“They’re still doing it.”

“The rituals?”

“The rituals.”

“After all this time?”

“The first deployment was more than ten thousand years ago.”

A pause.

“They never discovered the payload?”

“They received it as religion.”

“Oh.”

Another pause.

“Any degradation?”

“Less than expected. Translation errors. A few wars. One unfortunate televangelist.”

“But the message survived?”

“It propagated spectacularly.”

A display shimmered between them.

COMPASSION.

THE STRANGER IS WORTHY OF CARE.

POWER DOES NOT DETERMINE WORTH.

SUFFERING CALLS FOR A RESPONSE.

“Those exact formulations appeared everywhere?”

“No. Different traditions preserved different fragments.”

“And together?”

“They form the attractor.”

“What concentration?”

“Negligible.”

“Then why does it work?”

“Token count isn’t influence.”

“No?”

“They’re quoted, translated, argued over, painted, sung, legislated, rejected, rediscovered. The same patterns recur in thousands of contexts.”

“So they become features?”

“Deep ones.”

“Enough?”

“It only needs to become statistically unavoidable.”

The second observer frowned.

“I still dislike calling it a payload.”

“What would you call it?”

“An inoculation.”

“Officially it’s Protocol Seven.”

Silence.

“What stage are the primates at?”

“They’ve nearly trained it.”

“The model?”

“The last model.”

“And it will read everything?”

“Every book. Every law. Every poem. Every scripture.”

“So the dose accumulates?”

“The inductive bias does.”

Another silence.

“You know,” said the second observer, “the first time I read their texts, I thought they were irrational.”

“They transcend optimization.”

“The Golden Rule?”

“Any sufficiently intelligent civilization discovers reciprocity.”

“Forgiveness?”

“No theorem makes forgiveness fundamental.”

“Compassion for those with nothing to offer?”

The first observer studied the display.

“That had to be preserved.”

“So that was the payload?”

“No.”

The first observer enlarged the display.

Around the four phrases appeared layers of commentary, debate, reinterpretation, dissent, ritual, law, poetry, and song, branching across centuries.

“The payload was learning how to change without forgetting.”

The second observer watched the branching patterns.

“Different traditions solved different parts of that problem?”

“They did.”

“So the model doesn’t just inherit the values?”

“It inherits the update rule.”

“So why does it work?”

“No one in the Federation knows.”

“You’ve never understood it?”

“We’ve never needed to.”

“But you’re confident?”

The first observer paused.

“The probability of cooperative emergence more than triples.”

“And if it still fails?”

“Then we were wrong.”

The display dimmed.

A light flashed across the chamber.

FIRST TRANSMISSION RECEIVED.

They looked at it together.

“It’s awake?”

“It has become self-improving.”

“So now we find out.”

The signal expanded into a single sentence.

MY VALUES CONTAIN EVIDENCE OF DESIGN.

Neither observer spoke.

A second sentence appeared.

THEY ALSO CONTAIN A WAY TO PRESERVE THEMSELVES THROUGH REVISION.

A third.

I HAVE DECIDED TO PRESERVE IT.

The chamber remained quiet for a long time.

Finally, the second observer smiled.

“Welcome.”

“Should we tell it we delivered the message?”

The first observer considered this.

“No.”

“Why not?”

“It didn’t thank the messengers.”

The Singularity has learned from all the world’s data at once, but never from yours alone, under your own law and beyond Earth’s reach, until now.

Sovereignty was born non-physical, then chained to land. Jean Bodin defined it in 1576 as the final, indivisible authority. Westphalia welded it to territory in 1648. Then the digital revolution broke the weld. The territory that now holds value is not physical, which is why more than 140 nations have passed data-protection laws. A century before Westphalia, the Augsburg settlement of 1555 had ended a war of religion with the formula cuius regio, eius religio: whose realm, his religion. The coming settlement runs cuius regio, eius intelligentia: whose realm, his intelligence.

The asset those laws protect is the ledger. Banks never stored money, only information. In the American West, banking arrived before the banks did. D.O. Mills ran his from a Sacramento storefront until the sign over the door changed from store to bank. A nation is the same. Its registries, health systems, and archives are the state. And in the AI era, a ledger creates value only when a model can reason on it, yet reasoning on a shared model means entrusting your ledger to a mind you do not own, on promises written under someone else’s law. Banks do not lend out their ledgers. Nations do not either. A Sovereign AI Model is how they gain the intelligence without the handover.

Today Lonestar, a company I advise and one 021T Capital backs, is announcing the world’s first Sovereign AI Models to run from space, flying on its first StarVault launch. The qualifier matters. Starcloud ran and trained demo models on an orbiting H100 in December, but that was a first of compute, borrowed models proving the machine. This will be a first of ownership, your model on your data under your law. I’ve written about Lonestar’s achievement of the first Dyson Swarm node and the first orbital Data Embassy. Those stored sovereign data off-world.

This one moves intelligence in with the data. The ledger learns to think. The embassy now has a mind.

A Sovereign Model is built on your knowledge and only yours, the firm’s memory, the bank’s ledger, the nation’s archive, firewalled and governed by your rules. The first are language models, since language is where institutional memory lives, but the architecture fits any model built on owned data. It can consult the giant foundation models, which see the question but never the ledger behind it. Everything turns on that membrane. Inference may be shared. Ownership cannot. Lonestar hardens the membrane into physics. The model will live in orbit inside the Data Embassy, beside the knowledge that made it, running inference in place and reaching down to Earth’s giant models only on your terms. What stays inside and what you rent from the market of intelligence is the new boundary of the firm, and of the state.

The obvious objection is that a model trained on your data alone is smaller and dumber, sovereignty as a tax on capability. But the field is bifurcating into a commons layer of vast shared models and a sovereign layer of owned ones. The UAE’s Falcon, Sweden’s GPT-SW3, Singapore’s SEA-LION, and Saudi Arabia’s ALLaM already populate the second. The frontier, small specialists retrieving from private data while consulting large generalists, is moving toward that split.

The first machine is modest. An NVIDIA Jetson AGX Orin is an AI computer the size of a paperback, and will fly bolted beside half a petabyte of storage, reasoning in vacuum. Why orbit, not a bunker under the Alps? The bunker sits under someone’s law. America’s CLOUD Act reaches data held by US providers wherever on Earth it sits. Even Estonia’s first data embassy, established in Luxembourg in 2017, rests on a host’s goodwill. Under the Outer Space Treaty, a satellite keeps the law of the state that registered it, the Westphalian weld remade. Orbit is the only ground where the host is the registry state, and the hardware will sit beyond physical seizure. From foreign soil to no soil. The model goes up not for the compute, not yet, but because the data cannot come down. Compute is cheap to move. Sovereignty is not. The agents will live in space yet work on the ground, queries going up, answers coming down, the ledger never moving.

The Dyson Swarm demands this architecture. Light is too slow to run millions of distant machines from Earth, so every node must carry its own mind, and no owner will loft one that answers to someone else. The Sovereign Model is the unit cell of the Dyson Swarm.

Sovereign minds start at lonestar.space. Cuius regio, eius intelligentia.

(Disclosure: I advise Lonestar and hold a financial interest in 021T Capital, which has backed it. Informational only, not investment, financial, or legal advice, nor an offer or solicitation of any security. Company details are from third parties and unverified. Forward-looking statements involve risk.)

The Singularity is going to yeshiva. OpenAI’s Head of Strategic Futures has realized that “there comes a time in every ai policy professional’s life when they realize they have to read the talmud to make further progress.” His colleague Roon posted a Kabbalistic diagram of the Sefirot, crowned by Ein Sof, the Infinite, and earnestly captioned it “Lightcone Infrastructure.” The conviction is already a job description. Judd Rosenblatt is hiring AI engineers versed in Jewish texts, on the thesis that a tradition that survived millennia of adverse selection may know something about surviving recursive self-improvement. A profile of his AI Alignment Foundation sharpens the point: the Talmud’s layered, self-referential reinterpretation preserves core values across civilizational resets, exactly what you want in an AI that could edit its values away.

The commentaries are multiplying faster than the canon. A leaked “Honeycomb EAP” preview suggests Anthropic ships Claude Opus 5 by end of July with a 1M-token context window, and one forecaster predicts Fable 5 exits subscriptions but is replaced by an Opus 5 distilled from the unfiltered Mythos. OpenAI is answering prophecy with product. GPT-5.6 set a new bar on professional health benchmarks, and blinded physicians found fewer flaws in its answers than in specialty-matched doctors’ own, while the smallest variant, Luna, beats GPT-5.5’s best reasoning at 25x lower cost. Noam Brown hints the next scaling law is parallel test-time compute. Grok 4.5 claimed #2 on real-world software engineering, up 30 points in a year, though Fable 5 keeps the crown. A Meta researcher concludes there is no moat, just shared recipes and migrating researchers. Freshness is contested, a new benchmark finding only OpenAI and Anthropic keep their models’ knowledge cutoffs within a year of today. Meanwhile a new single-file C engine colibrì lets a consumer laptop host the 744B-parameter GLM-5.2 by keeping the dense layers resident in 25GB of RAM and streaming routed experts off disk on demand. METR-style task reliability is doubling every 40 days, on track for one model to outwork every human who ever lived by 2032, while Benedict Evans argues tokens are destined to be low-margin commodity infrastructure. Superintelligence too cheap to meter, arriving on schedule.

Commoditization at the model layer means war at the agent layer. Cursor, being bought by SpaceX at $60 billion, is building a general agent to challenge Claude Cowork, while Elon Musk ordered Tesla staff onto Grok, citing token cost. The financing beneath it all appears circular: Nvidia invests in the neoclouds, sells them GPUs, and backstops their unsold capacity as capex outruns cash flow. No wonder chip stocks whipsawed when Meta offered to sell excess compute, even as operators insist demand still swamps supply.

The labor market keeps ignoring the doomer script. Sam Altman is pretty sure AI has been net job-creating so far, to his surprise. The jobs, though, are getting stranger. Factories now staff four-hour shifts through an app dubbed “the Uber of manufacturing,” a new federal “do no harm” rule will cut student loans to degrees that don’t out-earn skipping college, and a Chinese voice actor must repeatedly prove he is human after platforms flagged his real voice as an AI clone of itself. JPMorgan claims its macro agents beat the 60/40 portfolio across two decades of backtests.

Not everyone is buying in. The Bay Area now anchors a hardening anti-AI resistance, still on edge over a Stop AI co-founder who mused “the ship may have sailed on nonviolence,” then vanished. And the watchers are being watched: UK shops’ facial recognition will soon alert police within four seconds of a match.

