Summary

Elon Musk’s DOGE is importing the Twitter takeover playbook into Washington, but the asset being “broken” now is the federal government itself. Roughly two million workers received a “Fork in the Road” resignation offer, nearly all USAID staff were put on leave, and a young, opaque technical cadre sought access to agencies’ systems. Jonathan Swan’s governing warning: when clean water, food safety, cybersecurity, air travel, and national security are involved, “the ‘break it’ part of it is pretty important.”
DOGE’s highest-leverage move is access to Treasury’s payment infrastructure, where more than $5 trillion a year and roughly 88% of federal spending flow. Officials insist the access is “read-only,” but Swan says former Treasury leaders had never heard of political appointees receiving it; the constitutional fight is whether a president can withhold money Congress appropriated. Musk’s operating premise is blunt: “The way to control government is to control the computers.”
Musk’s savings target has already fallen from $2 trillion against a roughly $7 trillion federal budget to “maybe” $1 trillion, which observers still consider implausible. DOGE is pairing zero-based budgeting with a “flood the zone” strategy: launch more aggressive actions than the media, Democrats in Congress, and outside nonprofits can fight simultaneously. With Congress showing little appetite to intervene, Swan says the practical limit is “the extent to which Trump tolerates him.”
Spotify’s first profitable year, after 18 years, rests on a platform with 675 million users and about 263 million paying subscribers—and increasingly on controlling what those users hear. The service evolved from a search-driven library into a recommendation engine whose playlists confer market power and can steer listening toward cheaper-to-license “perfect fit content.” The margin mechanism is invisible curation: production companies commission tracks for particular moods, and musicians sometimes produce “12 or 15 of these tracks in an hour.”
Spotify’s ghost-music model does not merely substitute cheap tracks; it rewires supply around whatever survives in a lean-back playlist. Liz Pelly traces lo-fi hip-hop from J Dilla- and Madlib-inspired experimentation on forums and SoundCloud to flattened study music optimized by prior streaming performance. Her disclosure critique is decisive: listeners cannot decide that they do not care if commercial recommendations are never identified as such.
OpenAI’s Deep Research is the first agent Casey Newton has used that feels capable of completing a substantive multistep knowledge-work assignment. A $200-a-month ChatGPT Pro subscription included 100 monthly queries; Kevin Roose’s AGI-history request took 10 minutes, consulted 36 sources, and produced seven or eight pages. Newton still finds “at least one mistake” in every report, but both hosts see potentially steep consequences for research-heavy journalism, consulting, and finance.
Granola’s strong meeting summaries expose a larger product gap: dependable AI that turns messy communications into decisions without surrendering decades of private data. Granola can structure meeting notes and surface action items, but its lack of recordings limits journalistic verification and its background operation raises consent questions. Roose’s desired email autopilot would scan three or four times daily, identify important messages, draft replies, and deliver an action digest—functionality he calls Gmail’s missing “failure of imagination.”

Deep dive

1. DOGE is replaying the Twitter takeover against a sovereign institution

Roose frames Twitter as the “warm-up act”: DOGE is reusing Musk’s tactics for seizing a hostile institution, cutting costs, challenging incumbents, and purging workers perceived as disloyal. The crucial change is scale—Twitter was a company Musk owned; the federal government derives spending authority from Congress and delivers services people cannot opt out of.
The opening balance sheet was already extraordinary: DOGE gained access to Treasury’s payment system, placed nearly USAID’s entire workforce on leave, and emailed roughly two million federal employees an option to resign and allegedly receive pay through September. Its subject line, “Fork in the Road,” repeated the one Musk used after buying Twitter.
Swan describes the career-service atmosphere as “terror.” USAID employees saw their website disappear, received after-midnight instructions not to report to work, and followed Musk on X as he called their agency evil. An unidentified DOGE visitor might arrive in a T-shirt and blazer and effectively demand: “You are a lazy, worthless, idiotic federal worker. Justify your existence to me, please.”
Newton’s distinction is the episode’s essential pushback: Musk largely had the legal right to reorganize a company he purchased, notwithstanding disputed agreements. In Washington, “the richest man in the world” was not elected and had not presented Congress with a plan or received legislative consent.

