Glossa is my editor. Every week it reads the writers, sites, podcasts and videos I follow, and on Sunday it writes back what happened, checked against other outlets, showing what is solid and what only one source claims. Paste an article or name a subject and it writes that one too, in the voice and the length I like to read, which I keep tuning piece by piece. No ads, no side taken.
Your sources, your subjects, and a voice you tune until it reads the way you like. Leave your email and I will write back when there is a place.
A retired colonel who once ran the Secretary of State's office lays out a country that buys its allies, cannot build enough artillery shells, and weighs the unthinkable over Iran. Lawrence Wilkerson's reading runs from Korea in 1950 to the Arctic now: a hegemon refusing to adapt to a power shift it cannot stop.
The people who grade artificial intelligence have quietly traded the quiz for the time sheet, the invoice, and the audit. Models now double every seven months the stretch of human work they can finish, yet they stay unreliable, sag when the briefing thins, and cannot clear a fifth of the marks when their research is checked. Three rulers now decide what counts as progress, and each one leaves something essential unmeasured.
Getting the right answer out of a language model has become an engineering discipline of its own: generate a million candidates and filter ruthlessly, as AlphaCode did in programming contests, or let the model pause mid-thought and search when it hits the edge of what it knows. Both approaches now define the front line of AI capability. Some of their numbers are documented in public records, and a few remain the lecturer's alone.
The smallest models in the room are beating GPT-3.5 on one of America's hardest math exams. The trick is not size but practice: the models generate their own reasoning, keep whatever survives the answer key, and train on it. Three papers built that loop, and the details decide whether it works.
The best language models still commit to the first plan they write, then follow it past the point where it stops making sense. Three frameworks now on the table — tree search over candidate plans, parallel execution of independent steps, practice data the model writes for itself — are the field's way of buying second thoughts, and every one of them is paid for in computation.
Language models were built to predict the next word, and they hit a ceiling: fluent, confident, and wrong with no way to catch it. Three research programs now grade them against something real, a search engine, a code execution, a written constitution. The gains are documented, and so is the finding that a small model grading itself can quietly get worse.
Language models still bluff their way through multi-step arithmetic, and the fix that is working is not a bigger model but a second one trained to grade the first. The newest graders watch every step, learn from cheap noisy labels, and get distilled back into a single small checker. Compute is quietly being rerouted from solving to judging.
The strongest results in AI no longer come only from building bigger models; they come from giving trained models more time and more tries. Two of the researchers behind that shift open their course by walking through both eras of the field. The catch, which they state plainly, is that somebody still has to pick the right answer out of the crowd.
The case that inference, not training, is the cheapest way to make a language model smarter: sampled hundreds of times and graded automatically, small open models beat frontier systems on math and on real software bugs. The catch is the grader. It comes free with code and mathematics, and almost nowhere else — which is where the fight over the industry's next compute budget begins.
Mexico was an A-rated borrower a few years ago and now pays more to borrow than Guatemala and Panama. Most of the gap traces to one line: roughly $130 billion spent rescuing Pemex, the state oil company, while filings and rating notes show a sovereign steadily absorbing a company's balance sheet. What the market has done in response is quieter, and more interesting, than panic.
Mexico's most famous recluse is back in the capital, and the reason runs through a courtroom in Brooklyn: a revoked visa, a general who turned, and a family that weighed exile in Russia or China before running out of doors.
Danny Moses, the FrontPoint trader of The Big Short fame, tells David Lin the Treasury's plan to fund bond buybacks from its $950 billion cash account is a distress signal, not routine debt management. He traces the strain from Tokyo's weak yen to private credit to NVIDIA's $500 billion AI facility, and sees one pattern: intervention stacked on intervention. Reporting from Reuters, Quartz and the New York Fed lets the reader check his numbers as the argument builds.
Six months into the war with Iran, the US Treasury has named five sectors of the Iranian economy and told their foreign suppliers to walk away or lose the dollar. Scott Bessent calls it an economic D-Day, but his own timetable, and his own question about the global financial system, suggest a weapon built not to be fired. The piece reads the warning against two decades of sanctions and one expert's verdict: the seal cannot hold.