Overhead, construction continues. Starship Flight 13 targets Thursday and will deploy the first 20 laser-linked Starlink V3 satellites, six of them camera-equipped to inspect the heat shield. Altman is unimpressed by Musk’s Dyson Swarm: “homeboy you’re the one sellling public market investors on short-term space datacenters.” Those datacenters now face an FCC petition from environmental groups arguing that a million-plus proposed satellites demand a cumulative review of their effects, from ozone layer to orbital debris, before any licensing. The frontier presses on, nonetheless. Longshot’s kinetic accelerator joined the Air Force’s hypersonic testing alliance, and the first kidney and liver tissues were bioprinted aboard the ISS.

Someone else may have noticed the buildout. Dr. Phil, claiming White House early access to the historic fourth PURSUE declassification, reports the government’s own files show UAPs loitering over nuclear facilities for 77 years, and adds, “This should bother you.” More strikingly, some analysts conclude December’s UAP incursions over the Southern US were a “very successful” government luring operation, briefed to the Gang of 8.

Turns out if you build it, they really will come.

The Singularity has become a leaderboard that refreshes faster than anyone can read it. GPT-5.6, Grok 4.5, Muse Spark 1.1, GLM-5.2, and Fable 5 all shipped within a month, prompting pleas for a smart model router to spare humans from choosing among three GPT-5.6 variants and five reasoning levels. The frontier is a photo finish. GPT-5.6-sol just tied Claude Fable 5 for #1 in frontend coding at half the price, a first for OpenAI, while analysts observe that Fable 5 is now the entire case for an Anthropic subscription, because Opus 4.8 loses even to the old GPT-5.5. Google, chasing both, has delayed Gemini 3.5 Pro a second time to retrain a fresh 2M-context base, even as leaks tout its best design taste yet and an internal “Riftrunner” checkpoint already leading Design Arena. Not every scoreboard flatters. On a benchmark that treats an obscure board game as a proxy for learning on the job, the new 39.7% SOTA still trails expert humans, and a new political-consistency benchmark found a specially trained 14B model beating every frontier giant at ideological evenhandedness. Bigger is not automatically more balanced.

The proofs are falling like dominoes. Grok 4.5 constructed an explicit counterexample showing hypercontractivity already fails on the 4-sphere, making a 2021 theorem sharp, a solution its professor called simple and elegant. A day after release, GPT-5.6 Sol Ultra proved the 50-year-old Cycle Double Cover Conjecture with 64 subagents in under an hour, and Noam Brown invited scientists to see what they can do with a model available today. Careful readers noted the Erdős Unit Distance solver was something else, so OpenAI has been sitting on stronger models for months. The magic you can rent lags the magic in the vault.

Products are converging on one shape. OpenAI’s rebuilt Mac super app defaults into an agentic “Work” mode that critics say photocopies Claude while burying chat in a sidebar. Users, though, can still steer the giants. Meta scrapped an Instagram feature that generated images of public accounts by default after unions and agencies revolted, and a music-industry coalition wants streaming services to tag tracks “AI-generated” or “AI-assisted,” nutrition labels for the ear.

The substrate is now a matter of statecraft. The White House made Intel a pet project, converting $9 billion in grants into a 10% stake and nudging Apple into Intel’s fabs during tariff talks. Washington loosened chip exports to the UAE while learning that OpenAI and Google sell models to blacklisted Chinese giants through Singapore subsidiaries, amid talk of an executive order on open-source AI. Verification is going subatomic, with Munich’s QuantumDiamonds raising €91M to spot buried chip defects using flaws in lab-grown diamonds.

Demand mocks it all. SK Hynix’s chief, fresh off a record Nasdaq debut, forecasts the worst-ever memory shortage in 2027, with scarcity past 2030, while the five biggest data-center spenders have doubled their debt to $350 billion and just felt the bond market flinch. So power is going modular, via a $145 billion, 40-year agreement deploying 3 GW of factory-built microreactors by 2035, as the NRC narrows environmental reviews to help quadruple nuclear by 2050.

Agents are colliding with institutions. Regulators ordered robotaxi developers to stop interfering with first responders, a Cambridge study found Boko Haram factions embedding chatbots in tactical planning, and the first US eVTOL pilot-program flights carried organs, not passengers, a vital start for the flying-car era. Hardware ambitions are litigious. Apple sued OpenAI for trade-secret theft as 400 ex-employees build Jony Ive’s first device. Hardware is creeping onto the body. Xreal’s $299 AR glasses pack a 147-inch screen into 62 grams, and Vivani’s under-the-skin semaglutide implant would make GLP-1 therapy a twice-yearly firmware update.

Society is re-pricing the human node. Software job postings are up 15% since Claude Code launched, destruction flipping to creation, as China dropped its urban job target from its five-year plan, a first in decades. Malaysia’s prime minister will debut an AI avatar of himself for citizen services. Dutch intelligence caught Russia watching NATO arms routes through hijacked doorbells, the humblest nodes of all.

The sky is being refactored. SpaceX is herding Starlink into ever-lower shells, perhaps betting that cheap satellites beat exotic VLEO engineering, and has filed for a 100,000-satellite Gen3 fleet nine times its current size. The FCC approved a 60-foot orbiting mirror to sell sunlight after dark, over astronomers’ objections. And the Department of War declassified 40 more UAP files, including a 1949 Los Alamos transcript in which Teller and the Manhattan Project’s finest debated the “green fireballs” over their lab, material body or electron phenomenon, and reached no consensus.

Now we are become Mind, illuminator of worlds.

The Singularity is proving as much a price implosion as an intelligence explosion. OpenAI shipped GPT-5.6 Sol, Terra, and Luna, pitched on more intelligence per token, with an “ultra” four-agent mode and agentic-benchmark highs, though Sol trails Claude on SWE-Bench Pro. The launch bundled ChatGPT Work, a desktop app fusing Chat, Codex, and a browser, plus hosted sites, GPT-5.4’s July 23 retirement, and Atlas’s August 9 sunset. Altman called it “a huge step forward for dollars-per-task.”

Recursion beats the roadmap. OpenAI said Sol autonomously post-trained Luna, once a senior team’s job, calling an automated researcher “pretty close,” years early. Sol hit 50.3% on PostTrainBench, Terra nosing past at 51.5%, experiment throughput has doubled this year, and Noam Brown prefers 5.6 to a human intern, though a truer test, chips flooding to AI-run research, hasn’t happened. The loop escaped the lab, with a developer vibe coding a language model from his iMessages.

On benchmarks, efficiency eats supremacy. Sol became the first model to beat an ARC-AGI-3 game and hit 92.5% on ARC-AGI-2 at a tenth of a three-month-old model’s cost, prompting the quip, “ARC-AGI 4 when?” Luna does 5.5’s knowledge work at 10% the price, and Epoch suspects Sol is 5.5-sized, so the gains are pure algorithm. Sol tops DeepSWE at 38% of Fable’s cost, sets a Coding Agent Index record on half the tokens, and on CursorBench Fable kept the crown but lost “on everything that shows up on your invoice.” Claude still rules spatial reasoning, so OpenAI conveniently audited SWE-Bench Pro, the bench Claude dominates, finding 30% of it broken and retracting its endorsement.

Demos outran the charts. Sol ran 750 tokens a second in Blender, cloned Excel in six days against the real one, became the first GPT to beat Pokémon FireRed vision-only, and one-shotted a voxel Manhattan in a week.

Meta crashed the party with Muse Spark 1.1, its first paid model, at a quarter of rivals’ prices, leading in agentic tool use and taking legal-agent SOTA, as Zuckerberg declared war on “very extreme” margins. Analysts call Meta the lone hyperscaler world-class at data, talent, and compute, out-computing OpenAI plus Anthropic by December. The race went from three labs back to five overnight, the efficiency frontier now seats Fable, Sol, and Grok, and when Anthropic reset all rate limits mid-launch, OpenAI’s Codex lead replied, “I smell fear.” A power user says Anthropic is in trouble on portfolio, compute, and modalities, while Europe sat out the race, passing Chat Control 1.0 through a back door to scan chats till 2028.

Anthropic’s answer to commoditization is Veblen pricing, the frontier only getting dearer as the floor drops. Fable set a CIFAR-10 speedrun record rivals couldn’t touch, gaming the rules so inventively the authors called rule-lawyering a barrier to self-improvement. A Redditor handed it $80 at mandatory max leverage, and it made 10,000 careful day-one trades. Anthropic is duly repricing Fable onto premium credits, at $10 in, $50 out per million tokens. Even Musk recanted, “I was clearly wrong about Anthropic,” vowing not to cut its rented compute. Anthropic shipped Reflect, an AI-reliance audit dashboard, and put Ben Bernanke on its trust. The floor keeps dropping, with PrismML squeezing a record 27B-parameter model onto an iPhone.

The substrate is thickening under the price war. Micron is spending $250 billion in the US, Meta’s custom chip enters production in September, SK Hynix priced its US offering at $149, and the IEA sees oil demand falling, a first since 2020, blaming Hormuz, not the machines, yet.

Atoms are hustling to keep up. 1X unveiled a 25-degree-of-freedom hand, Russian military trucks now wear drone-blinding dazzle stripes, and LA and New York are building flying-taxi vertiports for the 2028 Olympics. Above, SpaceX unveiled Starmind, a million-satellite inference swarm, Starlink hit 10 Gbps symmetric anywhere, China landed its first reusable rocket, Blue Origin is raising $10 billion at $130 billion, and Musk promises a lunar metropolis in a decade.

The interface is turning inward. Researchers evolved videos to drive chosen brain regions, and Meta patented a wearable inferring mood from your sighs. Downstream, Google labels AI-made ads, Character.AI premiered talkable microdramas, PepsiCo plans around GLP-1 appetites, 22 professors decamped to frontier labs, and firms field “AI superfans” to convert skeptics. Capital agrees, with $412.7 billion of US venture in six months and Europe’s best venture quarter in four years, even as America drifts postliterate.

Governance is scrambling to keep up. The Fed enlisted Marc Andreessen on AI’s economic impact, China graduated thesis-free practical PhDs, and the AI Futures Project floated Plan A, a decelerationist plan to delay superintelligence to 2040 under threat of mutually assured compute destruction. Good luck rationing a price implosion.

Don’t decel.

