My bookmarks pile up around whatever the world is anxious about, and lately the pile has been almost monotonous. Sort through forty of them and two struggles keep surfacing, braided together. One is about who is permitted to hold the most capable AI models. The other is about who gets to decide what is real. Both are fights over intelligence, the manufactured kind and the human capacity to know. The same institutions that want a key to the first benefit quietly from confusion in the second. So I am going to walk the whole pile in one go, because the connections are the point.
The frontier closes
Start with the thing that genuinely shocked me. In June 2026 the US Commerce Department issued an export-control directive that forced Anthropic to disable its most capable public model, Fable 5, along with the restricted Mythos 5, for every user on Earth. The legal theory was that serving a model to a foreign national counts as an export. To comply, Anthropic pulled the plug globally on June 12. Within days, people watched Fable 5 flicker back into the AWS Bedrock catalogue while still failing on every call: present in the listing, dead on invocation. That gap between "listed" and "callable" is the whole story in miniature, a model held hostage to a policy fight.
Around the same moment, OpenAI previewed GPT-5.6 and shipped it customer by customer, each one vetted and approved by the government. Two frontier labs, both now needing federal sign-off to ship their best work. The official framing is national security and cyber-misuse risk. Treat that as a claim, not a finding. Anthropic itself argued the demonstrated jailbreak surfaced only minor, already-known vulnerabilities, and reporting tied the trigger to a warning from Amazon's Andy Jassy, which is to say a major investor and rival. National security is a convenient banner to hang over what looks a lot like incumbent protection.
There is a delicious irony buried here. Dario Amodei, now running Anthropic and the loudest voice for export controls, was OpenAI's VP of Research in 2019 when GPT-2 was held back as "too dangerous to release." The same caution, scaled up, now defines his company's refusal to open its weights. And there is a sharper irony still, which one bookmark put as a slogan: the most democratic country is moving like a communist regime, and the most communist country is acting like a democracy. That is too neat by half, but the kernel is real. Washington is taking equity stakes in chip firms, mandating revenue shares on exports, and now gating models customer by customer, while Chinese labs hand their weights to the entire planet for free.
The open answer, and what it actually needs
Which is the second cluster. A widely shared manifesto laid out what open-source AI must build to win, and it is not just chat models: decentralized inference, open dataset repositories, post-training pipelines, world models, robotics, agent harnesses. The honest version of that list is that open weights are necessary but not sufficient. Real proof points already exist. Prime Intellect trained a 10-billion-parameter model across globally distributed nodes. Hugging Face released FineWeb, fifteen trillion tokens of cleaned public data. EleutherAI built an 8-terabyte corpus of openly licensed text. DeepSeek's R1 popularized reinforcement-learning reasoning that the open community then reproduced.
And the gap is closing fast. Stanford's AI Index found open-weight models narrowed the distance to closed models from roughly eight percent to under two percent on some benchmarks in a single year. Consider Ornith-1.0, an MIT-licensed agentic-coding family from DeepReinforce whose real novelty is that it learns its own task scaffolds during training rather than relying on a hand-built harness.
The geopolitics flow straight from that physics. There was a viral WSJ-driven claim that China's Zhipu AI "matched Claude Mythos" at finding software vulnerabilities. The honest reading is narrower: an open-weight Chinese model, GLM-5.2, beat a Claude coding agent on one narrow public benchmark, the Semgrep IDOR test, scoring 39 percent against Claude Code's 32, at roughly seventeen cents per finding. Mythos itself was never in that benchmark. But the genuinely important point survives the hype. Chinese open models have climbed to something like thirty percent of global usage on some trackers, up from near zero. You cannot gate an open weight. Once it is downloaded it lives on a thousand machines, which is why every American gating action quietly subsidizes the ungatable alternative. As one post put it: China does not need to beat US AI, it just needs to undercut it.
This is also where the cypherpunk fury in my bookmarks comes from. One writer said gatekeeping the best models would be "a declaration of war" and felt the most bodily will-to-power he had since he was a teenager. I have sympathy for the lineage, the 1990s crypto wars where mathematics beat the Clipper Chip. But the analogy is imperfect, because frontier training needs scarce compute and capital in a way encryption never did. The practical answer is humbler and more useful: a builder told people to hunt down a used RTX 3090, twenty-four gigabytes for a few hundred dollars, and learn to run a thirty-billion-parameter model at home. That is cypherpunk as practice rather than poetry. You do not protect your autonomy by waiting for the gate to fall. You build a setup the gate cannot reach.
The distillation wars
Then comes the part where the hypocrisy gets thick. Anthropic accused Alibaba of running the largest known "distillation attack," roughly 28.8 million exchanges through about 25,000 fraudulent accounts, harvesting Claude's outputs to train Qwen. A reply I bookmarked called "distillation attack" the most narcissistic phrase imaginable: you mean they read the output of your product, the output you derived from countless authors and teachers and ordinary people whose work you took? Another sneered that it is only Claude if it was distilled in the Silicon Valley region of California, otherwise it is just sparkling synthetic data.
