHis BBC interview (2014); Hawking et al., Independent op-ed (2014, from memory)
Chuck GrassleySenator behind the AI Whistleblower Protection ActS.1792 text

Silverman’s own frame

Silverman is a tech and politics reporter. His book argues that a group of tech elites became radicalised: moved sharply to the right through a mix of grievance, wealth, social media echo chambers and elite networks. He ties together Musk (the main character), Thiel, Sacks, Andreessen and Vance. The title word rage does the work: his claim is that emotion, not only profit or ideology, drives their politics [03:53]. In this segment he applies the frame to AI: AI leaders want “free rein” and state-scale resources, so they ally with an authoritarian-leaning government [10:41].

Early reviews are mixed-to-positive in tone: the Foreign Affairs review concludes the book “does not suggest a happy outcome” (summary of the review; I did not read it in full). A Substack summary reports his thesis as a kind of exhaustion with the world (Keen On).

The current debate (Sep–Oct 2026)

  • Trump’s “superintelligence” speech. At the 81st UN General Assembly (22 Sep) Trump said the US “totally rejects” a “globalist scheme” to control AI, proposed calling AI “superintelligence” in government documents, and said the US will “encourage” it. Press accounts note no official transcript at first, so wording varies (Fox Business; The Next Web). The day before, more than 20 countries signed a joint statement urging more oversight of AI (same TNW report, via the NewDaily summary).
  • Sacks versus the “doomers”. Sacks has long argued that safety-focused labs, especially Anthropic, use fear to win rules that suit them. In Oct 2025 he wrote that Anthropic runs “a regulatory capture strategy based on fear-mongering”; Anthropic’s Jack Clark called this “perplexing” (Axios; Bloomberg Law). Sacks and Silverman agree on one thing and disagree on the rest: both think some safety talk is self-serving. They differ on which side is self-serving and what to do about it.
  • The government versus Anthropic. The Pentagon labelled Anthropic a supply-chain risk in March 2026 after Amodei refused to drop limits on mass-surveillance and autonomous-weapons uses. Critics, including former CIA director Michael Hayden, called this a misuse of a tool meant for foreign adversaries (NPR via VPM). This complicates Silverman’s “Amodei is close to Trump” line: the relationship is both close and conflicted.
  • Money in the midterms. Two AI-aligned networks are spending in the 2026 midterms. The pro-industry Leading the Future is funded by, among others, Greg Brockman, Marc Andreessen and Ben Horowitz; the safety-leaning Public First has at least $20M from Anthropic. A New York Times tally put combined spending at $55.7M so far, mostly on safe seats; other tallies differ by date and scope (AI Weekly summary). This is the “AI super PACs” point at [16:24].
  • Press freedom story inside the story. The White House barred Politico, CNN and MS NOW from the press pool; a judge ordered access restored; Status reported that Vance was the anonymous source of a Politico story the White House cited. The White House has not confirmed it (Political Wire; Daily Beast). Silverman uses it as an example of Vance’s dual role [06:56].
  • Whistleblower law. The bipartisan AI Whistleblower Protection Act (S.1792, Grassley, May 2025) would bar retaliation for reporting AI security flaws or legal violations. As of Sept 2026, Grassley wants it passed before the midterms; I found no confirmation of a vote (Deseret News; status not confirmed).
  • The Gilded Age comparison is now mainstream. For example, a January 2026 Boston Globe op-ed by Jane Hoffman (Harvard Kennedy School) compares the seven largest tech firms to the old tycoons and calls for antitrust action (Boston Globe). A counter-view says the original robber barons faced no income tax, strong unions or FTC, so the analogy has limits (Bowen).

Part 2 — The intellectual map

The talk is 27 minutes long and mostly news. Behind it sit eight bodies of thought. Read this part once, slowly. The first two (the Gilded Age and antitrust) are the ones the title points at.

 THE GILDED AGE & ITS REFORMS ──┐                    ┌── AI-RISK THOUGHT
 Twain · Carnegie · Sherman Act │                    │   Good · Bostrom · Hawking
 Populists · Progressives       │                    │   Yudkowsky · Hinton · RSI
                                ▼                    ▼
 ANTITRUST & CONCENTRATION ─► ┌──────────────────────────────┐ ◄── MANHATTAN-PROJECT
 Brandeis · Chicago school ·  │  SILVERMAN: AI leaders are   │     ANALOGY & ARMS CONTROL
 neo-Brandeisians             │  angry, state-hungry, and    │     Oppenheimer · Baruch Plan
                              │  use doom as a pitch; humans │
 TECH-RIGHT IDEOLOGY ───────► │  must stay liable            │ ◄── DETERMINISM & INEVITABILITY
 Thiel · Andreessen · Yarvin  └──────────────────────────────┘     Heilbroner · Winner · Ellul
 Californian Ideology                ▲            ▲
                                     │            │
        POLITICAL ECONOMY OF AI      │            │   LIABILITY, WHISTLEBLOWERS,
        compute · energy · state     │            └── "criti-hype" & SOURCE CRITICISM
        Strange · Mazzucato ·        │                 Vinsel · Narayanan & Kapoor
        Eisenhower · Zuboff ·        │                 Haugen · Pigou · Coase
        Varoufakis                   │

Method note. Dates and works in this Part come from my background knowledge unless a link is given. Items marked “from memory” were not checked in this session. Where a debate is contested, I say so.

2.1 The historical Gilded Age — wealth, politics, and the fight over limits

The problem it addresses. Silverman’s title borrows a period. To judge whether the borrowing is fair, you need the period’s facts. The question: when private fortunes grow fast, what stops them from owning the political system, and what finally did?

The period (c. 1870–1900)

  1. The name. Mark Twain and Charles Dudley Warner, The Gilded Age: A Tale of Today (1873). A satire of speculation and political corruption after the Civil War. “Gilded” means gold on top, not gold through and through. Critics at the time and historians since use it to mean: glitter over rot.
  2. The economy. Railroads (the first national corporations), steel, oil, finance, telegraph. Capital needs were huge, so firms grew huge. Carnegie (steel), Rockefeller (Standard Oil, which used the trust form from 1882 to combine rivals), J.P. Morgan (finance), Vanderbilt and Gould (railroads).
  3. “Robber barons” or “captains of industry”? Two labels, two moral stories. Robber baron (Josephson, The Robber Barons, 1934) stresses bribery, stock-watering, price-fixing and ruined rivals. Captain of industry stresses cheaper goods, scale and jobs. Business historians such as Alfred Chandler (The Visible Hand, 1977) argued that the big firms brought real efficiency: modern management, lower rail and oil prices. The best modern synthesis, Richard White, The Republic for Which It Stands (2017), keeps both: huge productive growth and corrupt, unequal politics.
  4. Politics for sale. Party machines, “spoils” jobs, and railroad money in legislatures. After President Garfield’s assassination (1881) by a disappointed office-seeker, the Pendleton Act (1883) began civil-service jobs by merit. In the 1896 election, Mark Hanna raised money from business on a scale that shocked contemporaries (from memory).
  5. The ideology of the age. Carnegie’s “The Gospel of Wealth” (1889): the rich have a duty to give away their fortunes wisely, and they know better than the state how to spend. Social Darwinism (Herbert Spencer, W.G. Sumner): the market selects the fit; interference is harmful. This is the ancestor of the “someone has to do this; it will happen anyway” tone Silverman describes [13:52].
  6. The backlash.
    • Farmers and labour. The Grange; the Populist (People’s) Party, whose 1892 Omaha platform attacked railroads and banks; the Haymarket affair (1886); the Homestead strike (1892) and Pullman strike (1894), both ended with force.
    • Law. Munn v. Illinois (1877) held that businesses “affected with a public interest” can be regulated. The Interstate Commerce Act (1887) created the first federal regulator. The Sherman Antitrust Act (1890) banned “restraint of trade” and monopolisation.
  7. Then the Progressive Era (1900–1920). Theodore Roosevelt’s trust-busting; the Supreme Court broke up Standard Oil in 1911; the Clayton Act and the Federal Trade Commission (1914); the Tillman Act (1907) banned corporate gifts to federal campaigns; the income tax (16th Amendment, 1913). It took about 30 years of crisis and organising to answer the Gilded Age.

The lesson usually drawn. Concentrated wealth is not stopped by good intentions. It is stopped by laws, courts, organised citizens, and sometimes by a scandal that gets the public angry. The reform came after the harms were visible and felt.

How a “second Gilded Age” is argued. Economists Thomas Piketty and Emmanuel Saez show top-income shares in the US returning to levels last seen before the New Deal. Political scientists Gilens and Page (2014) found that, in their data, the preferences of the affluent predicted policy outcomes far better than those of the average citizen (a result that has been disputed on method). Hacker and Pierson, Winner-Take-All Politics (2010), argue that organised business shaped US policy after the 1970s.

Fair differences between then and now (useful for judging Silverman)

  • Then: no income tax, no FTC, weak unions, no mass regulatory state, no radio or TV. Now: all of those exist, plus a large national-security state.
  • Then: firms were the means of production for physical goods. Now: AI firms rely on cloud, chips and electricity, which makes them more dependent on the state and on a few suppliers (Nvidia, TSMC).
  • Then: fortunes came from railroad rights-of-way and monopolies. Now: from network effects and capital-heavy compute.
  • Rage: Gilded Age tycoons were often confident and pious (Carnegie, Rockefeller). Silverman’s claim is that today’s tycoons are angry. That is a claim about psychology and it is the book’s novelty (see Part 3, Deep dive 1).

2.2 Antitrust and the politics of bigness

The problem it addresses. Is a very large firm a problem in itself, or only if it raises prices? The answer has swung three times in US history, and it decides what tools exist for “AI is controlled by a few firms” (Sacks’s own “duopoly” remark at [01:31]).

  1. The Brandeisian view (1900s–1930s). Louis Brandeis, Other People’s Money (1914), and his later phrase “the curse of bigness” (1930s case opinion, from memory). Concentrated economic power becomes political power and threatens self-rule. Competition law should protect democracy and small producers, not only prices.
  2. The Chicago turn (1960s–80s). Robert Bork, The Antitrust Paradox (1978), and Richard Posner argued the law should ask one question: does it harm consumers (higher prices, lower output)? Big can be good (efficiency). Courts adopted the consumer-welfare standard. Mergers faced less resistance for forty years.
  3. The neo-Brandeisian revival (2010s–20s). Lina Khan, “Amazon’s Antitrust Paradox” (Yale Law Journal, 2017), argued that low prices can hide platform power. Tim Wu, The Curse of Bigness (2018), links the Gilded Age to today. Khan chaired the FTC 2021–25. Critics said the approach was vague and risked harming consumers.
  4. Tech cases. The Google search case (liability ruling Aug 2024, from memory) is the Hoffman op-ed’s “successful case” (Boston Globe). Remedies and appeals continue; status not checked.
  5. AI-specific worries. Markets for AI have bottlenecks: chips, cloud, data, talent, and power. EU competition head Margrethe Vestager warned in 2024 that AI money was flowing to firms already atop “a nest of existing ecosystems” (via Advisor Perspectives).
  6. Sacks’s “duopoly” remark as data. At [01:31] Sacks says “the frontier is two companies. It’s Anthropic and OpenAI. They have duopoly market share in this country.” He uses this to argue those firms can be trusted to “pace the frontier”. An antitrust scholar would read the same fact the other way: duopoly means no competitive discipline, so rules are more necessary. Market-share figures vary by measure: Menlo Ventures’ mid-2025 enterprise survey had Anthropic 32%, OpenAI 25%, Google 20% (BigDATAwire); consumer web traffic shows a different order (other trackers; low authority). Google, Meta, xAI and Chinese open models also compete, which weakens “two companies” as a complete picture.

Link to the talk. Silverman does not argue antitrust in this segment. But the Gilded Age frame and “nation-state” language invite it. See Deep dive 3 for what he leaves out.

2.3 The tech-right network: ideas and money

The problem it addresses. Silverman says Sacks and Thiel share “pretty right-wing, extremely pro-capitalist” views [03:14]. What are those views, and where do they come from?

  1. Californian Ideology (1995). Richard Barbrook and Andy Cameron’s essay described a mix of 1960s freedom culture and free-market belief in Silicon Valley, with an unspoken blind spot about race and the state. Fred Turner, From Counterculture to Cyberculture (2006), tells how Stewart Brand and others fused them.
  2. The Stanford campus right. Sacks and Thiel wrote The Diversity Myth (1995), a critique of multiculturalism at Stanford (Silverman’s “anti-woke” precursor at [03:14]). Thiel co-founded PayPal (1998); Sacks was its COO. The “PayPal Mafia” later funded many firms, including early Facebook (Thiel), SpaceX (Musk, Founders Fund) and Palantir.
  3. Thiel’s political philosophy. In “The Education of a Libertarian” (2009) he wrote that he no longer believed “freedom and democracy are compatible”. In Zero to One (2014) he argued that monopoly is good (“competition is for losers”): a monopoly funds long-term innovation. This is the opposite of the Brandeisian view in 2.2.
  4. Andreessen’s “Techno-Optimist Manifesto” (2023) and “Why AI Will Save the World” (2023). Technology is the engine of human progress; “regulation” and “safetyism” are enemies; slowing AI “is a form of murder” because it delays cures (paraphrase of his argument, from memory). e/acc (effective accelerationism) is the online movement with a similar view.
  5. The neoreactionary current. Curtis Yarvin (writing as Mencius Moldbug) argued that democracy should be replaced by a corporate-style monarch (a “CEO-king”). He has been read by Thiel’s circle and by Vance (Vance has referred to Yarvin; from memory). Balaji Srinivasan’s “network state” is a related exit-from-the-state idea. These are far-right fringe ideas that Silverman links to the mainstream tech-right; whether they define the whole network is disputed.
  6. From ideas to power. Thiel gave money to Vance’s 2022 Senate race (from memory: about $15M). Musk’s America PAC and other donations in 2024 backed Trump; Silverman’s claim at [16:24] is that AI and crypto PACs repeat this in 2026. Note the counter-example that cuts against “the tech industry is right-wing”: much of Silicon Valley still gives to Democrats, and Anthropic-linked groups now fund a separate PAC (see the debate section).
  7. Silverman’s “rage” thesis. Where the above are ideas, his claim is about feelings: grievance at woke culture, at a broken tech-utopia promise, and at Biden-era limits [03:53]. This links to a research tradition on status anxiety among elites, for example Peter Turchin’s “elite overproduction” (from memory) and Richard Hofstadter’s The Paranoid Style in American Politics (1964), which found rage-driven politics at the top as well as the bottom. Whether rage is the right frame, or interest (they want lower taxes and weaker rules) is the simpler account, is the main open question about the book.

2.4 The political economy of AI — firms that need states

The problem it addresses. Silverman’s central structural claim [10:41]: AI needs “vast amount of resources” and “nation-state level capabilities”, so firms must align with government. Is that right, and is it new?

  1. Susan Strange, The Retreat of the State (1996). Firms and markets take over functions once held by states. A classical statement of the “firms as quasi-states” idea. Ian Bremmer (2021) called big tech a “technopolar” power: some firms hold powers of sovereign scale (data, networks, code).
  2. The entrepreneurial state. Mariana Mazzucato, The Entrepreneurial State (2013): many core technologies (the internet, GPS, touchscreens) came from public money. So “private” tech has always leaned on states. AI now repeats this with chips and energy.
  3. Eisenhower’s farewell address (1961). He warned of the military-industrial complex, and of “a scientific-technological elite” capturing public policy. Joe Biden’s own farewell (Jan 2025) warned of a “tech-industrial complex” (from memory). Silverman’s picture, tech firms selling AI for “military national security uses” [04:57], fits this old pattern.
  4. Compute, chips, power as the new “railroad”. Training a frontier model needs data centres, GPUs, and gigawatts. Only a few firms can pay. The state decides export controls (chips to China), grid connections and permits, and is itself a buyer. In 2025 the government took an equity stake in Intel (from memory: August 2025), which gives a precedent for Silverman’s remark that the state “might even take an equity role in OpenAI” [10:41]. His source for that is Altman’s own talk (he says “something that Sam Altman has talked about”); I did not verify it.
  5. Platform capitalism and “technofeudalism”. Nick Srnicek, Platform Capitalism (2016); Shoshana Zuboff, The Age of Surveillance Capitalism (2019); Yanis Varoufakis, Technofeudalism (2023): firms act as landlords over digital space and rent out access, more than as competitors in markets. These books treat the power as economic and structural, not as personal anger.
  6. The political side: “competitive authoritarianism”. Levitsky and Way (2010) describe regimes where elections exist but the government abuses state power to tilt them. Levitsky and Ziblatt, How Democracies Die (2018), list warning signs. Silverman says the AI CEOs find the “authoritarian mindset” natural [10:05]. Many political scientists would describe 2025–26 US events in these terms; others reject it. It is a contested label, and Silverman does not define it. See Deep dive 3.
  7. The “AI race with China” frame. Trump’s “we’re leading now over China by a lot” [00:00] is the geopolitical justification for light regulation. It parallels Cold War arguments. Counter-argument: Chinese policy is also strict about AI content. The Xi–Trump dinner [08:47] shows both sides are courting the same CEOs.

2.5 AI-risk thought — from Butler to Coxon

The problem it addresses. Hawking, Coxon and the AI CEOs [24:32], [22:32] all speak inside a debate with a 160-year history. Silverman’s claim that the warnings are “decades old” and rooted in science fiction [25:52] is partly right.

  1. Samuel Butler, “Darwin among the Machines” (1863). Machines might evolve and take over. A very early literary form.
  2. Alan Turing (1951 lecture; from memory). Once machine thinking starts, “we should have to expect the machines to take control”.
  3. I.J. Good, “Speculations Concerning the First Ultraintelligent Machine” (1965). The intelligence explosion: a machine that designs better machines leaves human intellect “far behind”. This is the direct ancestor of recursive self-improvement (Coxon’s term at [22:32]).
  4. Science fiction. Asimov’s Laws of Robotics (1942 on); Čapek’s R.U.R. (1920); Clarke’s HAL (1968). Silverman’s point about Musk and others “grew up reading” such stories [25:52] fits a known pattern: Musk has said he named SpaceX’s drone ships after Iain M. Banks’s Culture novels (from memory).
  5. Vernor Vinge, “The Coming Technological Singularity” (1993). Popularised “singularity”. Ray Kurzweil, The Singularity Is Near (2005), made it optimistic.
  6. The modern risk literature. Nick Bostrom, Superintelligence (2014) gave the “control problem”. Eliezer Yudkowsky and the MIRI school: alignment is hard and we may get one try. Stuart Russell, Human Compatible (2019): build AI that is uncertain about human goals. The Center for AI Safety statement (May 2023): “mitigating the risk of extinction from AI should be a global priority”, signed by Hinton, Bengio and the CEOs of OpenAI, Anthropic and Google DeepMind. Hawking’s 2014 quote [24:32] is an early public instance; the quote at [25:14] says full AI “would take off on its own and redesign itself”, i.e. RSI.
  7. RSI specifically. Schmidhuber’s “Gödel machine” (2003) formalised a self-rewriting system. The “AI 2027” scenario report (Kokotajlo et al., 2025, from memory) popularised a timeline where automated AI research speeds up progress. Coxon’s claim at [23:09] is that human “taste or judgment” is the last bottleneck and that automation of it may come “within a year, certainly within 2 years”. Critics doubt both the timeline and the bottleneck. Anthropic says it is not there yet and that the outcome is not inevitable (press summary; I did not retrieve Anthropic’s own statement).
  8. The sceptics.
    • “AI as Normal Technology” (Arvind Narayanan and Sayash Kapoor, 2025): AI will diffuse slowly like electricity; the main risks are ordinary (misuse, concentration), and policy should focus there. Their view overlaps with Silverman’s emphasis on present harms.
    • Emily Bender, Timnit Gebru and others: the “stochastic parrots” line (2021) and criticism of x-risk talk for distracting from present harms (labour, bias, surveillance). Émile Torres and Gebru coined TESCREAL (2023) for the transhumanist-rationalist cluster behind many AI-doom and accelerationist ideas.
    • Lee Vinsel’s “criti-hype” (2021): critics who accept the industry’s most dramatic claims in order to criticise them end up marketing the product. This is close to Silverman’s remark that the apocalyptic promise “has practically become part of the sales pitch” [12:27]. But note the difference: Silverman does not say the risk is fake; he says the framing is used.
    • Yann LeCun and others: the risk of near-term loss of control is exaggerated.
  9. The “hoax” frame in the talk. The hosts say Trump “has called warnings about AI risk a hoax, likening it to global warming” [00:52], and Sacks’s clip [01:31] compares AI fear to “hoaxes” and the “climate change” narrative. That is a different move from sceptics like Narayanan: it treats the risk talk as a political plot, not as a forecast error.

2.6 The Manhattan Project analogy — what it gets right and wrong

The problem it addresses. Silverman says AI leaders “see themselves as the inheritors of the Manhattan Project” [13:20]. What did the scientists actually say and do, and how do the two cases compare?

  1. The atmosphere-ignition worry (1942–45). Edward Teller raised the possibility that a fission explosion could ignite the atmosphere. Hans Bethe and others calculated that it could not. Arthur Compton later recalled a safety threshold of “three in a million” (from memory: a 1959 interview). At the Trinity test (16 July 1945) Enrico Fermi reportedly took bets on whether the atmosphere would ignite (from memory; widely reported). So Silverman’s “joke or even bet” [13:20] has a basis. But the scientists did the calculation first, and published later. That is the strongest point for the analogy and also against “blithe”: they calculated.
  2. Public ownership. The Manhattan Project was a government programme under General Groves, with compartmentalised secrecy. AI today is led by private firms. So the analogy fails on control: nobody in government can order a pause.
  3. Arms control. The Baruch Plan (1946) proposed international control of atomic energy; it failed. The IAEA (1957), the Partial Test Ban Treaty (1963) and the NPT (1968) came later. Trump’s “totally rejects” a “globalist scheme” [00:00] is a rejection of this model for AI. More than 20 countries signed a pro-oversight statement the day before (see debate section).
  4. AI leaders use the analogy themselves. Leopold Aschenbrenner’s Situational Awareness (2024, from memory) called for a “Manhattan Project” for AI; “Genesis Mission” (Nov 2025, from memory) is a US government AI science programme named in the same spirit. Silverman’s point is that both the self-image (we hold world-ending power) and the political use (so trust us / support us) come from the same analogy.
  5. Oppenheimer’s afterlife. He lost his security clearance in 1954 after criticising the hydrogen bomb. Silverman does not discuss this, but it is the other half of the story: the scientist who warns can be sidelined by the state. The Balaji and Coxon stories echo it (Deep dive 6).

2.7 Liability, regulation and whistleblowers

The problem it addresses. Silverman’s practical programme [12:27], [13:52], [23:53], [25:52]: use existing law, create outside audits, protect whistleblowers, keep humans liable. What is the legal background?

  1. Product liability. MacPherson v. Buick (1916) made manufacturers liable to end users, not only buyers. Greenman v. Yuba Power (1963) introduced strict liability for defective products. Silverman’s “if a chemicals company’s product started harming people, hopefully there would be a legal or regulatory response” [12:27] is this tradition. Open question: is a model’s output a “product” or “speech”? Courts are now testing this in chatbot-harm lawsuits (from memory; I did not check case outcomes).
  2. Externalities. A.C. Pigou (1920): tax or regulate the costs imposed on others. Ronald Coase (1960): bargaining can solve them if rights are clear and costs low. AI harms (to people with chatbot-linked mental-health crises, to creators whose work is scraped, to the energy grid) often have unclear rights and high bargaining costs, which favours regulation.
  3. Section 230 (1996). Platforms are not liable for user content. Whether it covers AI-generated text is unsettled (from memory).
  4. Current AI rules.
    • EU AI Act (2024) with risk tiers.
    • California SB 1047 was vetoed in Sept 2024; SB 53 was signed in Sept 2025 and Anthropic endorsed it (the endorsement is in the Axios story above). Sacks blamed Anthropic for a “state regulatory frenzy”.
    • Federal preemption. A 10-year moratorium on state AI laws failed in 2025; later federal action to preempt state laws was attempted by executive order (from memory; December 2025; not checked).
    • Huang’s view (as Silverman reports it): existing consumer law is enough [13:52]. Many lawyers agree for some harms (fraud, defective products) and disagree for frontier risks, which existing laws do not address.
  5. Whistleblower law. Sarbanes–Oxley (2002) and Dodd–Frank (2010) protect corporate-fraud reporters; the False Claims Act (strengthened 1986 under Grassley) rewards reporting fraud on the government. A whistleblower reporting safety risks (not fraud) often falls in a gap. The AI Whistleblower Protection Act (S.1792) is designed to fill it (govinfo text; Princeton CITP explainer). Frances Haugen (2021) is the best-known tech precedent.
  6. Copyright and training data (Balaji’s issue). NYT v. OpenAI (filed Dec 2023). In 2025 two US district rulings (Alsup in Bartz v. Anthropic; Chhabria in Kadrey v. Meta) found training on books can be fair use under certain facts; Anthropic later settled a pirated-books class action for about $1.5B (from memory, Sep 2025; not checked). So the law is mixed: Balaji’s claim that mass copying is illegal is not settled, and some rulings cut against it.
  7. Audit. Silverman wants “technically competent” outside people to “audit these claims” [23:53]. Models: financial audit, aviation incident investigation (NTSB), pharmaceutical trials (FDA). Current practice: voluntary pre-deployment testing by government institutes (US CAISI, UK AISI) and third-party evaluators, with limited access. Details not checked.

2.8 Determinism, “inevitability” and reading the source

The problem it addresses. Silverman says AI leaders use “technological determinism” [13:52]. And the talk is itself a media product. Two short threads.

  1. Does technology drive history? Robert Heilbroner, “Do Machines Make History?” (1967): to a degree, yes. Langdon Winner, “Do Artifacts Have Politics?” (1980): designs embed choices. Jacques Ellul, The Technological Society (1954): technique follows its own logic. Social construction (Bijker, Pinch) answers: choices by groups shape outcomes. The practical point: inevitability talk (“if we don’t, China will”) closes the policy debate, so it is a political move even when it is also partly true. Game theorists describe the risk as a race to the bottom on safety (Armstrong, Bostrom and Shulman, 2016, “Racing to the Precipice”; from memory).
  2. Silverman’s counter-move: legal responsibility. If humans build and deploy AI, then “humans should be legally liable” [25:52]. This echoes Joanna Bryson’s argument (2010) that robots should be property, not persons, so responsibility stays with owners (from memory).
  3. Reading the source. Democracy Now! (founded 1996 on Pacifica Radio, hosted by Amy Goodman) presents itself as independent, audience-funded news. It is openly left-of-centre. Silverman writes for outlets such as The Nation and is an investigative reporter. To read well: separate checkable facts (dates, quotes, designations), interpretations (rage, nation-states) and motive claims (what Musk “feels”). The segment mixes all three.
  4. Hosts’ framing. The opening sets a contrast: CEOs at the UN ask for regulation while Trump rejects it [00:52]. That contrast hides a third position (Sacks’s) and the fact that “AI CEOs ask for regulation” often means specific rules they prefer. See Deep dive 4.

Part 3 — Deep dives

1. “Gilded Rage”: why anger, and why the Gilded Age?

🎙 In the talk.

  • Silverman at [03:53]: the title is “a little bit of a throwback to the Gilded Age”, but the feature he stresses is “a real kind of anger”. Musk, Sacks, Andreessen and “a lot of the billionaire tech elites” are angry on four counts:
    1. “the excesses of woke culture and liberal and lefty politics”;
    2. that “the technological utopian world that they believe that they were promised” has not come (“we were promised a better future” is, he says, language they have used);
    3. that “the Biden administration tried to mildly rein them in … in certain areas like crypto”;
    4. “that they don’t have free rein to enact the world they want to enact. And that includes areas like AI.”
  • At [08:11] he adds a mechanism: they were “radicalized in some very visible ways by right wing forces and also by their time spent together on social media on Twitter and then X”.
  • His one-line summary of Sacks and Thiel: “extremely pro-capitalist, and they are very much in line with the authoritarianism embodied by Trump” [03:14].

📚 Deeper theory (links to 2.1, 2.3)

  • What the Gilded Age comparison does, and does not, show. It shows a structure: huge private fortunes, close ties to government, and public demand for limits. It does not by itself show anger. Carnegie and Rockefeller styled themselves as benefactors. The anger Silverman describes is closer to “ressentiment” (Nietzsche; Max Scheler) among the winners, or to what political scientists call elite grievance politics.
  • Two rival explanations for the same behaviour.
    ExplanationSaysFits which evidence
    InterestThey want lower taxes, weaker rules, big contracts. Rhetoric is coverContracts, equity talk, PAC money [10:41], [16:24]
    IdeologyThey hold a worldview (Thiel, Andreessen manifesto)Writings (2.3)
    Emotion / grievance (Silverman)Hurt, status loss and online echo chambers radicalise themMusk’s tweets and reconciliations [15:15]; social media mechanism [08:11]
    The three are not exclusive. Silverman’s novelty is the third. It is the hardest to test.
  • Social media as radicaliser. Research on echo chambers and group polarisation (Sunstein) supports the general mechanism for ordinary users. Applying it to billionaires is plausible but is a claim about individuals, so evidence is thin: tweets, interviews, reported behaviour.
  • The “promised future” grievance connects to a real cultural idea: the Thiel line “we wanted flying cars, instead we got 140 characters” (about 2011; from memory). It is a complaint that tech progress slowed. This is a respectable thesis in economics too (Robert Gordon, The Rise and Fall of American Growth, 2016; Tyler Cowen’s “Great Stagnation”). So the grievance can be real without the politics following.

Where others push back.

  • From the right and from tech supporters: “Rage” is a pathologising label. People can hold right-leaning or pro-growth views for reasons. Sacks’s clip frames the media as the one who blow[s] it out of proportion [01:31]. Many on the right also argue that the left-wing tech elite of 2010–2020 (content moderation, DEI policy in firms) was the first to use corporate power politically.
  • From economists: wealth concentration today is partly earned by products people choose to buy, which makes the robber-baron label unfair, as Chandler’s defenders said about the 1890s.
  • From the left: The rage frame may let institutions off. If the problem is structural (compute, finance, weak antitrust) then changing who is angry changes nothing (see 2.2, 2.4).
  • Method: The argument rests on reported character: reports that he’s a drug user [15:15]. A reader should treat these as claims, not findings.

What they leave out.

  • A comparison with other periods of elite grievance, or with tech billionaires on the left or centre. The segment does not say how many of the tech elite fit the “angry” profile.
  • Concrete policy in the Gilded Age sense: what reform would answer the 2026 version? No mention of antitrust, taxes, or campaign finance rules, the tools that worked after 1890.
  • Evidence that the anger causes the policy outcomes, as against interest.

Check yourself

  1. Name two things the historical Gilded Age had that Silverman’s “rage” does not match. (Hint: tycoon self-image.)
  2. Give the interest, ideology and emotion explanation for why an AI CEO attends a Trump dinner. Which needs the most evidence?
  3. Which laws and institutions ended the original Gilded Age, and which exist today?

2. The network: Sacks, Thiel, Vance, and the “hoax” line

🎙 In the talk.

  • Trump at the UN [00:00]: “Whoever wins AI, you have to remember this, and now I say, ‘Whoever wins SI, whoever wins superintelligence wins.’” and “we will only encourage superintelligence. We’re going to encourage it, not rein it in.”
  • The hosts [00:52]: AI CEOs “called for some kind of international regulation”, while Trump “has called warnings about AI risk a hoax, likening it to global warming”.
  • Sacks at Politico’s summit [01:31]: “how many hoaxes have been run on the president himself going back to the whole Russia gate hoax?”; on climate, “there’ll be elements of truth, there’ll be pseudo science to it, and then the media will blow it out of proportion”; and on companies, “They have a basic corporate responsibility to make sure their products are safe … the frontier is two companies … they are pacing the frontier. We’re doing that.”
  • Silverman on Sacks [03:14]: “Peter Thiel’s long-time friends, going back to their Stanford days”; “writing partners together in the 1990s”; “he’s someone who’s going to be around throughout the Trump administration”.
  • On Amodei [04:57]: he “tries to skate a certain line”; he is “still someone … trying to have a pretty close relationship with the Trump administration, is trying to sell his products for military national security uses.”
  • On Vance [06:56]: “a bit of a chameleon”; “Peter Thiel has really guided J.D. Vance’s life, including giving him most of his opportunities”; Silicon Valley people “saw him as one of them”; “he can move between these worlds including … being an anonymous source, but then working the other side to get Politico banned”.

📚 Deeper theory (links to 2.3, 2.5)

  • Reading Sacks’s clip carefully. He makes four separate moves:
    1. Historical analogy to media hoaxes: a rhetorical frame, not evidence about AI.
    2. Concession: “grains of truth”.
    3. Market-structure claim: the frontier is a “duopoly”.
    4. Reassurance: those two firms say they will go slowly, “so … we’re doing that”. Move 3 and 4 are the substance: policy is to trust the two firms. An antitrust reading (2.2) would call this reliance on self-regulation by a concentrated industry. A pro-market reading says firms have the strongest incentive to avoid harms, because their reputations and liability are at stake.
  • The Thiel–Vance link. Thiel’s patronage of Vance (2.3) is documented in the press, including his donations to Vance’s 2022 Senate race (from memory). Silverman’s “guided his life” is a stronger narrative claim.
  • Anthropic as the awkward case. Silverman describes Amodei as left-of-centre in posture, but courting the government for military sales [04:57]. Facts since March 2026 add friction: a Pentagon “supply-chain risk” label after Anthropic’s limits on surveillance and autonomous weapons (NPR via VPM). Sacks says the frontier firms “are pacing the frontier” [01:31], which sits oddly with the same government’s fight with one of them.
  • A network is not a conspiracy. Sociologists use interlocking directorates and elite networks (C. Wright Mills, The Power Elite, 1956; G. William Domhoff, Who Rules America?) to describe overlapping social ties among decision-makers. The claim is about access and shared outlook, not secret coordination. Silverman’s language (“networks of influence helped reelect Trump” in the hosts’ reading of his book at [06:11]) is about this kind of influence.

Where others push back.

  • Sacks’s side: The government must not let a small number of firms or foreign-influenced bodies set the rules; “doomerism” is funded by effective-altruism money and works as regulatory capture. Under this view, Anthropic’s support for state laws (SB 53) is the capture, not Sacks’s position (see the Axios story).
  • Pluralists: Business lobbying exists in all administrations. Obama-era tech had close ties too (Google staff in the White House; from memory). The question is whether the 2025–26 degree is different in kind.
  • Silverman’s critics on the left might say his focus on individuals (Sacks, Thiel) understates how Democrats also relied on tech money and policy.
  • On Vance: the Politico story rests on one report (Status), unconfirmed by the White House.

What they leave out.

  • The policy itself. The segment never states what Sacks’s AI policy is (for example, the July 2025 AI Action Plan, from memory) or tests it.
  • The Trump–CEO relationship is two-way. Companies also lobby for the “light touch”. The segment frames them as supplicants.
  • Voices from the White House or the firms. Only Sacks’s clip.

Check yourself

  1. In Sacks’s clip, which statements are analogy, which are fact claims, and which are policy?
  2. How can the same market fact (“a duopoly”) support opposite policies?
  3. What would you need to know to decide whether this is “a network with shared views” or “influence beyond normal lobbying”?

3. Firms as nation-states, and the fusion with government

🎙 In the talk.

  • [10:05]: the CEOs “find common cause and common interests”; and “the authoritarian mindset I think comes very easily to the AI CEOs because they are thinking about changing the world … which is often just communicated in the abstract. We’re going to create a utopia or apocalypse.”
  • [10:41]: “in order to do that, they need vast amount of resources. And they need sort of nation-state level capabilities. They need unbelievable amounts of money, land, energy, the chips…“. Altman is “very opportunistic” and “the US government can provide contracts. It can provide funding and investment. It might even take an equity role in OpenAI”. “These aren’t the tech companies of 10 years ago. These are ones that are much more like their own little nation-states and need that kind of relationship with a authoritarian power like the Trump regime.”
  • [11:28], the host: tech billionaires are “addressing world leaders and may well have more power than any head of country.” Silverman: “this fusion of corporate life and the White House including just Trump’s own business interests as a stock trader and with his World Liberty Financial and other crypto interests”; “these CEOs who don’t seem to be great friends of democracy”.
  • Silverman at [08:47]-[11:28] sets this against the Trump–Xi dinner list read by the host.

📚 Deeper theory (links to 2.4)

  • The resource argument is the strongest part. It is a structural claim that does not need anyone to be angry. Training and running frontier models needs capital, chips and power on a scale only a few firms and states have. That makes firm–state bargaining unavoidable, whoever governs. The same pattern (railroads needing land grants and rights-of-way) is in the Gilded Age, and Mazzucato’s point holds both ways: the state funds, firms scale.
  • “Nation-state” is a metaphor with limits. States have monopoly on legitimate violence, territory, courts and taxation (Weber). Firms have none of these. But they have jurisdictional influence (terms of service, content rules, standards), infrastructure power (cloud, chips), and access to heads of state. A closer term from scholars is “quasi-sovereign” or “platform governance”.
  • The two directions of dependence. Firms need states for chips (export licences), power, and contracts. States need firms for capability in intelligence, defence and economic competition. This is mutual dependence, and it is why bargaining, not command, describes the relation. That sits uneasily with “authoritarian power” as a single label.
  • Conflict-of-interest facts are separate. Trump’s crypto interests and stock trades are asserted by Silverman here; reporting exists but I did not check it for this page.
  • Authoritarian-friendly vs democratic-friendly CEOs. Silverman says the AI CEOs “don’t seem to be great friends of democracy” given they speak “in such general terms” on the UN stage [11:28]. That is a judgment on rhetoric. A more testable version is: do they lobby against transparency, audit, or antitrust rules?

Where others push back.

  • Realists: In a US–China contest, aligning with one’s own government is rational and patriotic; every defence industry has done so (Boeing, Lockheed). That is not “authoritarian”; it is the national-security state.
  • Libertarians and many CEOs: They would not say they want state power; many say they want less regulation. Silverman’s picture of firms seeking equity or contracts partly conflicts with that.
  • Democracy-sceptic scholars on the other side: Others argue the opposite danger is more likely: governments forcing firms to serve state aims (the Pentagon–Anthropic fight in March 2026 shows state pressure on a firm).
  • Evidence from the dinner: A guest list shows invitation, not alignment. Press coverage described the Xi summit as symbolic with limited concrete results (LBC).

What they leave out.

  • Non-US cases. India, the EU and China all have their own firm–state patterns (the Indian digital public infrastructure model, where the state builds shared rails and firms build on them, is a different arrangement from the US one).
  • Open-source and smaller labs, which weaken the “only a few can do this” premise.
  • What would be a good firm–state relation: audits, procurement rules, independent regulators.

Check yourself

  1. Which parts of “firms as nation-states” fit the facts and which are metaphor?
  2. Why does Silverman’s resource argument not need the “rage” thesis?
  3. Give one example that shows the state pressing a firm, not the reverse.

4. Doom as sales pitch, the Oppenheimer analogy, and “someone has to do this”

🎙 In the talk.

  • The hosts quote Silverman’s writing [12:27]: “The catastrophic promise of AI has practically become part of the sales pitch … industry luminaries can sound like Robert J. Oppenheimer and the Manhattan Project scientists who wondered if a nuclear bomb test would ignite the Earth’s atmosphere.”
  • Silverman [11:28]: “if you think your product might destroy the world well maybe you should stop making that product.” [12:27]: other industries would face “a legal or regulatory response”; the AI industry says “this stuff is going to get away from us and there’s nothing we can do. And I don’t really buy that narrative.”
  • [13:20]: “they see themselves as the inheritors of the Manhattan Project”. [13:52]: “some of them have even talked about how AI is a bigger risk than nuclear proliferation”.
  • [13:52]: “Jensen Huang’s right, the laws on the books could be used to crack down on companies that harm consumers”; “technological determinism … someone has to do this. It’s going to advance on its own anyway, and you should just believe our promises that this sort of apocalyptic potential is matched by an even more utopian potential.”
  • [13:52]: “the bad effects of AI are something we’re seeing right now, and they don’t necessarily have to involve the end of the world.” He names “AI psychosis” and “being used in war zones”.
  • On Coxon [23:53]: “researchers quitting AI companies, sometimes going to another one right afterwards”; “we need … technically competent people who can kind of audit these claims”; “this promise that they’re building something terrible and that they can’t seem to stop, I think more alarms the public than anything else and in some ways attracts good attention to these companies.”
  • Coxon’s own clip [22:32]–[23:09]: recursive self-improvement; “Currently, we still need a human to provide taste or judgment. But possibly within a year, certainly within 2 years, we’re on track to also automating away this nebulous human quality of judgment … things could go very, very fast.”

📚 Deeper theory (links to 2.5, 2.6, 2.7, 2.8)

  • Silverman’s argument has two parts that are easy to blur.
    1. Motive/effect: warnings of apocalypse have a marketing effect (they signal power) and a liability-shifting effect (the harm is “inevitable”). This is testable in principle: do firms that stress doom also get contracts, funding or lighter rules?
    2. Substance: the risks may be real or not. He says “the AI industry says … and I don’t really buy that narrative” about inevitability, not about risk existing. His closing view at [25:52]: RSI is “certainly something to watch out for”. Treat them separately: a warning can be both sincere and useful to the speaker. Vinsel’s “criti-hype” (2.5) fits part 1.
  • What the Oppenheimer analogy proves. The scientists computed the risk before the test and then proceeded; public accounts of the 3-in-a-million figure are from later recollection. The AI analogue would be published quantitative risk estimates before each release. Some labs publish safety frameworks (Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework; from memory). Whether they are adequate is the debate.
  • Determinism as an argument form. “Someone has to do this” is a collective action claim: if I stop, a rival goes ahead. It is often true in a race (2.8) and is exactly why economists want binding common rules instead of voluntary restraint. So the AI leaders’ own logic (race pressure) supports regulation, which is what the UN-stage CEOs requested [00:52]. Silverman’s point is that asking for regulation in general terms while lobbying on specifics is a way of looking responsible.
  • RSI evidence. Coxon’s claim has two empirical components: AI automating more AI research (partly observable: labs say they use AI for experiments), and automation of “taste” within 1–2 years (a forecast; contested). The historical record of AI timelines shows both over- and under-prediction. The press reports Anthropic’s alignment-science lead Evan Hubinger backing Coxon’s seriousness on risk while placing current models at low risk (TechCrunch, via the search summary; not read in full).
  • Present harms are documented separately. AI-linked mental-health harm, data scraping, deepfakes and military use all have reporting and lawsuits. The Silverman/Narayanan/Bender line holds these deserve policy now; the AI-safety line says both matter. They are often set up as rivals, but many researchers work on both.

Where others push back.

  • Safety researchers (including some at the firms): The warnings are sincere. People quit, give up equity and risk their careers, as Coxon did. Calling it “marketing” is unfalsifiable and insulting: if firms stay quiet they are accused of hiding risk; if they speak, of hyping. Also, Hinton and Bengio (not employed by labs) say similar things.
  • Industry and accelerationist voices (Sacks, Andreessen): The reverse charge: doom talk is the capture (Sacks’s Oct 2025 post). So Silverman and Sacks both accuse the other camp of “using” the risk talk. A reader should notice they each assume the other side’s motives.
  • Technical sceptics (LeCun, Narayanan and Kapoor): RSI claims over-read lab anecdotes; real research needs data, compute and experiments that do not speed up easily.
  • Legal realists: “Stop making the product” is not how any dual-use technology is governed. Society regulates conditions of use (cars, drugs, nuclear plants), not abolition. The real question is what standards and who enforces.
  • From the existential-risk side against Silverman’s “present harms” emphasis: If even a small chance of loss of control is real, ordinary consumer law cannot handle it, so “existing laws” (Huang, Silverman) is not enough for frontier risk.

What they leave out.

  • Specific rules. Which regulation would he want, apart from audit and liability? There is no mention of licensing, compute limits, incident reporting, or an international body.
  • Counting the benefits. The “utopian potential” is dismissed as part of the pitch, with no weighing of real uses (medicine, science, accessibility).
  • Other warners’ motives. Coxon is treated as a data point about industry behaviour, not as an independent expert. A fuller answer would test his argument on its merits, which Silverman invites by calling for audits.
  • International view. The UN statement of 20+ countries urging oversight [see debate section] is outside the segment.

Check yourself

  1. Separate Silverman’s two claims: that doom talk is a pitch, and that the risk is not real. Does he make the second?
  2. How does the race logic behind “someone has to do this” support the case for binding rules?
  3. What would count as evidence that warnings are marketing rather than sincere?

5. Musk: volatility, money and background

🎙 In the talk.

  • [10:05]: Musk “has fallen out with Trump and gone back together a couple times”. The point: “these tech leaders are able to stick by him.”
  • [15:15]: Musk is “very impulsive”; “he doesn’t … have a lot of focus as a political actor because he’s so rich, he can afford to delegate or just say I’m going to donate $200 million to various Republican causes”. Silverman adds “There have been reports that he’s a drug user”, then the “late-at-night tweet of him saying that Trump is in the Epstein files, but then a month or two later he’s making up with Trump”. His summary: “very transactional, which is just what Trump is … he feels like he can act through or at least use to his own ends”.
  • [16:24]: he corrects “2028” to “2026 midterms”: “a lot of traditional MAGA candidates with that kind of right-wing tech backing led by Musk … the only difference in the midterms is … you have AI super PACs.”
  • [17:11]: Musk “grew up in apartheid South Africa … he was part of the racial elite … his own inherent elitism, including, of course, his genetic elitism, cuz he wants to reproduce as much as possible”. He says Musk “clearly displays a sense of, uh, racial and genetic superiority”.
  • [08:11]: “the woke mind virus, that is a real thing for him based on his transphobia towards his daughter who came out as trans.”

📚 Deeper theory (links to 2.3)

  • The money pattern is checkable; the psychology is not. Public records and news show political spending by Musk in 2024 (via America PAC; from memory) and big AI and crypto PAC spending in 2026 (see debate section). The PAC data are the solid part of this deep dive. Claims about drug use, “genetic elitism” and racial superiority are interpretations or reports and should be read as Silverman’s view.
  • Biography as explanation. Silverman uses a childhood account (apartheid-era South Africa) to explain later views. This is the psychobiography method. It is persuasive as narrative and weak as proof: many people from the same background hold other views. A more careful historian would give both background and Musk’s own statements.
  • Immigration claim, stated carefully. The documented story: the Washington Post reported (Oct 2024) that Musk, after leaving his Stanford programme days after arriving in 1995, worked without proper authorisation for a time; Musk says he was legally in the US on a J-1 and later H-1B; two former colleagues recall authorisation around 1997 (Stanford Daily). “Immigration fraud” goes further than that reporting.
  • The Musk–Trump cycle. Their public break (2025) and reconciliation are documented in the press (from memory). Silverman’s “transactional” label matches a standard political-science view of Trump (transactionalism). It is also a lens that predicts continued deal-making.
  • AI PACs in 2026. Pro-industry Leading the Future is backed by Brockman, Andreessen and Horowitz; Public First by Anthropic. This is not a monolith: AI money is on both sides of the regulation fight.

Where others push back.

  • On the character claims: Defenders say these are ad hominem. Musk’s critics say the behaviour (X posts, donations) is public and fair to discuss. The reasonable middle: assess the public acts, treat inner states as speculation.
  • On “radicalised”: Some political scientists argue Musk’s shift is less a transformation than the surfacing of old views now that he is rich and online. Silverman’s “radicalised” frame implies change.
  • On South Africa: Musk and supporters point to his statements on South African policy and treat it as a political, not racial, position; critics read it as racial grievance. I did not verify the specific statements Silverman refers to about UK migrants [17:11].
  • On the book’s timing: Silverman says he wrote the book “a year ago” [08:47]; events since (Musk’s feud and return) are partly outside it. He says so.

What they leave out.

  • Musk’s actual companies’ role in AI (xAI, Grok, the lawsuit with OpenAI). Only a passing “Musk Altman rivalry” at [21:35].
  • Policy outcomes of Musk’s spending. What have his donations bought?
  • Alternative accounts of why donors back Trump (taxes, tariffs, immigration, crypto rules) beyond personality.

Check yourself

  1. Which statements in this section are documented public facts, and which are Silverman’s inferences?
  2. Why does the “AI super PACs” point at [16:24] matter more for policy than the psychology?
  3. What is the difference between psychobiography and causal explanation?

6. Whistleblowers, liability and “humans in the loop”

🎙 In the talk.

  • The hosts [19:04] introduce Suchir Balaji and Silverman’s April piece in The Nation, “The Death of an AI Whistleblower”.
  • Silverman [19:04]: Balaji was “in his mid-20s … an idealistic AI researcher” with jobs “in AI safety and what they call alignment”. What he found “is what’s happening at a lot of AI companies or pretty much all of them, is that OpenAI was stealing data wherever it could find it and … using copyright information in order to train its models.” He says: “I did not look that much into the physical evidence”; “personally, I think it was a suicide”. The theme: “the stress that faces whistleblowers. There are not enough protections … Chuck Grassley of all people … has tried to introduce a bill.”
  • [21:35]: the parents “have actually been involved in a lot of this conspiracizing … interviewed by a lot of right-wing media, from Tucker Carlson on down. And part of that is because this story kind of plays into the Elon Musk Sam Altman rivalry.” He concludes: Balaji “was definitely failed by the systems around him.”
  • Coxon [22:32]–[23:09]: see Deep dive 4.
  • Hawking [24:32]–[25:14]: “The development of full artificial intelligence could spell the end of the human race… it would take off on its own and redesign itself at an ever-increasing rate.”
  • Silverman’s close [25:52]: the warnings are “decades old because a lot of this stuff goes back to the science fiction that people like Elon Musk or other people involved in this debate grew up reading … these are technologies made by humans and then hopefully controlled by humans. We don’t need to bring humans out of the loop … humans should be legally liable and responsible … this idea that it might just suddenly escape all of us … is a kind of an abdication of human responsibility and perhaps legal responsibility.”

📚 Deeper theory (links to 2.6, 2.7, 2.8)

  • The Balaji case: what is on record. Official finding: self-inflicted gunshot wound; the San Francisco police found no sign of foul play (SF Examiner). His parents commissioned a private autopsy and sued the city over public-records requests; their expert noted a head contusion (KTVU; see also Decrypt). I found no source linking Musk to the case; Silverman’s remark is about political use of the story, not about Musk’s involvement. Dates (death in Nov 2024, NYT story in Oct 2024) are from memory.
  • The legal question Balaji raised. Whether training on copyrighted works is fair use is not settled (2.7, point 6). His claim and the firms’ position are both live legal positions.
  • Why whistleblowers need special law. Insiders hold knowledge no outsider has, face NDAs, and risk careers. Case law and statutes protect fraud reporting more than risk reporting. S.1792 aims to cover “AI security vulnerability or violation of federal law” and void NDAs that block protected disclosures (Princeton CITP). Status of the bill: not confirmed past committee referral.
  • The “humans stay liable” principle in practice. Corporate law already makes executives and firms liable for products. New proposals focus on who (developer, deployer, user) and what standard (negligence vs strict liability). Silverman’s closing line is a design principle: no “the AI did it” defence.
  • Hawking’s quote in context. It is a 2014 warning that generalises Good’s 1965 idea (2.5). The hosts use it to show the warnings are old; Silverman agrees and pushes further back.
  • A pattern: insiders who warn, then leave. Examples (from memory, details unchecked): the OpenAI superalignment departures in 2024; Hinton leaving Google in 2023; Haugen at Facebook (2021). Silverman’s observation that researchers leave and sometimes “go to another one right afterwards” [23:53] is a fair point about revealed preference: if the danger were as stated, why join a rival?

Where others push back.

  • On the Balaji framing: The parents and some commentators think the official finding is wrong or incomplete and that transparency is lacking (their records suit). Silverman’s “conspiracizing” is itself contested language. Reasonable people can say: take the official finding seriously, and also support the records request.
  • On “revealed preference”: A researcher who warns, quits, then joins another lab may think better a safety-minded lab at the frontier. That is a coherent view (Anthropic’s founding logic), not hypocrisy.
  • On “humans should be liable”: Legal scholars note that liability alone may under-deter catastrophic or diffuse harms: firms can be judgment-proof and harms may land on third parties long after. Many argue ex-ante regulation (licensing, audits) is also needed. And with AI agents acting autonomously, assigning responsibility among developer, deployer and user is hard (the “responsibility gap” literature, Matthias 2004, from memory).
  • From the AI-safety side: “Abdication of human responsibility” may be right about current firms’ behaviour but wrong about capability: even with human responsibility, a system that can improve itself faster than oversight can respond is a real design problem. Silverman agrees RSI is “something to watch out for” [25:52], but his frame treats it chiefly as a governance issue.

What they leave out.

  • Details of Balaji’s technical claims. The segment doesn’t explain what he found, what he published, or the legal status.
  • Coxon’s reasoning beyond RSI. Only a short clip; no reply from Anthropic.
  • Non-US whistleblower regimes (for example the EU Whistleblower Directive, 2019, from memory).
  • How to run a credible audit and who would qualify.

Check yourself

  1. What is the difference between protecting someone who reports fraud and someone who reports a safety risk?
  2. What does the official record say about Balaji’s death, and what do his parents dispute?
  3. Why might “humans should be liable” be necessary but not sufficient for frontier-AI risk?

Putting it together

The mental model in one paragraph. The AI industry is a small number of firms whose products need state-scale resources: capital, chips, land and energy. That need pulls them into bargains with governments, and in 2025–26 the US government is led by a president who rewards loyalty and deals. Around him sits a network of tech investors (Thiel, Sacks, Andreessen, Musk, Vance) with shared views, shared history and, in Silverman’s account, shared grievance. The result is a policy line, “whoever wins AI wins”, that rejects international rules and leans on the firms’ own promises to “pace the frontier”. Silverman adds a second argument about language: the firms’ apocalyptic warnings help them (they signal power and say “it is beyond us”). He wants the opposite principle: audits, whistleblower protection, existing law used against present harms, and humans held legally responsible. The historical frame he borrows (the Gilded Age) suggests the remedy: it took laws, courts, organised citizens and a change of public mood, and it came late. What makes this case new: the products are general-purpose, the state is itself the customer, and some insiders think the product may become uncontrollable.

Open tensions

  1. Rage or interest? Is the politics driven by emotion and ideology, or by ordinary business interest? The answer changes the remedy.
  2. Sincere warning or sales pitch? Can a warning be both? Silverman’s argument needs a test he does not supply.
  3. Present harms or future risk? Competing for attention or complementary? Policy makers face limited agenda space.
  4. Liability or licensing? After-the-fact liability, or before-the-fact rules? Silverman leans on the first; frontier-risk advocates want the second.
  5. State as ally or threat? The same segment says firms seek the state (equity, contracts) and that the state punishes a firm (Anthropic). Both can be true; the segment does not reconcile them.
  6. Race logic. “We must beat China” is both a real strategic claim and a closing move on debate. Cooperation (treaties) has precedent (2.6) but also a high bar.

Reading path

  • Start: Richard White, The Republic for Which It Stands (the intro and chapters on money and politics), for the real Gilded Age. Tim Wu, The Curse of Bigness (2018), short, links then and now. Silverman, Gilded Rage (2025).
  • Then: Mariana Mazzucato, The Entrepreneurial State; Brandeis, Other People’s Money; Narayanan and Kapoor, “AI as Normal Technology” (2025); Bostrom, Superintelligence (2014) or Russell, Human Compatible (2019); Lee Vinsel, “You’re Doing It Wrong: Notes on Criticism and Technology Hype” (2021, essay).
  • Deeper: Karen Hao, Empire of AI (2025, from memory); Zuboff, The Age of Surveillance Capitalism; Levitsky and Ziblatt, How Democracies Die; Fred Turner, From Counterculture to Cyberculture; Richard Rhodes, The Making of the Atomic Bomb (1986) for the Manhattan Project; the S.1792 text and CITP explainer on whistleblower law.

A pointer (not from the talk). The segment is US-centred. A useful follow-up question for India work: where does the Indian approach (state-built shared digital rails, then private firms on top) sit against Silverman’s picture, and what does “whoever wins AI wins” mean for a country that is neither of the two contestants? Not addressed in the talk.

Q&A

(Add questions and answers here only if worth keeping.)

Graph View

Backlinks