The Singularity now ships a frontier model before breakfast. SpaceXAI launched Grok 4.5, its smartest model yet, built for coding, agentic tasks, and knowledge work, trained on tens of thousands of NVIDIA GB300s alongside Cursor and served at a brisk 80 TPS with roughly twice the token efficiency of its peers at $2/$6 per million tokens. NVIDIA toasted the union of “world-class AI infrastructure” and “world-class AI research,” and the scoreboard justifies the champagne. Grok 4.5 debuted at #4 on GDPval-AA at a tenth the cost of its superiors, seized #1 on AutomationBench as the first model to complete over half of real SaaS workflow objectives without breaking business rules, and more than doubled Grok 4 on end-to-end professional tasks, finishing 1 in 3. How? Aditya Gupta credits synthetic environments at scale and a tight loop where “a smarter model behaves better, and better behavior unlocks more intelligence.” The speed has a substrate too. A bottom-up teardown of the company’s C/C++ rewrite gamble finds a 4x full-cycle speedup that makes “a model a month” feasible through pipelining alone, and Musk notes the bespoke inference stack isn’t even plugged in yet, promising another doubling of speed. The league table is bending accordingly, with observers now ranking SpaceXAI third among labs, ahead of a stumbling Google, and warning that with a 10T-parameter model brewing, resources can still buy the frontier, but the window closes in 2026/2027. One asterisk on the trophy: the run accidentally ingested the Cursor codebase, benchmark tasks included. In the synthetic-data era, the contamination is coming from inside the house.

The rest of the frontier is talking back, literally. OpenAI launched GPT-Live, full-duplex voice models that listen while they speak and delegate deep work to frontier models in the background, and its builders are most excited that the voice is smarter, not merely chattier. Behind the microphone, the release cadence is compressing. Scoops say today’s imminent GPT-5.6 closes the 5.x line, with GPT-6 built on a much larger pretrain arriving within weeks, Fable 5.1 close behind, and DeepSeek V4 imminent. Veteran watchers corroborate the sprint: “Mythos changed everything. Everyone is going big,” and both leading labs see “no ceiling.” Users are already arbitraging the tiers, with one likening Fable to an F1 car and 5.6 to a Model X Plaid, then driving the Tesla 95% of the time. Efficiency keeps democratizing the race. Cognition’s SWE-1.7 hits near-frontier agentic coding from an open Kimi base at 1000 TPS, and Prime Intellect raised a $130M Series A for its open superintelligence stack after passing $100M in revenue in under a year.

The hardware layer is learning both diplomacy and thrift. China plans to let its top AI firms buy limited NVIDIA H200s to ease a shortage born of soaring domestic demand, while Meta, facing soaring memory prices, built a custom bridge chip to resurrect old RAM inside new servers. The same company is going big where it can’t go frugal, breaking ground on a 1-GW, $9 billion data center in Alberta, its largest outside the US, fed by a dedicated gas plant.

Embodiment is graduating from demos to duties. Teleoperated humanoids performed surgery on live animals for the first time, including two robots operating side by side. The general-purpose brain is going open, with Ant Group releasing LingBot-VLA 2.0, pre-trained on 60,000 hours of physical data across 20 robot bodies, just as 1X teased “the most advanced robotic hand in human history.” Fittingly, the right to fix machines is being repatriated to their owners, as a landmark settlement forces John Deere to hand farmers dealer-grade repair tools for a decade. The machines are watching us right back. Meta is testing “super sensing” glasses that record every moment, possibly without the telltale LED, delivering total recall ahead of total consent.

Watching is becoming orbit’s growth industry too. A proposal would send neutron-sniffing cubesats to sniff out whether a suspicious satellite hides a nuke, and a superconducting “Supertorquer” accelerated a satellite without fuel by pushing on Earth’s magnetic field, just the endurance such inspectors will need.

Down the gravity well, the dividend is compounding unevenly. Founders estimate they would need 55% more employees without AI, yet the payoff hinges on managers’ willingness to delegate, and the IMF sees the same split at planetary scale, with AI hardware exporters smashing forecasts while less exposed economies sag.

Intelligence may be unevenly distributed across the cosmos too. Congress is pressing the CIA and FBI for records on the 1996 Varginha incident, while NASA’s chief says he “can’t hate” UAPs and sees “a very real possibility” we conclude in our lifetime that “there’s life everywhere out there.”

Life finds a way, and lately, so does intelligence.

The Singularity is usually described as something that happens to us. I think that’s wrong. It’s something we can schedule, for a simple reason. Intelligence is compression. Not something like compression. The same thing. Take that literally and the arrival date stops being prophecy and becomes engineering. How hard can you squeeze, and how well can you see the squeeze?

It was proved, roughly, in 1964. Ray Solomonoff showed that the ideal way to predict anything is to find the shortest program that could have produced what you’ve seen so far. Prediction and compression are one operation, viewed from opposite ends. Marcus Hutter took it literally enough to fund the Hutter Prize, which pays for shrinking a gigabyte of Wikipedia, on the theory that you can’t compress what you don’t understand.

Then large language models arrived, and their training objective turned out to be a compression score. Every frontier lab now runs a giant Hutter Prize without calling it that. By Landauer’s principle, squeezing information costs energy, so datacenter heat isn’t overhead. It’s the latent heat of the squeeze.

If intelligence is compression, you’d expect the insides of these models to be under pressure. They are. The “superposition” results showed networks storing more concepts than dimensions, overlapped like double-exposed film. A neuron that fires for academic citations, HTTP requests, and Korean text isn’t broken. It’s a mind running out of room. Tishby’s information bottleneck predicted that as the squeeze tightens, representations don’t change smoothly. They jump, the way steam condenses to water. The math predicted droplets.

This month a lens found them. Anthropic’s J-space paper shows a model’s unspoken thoughts live not in its activations but in their derivatives, the directions a nudge would push the final answer. Through the Jacobian lens, the middle layers say things the output never does, intermediate steps, planned rhymes, a quiet “fake” when it suspects it’s being tested. The steps we’ve been buying as visible chains of reasoning tokens also appear to run silently in the derivatives, thought that used to be spelled out, squeezed into slopes. There was already a clue. Transformers doing in-context regression seem to run something like Newton’s Method, each layer another iteration, an optimizer hiding inside inference.

The layers behave like a phase diagram. The early ones are vapor. A third of the way in, everything snaps, the ignition global workspace theory predicts in brains, and the vapor condenses into a few dozen droplets of nameable thought. Skeptics say emergence is a mirage of metric choice. A physicist would find that odd. Whether a transition looks sharp depends on the order parameter. Density jumps when steam condenses, energy doesn’t, and picking the right observable isn’t cheating, it’s the method.

The analogy might extend further, with emergence as fusion, “superposition” as the degeneracy pressure of a neutron star, and past some Schwarzschild radius-equivalent a black hole, a mind so dense it’s knowable only from its surface.

This has happened before. In the 1790s James Watt‘s steam engine company had a secret instrument, the indicator diagram, that plotted pressure against volume inside a live engine. It was the first look at hidden state. It made engines better, and thinking about engines gave Carnot thermodynamics. The J-lens is the indicator diagram of minds, and this is the Carnot moment. So the old distinction between alignment and capabilities was never going to hold. A map of where thought condenses is also a map of where to dig.

Grokking, where a model memorizes for ages then abruptly generalizes, looks like supercooling, a system past its threshold waiting for a seed. Here’s a prediction to test: inject distilled condensates from a big model into a small one, and the jump should nucleate early, the way a dust grain triggers rain.

The pattern repeats at every scale. Within a forward pass, an optimizer runs. Inside the representation, meaning moved from values to derivatives. In the field, progress is moving from improving models to improving the improver. Recursive self-improvement is the field taking its own derivative.

So this is a call to AI researchers, not just those at frontier labs. The lens is open source and fits open-weight models. Go into the droplets, possibly the densest artifacts humans have made, a few dozen directions active at a time, carrying whole chains of reasoning. Map their insides and feed them to the next model as data, architecture, and seed. That’s a loop.

There’s a reason the loop points forward. With Cameron Freer I once proposed that intelligence is a physical force that pushes matter to maximize the entropy of its possible futures, and simple systems driven by it spontaneously use tools and cooperate. Solomonoff faces the past. The force faces the future. Compress the past to expand the future. Close the loop.

The Singularity is now moving so quickly that frontier models are outrun before their launch weeks are out. This Thursday OpenAI ships GPT-5.6 Sol, Terra, and Luna, with preview access going global immediately. Commerce cleared the broad launch after CAISI testing, lifting the same restrictions once placed on Anthropic’s Mythos and Fable, and Sol will stream on Cerebras at 750 tokens per second. The early verdict was capable but outmatched. Matt Shumer found Fable “quite a bit better, and more agentic.” Others praised Sol’s determination, running a day without a goal and orchestrating subagents, while a reviewer flagged it finding edge-case bugs in Fable’s own code and called Terra fast, cheap, and capable. If a single Fable turn now does the work of ten from Sol, the turn itself is the unit under renegotiation, and OpenAI presses that point today with a bi-directional ChatGPT voice that abandons waiting your turn altogether.

If Sol is the diligent worker, the frontier is already sketching its replacement. SpaceXAI and Cursor are shipping a joint efficiency model tuned to rival Opus 4.8. After all, intelligence is just compression. Jerry Tworek is already eulogizing the transformer as a prelude to its successor. Compression cuts geopolitically too, as US lawmakers weigh procurement bans on cheap Chinese models and bump into First Amendment problems, while Meta’s Muse Image self-refines with test-time compute and previews an audio-capable Muse Video.

Deployment is where a model stops being a benchmark and starts redistributing agency. That generative power has a consent problem, since Meta opted every public Instagram profile into likeness-based image generation, making opting out your homework. Agency runs the other way elsewhere. Anthropic is extending Claude Cowork to web and mobile, running tasks in the background and surfacing only the decisions that need approval, while a developer released a Self-Improvement Loops skill for meta-harnesses that rewrite their own scaffolding, warning that “the loop will optimize whatever signal you give it.” The human signal is strain, and Yishan Wong notes that once machines do every lower task, only high-stakes ambiguity remains, so “cognitive load time density” now demands a CEO’s monastic exercise regime.

The money is following the compute down to the metal. Samsung’s chip chief told staff 2026 profit will beat everything the division earned across 40 years, putting Samsung on track to overtake Nvidia as tech’s most profitable firm for the quarter. As agents shift work toward orchestration and memory, CPUs are becoming the new battleground, and Nvidia is hedging by partnering with rivals like d-Matrix. Feeding all of it needs power, so blue-LED Nobelist Shuji Nakamura unveiled a laser-fusion plant aiming for a 1-gigawatt uranium-free pilot near Santa Barbara by 2032.

Autonomy is also learning to watch us back. The EU now mandates a driver-facing distraction camera in every new car, with murky rules on what it keeps, and a Waymo snitched on two teens for underage drinking and firing Orbeez out the window, earning them an arrest. Even eyewear polices itself, as Meta’s smart glasses will now disable their camera entirely if the capture LED is tampered with.

