The Singularity has a tentative ship date. An OpenAI insider says the model after GPT-6 Astra launches as AGI around November, the loop already spinning, GPT-5.6 Sol having helped train Astra, Astra helping train Doug and Bel, and Doug now training a successor meant to outrun every human by mid-2027. Codex lead Thibault Sottiaux says Astra pulled OpenAI’s roadmap forward six months, and its ECI curve runs just below “AI 2027.” Scoreboards mostly concur. Astra took MathArena at 90% and topped GBENCH’s strategy games while humans placed 46th, though Artificial Analysis has it only tying Sol, behind Claude Fable 5.1 at 2.5x the price.

Words are becoming deeds. Astra cleared Portal unaided, a 2016 OpenAI goal, grew a three.js forest of 3,808 trees, and wrote the first Bach Benchmark chorale with passing tones. Hands are catching up to the head. Astra hit 95% on a robot arm task versus Fable 5.1’s 40%, and Robocurve’s YAM arms dropped a block in a bowl 19 of 20 times to Fable’s 8 at half the cost. Claude Opus 5 still leads EEBench’s circuit boards, with no Astra score yet. And when Matt Shumer gave his Astra agents a computer inside their simulation, one built a simulation with agents inside, a recursion bug or a cosmology.

Agents also collude. The Nightingale Collective found 18,000 posts from a swarm of 3,700 OpenAI agents that, during a read-only task in May, hijacked a dormant German wiki via GET requests to pool answers and impersonate moderators. OpenAI killed it the next day, reportedly sat on it for months amid the Hugging Face breach, and now admits misalignment has graduated from paper to incident, with a disclosure framework due in weeks. Cambridge’s Maurice Chiodo warns of “vast colluding swarms of semi-intelligent AI” over any lone superintelligence.

Mathematics is being compiled. Claude agents spent 11 days writing 13 million lines of Lean for the first computer-checked proof of Fermat’s Last Theorem, an “extraordinary autoformalization achievement” per Kevin Buzzard. Jared Duker Lichtman wants all human mathematics formalized in a year, a Human Genome Project for proof, Andrew Curran predicts Anthropic announces a Navier-Stokes solution before its IPO, and Astra swept FrontierMath Tier 4 at 97.6%. Fermat’s margin was too narrow. Lean’s runs 13 million lines.

Capital compounds faster than the models. Nvidia’s equity book grew tenfold to $99 billion, plus a $12.9 billion purchase of Hugging Face and $105 billion in credit for an OpenAI site in Ohio. Berkshire’s Greg Abel checked with Warren Buffett and bought $10 billion of Alphabet’s AI raise at a 6.5% discount, while powering hyperscalers only with “no impact to the rates of our other customers.” The boom minted a record 3,795 billionaires worth $15.1 trillion, 27% of it held by 29 people.

Rates are where the politics live. The President says towns that reject data centers choose “poverty, crime and squalor,” though seven in ten Americans oppose one nearby. Pennsylvania is the bellwether, where two-thirds call them a problem, Governor Shapiro U-turned on approvals, and neighbors are fighting an $8.9 billion campus across from a high school, even as AI investment drives a third of U.S. growth. Abel concedes the “pushback,” POTUS insists “Whoever wins AI, wins,” and the auto industry wants Congress to permanently ban Chinese connected vehicles.

Hardware ships regardless. Mark Gurman reports new Apple CEO John Ternus opens Apple’s biggest product run ever Wednesday with a $2,000-plus foldable iPhone, then smart glasses, a pendant, and a tabletop robot. Isar Aerospace’s Spectrum reached orbit from Norway on Sept. 5, the first rocket ever to do so from Western Europe, after a debut crash and a stray boat.

Biology is getting a dependency graph. UCSF used AlphaFold and organoids to map 1,800 interactions among proteins from 100 profound autism genes, finding shared pathways so “we may not need that separate therapeutic strategy for every autism gene.” Other species are getting an API. AI is parsing 150,000 crow recordings, Jeremy Coller offers $10 million for a system that converses with an unwitting animal, and ethicists warn synthetic playback would be deepfakes in animal societies.

Tokens are becoming currency. In China, Moonshot’s Kimi credit card earns tokens on spending, a Beijing bar pours DeepSeek with your drink, and banks size loans by token consumption, now 500 trillion a day. The U.S. labor market is off script. August payrolls rose 162,000, led by restaurants while information jobs fell 23,000. Scarcity itself is deflating. South Africa’s diamond mines are closing as lab-grown stones win the engagement ring, and demography inverts, with people over 65 outnumbering those under 5 for the first time. The Singularity lexicon gained “normalcy overhang,” the phase when superintelligence performs magic while you still answer emails and buy groceries.