In a wide-ranging interview, Zulfiqar, joined by the analyst Parveen Sawhney, argues that a Saudi-Pakistani defense pact now courting Bahrain and inviting Iran, a Fifth Fleet in diplomatic motion, and Tehran's uranium sprint add up to the end of American primacy in West Asia and of the petrodollar beneath it. Much of the scaffolding is his own assertion, and his most alarming number, ninety percent enrichment, is contradicted by the IAEA's published ledgers, which say sixty. The piece lays out the argument as he made it, then holds it against the public record.
In July 2025 the director of OSTP sat down at CSIS and articulated, with unusual clarity, why the United States will not converge with the EU AI Act: not because the crosswalk is technically hard, but because Washington rejects horizontal regulation as a category. A year later the doctrine has an executive order, a litigation task force, and a legislative blueprint — and, strangely, the NIST AI Risk Management Framework keeps winning anyway. A reading on what political non-convergence means for a thesis about technical interoperability.
NeOn is the methodological spine of a thesis that formalizes the EU AI Act and the NIST framework as structured knowledge. This reading traces what the method actually asks — reuse over creation — and why that doctrine now sits between the models that automate it and the tools that audit it.
Every serious comparison of American and European AI governance opens with a crosswalk: a table setting the NIST Risk Management Framework's four functions beside the EU AI Act's articles. The mapping mostly works — the same evidence fills both columns — which is exactly why the places it fails are easy to miss. This reading walks the table down to the rows where the American column goes blank, and argues that the blanks are the point.
A single algorithm sorts almost everyone in U.S. federal prison into a risk category that helps decide when they go home. It is audited by the GAO, revalidated by the Justice Department's own researchers, and challenged in a handful of lawsuits — and still there is almost no built-in way for the person scored to contest the score. This reading uses PATTERN as the worked example of a thesis-level claim: that documented compliance and contestable compliance are different things, and that the distance between them is exactly what a high-risk regulatory regime is built to close.
Method, language, and tool are three different things, and confusing them is the most common mistake in ontology work. This reading sorts the landscape of ontology engineering methodologies, explains how the main ones actually work, and shows why a regulatory-comparison thesis should reuse rather than invent — and which method makes reuse the whole point.
The XAI field thinks producing a decision tree or a counterfactual is the end of explanation. Two leading ontologists say it is only the beginning — and the confusion is causing real problems for data integration, AI transparency, and the compliance frameworks that depend on both.
A follow-up to N° 10. Karaganov, now joined by Mearsheimer, sharpens his case for using a small number of nuclear weapons against European targets to coerce the West into backing down — and reveals, in passing, a position on Germany that Mearsheimer correctly identifies as essentialism with a longer pedigree. A guided reading of where the two analysts agree, where they break apart, and what each side keeps invisible.
The Arctic corridor is open four and a half months a year, carries nearly 40 million tonnes of cargo, and is gaining new port investors, shipping partners, and energy customers at a pace that has little to do with geopolitics and everything to do with cost.
Bond yields, a $1.9 trillion deficit, and a savings rate at 2.6%. David Lin walks through the mechanics of how America's debt problem is quietly becoming everyone's cost-of-living problem — regardless of whether the Strait of Hormuz stays open or not.
Two Marxist economists — one in New York, one in Athens — work through the financial plumbing of US empire: the petrodollar, the recycling mechanism, Scott Bessent's panicked swap line, and why the war with Iran is unlike any crisis since 1945. A guided reading of a wide conversation.
Art Berman breaks down why the US imports 6.5 million barrels a day despite being called energy-independent, why the two oil prices on your screen describe completely different realities, and why the consensus among every credible energy analyst is that nothing is going back to normal.
Petroleum geologist Art Berman explains why the US can't simply replace Persian Gulf oil with its own, why the strait cannot reopen quickly by any realistic measure, and what the futures market's $89 July contract actually means when spot oil trades at $140. The answer to each question is the same: the administration's public claims are false.