2. Controlling the computers is DOGE’s theory of state power

Swan’s reporting captures Musk’s premise: “The way to control government is to control the computers.” Rather than accept career officials’ explanations, DOGE wants source code and direct access to the “pipes of government,” applying Musk’s factory-floor habit of questioning every component and publicly spotlighting spending he considers absurd.
Treasury is the highest-stakes implementation of that idea. Its payment system distributes more than $5 trillion annually—about 88% of federal spending—including Social Security payments. Former officials told Swan it had historically been managed by a small group of trusted, experienced civil servants; they had never encountered political appointees requesting, much less receiving, such access.
The administration says DOGE’s Treasury access is “read-only,” meaning its personnel cannot alter payments, but Swan emphasizes that even this is extraordinary. The cadre itself remains opaque: about 40 people were reportedly present around the inauguration, ranging from seasoned Musk allies to participants associated with the ancient-scroll effort, with one participant reportedly around 19.
DOGE’s budgeting doctrine starts every account at zero and forces agencies to justify each dollar added. Musk reportedly measures success in dollars saved per day, while teams inspect Treasury payments and USAID and examine the government’s real-estate portfolio rather than merely trimming an inherited baseline.

3. The $2 trillion promise meets arithmetic—and a flood-the-zone strategy

Musk initially said he wanted to remove $2 trillion from a federal budget of roughly $7 trillion. Swan calls implementation “almost impossible to imagine”; Musk subsequently reduced the target to perhaps $1 trillion, itself a figure observers still regard as astonishing and implausible.
The ideological allocation is clearer than the arithmetic. USAID embodies what the Trump movement sees as a leftist “deep state” institution spending abroad rather than following “America first”; the resignation email similarly treated federal employment itself as lower-productivity work, without distinguishing expertise or experience.
Swan rejects the idea that every DOGE action is expected to survive. The learned strategy is to “flood the zone”—advance aggressive policies faster than the media, congressional Democrats, outside nonprofits such as the ACLU, and the legal system can fight them. Limited institutional bandwidth forces resistance to choose targets while other measures become facts on the ground.

4. Courts may arrive after agencies and programs have already changed

Congress unquestionably holds the power of the purse, while the White House characterizes halted spending as a temporary review for conflicts with Trump’s priorities. Swan says the actions may seed a Supreme Court fight over presidential authority to refuse expenditures that Congress has appropriated, a power Trump’s aides have long wanted to reclaim.
Litigation is structurally slower than DOGE. Musk and Trump are taking so many actions so quickly that the legal system struggles to catch up; meanwhile, foreign-aid projects in Sudan and elsewhere have already stopped. Swan’s forecast stays properly mixed: “Some of this won’t fly ultimately, but some of it will.”
The legislative check looks weak because Republicans control both chambers and Congress has shown little appetite to assert its authority. For Musk personally, Swan sees one operative constraint: “The limit on him is the extent to which Trump tolerates him. That’s the only kind of limiting principle.”
Defense may eventually become the unavoidable target. Trump has promised not to cut Social Security or Medicare, leaving less room to reach DOGE’s savings ambitions without confronting Pentagon spending. Swan flags the accompanying conflict: Musk’s SpaceX holds major federal contracts, yet meaningful government-wide cuts cannot indefinitely ignore the Defense Department.