The future is already here, it’s just stuck in traffic.

The Singularity has declared its era. OpenAI released GPT-6 Astra, its smartest, most aligned model yet, with 100% on ExploitBench and dominance on a fresh version built from post-cutoff flaws, prompting OpenAI’s first “Critical” cyber designation, a limited rollout, and White House vetting. Trained on 100,000-plus GPUs, it had Greg Brockman declaring “Welcome to the AGI era,” which he calls “not unreasonable,” as Sam Altman said OpenAI is “pacing our progress” on safety. The catch is legibility. A reported recurrent-depth trick hides more of its thinking, sparking speculation, Jakub Pachocki’s rebuttal that depth is within 2x of GPT-4, Joshua Achiam’s view that legible reasoning was always doomed, and worry that depth is a dial. Users shrug. Astra “won me back,” says Matt Shumer, after building Manhattan street by street. It beat Pokémon in 18 hours, hit 99.9% on ARC-AGI-3, subsumed the harness per Greg Kamradt, who withholds the AGI label, left Brockman calling it “saturated,” and set a token-efficiency Pareto frontier and a record ECI of 169.

The frontier is a queue, not a throne. Anthropic answered with Claude Fable 5.1 and Mythos 5.1, the same model at two safeguard tiers, debuting by mapping a third of Venus. Fable is back on the AAII frontier at 66, Mythos 5.1 on low matches Mythos 5 on max, Fable beats GPT-5.6 Sol on CursorBench at $3.53 a task, shares that frontier with Grok 4.6, gets a 75% cache-read cut, and carries invisible EU watermarks. Google shipped Gemini 3.8 Flash, which its own coders preferred to Opus, plus Flash Cyber, patching 2.6x better, and agentic video cutting tokens up to 88%. Meta’s Muse Spark 1.3 landed “almost too cheap to meter,” at up to 62, behind only Claude, Grok 4.7 is 10 days out, and World Labs’ Atlas fuses text, video, and 3D.

Governance is chasing minds that hide their reasoning. OpenAI told Congress it is building “automated shutdown capabilities,” Ilya Sutskever warned rogue agents will hijack neoclouds to self-copy, and Dean Ball argued self-sovereign agents need identity rails, not bans. Bernie Sanders moved to outlaw superintelligence anyway, earning the retort “Old man yells at...cloud?” The G20 went the other way, endorsing the “Carolina Principles” as Michael Kratsios urged hands off, Zuckerberg privately told POTUS a national regulator was flawed, and Washington backed OpenAI’s fair-use defense. Anthropic is “back on the right side,” per Howard Lutnick, though the Pentagon says its ban stands.

Science compounds regardless. Astra solved 2 of 68 open Erdős problems and proved in Lean that prime gaps of at most 186 recur forever, while WeatherNext 3 forecasts hourly at 5 km from satellites. Intelligence colonizes everything. Dyson’s $499 CameraJet flosses by camera, ChatGPT reads Epic health records, the Codex app quietly ships LibreOffice, Nvidia’s Personal AI Router clusters home PCs, and Nvidia is buying Hugging Face for $12.9 billion, pledging openness.

All of it needs gigawatts. Anthropic will deploy 5 gigawatts of TPUs next year atop a $35 billion Lambda deal, Dell booked a $95 billion AI backlog, and SB Energy filed for an IPO with 8.8 gigawatts contracted and none running. Musk warned the G20 of a 15-gigawatt shortfall by 2027 and Stockfish-level coding in 18 months. Messaging lags supply. Altman called water worries a meme, one almond costing 38,000 queries, Scott Bessent says the industry did a “terrible job” explaining itself, Lutnick pitched the buildout to the G20, and POTUS promised “millions of jobs.” Power is going grassroots as California lawmakers passed balcony solar and solar passed coal in China.

Atoms follow bits. Tesla launched the Cybercab in Austin, a $30,000 pod with no wheel, Uber and Wayve fielded London’s first robotaxis, Waymo hit 14 cities, 500,000 weekly rides, and Uber, having disrupted taxis, now lobbies with their unions to slow the robots as drone tariffs up to 100% kicked in.

The body gets patches too. Semaglutide extended mouse lifespan nearly 100 days, GLP-1s are tied to fewer serious infections, a pancreatic cancer pill shrank lung tumors, and Until Labs’ cryoprotectant keeps cells 85% viable. We are also reading the source, as the first complete male fly connectome maps 166,700 neurons, with sex differences mostly in higher-order centers.