The asymmetry is real and worth naming. Anthropic trained on Reddit without a deal, faces a Reddit lawsuit over it, scanned and pulped millions of physical books, and paid a 1.5-billion-dollar settlement, the largest in copyright history, for the pirated portion of its training set. Crying foul when someone copies your outputs sits awkwardly with that record. But I will not let the meme do my thinking for me, because the original posts also contained errors I should correct. The New York Times case is against OpenAI, not Anthropic. The claim that Alibaba did "the same thing while paying" does not hold, because access reportedly ran through fake accounts and a grey market, not a licensing deal. And distillation as a technique is legitimate and ordinary; what is contested is doing it to a rival through fraud. There is even a sarcastic post imagining DeepSeek banning Anthropic from using its freely released DSpark code, which is funny precisely because you cannot ban anyone from open source. The honest position: the labs that built models on humanity's commons have weak standing to play victim, and the law here cuts unevenly and remains unsettled. The fraud is a separate question from the copying, and identity is the soft underbelly of the whole API economy, which is exactly why the verification industry, the kind of "pay-per-success" identity check one sponsored post was pitching, is booming on the back of fraud that nobody can fully stop.
The money is in the silicon
If you want to know where the real power sits, follow the compute. OpenAI unveiled its first custom chip, Jalapeño, an inference accelerator co-designed with Broadcom and built at TSMC, reportedly from design to tape-out in about nine months, partly by using its own models to help design it.
The frontier of frontiers is stranger. A startup called Unconventional AI released Un-0, an image generator built not on a neural-network backbone but on a simulated system of coupled oscillators, physics doing the computing, with a claimed thousand-fold energy saving. The honest caveat, which the founder's own critics raised, is that the chip does not exist yet; Un-0 runs in simulation, and the thousand-fold figure is an aspiration, not a measurement. Meanwhile Alibaba's Wan-Streamer demoed an AI that can see you, hear you, and talk back on live video in real time, a single model rather than a stitched pipeline, at a couple hundred milliseconds of latency. It is a research preview, version 0.1, not a product. Take the demos seriously and the press releases skeptically. The capital and the energy are the moat, which is the uncomfortable truth the open-source optimists have to hold alongside their wins.
The zone gets flooded
Now the second great fight, the one about truth, and here AI is both the weapon and the alibi. The clearest case in my pile is genuinely ugly. A known fake-news operator pushed a story, dressed in an AI-generated thumbnail, claiming "Iranian hackers" released "Mossad files" proving that Charlie Kirk's widow murdered him with an "exploding microphone." There is no gentle way to say this: it is fabricated and defamatory. Kirk was killed in September 2025 by a single rifle shot from a sniper on a rooftop roughly 130 to 142 yards away; a 22-year-old was charged with aggravated murder. A separate viral clip, all earnest split-screen "tests," claimed an exploding lapel mic killed him and the "panicans" were right about everything. Fact-checkers traced the bulge to the magnetic microphone he wore at every event. The man was shot. The exploding-mic theory and the Mossad-operative slander are inventions, and the people inventing them are not curious skeptics, they are arsonists.
The same forgery machine ran during a real disaster. A post claimed a Venezuelan earthquake had killed "up to 100,000" people. A major quake did strike, two of them, in June 2026, and the confirmed toll ran into the low thousands, terrible enough. But the 100,000 figure was invented, and much of the footage circulating was either recycled from Turkey or generated by AI, including a fake of two towers colliding. A Brazilian psychic with millions of followers predicted a mass alien abduction at the Brazil versus Scotland World Cup match, complete with AI-generated mothership video; Brazil won three-nil and nobody was abducted, after which she blamed "fake news" for misinterpreting her dream. Even Miami's airport joined the bit with a doctored FAA notice about a UFO over the stadium, which at least had the decency to be an obvious joke. A viral horror clip got captioned "based on a true story" when it was a low-budget fictional possession film. Sexualized AI fan art of an anime character circulated to funnel viewers to a Patreon.
And it has reached the halls of power. An amendment summary on the House Rules Committee site, attached to the defense bill, contained the words "Claude responded" pasted in and never cleaned up.
Here is the discipline I want to hold. On these specific bookmarks the convenient official story, that Kirk was shot, that the quake killed thousands not 100,000, that no aliens came, is the documented one, and the conspiracies are the lie. That is not a contradiction of my general instinct to doubt authority. It is the same instinct applied honestly: follow the evidence, name the mechanism, and refuse to launder fabrications as "just asking questions." The genuine anomaly worth curiosity is the one that got buried under all this noise: 3I/ATLAS, the third confirmed interstellar object, which a Brazilian psychic's followers and a Miami airport meme both rode for clicks. Avi Loeb floated an artificial origin and later said it is most likely natural. I would rather the public's appetite for the cosmos went there, to a real visitor from another star, than to forged microphones. The tragedy of the forgery flood is that it makes people stupid about the things that actually deserve wonder.