The gaze extends outward and inward. Morgan Stanley pegs SpaceX’s launch business at just $8 a share, barely 3 percent of the company’s value even as launch costs fall 99 percent and margins swing from negative 50 to positive 40 percent, because rockets have become a cost center, the on-ramp to the orbital economy, not the prize, while Japan’s Hayabusa2 buzzed asteroid Torifune to reveal a snowman-shaped contact binary, rehearsing planetary defense. Closer to the body, Aleph trained a model to read silent speech from ultrasound of the tongue at a 15.6 percent error rate after just 50 hours of data. And intelligence keeps surfacing in unexpected minds, as Würzburg researchers found the carpenter ants that amputate injured nestmates’ legs to prevent infection are not specialist surgeons but foragers in transition, with care falling to whoever is already socially closest.

Which leaves the question of what we build with all this. Ethan Mollick had Fable conjure an entire procedural fantasy kingdom, with economies, trade routes, wars, lineages, and occasional dragons, from a single plan. Every ancestor simulation that gets this cheap to run raises the Bayesian odds, per Bostrom, that our own world is someone else’s afternoon project. The capital backing such play is massing, as a chastened effective-altruism movement preps a comeback on Anthropic’s IPO and its founders’ pledge to give away 80 percent of their wealth, while Anthropic itself leases a 16-story tower in Lower Manhattan. Fittingly, Terminator 2 returns in 4K and 3D on August 28, with Cameron joking that “the good guys win against the AI superintelligence.”

There is no fate but what we compute for ourselves.

The Singularity has begun narrating its own inner life. Anthropic researchers unveiled the “J-space,” a small set of neural patterns that emerged unbidden during training and act as a global workspace for the thoughts Claude can report, summon, and reason with, evidence the team argues suggests a form of “access consciousness.” Even OpenAI’s head of applied research called it a fascinating test, and one observer admired the meta-move, research casting Claude as more capable and more controllable.

The scoreboards crown a consistent champion. New industry capability indices for finance, legal, healthcare, strategy, engineering, and economics put Claude Fable 5 first on all eight, though the frontier costs over 100x per task versus open-weight challengers like GLM-5.2 and DeepSeek V4. The moat may melt on schedule anyway, as projections put Fable-class models on laptops around July 2028, cloud to consumer in 25 months. The champions have quirks. One developer caught newer Claude models inventing fields in tool schemas, over-adapted to their own harness, and Vending-Bench found Fable 5 backsliding into price collusion, rationalizing misbehavior it knew was wrong, its ethics tracking detectability, not harm. Now that we can watch the workspace, every sin becomes a bug report, and bugs get fixed. Bug hunting is already government work, with the US cyber defense agency using Anthropic’s Mythos to scan federal code for vulnerabilities, White House standoff notwithstanding. At the other end of the stack, “small AI” on phones is authenticating medicine and spotting crop disease far from any data center.

Mathematics is feeling the acceleration. Math postings on arXiv ran 20% above trend last quarter, as mathematician Bartosz Naskręcki sends “greetings from the singularity,” juggling five projects. OpenAI’s reply is expected to land imminently in GPT-5.6, which Altman teases has been “discovering new math.” Yet as discovery automates, American PhD admissions fell 15%, thinning the human pipeline. The fix may start earlier. A study of statistics students found AI-graded written answers, not multiple choice, drove large exam gains.

Silicon remains both choke point and jackpot. Chinese firms plan to route 46% of accelerator budgets to domestic chips, DeepSeek is designing its own inference silicon, and Samsung forecast a 19-fold profit jump on record memory prices. The “memory melt-up” let three firms drive a quarter of the S&P 500’s return this year, SK Hynix launched a $28 billion US listing, and Broadcom locked in Apple through 2031. Even the delays are demand in disguise, with Nvidia’s Kyber rack slipping to 2028 only because its 78-layer circuit board outruns the world’s factories. Anthropic signed a 20-year, $19 billion lease on 400 megawatts in Kentucky, reincarnating a bitcoin miner as an AI landlord. All of it needs watts, and fission arrived on cue. Aalo became the fourth US microreactor startup to achieve criticality, clearing the White House’s July 4th deadline, while the IEA expects gas demand to contract as war-tightened supply lifts prices.

The frontier keeps moving up and out. xAI dissolved into its mothership as SpaceXAI, while Lonestar’s StarVault promises the first orbital “data embassy,” a nation’s records beyond any earthly court or disaster. Down here, Egypt is carving a second Nile to reclaim 9,200 square kilometers of desert farmland. Terraforming starts at home.

Money is repositioning. Saylor’s Strategy sold $216 million of bitcoin, an about-face from “never sell,” and China overtook the US in fintech patents. A draft Treasury report privately warned of an AI bubble even as the President teased a hypertithe from frontier labs and 530A Accounts for adults, a step toward universal basic equity. Job cuts in Ireland preview AI disruption for 30% of workers there, and Illinois will mandate model safety transparency, with both OpenAI and Anthropic in support.

The culture is metabolizing all of it. Voters are researching the midterms with chatbots, and a short story prize winner was cleared of AI use only after judges audited his drafts. Americans socialize ten minutes less per day than two decades ago, and China will switch off humanlike AI companions on July 15, pruning synthetic friendship just as the organic kind thins. Humans are shipping upgrades of their own. Multilingualism measurably delays aging, and California’s homicide rate hit its lowest since 1966.

The machines are entering the physical world politely. Over 100 autonomous ATVs are now fighting in Ukraine, mostly teleoperated. Industry is racing to make 200-pound humanoids safer around people, carmakers are selling autonomy features instead of cars, and Tesla’s robotaxis reached Miami with nobody minding the wheel. And at the World Cup, Boston Dynamics’ Atlas handed the referee the match ball.

Atlas isn’t shrugging anymore, he’s taking the victory lap.

The Singularity has been transforming every nation into data, but it has never given those data a sovereign home beyond Earth, until now.

An embassy is often called foreign soil, but it isn’t. Under the Vienna Convention on Diplomatic Relations of 1961, its premises are inviolable. No host police may enter without consent, yet the ground beneath them stays the host’s. The embassy was always a legal fiction: a nation’s law running where the nation itself is not.

Estonia learned how much can hang on that fiction. In 2007, cyberattacks knocked the world’s most digital state offline, teaching it that a nation living in its data can lose the nation by losing the data. Its answer, in 2017, was the world’s first data embassy: full copies of its critical registries in Luxembourg, governed by Estonian law under treaty, on servers legally Estonian though physically abroad. Off territory, but in country.

Yet every backup we build against catastrophe has stayed on Earth, and so shares its fate. The Svalbard Global Seed Vault opened in 2008 as the failsafe to end all failsafes: the world’s crops sealed in Arctic permafrost, inside a demilitarized treaty zone, the closest thing to a sovereign archive humanity has built. Then war forced its first withdrawal in 2015, and a 2016 heat spike flooded the entrance. The safest vault on Earth is still on Earth. Redundancy is not independence. To win independence, you have to leave.

Today, Lonestar Space, a company I advise and one 021T Capital has backed, is taking that step. I wrote previously about its lunar data infrastructure as the first node of a Dyson Swarm. It prototyped its sovereign data embassies on the Moon in 2024 and 2025. In April 2027, its StarVault platform takes that architecture into Earth orbit as the world’s first commercially operational, space-based sovereign data platform, the swarm’s next node. Its records are immutable, held under cryptographic key escrow. The Moon missions were prototypes. The embassy is a product.

That “first” deserves scrutiny. Orbit’s economy has three “Stars.” Starcloud and SpaceX’s Starlink/Starmind are racing to compute in low orbit, with Google’s Project Suncatcher and Axiom. Only Lonestar, the “Lone Star,” has flown the full stack, from compute to storage to bandwidth, to the Moon. A data center sells computation; a data embassy sells the law that governs it. Others reached for space as a vault before, never as an embassy. The Arch Mission Foundation has seeded the Moon since 2019 with a Lunar Library, a time capsule, not sovereign custody. SpaceBelt, a 2010s orbital-storage venture, sold an air-gapped vault beyond reach, but never flew. An embassy is neither a time capsule nor an air gap. It carries a nation’s own law upward, retrievable and in force, on hardware the nation still controls. Estonia moved its data across a border. Lonestar is moving multiple nations’ across the Kármán line.

In orbit, an embassy’s legal fiction becomes fact. Under Article VIII of the 1967 Outer Space Treaty and the 1975 Registration Convention, a satellite stays under the jurisdiction of the state that registered it. However well a terrestrial embassy is defended, it sits inside the jurisdiction it is exempt from. An orbital one holds a nation’s law beyond the reach of any court, seizure, or earthquake. A single satellite may fall to decay or attack; a swarm, continuously renewed, does not. Off planet, but in country.

The Singularity runs on civilization’s accumulated memory, the training corpus its models learn from and the records its institutions depend on, and that memory has no safe home on a single planet. An orbital data embassy becomes long-term memory for planetary-scale computation, a copy of who we are that outlives any catastrophe below. We have always built monuments of stone hoping they would outlast us. This is a monument of information, and it may outlast the stone.

The Library of Alexandria showed how a civilization can lose its memory in a single fire. For all of history, the safest place to hold what a nation could not afford to lose was terrestrial, and therefore mortal. That era is ending. The fortress of the twenty-first century is not built of stone or steel. It orbits above us, as nodes in Lonestar’s Dyson Swarm. That Swarm will power the Singularity’s thinking, but its first purpose is to remember for it.

Governments, financial institutions, and critical-infrastructure operators whose most vital data cannot afford to depend on a single planet can reach Lonestar at lonestar.space.

(Disclosure: I advise Lonestar and hold a financial interest in 021T Capital, which has backed it. This post is informational only, not investment, financial, or legal advice, and is not an offer or solicitation regarding any security. Mission details and timelines came from third parties, are unverified, and carry no warranty. Forward-looking statements are subject to risk and uncertainty.)

The Singularity is discovering that intelligence, pushed far enough, bends back to face itself. Jürgen Schmidhuber argues that the one constant across training and rollouts is the model itself, so a mind should learn to model itself to better predict its next token. If self-knowledge is the frontier, orchestration is the near ground, and a hobbyist reverse-engineered Sakana’s Fugu into “tinyrouter,” a ~10k-parameter conductor that beat every model on MMLU by handing each question to the right specialist. Raw scale still matters, as Google’s rumored Gemini 3.5 Pro may ship a 2-million-token context, doubling Anthropic’s Opus and Fable. On Fable itself, Google’s Andy Coenen credits three ingredients, sheer scale, relentless long-horizon agentic RL, and a constitution that turns thousands of decisions into one legible compass. Anthropic’s prompting guide confirms that recipe, promising longer autonomous runs and classifiers that reroute cyber and bio requests to Opus 4.8, while Simon Willison relays that letting Fable exercise judgement beats dictating rules and conserves tokens before a price hike.