The Brussels Effect framework asks whether a jurisdiction can globalise its standards unilaterally through market force. Applying Bradford's five conditions to the EU AI Act yields a mixed verdict — and a precise diagnosis of why interoperability between the AI Act and the NIST RMF is structurally harder than it looks.
Michael Hudson argues that the Iran war is not a geopolitical accident but the deliberate culmination of a US strategy to control the world's oil supply — and that the resulting financial shock, layered on top of already over-leveraged economies, will be irreversible.
David French, Jamelle Bouie, and Michelle Cottle dissect how Trump turned a lawsuit against his own IRS into a $1.776 billion patronage fund, immunized his family from tax enforcement, and handed the Republican Party an anchor headed into the midterms.
Trump v. IRS was a legal fiction: Trump sued his own agency, his DOJ defended it, and no genuine adversary existed. When the case collapsed, what replaced it was $1.776 billion with zero judicial oversight and a sweeping release of civil liability for Trump and his family. David French and Jamelle Bouie explain why it may be the most corrupt act of either Trump term — and why the administration cannot see the political cost coming.
On May 20, Putin and Xi signed a four-principle joint declaration in Beijing. Pepe Escobar reads it alongside Judge Napolitano. This piece works through the document itself — its lineage, its architecture, and what it means that it names neither the United States nor Iran nor Ukraine.
Russia's top diplomat at the OSCE argues that the margin between indirect and direct NATO-Russia war has dissolved into daily events. Latvia's drone crisis, the Baltic Sea, and a Russian expert community where tactical nuclear use is no longer taboo.
Trita Parsi argues the war the US fought against Iran in June was a strategic defeat already delivered — every escalation triggered the next failure rather than resolving the last. The second round, should it come, would target US energy infrastructure in the Gulf, undersea fiber cables, and businesses connected to Trump himself. A reading on what the war cost and what a return would cost more.
Anu Bradford's 2012 Northwestern University Law Review article identified the precise conditions under which a single jurisdiction can set global standards without treaties, coercion, or consent. The framework has become the dominant lens for analysing EU regulatory power — including, now, in the architecture of the AI Act. A reading on market power, regulatory capacity, the mechanics of nondivisibility, the politics of EU motivation, and the structural limits that constrain the Effect.
Anu Bradford's book-length argument: five conditions, two mechanisms (de facto, de jure), four case studies — competition, digital economy, consumer health, environment — and an assessment of whether the EU's regulatory hegemony will outlive its economic decline. A reading on the framework the AI Act now inherits.
The US pours hundreds of billions into the pursuit of artificial superintelligence. China builds delivery robots, humanoid factory workers, and cheap open-source models that half the world downloads for free. Kyle Chan maps the divergence and argues the American obsession with AGI may be obscuring the competition that actually matters.
Mercouris and Diesen map the legitimacy crisis spreading across Europe — from Starmer's Britain to Merz's Germany — and argue that the continent's foreign policy obsession with Russia is accelerating its own decline. A reading on rigidity, anger, and the closing window for self-correction.
After Trump's two-day summit in Beijing, the Marxist economist Richard Wolff sits with Glenn Diesen and treats the meeting as a transition marker. China, he argues, has won the developmental race the American model was supposed to win — and the Iran misadventure has made the new alignment visible.
Richard Wolff returns to a different host after the second Trump–Xi summit and offers two analytical moves: the 'hybrid' that names what China actually is, and the 'reaction formation' that names what the empire is doing instead of what it cannot do.
The companion piece to the empire-decline reading. Wolff is interested in what fills the conceptual space the Washington Consensus used to fill — and in a third option, multilateral and unrehearsed, that he says is being discussed in Beijing and not in Washington.
Colonel Lawrence Wilkerson, former chief of staff to Colin Powell, dissects the Beijing summit: what the choreography revealed, why China owned the meeting, and what the Kagan mea culpa tells us about neoconservative reckoning with imperial decline.
A 341-page open-access textbook from Oxford turns out to be a foundational argument for anyone studying transatlantic regulatory comparison. Hildebrandt claims that legal protection is contingent on the affordances of text, and that the move into computational infrastructures opens two distinct design paths — only one of which preserves what makes law protective in the first place. The distinction reads as a diagnostic for everything that comes after, including the AI Act and the NIST AI Risk Management Framework.
A 90-page article in the Columbia Law Review maps four ways a right to contest AI can be designed, finds that every major instance fails differently, and argues the United States has missed something the GDPR, Brazil, Canada, and the OECD have all reached for. Four archetypes, four case studies, one diagnostic for American algorithmic law.
You can contest a wrong credit card charge and win 90% of the time, for free. You cannot contest an algorithm that denies you a kidney transplant. Kaminski and Urban map the gap, build the case from due process theory, and propose four archetypes for a right the United States still lacks.
The Ghent clinical psychologist whose mass formation theory went viral in 2022 returns to Glenn Diesen's podcast to argue the Soviet Union and Nazi Germany were not classical dictatorships at all, and that the version Hannah Arendt warned of in 1953 — led by bureaucrats and technocrats — is now arriving. A guided reading of his core argument and its quieter contestants.
Pepe Escobar argues a new triangle of Russia-China-Iran is rewriting Eurasian power, that Hormuz is the choke point Washington has lost the ability to dictate, and that the US-Israel axis is cornered, not ascendant.
A January 9 memorandum from the Attorney General establishes a Justice Department task force whose sole purpose is to challenge state AI laws. The text runs three pages and is, in operative terms, a transcription of the executive order it implements. A guided reading of what the brevity does — and what the institutional commitment signals.
Sergey Karaganov, in his latest conversation with Glenn Diesen, argues Russia is inside a fourth world war and must climb the nuclear escalation ladder against European elites. A guided reading of what he wants, what the Kremlin has actually adopted, and where the limits of his influence lie.
A University of Chicago strategist who has advised every White House since 2001 explains why the war on Iran is following the playbook he modeled for two decades. Tactical wins, in his account, are not strategic wins — and the gap between them is what closes the trap.
An executive order to dismantle state AI regulation that explicitly defers the framework it claims to provide. A guided reading of what the order actually directs, what its mechanisms can and cannot do, and the distance between the announcement and the architecture.
The Marxist economist reads the post-Hormuz oil shortage as more than an oil story. Forty years of just-in-time production are giving way to just-in-case storage; the 1970s petrodollar bargain has been hollowed out by missile range; and an American empire that cannot tax its way through a shock has run out of room to borrow.
Brussels in Denver: Colorado's AI Act faces a federal preemption challenge from xAI, with the DOJ stepping in. A guided reading of what's actually being decided and why it matters beyond one state.
A Carnegie Europe fellow argues the EU's deregulatory turn on AI is solving the wrong problem. The real obstacles — underfunding, fragmented markets, foreign infrastructure dependence — don't respond to lighter rules. A close reading for anyone tracking the Brussels Effect in retreat.
The FBI director hands out engraved bottles of Woodford Reserve, transports cases on DOJ planes, and threatens to polygraph staff over a missing one. Sarah Fitzpatrick's follow-up reads as a study in the personalization of an institution whose authority depends on impersonality.
The director of Harvard's Carr-Ryan Center argues that AI governance needs three things at once — governance reach, technological power, and rights commitment — and that no major regime has all three. A guided reading of the trilemma, the gaps, and the closing window.
The Financial Times unpacks why the UAE really left OPEC after fifty-eight years — quotas that no longer fit, a Saudi partnership fractured over Yemen, and the 400-kilometer pipeline that made walking away possible.
The EU's AI Act trilogue is stalling on general-purpose AI obligations as the Digital Omnibus reopens settled questions. A guided reading of where Brussels is and what it means for who writes the rules.
Donald Trump arrives in Beijing on 14 May for a summit twice postponed by an Iran war he did not plan. While Washington depleted Pacific munitions and dispatched the Vice President to Islamabad, Xi quietly built the leverage he now intends to charge for.
The founder of the world's largest hedge fund argues the United States is moving through a 500-year empire-cycle script — peaking, indebted, dollar-strained, and culturally divided. A guided reading of his case.
No pieces under that topic yet.