5. Government AI could find waste—or compound opaque access risks

Musk’s allies want to centralize federal contracts and analyze them with AI to recommend cuts. Thomas Shedd, a former Tesla engineer leading a General Services Administration technology team, reportedly described that plan to staff; Swan had separately heard Musk discuss using AI to identify waste for months.
Swan does not dismiss the concept: using better tools to find waste “doesn’t seem like a crazy idea” in principle. His problem is the lack of visibility into which tools will be deployed, how they will work, what information they will receive, and what safeguards will govern their recommendations.
Newton distinguishes DOGE from established modernization efforts such as the U.S. Digital Service and 18F. Prior administrations also recruited technologists, but introducing tools into sensitive systems normally required procurement, security, and privacy review—not engineers declaring, “We’re going to use these tools, whether you like it or not.”
Newton frames Musk’s approach as an extension of “founder mode”; Swan’s response emphasizes consequence, not efficiency. The government protects water, food, infrastructure, cybersecurity, aviation, and national security, so the “break it” part carries country-level stakes.
A former senior official highlighted counterintelligence exposure: young private-sector personnel moving quickly through sensitive systems create new vulnerabilities for foreign governments already targeting the federal workforce.

6. Spotify converted universal access into curation power

Spotify reported its first profitable year, prompting CEO Daniel Ek’s line: “It only took 18 years for us to get here, but we’re here.” The platform reached 675 million users and roughly 263 million premium subscribers, with ad-supported revenue also rising.
Pelly traces the product’s initial appeal to the post-Napster and post-Pirate Bay listener: Spotify resembled a search bar attached to a vast digital library. Through about 2012, its own branding emphasized “instant, simple, free” access rather than a service choosing what listeners should hear.
Research then showed that users wanted recommendations and an immediate feed of appropriate music, not merely access. By late 2012 and early 2013, Ek acknowledged he had been too precious about non-curation, the homepage was redesigned, and Spotify leaned decisively into playlists.
Playlists such as RapCaviar became promotional infrastructure with radio-like gatekeeping power. Spotify marketed a supposedly democratizing pyramid: independent artists would enter feeder playlists, generate strong completion or reaction data, and rise through the system. Pelly calls that meritocratic pathway a “myth” many independent musicians never experienced.

7. “Perfect fit content” turns recommendation control into margin

Pelly’s central business mechanism is a series of cost-saving initiatives that nudged users toward music cheaper for Spotify to license. Around 2017, listeners noticed study, sleep, chill, and relaxation playlists filling with tracks from artists who did not seem real—material resembling royalty-free stock music.
Internally, this was called “Perfect Fit Content,” or PFC: music commissioned for particular playlists and moods at improved margins. Rather than simply curate existing recordings, Spotify worked through privileged production companies that hired producers and composers to manufacture music suited to its lean-back environments.
Musicians told Pelly they could crank out “12 or 15 of these tracks in an hour.” Production companies supplied examples of songs already performing well on Spotify, and composers reproduced similar styles—simple enough to remain unobtrusive in the background and abundant enough to optimize streams and cost.
Roose supplies the demand-side defense: as a heavy user of sleep and study playlists, he cares more about the appropriate sound than the artist’s identity. Pelly’s answer is disclosure: some senior executives reportedly reasoned that “most people don’t know, and also they don’t care,” but users cannot decide whether they care when the commercial arrangement remains invisible.

8. Playlist economics flattened lo-fi before generative AI arrived

The “lo-fi hip hop beats to study and relax to” genre had a more experimental prehistory. In early-2010s forums and on SoundCloud, artists made J Dilla- and Madlib-inspired beats, flipped samples, competed over impressive drums, and created music that was neither exclusively mellow nor designed as background.
As the scene moved to YouTube and Spotify, playlist curators placed portions of it into study contexts. Tracks suited to those playlists performed financially, so creators made more of that type—a feedback loop in which distribution incentives progressively flattened the culture supplying the feed.
Newton keeps the ambiguity intact: popular playlists encouraged much more production of something listeners demonstrably enjoyed. Pelly’s objection is not that the resulting music is simply “fake”; it is that independent ambient, jazz, classical, and lo-fi artists—of whom there is no shortage—lose discovery and income when editorial-looking placements instead favor stock catalogs.