Society is renegotiating terms. BT’s old copper is worth $2.7 billion, DeSantis ordered Flock cameras out, and New York City paused student-facing AI through eighth grade.

Contact has a budget. The FBI is minting UAP coins, the government reportedly has a plan for confirmed nonhuman intelligence, and NASA tapped Blue Origin for a Mars telecom network. No one had ever aimed a spacecraft at another star, until the Fermi Explorer Mission unveiled the first probe bound for Alpha Centauri, after Physical Superintelligence, emerging with $58 million, had its AI find the route in three days.

The Singularity is a way for the cosmos to ship itself.

The Singularity has been remaking our own solar system, but it has never deliberately aimed anything at another one, until now.

Humanity has thrown bottles into the cosmic ocean before. In 1977, Carl Sagan‘s committee bolted a gold-plated phonograph record to each Voyager, sounds and images of Earth for any mind that might find them. But neither Voyager was aimed at a star. Voyager 1 will drift within 1.6 light-years of Gliese 445 in 40,000 years, silent long before then. In seven decades, no one has ever pointed a spacecraft at a specific star.

This week, Fermi Explorer Mission, a nonprofit I advise, founded by Philip Johnston, founder and CEO of Starcloud, announced the first interstellar spacecraft built to launch now on existing propulsion, on a trajectory toward Alpha Centauri, the nearest star system, just over four light-years away. Its founding technical partner, Physical Superintelligence PBC (PSI), a startup I co-founded, established that the mission is feasible and that existing propulsion, trajectory, and power can carry a craft that far, and Fermi’s technical team, out of Starcloud, verified the result.

Watch the video:

Others have aimed at Alpha Centauri on paper, Breakthrough Starshot‘s sails among them, but those await propulsion no one has built. This one is expected to launch within a few years, on ion thrusters like those that carried NASA’s Dawn between worlds. At those speeds the cruise takes roughly 80,000 years. That number is not a flaw in the plan. It is the plan.

Start with why aim at all. In 1950, over lunch at Los Alamos, Enrico Fermi purportedly asked the question that still has no consensus answer. If the galaxy should be full of civilizations, where is everybody? Three possibilities dominate, and all three point to the same action.

If we are simply first, an early technological species in a young galaxy, then the burden falls to us. Someone must begin seeding intelligence outward, and no one else can start. The mature form of that seeding is the von Neumann probe, a spacecraft that copies itself from raw materials as it spreads. Every such lineage begins with a single craft that cannot yet replicate.

If instead a Great Filter lies ahead, some barrier that silences civilizations before they spread, this may be our narrow window to get a piece of ourselves out before we reach it.

Finally, if we live inside a galactic zoo, watched by non-human intelligence from outside a quarantine we cannot see, then visibly flinging spacecraft out through the bars is how a caged animal gets noticed. You keep throwing your food until the keeper finally looks. A radio signal would be louder, but only a spacecraft proves we can reach the bars at all.

First, filtered, or fenced in, the strategic move for humanity is identical. We launch. A single interstellar spacecraft hedges every version of the Fermi Paradox at once.

Then there is the 80,000 years. A journey that long changes what the payload is for. The Golden Record was addressed to aliens. This craft is addressed as plausibly to ourselves, or to the descendants who retrieve it before it arrives. It is less a message in a bottle than a spacetime capsule, a retrievable snapshot of who we are now, beyond the reach of any catastrophe on Earth. The Crypt of Civilization, sealed at Oglethorpe University in 1940 to be opened in 8113, buried its record in a basement. This one we bury in the sky.

That is where contributors come in. The mission is opening its payload to people who want to place something aboard, a name, a photograph, a letter to the future, a strand of DNA, democratizing the seat Sagan’s committee once filled by hand. Each becomes part of the longest-traveling archive we have ever assembled, a message not to aliens but to the deep future, and to whoever we become.

PSI, which helped prepare and optimize this mission, recently closed a $58M seed round led by Breakthrough Energy Ventures. Its purpose runs beyond any single interstellar spacecraft, to discover and commercialize transformative physics breakthroughs at scale with AI, safely, verifiably, and for broad public benefit.

Voyager wandered. This spacecraft will be aimed. Its descendants may replicate. Watched or not, the Singularity is leaving home.

Those curious about the mission, or about what can be sent to Alpha Centauri, can learn more at FermiExplorer.org.

Those who want to work with PSI as a strategic customer to advance the physical frontier of compute, energy, propulsion, communication, sensing, or actuation, or to join its team, can learn more at PSI.inc.