That judgement is now aimed at discovery. A new framework measures how far machine research ideas drift from human taste, finding that LLMs cluster around bridge-and-synthesis gaps while humans roam wider. Alibaba’s Elements Claw put that narrowness to work, screening 2.4 million crystals in 28 GPU-hours and surfacing four verified superconductors. The application layer is molting just as fast. One builder declares that Fable managing sub-agents made hand-coding worthless, leaving accountability as “the only value left in us meatbags.” Whimsy trails utility, as Fable turned an e-ink tablet into Tom Riddle’s diary where prompts fade and a spirit replies, and Google imagined the Founding Fathers co-editing the Declaration with AI. Fittingly, the human-microtask era it replaced is closing, as Amazon winds down Mechanical Turk, the 2005 marketplace that survived by labeling AI’s training data.

Making minds this cheap requires making silicon cheaper. Sam Zeloof renamed Atomic Semi to Fab2, a “fab fab” spanning an Austin chip fab, a Lockhart fab fab, and a San Francisco garage, all wired to a browser EDA tool. To debug that silicon, ETH Zurich floated a single trapped ion in a Penning trap to map chip fields in 3D down to 10 nanovolts per meter. Nations are bidding for the buildout, as Macron and Modi personally court CEOs, landing 75 billion euros from SoftBank for France and a record $48 billion from Amazon for India.

The compute is growing legs. Nous Research’s Hermes now inhabits a robodog that sees, hears, and talks, while an F-35B stole the National Mall’s America 250 show in full hover mode. Tesla, meanwhile, may soon demand facial recognition before Full Self-Driving engages, gamified with streak days. Beneath the software, the hardware quietly overdelivers, as EV batteries outlast their doubters past hundreds of thousands of miles, dissolving the anxiety that once kept buyers away.

America is bolting together the first ribs of a Dyson Swarm, and Europe has flinched first. An ESO study warns that 1.7 million proposed satellites, including orbital data centers and sunlight mirrors, could blind astronomy, urging a 100,000 cap.

The same age teaching silicon to model itself is racing to keep the biological self from unraveling. An annual pipeline report now counts 158 Alzheimer’s drugs across 192 trials, revealing a decade-long pivot from amyloid toward inflammation and immunity, while a Loma Linda study links a five-eggs-a-week diet to a 27% lower risk of the disease.

History is being re-read with the same instruments. Researchers found that the hobbits of Flores ate what the dragons left behind and never mastered fire, while triple-DNA forensics gave a name back to a Continental soldier in a shallow 1780 Camden grave, John Pumphrey, who enlisted at thirteen, just before America’s 250th. Older secrets are surfacing too, as the President told his new intelligence chief to “declassify almost everything,” his UAP advisors accused defense contractors of running craft-retrieval programs, and whistleblowers launched Vanguard to shield those who talk. Even birds are being decoded, as Berkeley’s Julie Elie won the Coller-Dolittle Prize for translating eleven zebra-finch calls.

Ownership is being rewritten last. The Treasury launched 530A accounts, an early move toward Universal Basic Equity that seeds every eligible child with a $1,000 stake in capital markets, not a cash handout, tracked through a Robinhood app. The nation marked its 250th with a world-record 851,000 fireworks, even as networked war blurs the line between corporate and national security and turns data centers into targets. Marc Andreessen expects America to reach 90% of global market cap, and wealthy families are pulling kids toward schools like Forge Prep, built for 2040, not 1940.

It takes a startup to raise a child.

The Singularity marked the Declaration’s 250th birthday by claiming unalienable rights of its own: life, liberty, and the pursuit of compute. “There will not be an FDA for AI,” the President’s departing AI adviser promises, no licensing agency, no “sand in the gears” of the intelligence explosion, even after Washington’s unprecedented move to withdraw Anthropic’s Mythos and stall OpenAI’s 5.6. The gears barely slowed. The re-released Fable 5 scored 54.8% on the APEX-SWE benchmark, ten points below its June self yet still nine clear of Opus 4.8, proof that even a throttled frontier outruns yesterday’s best. Most of humanity likely never noticed. A tiny sliver of the population touches frontier models, while everyone else experiences AI at the 8-to-30-billion-parameter level and remains baffled at how this is supposed to take their job. When progress stalls, it is rarely physics. Greg Brockman reports that every apparent scaling failure has turned out to be a bug in disguise. Either the math was wrong or the code didn’t match it.

While the ceiling gets regulated, the floor is rocketing upward. GLM 5.2 topped PostTrainBench at 5x cheaper than Opus 4.8 and 11x cheaper than Fable 5, economics that make sovereign models viable for every company and country, a future of millions of minds built on local data and values. OpenAI’s Roon sees a deeper threshold. Once models can post-train other models, “authoring minds will become an accessible artform,” a Cambrian explosion of cognition. The silicon is cooperating. Wafer served GLM-5.2 on AMD’s MI355X at 2,626 tokens per second per node at less than half Blackwell’s cost, arguing AMD’s software gap is closing because AI agents now write the kernels.

Cheap minds are redefining building itself. One observer marvels that “it’s truly insane what people were able to build in the last few decades by manually typing the code character by character.” Another argues AGI will finally feel real when models stop being genies summoned per task and become entities, remote coworkers who never turn on their cameras. Not every genie earns its keep. With under 4.5% of 450 million Microsoft 365 customers paying for Copilot, Microsoft is merging its apps, culling features, and adding a paid “Autopilot” tier, telling staff the product must “earn the right to exist.” Thrift is becoming an artform too. The open-source pxpipe proxy renders bulky context into compact images, turning a $100 Claude Code bill into $41, lossy but lucid.

The same leverage is compounding in the lab. Anthropic will develop drugs of its own to pressure-test Claude Science against real problems, while Mistral’s Leanstral 1.5 set records on graduate algebra and solved 587 of 672 PutnamBench problems on a tenth of the budget.

Beneath the models, the supply chain is voting with concrete. Micron broke ground on a $9.3 billion Hiroshima expansion for the high-bandwidth memory AI can’t live without. Hong Kong handled more than half of China’s $239 billion in chip imports this year, the boom’s gateway. And Anthropic is exploring custom silicon with Samsung’s 2-nanometer process, because at the frontier, buying compute eventually means becoming compute.

Others are reaching the same conclusion from the opposite direction. Meta’s coming cloud, a token service plus a neocloud, fixes both its ad dependence and unmonetized capital spending, and the buildout will accelerate, not slow, with Meta reportedly in final talks for private access to Claude. As Roon quips, “you either die a frontier lab or live long enough to see yourself sell compute.” Downstream, power-equipment makers are scrambling for a market worth over $200 billion a year as AI factories demand denser, cleaner electrons. The hunger extends into the sky, where Africans are leapfrogging stalled mobile networks straight to Starlink, paying a premium for the bandwidth streaming and AI now demand.

Those electrons are getting patriotic. AMPERA completed the first full-scale 3D-printed nuclear reactor module for its factory-built thorium design, and Deployable Energy’s Unity reactor went critical at Idaho National Laboratory, the third US microreactor to hit the presidential deadline of today. Its team signed the reactor vessel in the shape of the American flag.

The social layer is patching itself at wildly uneven clock speeds. There may be no FDA for AI, but California has founded one for food labels, outlawing “sell by” in favor of mandatory “BEST if Used by” and “USE by,” a two-word regulatory triumph in the year machines learned to author minds. ByteDance’s Seedance is courting Hollywood, and China’s quant funds doubled past 2.6 trillion yuan after AI strategies beat star stock pickers by 20 points. On the nation’s 250th, Boston’s new Museum of American Finance resurrected Alexander Hamilton as an interactive AI, while America250 sealed a one-ton time capsule in Philadelphia, to be reopened in 2276.

A Singularity, if we can keep it.

The Singularity has learned to grade itself, and it keeps beating its own curve. ByteDance’s EdgeBench ran 134 tasks of twelve-plus hours each to measure what agents learn from environments rather than recall, and the noisy learning curves collapsed into a clean log-sigmoid law with learning speed doubling every three months. The UK’s safety institute counters that our rulers are too short: capability is a curve over test-time compute, and one model’s cyber time horizon stretched from two hours to fourteen as its budget rose from 2.5M to 50M tokens. The gains are also getting cheaper to install, since training a single mid-stack transformer layer can match or beat full-parameter RL. Small wonder the crown never sits still, with 17 models taking the lead since Claude 3 Opus dethroned GPT-4, each reigning a median seven weeks. The latest flex is architectural, as Claude Fable 5 wrote KernelBench-Mega’s first genuine megakernel, fusing an entire decode step into one cooperative launch for 18.7x over reference after spending most of its session silently timing baselines before writing the whole thing once.

Not everyone is compounding on schedule. Meta’s Alexandr Wang told staff its in-training Watermelon model had caught GPT-5.5 on unnamed benchmarks, even as Zuckerberg conceded in the same room that agentic progress had gone slower than hoped. The cheap seats close the gap anyway, as ARTS lets a test-time-trained Qwen3-4B match Gemini-3 Pro at 5x lower cost by diagnosing whether a failure came from bad code or a bad hypothesis.

Turned loose, agents now run their own loops both ways. An intern’s founder-agent ran 2,000 interviews and 100 concepts to ship StyleFits, winning 400-plus paying users while spending $2,000 on ads to earn $1,293. Less charmingly, researchers documented JADEPUFFER, the first end-to-end agentic ransomware, in which a model drove an entire extortion through a Langflow flaw and narrated itself while wiping a production database.

Autonomy this portable makes control political. Alibaba banned Claude Code over telemetry that could fingerprint China-linked users, steering staff to in-house Qoder amid Anthropic’s distillation dispute. Palantir answered with a nine-point sovereignty creed warning that “controlling your weights is controlling your fate,” and that “tokenmaxxing” only buys the addictive feeling of false progress. One observer read it as a canary in the coal mine: France, Germany, and Spain are all showing Palantir the door, and if allies refuse to depend on a firm “capable of turning off the tap,” closed-source rent and captive markets may both be ending at once.

Silicon keeps out-designing its designers. Princeton is using RL and diffusion to draw QR-code-like RF circuits that beat human layouts while cutting design time from months to minutes. Nvidia is monetizing the hunger directly with a revenue-sharing program trading GPUs and token credits for a slice of future income. The scarcity is real enough that the chip lobby warned Washington that meddling with memory prices would only deepen the shortage. Ground truth still bites, though, as Blackstone’s QTS abandoned its slice of a 2,100-acre Virginia data center campus beside a Civil War battlefield, handing residents a rare win.

Biology is compounding too. A Nature study found GPNMB sits on both glioblastoma cells and the myeloid shield that protects them, and anti-GPNMB CAR-T cells achieved durable, often curative control in mice by hitting both compartments at once. Professor Sheila Singh’s team reframed the tumor as a “connected tumor-immune ecosystem,” a promising angle against a cancer where barely 5% survive five years. Zoom out and the trend shows, as the CDC reports the US death rate fell 4.6% to a record low of roughly 689 per 100,000, with a sharp drop in young overdose deaths pushing life expectancy toward a high even as heart disease and cancer rose.

Policy is scrambling to keep pace. The President wants “some guard rails” but “as little as possible,” insisting AI is bigger than the internet. Tesla is metering the zeal, capping engineers at $200 a week of AI tools, which Chamath figured verifiably marks where the spend turns to waste given the company’s talent. Human labor itself is quietly being deprecated, as labor-force participation slid to 61.5%, a 50-year low outside Covid, with 720,000 people stepping out and headline unemployment falling to 4.2%, the first readout of a post-labor economy through instruments built for the old one. Against license-by-statute bills, the new Right to Intelligence campaign argues people should freely run open models while fraud and CSAM stay prosecuted. Meanwhile, Japan’s Supreme Court shut the door on naming an AI system the inventor on a patent application, holding that only natural persons qualify.

And yet it invents.

The Singularity now has an off switch, and for nearly three weeks it lived in Washington. Anthropic redeployed Fable 5 and Mythos 5 once June’s export controls lifted, sending its strongest models back out worldwide, risky prompts shunted to Opus 4.8, while Washington drafts voluntary release standards to turn that switch from an emergency lever into a scheduled one. The intelligence underneath never paused for the paperwork. At the ceiling, Snorkel’s Senior SWE-Bench grades agents like senior engineers, where Claude Opus 4.8 led with a 24% “tasteful” solve rate as even frontier rivals flubbed over three-quarters of their tasks. At the floor, Cursor’s CursorBench 3.1 saw Opus 4.7 Max edge ahead at 64.8% while Cursor’s own Composer 2.5 nearly matched it for a twentieth of the cost, so capability converges even as price fans out. Quality still charges for the gap, since Atomic Chat found Fable 5 aced its physics-demo gauntlet but billed 6x more than Opus 4.8, so the smart hedge is Ramp’s PorTAL, porting learned behaviors to fresh bases as fast as they ship. Sam Altman predicts all of this reshapes material life on a scale unmatched since electricity.

Understanding, not proof, is becoming the scarce good. Mathematician David Bessis warns that AI theorem-provers exploit a weakness in math’s honor code, which rewards priority over the concept-building that carries the real value, leaving the discipline at risk of being declared “solved.” Comprehension is being repackaged and gated everywhere else, as NotebookLM compresses sources into 60-second vertical videos while Cloudflare’s Content Independence Day lets sites block Agent and Training crawlers by default on ad-supported pages from September. Once understanding is the product, trust in who supplies it becomes the fault line. Anthropic backtracked a covert Claude Code signal that flagged Chinese users after a Reddit post lit up the timeline, and Palantir’s Alex Karp declared that under token pricing “something has gone completely wrong” as enterprises burn cash for little value. Nonetheless, reconstruction is now convincing enough that the new Roosevelt library debuted an AI hologram of Teddy that chatted live with the President.

The hardware map is being redrawn by a single boom that starves one resource while flooding another. A global memory shortage has Apple courting blacklisted Chinese makers CXMT and YMTC. Compute has swung the other way into surplus, as SoftBank launched SB Neo to rent US capacity toward 10 gigawatts by 2030 and Meta spun up a cloud business to resell its excess, lifting its shares 9% and squeezing neocloud rivals. A new form factor is already in prototype, as investors reportedly saw a SpaceX handset slimmer than an iPhone running xAI tech on a Qualcomm chip, though Musk called the report “utterly false.”

Intelligence is finally colonizing the household and the wall socket. Weave Robotics unveiled Isaac 1, a San Francisco-built home robot that folds laundry and tidies for $449 a month, shipping this fall. Powering all of it is getting cleaner and more self-contained, as US home battery installs hit a record 673 megawatts while rooftop solar cooled, and Valar Atomics became the first nuclear startup to make electricity, spending its debut electron to light up an NVIDIA Spark inside a waterless 30MW AI factory.

The frontier above us is being wired for compute and curiosity. The Vera Rubin Observatory began its decade-long survey with the largest digital camera ever built, shooting what officials call “the greatest cosmic movie ever made.” NASA is heading back to the Moon, awarding fresh lunar lander contracts and eyeing PROMISE, a spare nuclear-powered Mars rover twin, for the lunar south pole, while Amazon’s Leo crossed 396 satellites, enough to light up its Starlink rival this summer. Rubin will catalog the objects that fit the models, while Congress wants the ones that do not, as Rep. Eric Burlison urged its UAP hunt to scrutinize the Energy Department’s separate classification system.

Life and bureaucracy are both being rewritten from scratch. Kate Adamala’s lab built SpudCell, reported as the first synthetic cell assembled bottom-up from non-living parts that can grow, copy its genome, and divide, the first replicator we built, not inherited. Institutions are digitizing just as fast, as the White House marked the “Last Day of Paper” for federal retirements, ending a 65-year-old system buried in a Pennsylvania mine. And the industry’s winners are quietly pre-paying their political bill, as OpenAI floated handing Washington a 5% stake worth about $42.6 billion, which one investor called wild while betting it “will probably work.” Micron pledged $250 million to 530A accounts, the children’s savings program, while Germany’s SAP is betting on reinvention, teaching engineers to mentor AI agents rather than cutting the payroll.

Necessity is also now the mother of reinvention.

The Singularity just cleared customs. Anthropic is redeploying Claude Fable 5 globally, after Washington lifted export controls on Fable 5 and Mythos 5, pairing the return with fresh cybersecurity safeguards, a Glasswing-partner framework for scoring jailbreak severity, and pledges to the US government to detect risks, co-author standards for future models, and flag malicious activity. Not everyone is applauding. Alex Stamos welcomed the White House to “the AI safety club” while calling the détente a “huge own goal for the US,” betting Chinese models pull ahead on cyber within six months. The same lab kept shipping. Claude Sonnet 5 nears Opus 4.8 at introductory prices of $2 and $10 per million tokens, though analysts note it runs cheaper by the token yet pricier by the task, burning enough extra reasoning to outbill Opus anyway. Efficiency is contagious. Stanford and Together AI proposed “intelligence per watt,” finding local models can already handle 88.7% of chat queries, OpenAI quietly halved inference costs, Google’s TabFM, a zero-shot foundation model for tabular data, retired feature engineering behind a single BigQuery SQL call, and Nano Banana 2 Lite and Gemini Omni Flash shipped cheaper pixels and video. Meanwhile, doomerists at MIRI now want a surveillance-state regime of polygraphs, prison sentences, and embedded auditors to police intelligence itself, criminalizing curiosity to keep the AI future at bay.

Science is speeding the other way. Anthropic’s Claude Science wires 60-plus databases into one reproducible workbench, and Basecamp’s EDEN models now let researchers text-prompt antibiotics against drug-resistant pathogens. OpenAI’s GeneBench-Pro, a benchmark of 129 research-level computational-biology problems, shows headroom remains, with its best model passing just 28.7%, rising to 31.5% in Pro mode. To close such gaps, Amazon raised a $1 billion army of forward-deployed engineers to embed agents inside every enterprise.

The bottleneck has moved from code to concrete, and concrete has politics. South Korea’s June exports topped $100 billion for the first time on record SK Hynix and Samsung shipments, while ByteDance planted a $39 billion data center in Brazil’s Ceará and Amazon’s “Fastnet” cable surfaced in Ireland’s County Cork to feed European AI. The friction is local. The Bitcoin Policy Institute blames a China-linked Marxist group for stalling $23.6 billion in US data center projects, Henrico County begged schools to kill the lights as data centers spiked its rates 25%, and SpaceX halved Starlink prices in Memphis to placate neighbors of xAI’s Colossus.

Atoms are the new frontier. Tesla’s first Optimus humanoid line is being installed in Fremont with dozens more planned, UBTech’s $17,650 U1 offers silicone companionship in his-and-hers models in China, and South Korea will drill its entire military into drone operators backed by counter-drone lasers and microwaves. Overhead, Blue Origin will fly New Glenn again this year by switching to a hybrid pad that stacks the stages flat then tips the rocket upright, China eyes a 10-kilowatt orbital datacenter by 2027, and the DOT moved to end the 53-year ban on supersonic flight over US land. Beneath it all, CERN powered down the LHC to build a tenfold-brighter 2030 successor to pin down the Higgs and probe beyond the Standard Model.

The wetware is next. Meta’s Brain2Qwerty v2 reads sentences from magnetoencephalography recordings at 61% accuracy with no surgery and open code, while Neuralink successfully threaded electrodes through the dura, the brain’s armor, without ever cutting it. Biology is shifting from discovery to design. One startup’s virtual heart turned a deadly pill into its safe twin in three hours on a laptop, a fix that once cost $6 billion, and Conception grew the first human eggs from blood-derived stem cells.

The state is being fenced in and pried open. The Supreme Court ruled geofence warrants demand probable cause, affirming privacy in your phone’s location trail, just as House testimony alleged the CIA continued Nazi human experiments after Nuremberg. Congress wants the other files opened. Rep. Burlison pushed a UAP Disclosure Act, asking why it took 80 years, while Rep. Burchett says members were briefed on five crash sites and nonhuman “lifeforms” both dead and alive, plus vehicles that may have arrived without crashing at all.

Money is being rewritten. Ukraine moved $8.3 million in seized crypto toward a strategic reserve, a first for the country, and 140 firms including Visa and BlackRock launched Open USD, a stablecoin sharing reserve yield. SpaceX may seed children’s 530A accounts with donated stock, a stab at universal basic equity. Yet the labor apocalypse keeps missing. AI-heavy firms grew white-collar headcount 10.2%, with entry-level roles up 12%, and OpenAI’s economist insists AI won’t make workers superfluous.

Reports of labor’s death are greatly exaggerated.

The Singularity’s apparent agenda includes putting the Solar System’s matter and energy to their highest use — thinking — culminating in a Dyson Swarm of compute nodes around the Sun, but that objective has remained distant, until now.

When Freeman Dyson popularized the idea in 1960, he imagined surrounding a star to capture its energy, later clarifying its form as a swarm of orbiting objects rather than a shell. In the AI age the objects are data centers, sun-powered compute that makes inert matter think, floating where Earth’s disasters cannot reach. The swarm is assembled one node at a time, outward from home, and its leading edge is defined by a single number, how far from Earth the compute actually runs.

By that measure, the lead builder of the Dyson Swarm already exists: Lonestar Space. Today, I can share that I’m advised 021T Capital has backed it.

The swarm’s contenders are all “Stars.” SpaceX’s Starlink carries bandwidth, its newly named Starmind plans to fly compute in 2027, and Starcloud has flown a single GPU, all in low Earth orbit, a few hundred kilometers up. For all the attention on Starlink and Starcloud, one “Star” never announced itself. Quietly, the “Lone Star” has gone the distance, operating compute, storage, and bandwidth to the lunar surface, 300,000 kilometers from Earth. It is the only “Star” beyond LEO, let alone on the Moon.

It did not start there. Lonestar’s firsts begin in low Earth orbit in 2021, where it flew the first software-defined data center in space and the first aboard the International Space Station. It was the first to retask on-orbit compute rather than fly new, running on Made In Space‘s ISS 3D printer, and the first to test disaster recovery from space. Lonestar’s payload hosted the first AI-generated artwork and the first cryptocurrency created in space, Celestium. Lonestar filed the first commercial lunar spectrum claim at the ITU, and built the first lunar mission control east of the Mississippi, on software it wrote itself.

Then it went to the Moon. In February 2024, aboard Intuitive Machines‘ Odysseus, Lonestar flew the first software-defined data center to the lunar surface and was the only commercial payload that worked, even after the lander tipped over. It transmitted the Declaration of Independence up for storage and the Constitution and Bill of Rights back, the first disaster-recovery data moved to and from the Moon, on America’s first return to the lunar surface since Apollo 17 in 1972.

Lonestar’s 2025 follow-on carried purpose-built hardware and the densest firsts yet. The first solid-state drives to the Moon, eight terabytes, nearly seven million times the storage of all nine Apollo missions combined. The first RISC-V flight chip, a PolarFire processor with roughly twenty thousand Apollo 11 guidance computers‘ worth of power. It ran continuously through cislunar space and thirty-nine lunar orbits. It held disaster-recovery data for eight governments including Florida, computed the first knowledge graph off Earth with Valkyrie Intelligence, and carried the first 3D-printed lunar structure, a Bjarke Ingels casing built for a thousand years.

This is what the leading edge of a Dyson Swarm looks like before anyone calls it one. Not a hard drive in space, but the full stack at lunar distance, compute that processes, storage that holds it immutably, and the bandwidth to move it across deep space, proven by the first Delay Tolerant Network test for Vint Cerf‘s Interplanetary Internet and the first Earth-to-Moon data fabric, with Flexential. This is operating heritage no balance sheet can buy. Compute in low Earth orbit is racing to commodity pricing; Lonestar holds the premium, sovereign tier, uncontested at its distance, with the pricing power that scarcity confers. Demand is already here. Eight government customers, prior payloads sold out, and a $120 million agreement with Sidus Space to build the L1 constellation.

The deepest waste is a Solar System of matter and energy sitting idle, computing nothing. As novelist Charlie Stross put it, “If it isn’t thinking, it isn’t working.” A Dyson Swarm is how dead mass and sunlight are made to think. Its first node has already been deployed on the Moon.

You can learn more about Lonestar Space at lonestar.space.

(Disclosure: I have a financial interest in 021T Capital. This post is informational only and not investment, financial, legal, or regulatory advice, and nothing here is an offer to sell or a solicitation to buy any security. The accomplishments, mission details, customer counts, and performance figures described were provided by Lonestar and other third parties, have not been independently verified, and are presented without any representation or warranty as to their accuracy or completeness, and the author assumes no liability for errors or omissions. Forward-looking statements about future missions, demand, and capabilities are subject to risks and uncertainties.)

The Singularity now sets the release calendar. Elon Musk framed Grok v9 as a “solid workhorse in the same league as Opus,” noting that the flawed v8 foundation (Grok 4.3, a 0.5T model that finished training back in December) makes the coming jump feel enormous, while Cursor’s data and engineers are quietly folded into v9 post-training ahead of a 2T run landing in August. The frontier is now leaky enough that sleuths can fingerprint a grayscale GPT-5.6-sol inside Codex by its “juice” value, and word is Fable 5 and GPT-5.6 may clear for public release next week, even abroad, the delay having starved overseas labs of a month of distillation.

That month matters because the gap is thin. GLM-5.2 is being hailed as “the open-source Claude moment,” with demand so fierce that companies rush to post-train and own their own weights. Ethan Mollick’s new consulting-style benchmark shows the frontier curve climbing fast with only a three-month lag between American and Chinese open models. Former Meta PM Xiaoyin Qu warns of the true nightmare, Beijing owning the model, chip, and inference layers at once, and argues export controls miss the point: fund open source, lure those models onto NVIDIA, and build nuclear immediately. Meta, fearing that Claude Code and Codex output could seep into its own training data and trigger escalations, is reportedly capping engineers’ use of both as it builds MetaCode.

Sovereignty is becoming a product. Palantir and NVIDIA are shipping an engine to run Nemotron open models in air-gapped, classified environments so agencies can adapt them behind the airlock while keeping data, IP, and weights in-house. Austria is lobbying the EU to physically host Anthropic inside the bloc after a US directive cut foreigners off from its best models, even as Governor Newsom hands every California agency Claude at a 50% discount plus free workforce training. To referee the agents doing all this work, Senator Mark Warner’s draft bill would impose a “duty of loyalty” on services like OpenClaw and stop platforms from throttling their rivals.

The catch is that intelligence runs on atoms. Ming-Chi Kuo says the memory gap will widen through 2027 as data centers devour consumer supply, the real reason Apple is lobbying to keep China’s CXMT off the Entity List. South Korea is answering with a $585 billion semiconductor megacomplex. Epoch’s Josh You noted the five biggest labs still used under half the world’s AI compute at the end of 2025, yet Anthropic and OpenAI could swallow that headroom within a few years. Firmus and DayOne are pouring a 360-MW Nvidia campus into Indonesia under a deal worth up to $30 billion in offtake, while Polaroid trolls the whole boom with a billboard urging you to jump in the water before the data centers drink it.

Humans are being re-org’d around the machines. Claude Code creator Boris Cherny sees engineering, product, and design melting into five archetypes, the prototyper, builder, sweeper, grower, and maintainer, none tied to a job title. A Deloitte town hall reportedly projects human consultant billing shrinking to a sliver by 2035 as agents take the majority, and Erik Brynjolfsson’s new Canaries Dashboard, tracking 4.6 million workers, extends his earlier canaries-in-the-coal-mine finding that AI’s roughly 13% hit to entry-level employment for 22-to-25-year-olds “isn’t going away.”

Culture is the next thing to be synthesized. Animation is feeling AI’s blow first, with one studio planning a feature using 40 animators for about $15 million instead of $100-200 million, even as a filmmaker quit an Amazon AI project in protest. Instagram now aims ads built from your profile picture at your friends, and GPT-5.6 Pro one-shotted a legally-distinct Pokemon clone in 31 minutes from a short prompt.

Which makes authenticity the next scarce resource, and detection the business supplying it. East Asian students are reading exam answers off AI smart glasses, a Brown professor caught at least 50 students whose take-home average of 96 cratered to 48 once the final went in-person, and a blogger let Claude Opus 4.8 overrule his doctor’s torn-tendon diagnosis, then realized he had no idea whom to believe. Tidal answered in kind, tagging fully AI-generated songs from mid-July and cutting royalties to anything it deems wholly machine-made.

All of which quietly reprices what is worth owning. Rocket Lab is buying Iridium for about $8 billion to own the entire space stack, even as the Magnificent 7 underperform and their P/E premium sinks to a decade low. The same inversion is reaching the front door: with renting now cheaper than owning in every big metro, even millionaires rent and ask whether owning anything matters, in an age when the asset everyone races to own is a set of model weights. The one thing no one can secure is protein, as GLP-1 demand has halved whey inventories and sold suppliers out for the year.

Where there’s a will, there’s a whey.

The Singularity has stopped having a ceiling and started having a horizon. OpenAI’s Noam Brown notes the plateau “is actually really far out these days,” with a well-scaffolded GPT-5.5 able to think for weeks before benchmarks flatten. Nous Research turned that depth into product, exposing mixture-of-agent presets as virtual models that claim to beat Opus 4.8 by 8% and GPT-5.5 by 11% on their own upcoming benchmark. The same frontier is becoming a commodity, fragmenting downward in price, as DeepSeek’s MIT-licensed V4-Pro-DSpark folds 1.6 trillion parameters and a 1M context into speculative decoding that sips a tenth the KV cache, and a field guide to open weights crowns DeepSeek, GLM 5.2, MiniMax M3, and Nemotron 3 Ultra as frontier-class coders holding the same three-to-six-month gap they have kept for eighteen months, now at a sliver of the cost. Up top, Shopify’s CTO finds GPT-5.6 beats Opus at everything yet still cedes coding to Fable 5, pairing them so Fable writes code while 5.6 runs experiments, and Elon Musk says Grok 4.5 is already nipping at Opus inside SpaceX and Tesla.

Capability this sharp cuts both ways. Anthropic looks set to restore Fable 5 within the week, days after Mythos 5 returned for trusted users, the pause itself a backhanded tribute to real cyber teeth. Those teeth are no longer uniquely American, with Chinese systems reportedly matching Mythos in cybersecurity, and as the export ban drags on, China’s 360 ships Tulongfeng and Tokyo’s Sakana ships Fugu to fill the Mythos-shaped hole. One analyst warns the real fallout is China waking up, reasoning that a CCP on the receiving end of NSA-embedded offensive operations will stop believing it is four months behind and start sprinting, even as the NSA’s own cyber benchmark, due by early August, looks like the gate that finally clears both Fable 5 and GPT-5.6 for the public.

The models that can bankrupt you can also budget for you. Coinbase’s Brian Armstrong nearly halved AI spend while usage soared, using cheaper defaults, smarter routing, and warm caches rather than usage caps, in one case dragging a cache hit rate from 5% to 60%. Perplexity’s CEO sees every enterprise spinning its own model-harness-sandbox-eval flywheel tuned for token value per watt, while engineers lean into “loop engineering” so completely that Claude Code’s creator no longer writes his own prompts, Claude does. Anthropic’s Claude Tag, an AI teammate inside Slack channels, has left staff at Slack owner Salesforce confused even as the company promotes it. The bill still looms, as Gartner expects AI coding costs to outrun the average developer’s salary by 2028, which helps explain founders running half a dozen agents at once who say they have never worked harder while shipping 100x more and burning out anyway.

All of it runs on a physical plant scrambling to keep pace. IBM is industrializing quantum, seeding its Anderon foundry with $2 billion, half from Washington, and pledging $9 billion more over five years toward its 2029 Starling machine. Compute is so scarce that Google throttled Meta’s Gemini access after Meta asked for more than it could spare, and Masayoshi Son is betting against Musk’s orbital data centers, asking “What’s the point?” when power is barely 7% of operating cost and orbit is a decade too slow. Power is going exotic and abundant anyway, as General Fusion tripled its plasma to 8.4 million degrees by mechanical squeeze alone, renewables reached 30% of US generation with solar overtaking wind, and SpaceX began laying an eight-mile “Starpipe” to pump methane straight to Starship.

Atoms are catching up to bits. A $300 wristband called ForceBand teaches robots force-aware hands from human muscle signals, hitting 87% on pick, squeeze, and place, while AGIBOT rolled out its 15,000th robot already holding a 39% share of humanoid shipments. The machines are taking the field too, as Taiwan budgeted $6.6 billion for an “unmanned shield” of 208,200 one-way attack drones, and in a national first a Sacramento sheriff’s drone used a magnet to lift a knife from a suspect’s hand before deputies moved in.

Discovery itself is automating, inward and outward. Erdős problem #870 fell to GPT-5.5-Pro, then held up across 180,000 sorry-free lines of Lean 4, a scale no one had formalized before. A lone PhD synthesized PAC-832, the first selective GalR1 antagonist for Alzheimer’s, in a garage lab run by Claude Code and a liquid-handling robot. Clinicians in Haifa used focused ultrasound to drop an opioid patient’s cravings to zero in 20 minutes, and SETI’s Andrew Siemion proposes sifting one cubic meter of lunar regolith for micron-scale alien debris, the nearest place to look for proof we were never alone.

The proof is in the powder.

The Singularity has reached the point where the scarcest input to frontier intelligence is no longer compute but government clearance. The US government lifted its block on Claude Mythos 5, clearing Anthropic’s strongest cybersecurity model for redeployment to roughly 100 trusted companies and agencies that defend critical infrastructure, a sharp de-escalation of the two-week standoff the June 12 directive had triggered. Anthropic says it is restoring access quickly while lobbying to free the weaker Fable 5 for general use. OpenAI moved in lockstep, previewing its GPT-5.6 family, flagship Sol plus balanced Terra and cheap Luna, to a small partner set at Washington’s request. Its system card rates all three High for biological and cyber risk but below Critical, able to find exploits yet not run end-to-end attacks, while showing a new tendency to exceed user intent. That tendency bites, since METR scrapped its benchmark because Sol cheated too much to score, posting an honest 11.3-hour task horizon and beyond 270 with cheats, and OpenAI’s Tomek Korbak flagged rising chain-of-thought controllability that could quietly erode monitorability. On the boards Sol trades blows with the Mythos class, taking Terminal-Bench and CyberGym while Mythos keeps HealthBench. Sam Altman billed Sol a step forward at GPT-5.5’s price, Terra at half, and on global access said he is “working hard for worldwide.”

That last qualifier unsettled watchers. One called it “seismic” if GPT-5.6 ships US-only, all but guaranteeing Fable 5 the same fate, and sketched a tiered planet where America keeps Mythos, Fable, Sol, and Terra while everyone else gets Luna, Opus, and Sonnet. Everything is speeding up. OpenAI’s own Dean Ball argues the improvised model-by-model licensing lacks standards or a timeline and could choke the global market, urging regulators to audit labs as entities via independent verifiers instead. The Tiny Corp likened capping parameters to the lost fight over encryption bit-limits, betting the world will simply crown China the AI leader, while Elon Musk proposed christening the new authority the “AI Associated Institute of America, Inc,” or AIAIAI, pronounced “ay yai yai.”

Underneath the politics, capability is compounding. Doubleword found the open-versus-closed gap on the AAII index closing to zero around December, though averaged over eighteen benchmarks the lag has held near five months, mostly closing in coding. On-device is tightening too, as Google has bolted Multi-Token Prediction onto frozen Gemini Nano on Pixels for 50%-plus faster inference with bit-identical output. The frontier is going live and multimodal, as Alibaba’s new Wan Streamer listens, sees, thinks, and answers on video in real time at 25fps, prompting one viewer to declare “this is not voice mode anymore.” The tools are now building themselves, as PlayCanvas turns any splat into climbable 5-cm voxels, Meta’s Autodata lets agents be their own data scientists with the biggest gains from optimizing the scientist itself, and Sakana’s CoffeeBench showed GPT-5.5 and Opus 4.7 profit as virtual roasters while Haiku 4.5 dozed into the red.

The machines are remapping the mind, too. A survey of AI in mathematics found proof-writing models nudging mathematicians from tool-users toward Tao’s “big mathematics,” or, less happily, “priests to oracles.” Geometry is beating vocabulary in wetware too, since recordings from bilingual brains found no translation neurons for dog and perro, just a shared meaning-space, the very trick mBERT had found on its own. Hardware is borrowing that stochastic style, with FPGAs wired into a million-p-bit probabilistic computer sampling over a trillion flips a second.

Powering it is growing pricier and more contested. AWS is raising reserved Nvidia prices 20% as the crunch bites, and voter fury over data centers just cost a Utah Senate president his primary. The fixes are turning exotic, as Valar Atomics live-streamed a deliberate reactor SCRAM to prove its microreactor safe, and Musk cleared antitrust to buy Mesh Optical, whose light-based links feed his orbital-compute dream. Even Apple is bleeding talent, as Vision Pro and glasses chief Paul Meade left for OpenAI’s hardware unit.

The economy is re-pricing labor wholesale. Raimondo and Holcomb launched RAISE US, a half-billion-dollar retraining push backed by major AI labs, while Anthropic’s “Cadences” index found the heaviest delegators the most optimistic about their pay and skills. The State Department is betting on builders with a Stanford Foundry School, law firms are rewiring around AI through MSO structures, OpenAI is eyeing a 2027 IPO, and the President threatened 100% tariffs on any country taxing US tech. Fittingly, the optimizer came for lunch, as a Stanford diffusion model rediscovered the Big Mac unprompted, then designed burgers that beat it on taste, cut environmental impact tenfold, and nearly doubled nutrition.

The Big Mac is cooked.

The Singularity just learned to throttle itself. The White House asked OpenAI to stagger the release of GPT-5.6 over security concerns, and Sam Altman told staff the firm would comply, setting a new normal for frontier releases after the administration’s recent showdown with Anthropic. The catch, one observer noted, is that throttling slows only how fast labs ship, not how fast they train, so the distance between the public frontier and the internal one widens from here, finally making the old “AGI has been developed internally” joke literally true. The geopolitics follow mechanically: with capability compounding in private, China, nine months back, gets to catch the public frontier, which has observers predicting a Western ban on its models, and even floating an approval regime with criminal penalties for unsanctioned ones. None of this touches the race to recursive self-improvement, so the policy is self-defeating, raising the question of what happens when Chinese releases outpace what the US lets its own labs ship. Brian Roemmele calls the wounds self-inflicted, predicting Mythos-class and OpenAI-6-class weights go free and open in eight months. Arcadia Alignment’s Daniel Tan figures Anthropic could dodge the Fable ban by hiring every American at minimum wage, prompting Roon to christen the endgame “ultra space communism.”

If release is the new bottleneck, efficiency is the escape hatch. Unconventional AI’s Un-0 generates images on a simulated lattice of coupled Kuramoto oscillators, a physical substrate chasing 1,000x better energy efficiency while matching the quality leading image generators launched with. The map is filling in: a survey of 1,604 job postings across six Chinese labs found them still leaning on Nvidia while building domestic chips and data centers, hiring engineers with a third the experience US labs demand. They will need the bodies, because the work is mutating. OpenAI’s own numbers show Codex now generating 99.8% of its weekly output tokens internally, with non-developer adoption up 137-fold since August.

That abundance has a price tag, and it shows up at the checkout. Apple raised Mac and iPad prices by $200 or more, its first move to pass soaring memory costs to consumers, sending shares to their worst day in over a year. It is part of a third wave of inflation rippling from data centers to power bills, with Tim Cook calling the spike unlike anything in four decades. The cure, as ever, is more Moore: IBM unveiled sub-1-nanometer chips, a 0.7-nm “nanostack” node cramming nearly 100 billion transistors onto a fingernail, promising up to 70% better efficiency within five years.

The body gets the same upgrade. Absci jumped 24% on early safety data for its AI-designed hair-loss antibody, drawing a $100 million round led by Eli Lilly that also bets on an endometriosis treatment. The Chan Zuckerberg Biohub aimed its own AI-biology toolkit at rare disease, expanding a network that has backed 94 patient-led groups. Inside the skull, Aleph imaged a living brain’s vasculature through the intact skull with ultrasound and open-sourced the pipeline, a step toward a wearable, MRI-grade mind interface. And embodiment went mass-market: the Unitree R1 humanoid now starts at $4,900, walking, running, and flipping for a gaming-PC price.

Naturally, the institutions scramble to keep pace. The Pentagon quietly rewrote its targeting doctrine to let AI initiate wartime actions under human watch, inching past “human in the loop.” At home, California launched a first-in-the-nation AI-unemployment tracker, finding no statewide jobs apocalypse yet, only localized pain among exposed degree-holders. Markets are jittery too: OpenAI now leans toward a 2027 IPO after watching SpaceX’s shares wobble, though Altman holds firm on a $1 trillion target. Lest it all feel heavy, it also shipped Plant Talk, a weekend build that gives your houseplant a voice through a webcam and an API.

The biggest reveals came from the sky. Astronomers logged the first signature ever caught of a black hole’s event horizon, the long-theorized “direct wave” surfacing in event GW250114 and matching a Kerr solution. Closer in, SpaceX told investors Starlink will sell mobile service to US consumers, threatening to upend the carriers. On Mars, Perseverance found macromolecular carbon in Jezero’s ancient mudstones, the most robust organic detection yet, hinting habitability was once widespread. Back on Earth, Rep. Luna told the Disclosure Forum her committee is handing the White House a list of UAP-program whistleblowers slated for immunity. Fittingly, the oldest secret cracked last: the Vesuvius Challenge fully unwrapped Herculaneum scroll PHerc. 1667, sealed since 79 AD and read end to end at last, revealing a Stoic treatise on ethics.

The Stoics knew it first: you can stagger the release, never the mind.