9. Flat-fee production deals can “buy composers out of their luck”

Pelly rejects any suggestion that she is shaming musicians for taking production gigs. The deception can extend to them: composers submitted tracks, received payment, and often knew nothing about Spotify’s broader PFC arrangement or what happened to their work afterward.
Contracts varied by production company. Some musicians received a flat-fee buyout for the master, perhaps retaining other royalty rights; composers associated with the U.K.’s Ivors Academy argued that companies such as Epidemic Sound were trying to “buy composers out of their luck.”
Their reasoning is specific to portfolio economics: production composers make large volumes of music without knowing which recording might go viral, land in a commercial, or generate sustainable income for years. A buyout transfers that uncertain upside away from the creator precisely because any single track’s eventual value is unknowable.
Generative AI could intensify the model, but Pelly marks the evidence carefully. Ek has praised its cultural and engagement potential, and Epidemic Sound has expressed interest in composers using such tools; she did not directly observe AI inside PFC, even though AI-generated music was already flooding streaming services daily.

10. Frictionless listening trades active taste for retention

Pelly separates access from passivity. She favors universal access and credits Napster-era file sharing with her own musical formation; Roose similarly remembers saving $18 for a single CD. The concern is that streaming makes lean-back listening the most frictionless path, devaluing music culturally as well as financially.
A former Spotify staffer described curation’s goal as reducing the “cognitive work” users perform after opening the app. The long-term ideal is a perfect recommendation at the perfect moment, requiring no decision, choice, or thought—the TikTok engagement model applied to sound.
One former Spotify engineer called TikTok “the ultimate distillation of lean-back listening”: the user supplies no explicit input beyond how long they linger. Pelly argues that listening, rejecting, asking why, and encountering music outside one’s comfort zone are culturally important activities, not inefficiencies to optimize away.
Roose accepts the critique personally: he used to seek specific artists but had become lazy, losing agency over a taste that felt less his own. He resolves to choose more intentionally; Pelly’s human recommendation is ambient composer Emily A. Sprague, also the singer of Floras—ideally heard, Newton jokes, “while you’re awake.”

11. Deep Research clears the agent bar; communications AI still does not

OpenAI’s Deep Research, available through the $200-a-month ChatGPT Pro plan with 100 monthly queries, asks clarifying questions before using the unreleased o3 model. Roose’s AGI-history assignment ran for 10 minutes, consulted 36 sources, and produced seven or eight pages spanning the 1956 Dartmouth workshop back to Thomas Hobbes in the 17th century.
Newton calls it “the first good AI agent” he has used: every report has contained at least one mistake, but the tool builds timelines, organizes conceptual buckets, and supplies citations at a speed he doubts an editorial assistant could match. Compared with Gemini Deep Research, Roose found OpenAI’s reasoning-model output at least twice as long, more comprehensive, and more plausibly insightful.
The labor implication is broad because research underlies journalism, consulting, finance, and other knowledge work. Deep Research rescued Newton’s Pro subscription after Operator disappointed him; if limited to one subscription, he would now choose ChatGPT Pro, while Roose prefers Claude’s $20 Pro plan as a strong daily driver despite lacking web browsing and comparable research features.
Granola already handles a narrower workflow well: it transcribes computer audio without retaining a recording, produces structured summaries, supports VC-pitch and one-on-one templates, surfaces quotations behind bullet points, and answers questions about meetings. Its privacy design limits journalists who need recordings for quote verification, while its invisible background operation makes prior consent essential.
Email remains the conspicuous failure. Newton sees only nine or 10 recurring message types, yet Notion Mail’s beta and Shortwave could not reliably extract action items; Roose will not hand 20 years of Gmail to a company he does not necessarily trust, and his locally built Claude prototype began replying to spam.
Roose’s specification is modest but unmet: scan the inbox three or four times daily, identify important messages, draft replies with send and edit buttons, and issue a daily digest of action items, optionally linked to the calendar. His verdict to Google—despite its new “2.0 flash thinking experimental apps” model—is that Gmail’s missing autopilot represents “a failure of imagination.”

Boldstart Ventures' Shomik Ghosh on Kelly Partners' $KPG.AX unique accounting firm roll-up strategy No Priors Ep. 100 | With Sarah and Elad