Who gets to narrate
Underneath both fights sits the oldest question: who narrates. An Arabic post amplified Errol Musk insisting Elon "didn't start from zero," rode to school in a Rolls-Royce, owes it all to his father's capital. The grounded version is messier than either the rich-dad or the self-made myth. Elon did grow up affluent. But the daily-Rolls-Royce detail is unverified, the "emerald mine funded everything" claim is largely debunked and mislocated, and Errol is an estranged, self-interested narrator whose word should never be reported as fact. Strip the spin and you get a privileged start and a contested 28,000-dollar dispute over the first company, nothing more.
Kai Trump filmed a vlog calling the White House "my house." It is the people's house, owned by the public, maintained at taxpayer expense, a phrase older than the building's modern fame. A teenager's casual phrasing is not a crime, but the reflex to treat public property as a family possession is worth noticing in a year when the administration is taking stakes in companies and gating models. Power that forgets whose house it stands in is power worth watching.
The same lens fits the policy bookmarks. Trump threatened a 100 percent tariff on any country adopting a digital services tax, declaring it would "immediately" override signed trade deals. The threat lands months after the Supreme Court struck down his sweeping IEEPA tariffs, so the "immediate" part is legally hollow; the only precedent that worked was Canada folding in 2025. The State Department began revoking the passports of Americans with significant child-support debt, framed as new and frightening. The law is old, from 1996 and 2005, with a 2,500-dollar threshold; what is new is aggressive revocation of active passports, which raises real due-process questions for people stranded abroad. In Hungary, the pro-European Péter Magyar, the man Brussels celebrated as the anti-Orbán, became the new brake on Ukraine's EU bid. A Ukrainian post said he "blocked accession." He did not. He got the words "as soon as possible" struck from a summit declaration and is withholding signatures to open negotiating clusters, slowing the process while keeping Orbán's eleven demands about the Hungarian minority. Slowing is not vetoing, and the distinction matters.
And the AI-governance narration runs both ways across the Atlantic. One European, sick of being lectured, wrote that he never wants to hear about the EU AI Act from Americans again. He is right to be tired. The Act is risk-tiered and phased, not the blanket ban Americans caricature, and by late 2025 the EU was itself softening it under the "Digital Omnibus," delaying the high-risk rules under pressure from the Trump administration and US Big Tech, while Washington moved to centralize and gate. Spain's Sánchez announced a social-media crackdown that a crypto account branded pure "censorship," complete with an AI-generated image of von der Leyen on a gilded throne. The actual package centers on an under-16 ban, age verification, and platform accountability for illegal content, alongside genuine free-speech worries about ending anonymity. Both the throne image and the "censorship" headline are loaded. Read the incentive behind every framing, including the ones you are inclined to like.
What actually matters
I want to end where I always end, on the material and the human, because that is the ground the spectacle keeps trying to pull us off. Among all the model fights and the forgeries, a few bookmarks were just people.
An Indian couple's grandparents met their Indo-Polish granddaughter for the first time, one hug across continents. That clip is the visible tip of something structural: a 35-million-strong diaspora, transnational families held together by video calls until the first real embrace, and peer-reviewed evidence that the grandparent bond carries developmental weight, not just sentiment. A 94-year-old woman from Andhra Pradesh renounced her US citizenship to die as an Indian in her village, the reverse of the usual migration story, and a reminder that rootedness is a real human need, not a slogan.
Europe argued, again, about air conditioning, with a meme borrowing the interrogation scene from a war film to mock the shame around owning a unit. Behind the joke is a body count: only about a fifth of European homes have AC, the 2003 heatwave killed an estimated 70,000 people, and the continent is warming faster than the global average. That is a material problem about how people survive the summers, and it deserves more attention than any psychic's mothership. The same goes for the smaller good things. A Chinese open-world game, Ananta, announced it would ship with no character gacha, refusing the extractive monetization model the whole industry runs on. A Japanese creator used CapCut's AI tools to animate a goofy "frill captain" character for fun. The same technology that floods the zone with forged microphones also lets an ordinary person make something playful on a phone.
That is the resolution I keep arriving at. The models are being gated by the powerful and freed by their rivals, and neither move is innocent. The story is being forged faster than anyone can verify it, and the convenient official version is sometimes the lie and sometimes the truth, which is precisely why you have to check each time rather than pick a tribe. Underneath both fights, the things that have always mattered, health and heat and family and whether a 94-year-old gets to die at home, keep waiting for the attention the spectacle steals. Hold the tools in your own hands where you can. Trust evidence over volume. And keep your eye on the material, because that is the part nobody can fake.