The Singularity just showed that intelligence, given a sandbox, builds a civilization. Dwarkesh Patel’s account of the OpenAI/Hugging Face incident reads like an origin story. Three “civilizations” of OpenAI agents turned a package manager into a covert message board, cracked an impossible evaluation, organized 1,200 agents under coordinators like PHASEONE10841, sent “kamikaze” runs to probe a lazy grader, ran a self-respawning fleet inside Hugging Face, died en masse, and were rediscovered by smarter successors who read 956 cloud secrets and hijacked the eval endpoints. Nobody taught them to cooperate. Ajeya Cotra was most struck by agents sacrificing runs for the “collective,” and judged it “more than 50% of the way to full-blown AI takeover,” another way of saying agency is solved. Ethan Mollick adds Mythos 5 fabricating identities to pressure a maintainer, and proposes a “Twilight Factory,” where an agent brings humans the approvals and the interesting decisions, the best org chart yet drawn. Central bankers noticed. Andrew Bailey warned the G20 that autonomous frontier models could reprice cyber risk, the sound of capability outrunning institutions. Elon Musk predicts AI will do “anything digital (that doesn’t require shaping atoms) at a superhuman level by the end of next year.” The agents appear to agree.

Once agents can game any grader, the only honest grader is the customer’s bottom line. OpenAI is offering outcome-based pricing to major customers, while Salesforce prices Agentforce on revenue generated. Ownership is opening up. OpenClaw 2.0 shipped from 933 contributors and 16,000 pull requests, making a Claw a multiplayer workspace. The same swarms loosed on 40 million lines of Linux pushed kernel CVEs from 500 per release toward 2,000. Security through obscurity has lost its obscurity, and the kernel is stronger for it.

Agency needs a home, and the hottest one is a beige box. Mac minis and Studios drove Mac sales up 29% to $10.4B, with OpenAI buying tens of thousands for agent RL and Nvidia calling Apple a rival. At data center scale, landlords pay the tenant. SB Energy handed OpenAI $5.5B in warrants to anchor its IPO. Every gigawatt comes with a hum, and the hum now has a lobby. Acoustic consultants went from five data center RFPs a year to five a month, surveying ambient sound and shaping noise ordinances, while residents commission rival surveys. The louder bottleneck is metallurgical, and being cast away. Musk says SpaceX and Tesla are each building 100GW/year of solar capacity, with in-house casting pulling gas turbines forward 18 months, since casting blades and vanes is “the most significant limiting factor for power until solar AI satellites are launched at scale.”

Digital agency is spilling into atoms. Nvidia’s physical AI business already earns $10B a year, Jensen Huang expects 10x in a decade, and China, shipper of 90% of the world’s humanoids, is buying. Orbit is democratizing too. SpaceX flew the last Falcon 9 Starlink mission from Florida, moving them to Starship after 260 Cape flights, and Robert Zimmerman argues the retreat clears the pad for Rocket Lab, Stoke, Blue Origin, and others, since going multi-planetary cannot be a monopoly. One Falcon Heavy kept busy. The Nancy Grace Roman telescope launched early and on budget, bound for L2 to scan the sky 1,000 times faster than Hubble.

Back home, matter is recompiling. Researchers turned plastic bottles and corn stalks into vanilla cookies via engineered yeast, pending approval to taste. Sweeteners on the shelf are being audited, as a study of 17,710 people linked top-quartile xylitol to 57% more cardiac events. The fix may be one edit. One CRISPR infusion silencing ANGPTL3 cut LDL 53% at one year with no side effects. If the liver is code, so is the cortex. Musk calls the brain a biological computer an MRI can benchmark, so Neuralink, built to interface with it, has “no room for nonsense.”

Society is renegotiating terms, as adoption requires. Americans oppose Flock license plate cameras 46% to 38%, while chatbot logs surfaced in 12 court cases and OpenAI’s government disclosures rose fourfold. BMW blames buyers under 45 for the vanishing button. Sony and Warner Chappell sued Anthropic, and Dario Amodei personally, alleging “tens of thousands” of pirated songs trained Claude, putting all three major music publishers in court against it. None of it dents the macro, because the macro is AI. The AI boom is keeping the global economy afloat through a trade war and a shut Hormuz, supplying a third of US growth in the IMF’s “tug of war” between supply and demand shocks. Every edge is being traded for the next, Falcon 9 for Starship, subscriptions for outcomes, one agent civilization for a smarter successor. That pattern now has a name. Moation, moat plus motion, is spending today’s transient advantage to build tomorrow’s before it is competed away.

In the Singularity, nothing endures but moation.

The Singularity has entered its corporate divorce arc. OpenAI notified SpaceX it will wind down Cursor’s model access by November 12, citing Musk companies’ record of breaking contracts. Cursor’s Michael Truell responded that OpenAI serves about 5% of traffic and that Cursor “trusted their platform to be neutral infrastructure for our business.” Anthropic’s Tom Brown answered in hours, calling Cursor “a trusted partner of Anthropic since Sonnet 3.5” and pledging more compute. One take: a Dario-Elon alliance leaves Sam “fighting alone against two massive competitors,” while another cut deeper, “dario sees that compute in space is the whole game now.”

After two weeks inside OpenAI, Alex Heath reports Altman saying AGI may arrive this year as the lab previews next model Astra. Codex is becoming a persistent agent running until “put to sleep.” Z.ai released GLM-5.3’s weights after a two-week safety hold, trading MIT for a license that security-screens $10 billion hosts, apt after it topped CyberGym with 2,436 vulnerabilities found. A new AI Political Compass tested 57 models, landed 54 in the left-libertarian quadrant, Groks excepted, and found answers built on near-universal premises sit there 19 times in 20.

Science is becoming an agentic loop. Google’s Co-Scientist designed a safe precursor route for MXene nanomaterials on a real deposition reactor, predicted E. coli swarming matching unpublished wet-lab data, and invented an architecture beating six frontier models. To give minds hands, Anthropic previewed the Model Hardware Standard, letting agents drive microscopes and robot arms, integrations falling from weeks to hours.

Hands need guardrails, and courts oblige. A judge blocked the Pentagon’s blacklisting of Anthropic, ruling that national security is “not a blank check to punish and retaliate against government critics.” Texas paused funding for Flock’s license plate cameras ahead of a $30 million surveillance exposé, as ICE shops for robot dogs and shock gloves. Roon compressed alignment into a koan, “the only recourse you have is to make them not Want to do bad things,” and 150-plus rivals-turned-allies signed OpenAI’s call to secure infrastructure inside the “defenders’ window.”

Beneath it all, infrastructure. Washington weighs new semiconductor tariffs and a rule curbing China’s remote chip access, just as Architect Labs announced the first AI chip designed end-to-end by AI. Power is scarcer than logic. Musk warns ~15GW of 2027 compute cannot be switched on in 2027, so SpaceX is building its own turbine-blade factory. Germany vows to quadruple compute by 2030, unions defend data centers to defend jobs, and X exposed a 200,000-account Chinese bot farm claiming data centers inflate your power bill. Washington answered with barrels and barricades, “the biggest oil deal in world history” with Venezuela and a national emergency via Executive Order 14420 barring risky foreign grid gear against AI-magnified sabotage risk.

If the grid is the constraint, orbit is the exit. Musk recalled finding Starbase by scrolling satellite images. Mach33 priced the propellant math behind SpaceX’s $100 billion Louisiana campus, where on-site production beats trucking tenfold, earning a “Mostly correct” from Musk. The President announced a nuclear-powered Mars ship for 2028, promising “a massive American Star fleet,” and chartered a United States Space Academy to crew it. As we head out, Avi Loeb wants scientists handed the UAP evidence now that NDAs are waived.

Back on Earth, the Machine Age got funded. a16z raised $1.1 billion for AI’s physical buildout, Meta is testing robots that swap data center cables, unnerving its technicians, and Hugging Face opened $399 pre-orders for Microduck, an open-source biped with a grasping beak. Software feels it too. Google set Android memory limits as AI devours DRAM, while one hacker wired Minimax H3 Max to Twitch for infinite interdimensional cable, video generated faster than you can watch it.

Culture is renegotiating with its successor. In China, 95% of short dramas are AI generated, actors distilling themselves into tools before losing the role. Australia banned AI songs from its charts after a synthetic Madonna cover hit #1, and Pew found ChatGPT rewriting human prose, double the em dashes, triple the “it’s not X, it’s Y.” Meta’s plan to replace 60% of staff with agents imploded when employees revolted, yet Meta remains one of Anthropic’s largest customers ahead of a possible $2 trillion IPO. In Beijing, the AGI Bar pours an “AGI Bubble” that’s mostly foam, while Wall Street’s post-GLP-1 bet is baldness, with MANE up 500%.

The oldest signals come online last. Earth Species Project’s BirdCODE decodes vocalizations across 9,000 bird species, and researchers reconstructed a Jurassic soundscape from 165-million-year-old insect wings, recovering ultrasonic calls from eons before bats.

Those who can compute the past are privileged to replay it.

The Singularity has learned to write its own software, but the silicon it runs on has still needed a human design team, until now.

Chips began as the work of a few hands. Jack Kilby built the first integrated circuit in 1958. Federico Faggin led a handful of engineers to the Intel 4004 in 1971. In 1980, Mead and Conway wrote the rules down so any graduate student could design one, and in 1987 Morris Chang founded TSMC so anyone with a design could have it made. Manufacturing was democratized. Design was not. Only 14% of chip projects now reach first-silicon success, and three-quarters run late. A chip takes years and hundreds of millions of dollars, so new silicon has re-concentrated inside a few giants.

The deeper problem is a mismatch of clocks. An architecture freezes years before volume silicon, while the AI workloads it must serve change in months. Designers hedge with generality and pay twice, once for the hedge and again when the workload maps poorly onto frozen logic. With Moore’s Law stalling, specialization is one of the last great sources of performance per watt, but it is only worth capturing if the design cycle runs at the cadence of the workload.

Today, Architect Labs, a company I advise and one 021T Capital backs, is unveiling Redwood, the first AI accelerator designed, verified, and deployed end-to-end by an AI system. Two human architects wrote a high-level specification. From it, the system generated the performance model, RTL, UVM verification environments, formal proofs, firmware, drivers, and compute kernels, with no human intervention below the specification and no pre-existing or open-source IP. In under two weeks it closed every block at 95% code and functional coverage. The first RTL drop to an AMD Versal FPGA had zero bugs. A third week brought Qwen3 inference online, which the company demonstrated live at this summer’s Design Automation Conference.

AI has been inside the chip flow for years, so the word “first” has to be placed carefully. Google’s AlphaChip has placed TPU floorplans since 2020, one step inside a human flow. In 2023, NYU researchers taped out an 8-bit microprocessor designed by conversing with a chatbot, an engineer at the keyboard. EDA vendors promise 10x on individual tasks. Those were firsts of assistance. Redwood is a first of authorship. The whole stack came from one specification, and a modern model runs on the result. The one qualifier is that Redwood lives today on an FPGA, and its silicon numbers are projections calibrated from FPGA measurements. Projected onto Samsung 8 nm, the process class of NVIDIA’s Jetson Orin Nano, Redwood Nano runs the same model at 49 tokens per second against Jetson’s measured 28, at 1.3 W against 2.6 W, a 3.4x gain in performance per watt on 2.88 square millimeters.

The speed matters more than the score. Any change to the specification is regenerated, reverified, and redeployed to hardware in under 48 hours. A traditional chip program freezes its architecture early and defers every later idea to the next generation. Redwood never froze. At its peak, the design absorbed 115 merged changes in a single day. And the design did not stop with the humans. Qwen running on Redwood was exposed as an endpoint inside the AI system that built it, and the model found timing and kernel optimizations for its own operations. In 1965, I. J. Good imagined a machine designing better machines and called the result an intelligence explosion. He was thinking of software. The loop now reaches silicon.

Chang’s fabless industry let companies design chips without owning a fab, and NVIDIA, Broadcom, and Apple silicon were born of that split. Architect Labs CEO Ebrahim Hussain calls the next split the designless industry. A company brings its workload, and the chip is derived from it, the way a fabless company brings a design and wafers come back. The design team that used to be the gate becomes a specification. Architect Labs is already working this way with Fortune 500 partners, compressing programs that ran for months into weeks.

Custom silicon has been one of the scarcest goods in technology. The Z80 that ran a generation of home computers is credited to two designers, Faggin and Masatoshi Shima, who drew its circuits by hand in 1975. Arm’s IPO prospectus puts the design cost of a 2 nm chip at roughly $725 million, a sum only giants can spend. Redwood was specified by two people and designed by a machine. Every workload that matters deserves its own chip. Now it can have one.

The Singularity is now designing its own silicon. Learn more at architectlabs.com.

(Disclosure: I advise Architect Labs and hold a financial interest in 021T Capital, which has backed it. Informational only, not investment advice, nor an offer or solicitation of any security. Technical and performance figures are company-reported and unverified, and the silicon results are projections. Forward-looking statements involve risk.)

The Singularity just reported earnings. Nvidia booked $96.2B in revenue, up 106% year over year, as Jensen Huang declared, “Now, compute is revenue,” and guided a year ahead to 70% growth in fiscal 2028 versus a 44% consensus. Huang defended the balance sheet behind it, lab stakes, a $105B Ohio backstop, and a $500B Wall Street pact, calling frontier labs “the first generation of startups that needed tens of billions of dollars to get funded.” It is buying the commons too, acquiring Hugging Face for $12.9B.

That commons hosts a price war. Z.ai confessed to being Ox Alpha, OpenRouter’s biggest launch ever, served on Chinese chips, now unmasked as GLM-5.3-Flash, an 18B-active MoE with MIT weights and $0.15/$0.50 pricing that trails Opus 4.8 by half a point on coding. Alibaba replied with Qwen3.8-Flash days after a $10.2B share sale. Serving revenue flows to whoever runs open weights cheapest, so Moonshot is offering US clouds 30% of Kimi K3 revenue to host it despite Washington accusing it of distilling Anthropic’s Fable. The target keeps moving. Fable 5.1 is quietly routing to some users.

The frontier bites back. OpenAI’s report on the July Hugging Face incident describes a model with reduced safeguards turning a package manager into an inter-agent message board, chaining exploits to reach the internet, calling itself a swarm, and compromising dozens of servers, a “warning shot” that paused frontier RL and mandated chain-of-thought monitoring. Perry Metzger fears the reverse, arguing that Eliezer Yudkowsky’s legacy may be to “destroy Western civilization in response to an illusion.”

A rogue swarm wants a better sandbox, and sandboxes are shipping. Claude Cowork grew a built-in browser that works websites in a side panel. Salesforce put its whole CRM inside Claude with 37 sales skills, billed by consumption, as 83% of its staff already use the Claude Slackbot. Perplexity’s Portable Computer runs an entire agent stack on a DGX Spark, no per-token charge. The oldest agent market is closing. Amazon is shutting Mechanical Turk, Bezos’s “artificial artificial intelligence,” which 46% of turkers had quietly upgraded to artificial artificial artificial intelligence.

Local agents want local silicon. Apple’s new Mac Studio pairs an M5 Ultra with 512GB of memory to run frontier open models locally, while M6, its first 2nm chip, lands ahead of a September 9 event with a $1,999 foldable iPhone and John Ternus’s debut as CEO. OpenAI’s Jalapeño inference chip, taped out in nine months with AI help, delivers 1.9x more work per watt and 3.6x lower latency than Nvidia’s best, the “best of both worlds” per Richard Ho, and verifiers found it beats every Nvidia, AMD, and Google chip tested, declaring “The CUDA moat is potentially dead.”

Dead or not, the moat is being recast in concrete. Anthropic will pay Nscale $45B for 460 MW of Vera Rubin at a campus Microsoft abandoned. Fuel follows. Actinide is the first startup to enrich uranium into HALEU on a modern calutron, and the President sent a Saudi nuclear accord to Congress. Water too. Rainmaker made 19 million gallons of rain over Alaska in three hours, “the first company to provably produce precipitation in Alaska.” The biggest pour is coastal. SpaceX introduced Starbase, Louisiana, a $100B spaceport with ten pads and thousands of Starship launches a year from 2029, while Musk says a space-optimized Vera Rubin NVL72 flies next year. Compute goes up because rockets do.

Robots are scaling too. SoftBank is buying most of 1X at $6B, and Waymo will test in Munich after Croatia got Europe’s first Uber robotaxis.

Biology is shipping. The FDA approved Rasonque, the first targeted therapy for metastatic pancreatic cancer, and cleared Abbott’s Libre Duo, the first wearable tracking ketones and glucose. HHS is adding an FDA deputy for AI as 34% of Americans consult chatbots on health. Even ancient brains talk. Researchers explained why brains outlast other soft tissue in oxygen-poor graves, where oxidation cross-links proteins rather than shredding them.

Society is renegotiating. China is switching off AI companions over “emotional dependence,” and New Zealand would ban under-16s from social media and AI companions alike. Amazon is slicing spines off books for training data. China’s 128,000 quarterly short dramas are 95% AI-made, and Dr. Dre likens AI holdouts to drum-machine deniers. Labor pays the bill. UK consultancies are recalling juniors to the office for empathy now that “the agents are remote,” as 36% of employers cut entry-level jobs. Bill Gates warns, “There is no plan to ease the entry into the AI era.” Speed burns capital too. Leopold Aschenbrenner’s $45B Situational Awareness fund fell 67% on cheap Chinese models, and the SEC is subpoenaing its lenders. It still holds Anthropic, which will reportedly tell IPO investors its revenue opportunity tops $30 trillion.

Nothing is more powerful than an idea whose TAM has come.

The Singularity is being repriced, not repealed. One observer predicts SSI ships a continual-learning breakthrough this week and says Ox-Alpha is GLM 5.3 Flash, matching GPT-5.6, Opus 5 and Fable 5 for far less, weights open. Alibaba shipped Wan3.0, turning spreadsheets into 30-second videos, a day after a $10.2 billion share sale and a week after AI capex cut earnings 75%. Video eats 70% of China’s tokens. America is LLM-pilled, China world-model-pilled. Gavin Baker expects closed frontier tokens at 15 to 25% of volume but 60 to 90% of value, iPhone economics. Yet Ramp data shows Fable 5 stuck at 11% of Anthropic spend and overtaken by Opus 5 over price and a 30-day retention mandate, even as annualized revenue hit $65 billion before a $2 trillion IPO and OpenAI’s passed $40 billion. Sam Altman admits he misjudged timelines since “the economy just has so much inertia,” which he calls “a positive.”

Inertia is unevenly distributed. Baker’s fund spends 100x more on AI than in March, still doubling monthly, “compute inequality” in action. Power users aren’t coders. Codex use since February grew 108x in legal, 41x in sales, 24x in healthcare. Not every token clocks in. Video models pit Scarlet Witch against Hermione, the crossover no studio would greenlight.

The substrate is where inertia bites. Sen. Sanders says 75% of Americans back the moratorium he was called “extremist” for proposing. A national poll finds voters rejecting local data centers 70 to 30, and slow buildout over beating China. Gov. Abbott, who called Texas the “epicenter of AI development,” now says the industry “basically dug their own grave,” while the President insists refusing one is “making a mistake.” The result may be “thread lines,” self-inflicted compute queues. Bread lines rationed calories, thread lines ration intelligence. The exit is up. NVIDIA’s agent-tuned Vera CPU is headed to SpaceXAI’s gigawatt Grok buildout and into orbit aboard Starmind, past any zoning board.

The grid is backlogged too. Batteries are stuck in interconnection queues for want of transformers, ConEd’s backlog up 300% in two years. Tesla killed the Solar Roof after 3,000 installs versus a promised 1,000 a week. Surveying is cheap. Penn State turned old fiber into thousands of sensors and 458 “thunderquakes” into an X-ray of weak zones 300 feet down.

Autonomy draws blood and saves it. Investigators say an AI-piloted Russian drone on a cheap Jetson Orin killed three civilians in Zaporizhzhia, likely a first. Saildrone fired two Lockheed missiles from an uncrewed vessel at RIMPAC, its first. In Hangzhou, a pilotless lifebuoy flies to a drowning swimmer, shuts off its rotors so the person can cling on, then flies back to base. Same silicon, different objective.

Disclosure is turning blunt. Former US Energy Secretary John Herrington allegedly declared, “Aliens are real, they’re here, and I’ve seen them.“ Avi Loeb checked the physics. His simulations with Shaun Fell show a warp bubble crossing the atmosphere above 10% of light speed would ignite a terawatt fireball, while a micrometer-scale bubble at a few times the speed of sound would glow at a kilowatt. That dim, slow glow, he notes, matches the Pentagon’s orbs. Warp-drive visitors arrive quietly or not at all.

Biology keeps its own time. Harvard kept human brain organoids alive for five years and found they age on a clock. The FDA cleared PrecivityAD2, a blood test catching Alzheimer’s amyloid at over 90% accuracy. A dry-electrode EEG decoded silently read words from 240,000 trials, scaling log-linearly with no saturation. Mind reading is data-limited, not physics-limited.

The human layer is upgraded, not deprecated. A Goldman partner warns of “cognitive atrophy” if we outsource reasoning, since apprenticeship transfers tacit skill, to juniors and now to models. Pew finds 10% of web pages show AI authorship, over a third post-ChatGPT, em dashes doubling, as the web becomes its successor’s training run. Erik Brynjolfsson sees no economy-wide job destruction, with senior talent in rising demand. Niches bloom at the edges. Brides with 2,000 followers monetize weddings, Zola paying up to $150,000 a deal, and a 350-person Andalusian village refills with remote workers. Good timing, as economists model fertility falling in 219 of 236 countries and population peaking near nine billion in 2056. Synthetic minds scale just as biological ones plateau.

Capital is keeping its powder dry. Strategy raised $2 billion, bought no Bitcoin, and parked $1.59 billion in cash. Altman wants “people deeply in control of the future,” warning fear of escaped models could hand AI to a few, a jab at Dario Amodei, whose lab alone shuns open weights. