Summary

OpenAI’s Pentagon deal turned a government-contract win into a trust, workforce-retention, and data-center-opposition liability. Sam Altman conceded the rollout looked “opportunistic and sloppy,” then added language barring deliberate domestic tracking of US persons, but OpenAI released only the supposedly relevant contract portion and employees called it “window dressing.” The near-term subscription boycott may fade; the durable risk is Meta-style reputational decay that mobilizes opposition to OpenAI’s planned data centers.
Anthropic is simultaneously facing an existential government fight and possibly the fastest revenue ramp in American tech. Bloomberg put the company on track for $20 billion in annualized revenue, versus roughly $1 billion at the start of 2025 and $9 billion around year-end, with Claude Code and enterprise adoption driving a doubling in barely over two months. Yet the Pentagon’s supply-chain-risk designation is official, and the State Department is replacing Claude with OpenAI and reverting to GPT-4.1—leaving Anthropic “printing money” while preparing for costly litigation.
Scarce frontier-model talent still has unusual leverage over AI-company strategy. Casey Newton divided OpenAI into a mission-focused, three-plus-year technical core and newer “Meta people” who may tolerate more commercial flexibility; leadership needs the former to build GPT-6 and GPT-7. Post-training lead Max Schwarzer’s departure for Anthropic, alongside public employee distrust, makes military policy a human-capital issue rather than merely consumer communications.
The Pentagon dispute may be an early rehearsal for “soft nationalization” of frontier AI. Kevin Roose argues that pressure is more likely to arrive incrementally—remove safeguards, alter model behavior, or let government help write “the clauses in the constitution of Claude”—before any outright takeover. If systems become a “country of geniuses in a data center” embedded in military command and control, private ownership may matter less than who effectively dictates development and deployment.
War markets exposed prediction platforms’ core informational paradox: their edge may be insider access, but that same edge creates dangerous incentives. More than 150 accounts reportedly placed hundreds of bets of at least $1,000 shortly before the US strike on Iran, while one market still showed roughly 17% odds an hour before the attack. Casey’s verdict—“a bunch of vibes plus some insider trading”—undercuts the price-discovery pitch while sharpening regulatory, reputational, and political risk for Kalshi and Polymarket.
YouTube’s children’s-content problem is becoming an AI-scaled sequel to Elsagate. In one controlled 15-minute recommendation session, more than 40% of videos were judged AI-generated; examples ranged from alphabet animals squeezed from goo to violent or bizarre knockoffs of familiar characters. There is no AI-slop filter, labeling generally covers only realistic synthetic media, and cheap generation lets content makers exploit the same engagement algorithm at far greater volume.

Deep dive

1. OpenAI’s Pentagon deal created a trust crisis it could not amend away

Kevin Roose’s starting point: OpenAI said its Pentagon agreement preserved prohibitions on domestic mass surveillance and autonomous weapons, essentially the red lines that had brought Anthropic into conflict with the government. Instead of looking principled, the late deal provoked one of the largest backlashes in OpenAI’s history, including highly upvoted condemnation, ChatGPT cancellations, and users switching to Claude.
Casey Newton reduced the disclosure problem to a bleak exchange: both OpenAI and the Pentagon are effectively saying, “You’re just going to have to trust us,” while the public answers, “Well, we don’t.” OpenAI published what it called the relevant contract language, but procurement experts warned that nobody could know whether it was truly sufficient without seeing the entire contract.
Altman later admitted, “We shouldn’t have rushed to get this out on Friday,” and said the move looked “opportunistic and sloppy”—or, in Casey’s coinage, “slopportunistic.” OpenAI then amended the agreement to prohibit deliberate tracking, surveillance, or monitoring of US persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information.
Casey’s warning was that surveillance disputes often turn on “Jedi mind tricks”: governments can approach the legal limit, then rename surveillance as intelligence gathering. “Whether or not you personally are surveilled will come down to semantics,” which is why a seemingly explicit contractual restriction still leaves material uncertainty about enforcement and interpretation.

2. OpenAI’s hardest constituency is its irreplaceable technical core

Employee dissent moved into public view. Leo Gao called the contract language “window dressing,” arguing that it still appeared to leave the Pentagon control over autonomous-weapons deployment and failed to close other loopholes. On Tuesday, post-training lead and research vice president Max Schwarzer announced his departure, praised Anthropic’s values, and said he was joining the rival lab.
Casey did not think OpenAI had “stemmed the tide.” The AMA, blog post, partial contract disclosure, and amendments showed that leadership was “really going for it,” but also that it was “really scared”: AI was already associated with slop and workplace coercion, and adding government spying or “a murder bot” made public rejection almost automatic.
Kevin had assumed Silicon Valley’s worker-empowerment era had ended. Casey argued that frontier researchers possess scarce knowledge about training next-generation systems; because they are difficult to replace, their ethical discomfort retains strategic force.
Casey’s deliberately sweeping taxonomy separated a mission-driven cohort with three-plus years at OpenAI from newer “Meta people” who may be more flexible about company conduct. Leadership can tolerate the second group’s acquiescence, but it needs the original core to make GPT-6 and GPT-7 “blow everybody’s minds,” so much of the damage control was aimed internally.

3. Anthropic is “printing money” while the government attacks its distribution

Casey’s two-word summary of Anthropic’s commercial position was “printing money,” followed by the larger paradox: “Has an American technology company ever had such a good week and such a bad week at the same time?” The Pentagon fight could still involve the Defense Production Act, potentially compelling Anthropic to provide a version of Claude it does not want to build.
The bad side became concrete when the Pentagon formally notified Anthropic of its supply-chain-risk designation. Casey expects a long, expensive legal battle over whether American companies may continue using Claude for nonmilitary purposes, with “an existential threat to the company” still buried inside the dispute.
The counterweight is extraordinary growth: Bloomberg reported that Anthropic was on track for $20 billion in annualized revenue, up from roughly $1 billion at the start of 2025. Anthropic was on pace for about $9 billion by the end of 2025, then more than doubled in barely over two months as Claude Code and enterprise Claude adoption accelerated.
Federal displacement is already visible. Reuters reported that the State Department replaced Anthropic in its internal chatbot with OpenAI and went back to GPT-4.1, an early-2025 model; Kevin’s comparison was that “the average college freshman with a ChatGPT subscription” now has substantially better tools than the State Department.

4. Consumer anger may fade, but institutional distrust compounds

Altman has repeatedly urged the Pentagon to offer Anthropic the same terms it gave OpenAI. Casey saw sincere de-escalatory intent—Altman does not want the government nationalizing AI companies “at least not right now”—but also obvious competitive cover: if Claude accepted identical terms, OpenAI could neutralize “quit ChatGPT” campaigns.
Kevin doubted consumer boycotts would matter, recalling repeated promises to delete Uber or quit Facebook. Casey agreed that cancellations usually blow over, but invoked Meta’s more durable lesson: users stayed while learning to hate and distrust the company, weakening it politically even when most people did not quit.
The physical pressure point is data centers. OpenAI intends to build many across the country, while local opposition is already entering politics; if the company cannot persuade communities that AI will improve their lives, resistance to nearby infrastructure can become a proxy for broader distrust.

5. The Pentagon fight looks like a rehearsal for AI nationalization

Kevin framed the long-run question through Dario Amodei’s “country of geniuses in a data center”: once such systems carry decisive geopolitical and security value, can they remain inside private corporations? Casey said models already embedded in military command and control are becoming weapons, and that he absolutely believes the government will take an interest in systems three, four, five, or 10 times more powerful and potentially oversee their development and deployment.
The outcome depends on “the quality of the government” doing the overseeing. Casey contrasted a state seeking opportunity, safety, and democracy with one seeking an authoritarian global takeover; the technology’s concentration makes governance quality part of the asset’s fundamental risk, not an external political footnote.
Kevin wondered whether the rational response was to stop dealing with an untrustworthy federal counterparty. Casey noted that even after the Anthropic confrontation, Amodei said the company had nearly reached agreement, liked working with the military, and wanted to resume—an effort to “keep the tigers at bay” while buying time.
Kevin’s view was that “soft nationalization” was the more likely direction: incremental demands to remove safeguards, modify models, or let officials shape “the clauses in the constitution of Claude.” He would not dismiss brute-force takeover, but expects control to shift through pressure before ownership formally changes.

6. The Manhattan Project analogy reveals where scientific control ends

Rereading The Making of the Atomic Bomb, Kevin focused on scientists’ unsuccessful 1945 petitions against using the weapon on a city as the first act of war. Government effectively replied: “You’re the scientists. You’re the geniuses who made this all work. But now you’re playing in our turf,” then took control of deployment.
He also marked the analogy’s limit: the Manhattan Project was government-funded and staffed, whereas today’s frontier systems were built by private companies. Still, once the capability becomes too useful to governments and militaries, even a different ownership history may not preserve corporate autonomy.
Casey added that OpenAI was initially imagined as a government-funded project, but its founders correctly concluded no government would supply the required capital. Private financing solved the development problem without eliminating the expectation that the state would eventually take an interest.
The political reversal struck Casey as “the whiplash is insane.” President Biden’s order gently requested safety-testing information and agency preparation, yet critics called it anti-capitalist; the same people who objected to that approach later came to power and said, in effect, build models for the military as directed “or else we will destroy you.”

7. Iran bets turned prediction markets from abstraction into direct incentives

Kalshi bars explicit war and assassination bets but listed “Khamenei out as Supreme Leader,” a proxy whose wording allowed multiple interpretations even though Casey assumed most traders expected death in war. Kalshi later voided the market and said it would reimburse anyone who may have lost money, angering bettors who believed they had correctly predicted Khamenei’s death and deserved their winnings.
Casey’s response to disappointed traders was categorical: “I don’t care. And it doesn’t matter.” Polymarket was still more permissive, offering wagers on strike dates and other operational details, though it drew one boundary at nuclear detonations by specified dates.
Senator Chris Murphy called the situation “insane,” argued that people around Trump were profiting from “war and death,” and proposed legislation. The concern exceeds bad taste: Israel had recently arrested people accused of using classified military information to place Polymarket bets, showing that the incentive problem was already starting to happen.
The Times reported that last Friday more than 150 accounts placed hundreds of bets of at least $1,000 predicting an American strike by Saturday—an unusual concentration of correctly timed money. A service member with advance orders could treat the platform as a “little Kalshi bonus,” converting classified operational knowledge into personal profit.

8. Prediction-market prices are “vibes plus some insider trading”

Kevin had accepted the theoretical case that money improves forecasts: traders may build better polling, reveal beliefs they would hide from pollsters, or transmit insiders’ superior information. But Kalshi officially prohibits inside-information betting, the CFTC is small, and categories of public versus private information are much less settled outside securities markets.
Casey’s challenge went to utility: beyond a brief warning that something horrible might happen, what did the Iran contracts help the public understand? An hour before the strikes, one market reportedly showed only roughly 17% probability—suggesting prices aggregate vibes until an informed trader arrives and captures the edge.
Kevin steelmanned the defense that people already profit from wars through defense stocks and oil. Casey’s distinction was directness: buying oil shares is meaningfully removed from assassinating a leader, while a dated removal contract could function like a bounty and create “the same horribly grim incentives.”
Regulation faces a time race. Casey fears Kalshi and Polymarket could become crypto-like entrenched interests with enough money to secure bipartisan protection; Kevin’s deliberately friction-heavy alternative was requiring war bettors to visit “a seedy, like, OTB betting place” instead of wagering instantly by phone.

9. YouTube’s AI-slop pipeline scales the same old engagement failure

The review showed alphabet animals squeezed from paint tubes, injected by an animal doctor until they changed color, or transformed into armored vehicles. Casey identified both the attention hack—needles frighten children—and the production logic: alphabet videos let creators stitch together the few-second clips current AI video generators produce into something superficially coherent.
Kevin pushed back on claims that hyperstimulation began with AI: CoComelon and other established children’s shows already use rapid songs and cuts, meaning “we crossed the Rubicon a while ago.” AI’s decisive change is economic—making endless surreal variations easier and cheaper—while Casey worried that raw visual bombardment replaces calm, story, and narrative development.
Ariella Leka’s investigation began by looking at recommendations around channels parents approve of, including Ms. Rachel and Bluey, then testing what appeared next to CoComelon videos. In one 15-minute session, more than 40% of the videos were AI-generated; her team logged recommendations without interacting, then inspected frames for morphing objects, distorted text, synthetic labels, and abrupt changes in longstanding channels’ production style.
YouTube requires disclosure mainly for realistic synthetic media, leaving animated AI content largely unlabeled and parents without a filter. Ariella cited cognitive overload, weak narrative arcs, and rapid changes for under-fives; worse, searches surfaced Masha’s stomach being cut open and pregnant KPop Demon Hunters characters—a cheap, scalable sequel to Elsagate. Shorts time limits may help, but the closing discussion broadened the concern to older teens and adults: TikTok and Instagram Reels can likewise deliver raw, hypnotic stimuli across a lifelong pipeline of engagement-optimized slop.

Atlassian CEO on the SaaS Apocalypse, AI Agents & What Comes Next Thomas Laffont, Coatue - Anthropic, Citrini Paper, AI Volatility & Next Mag 7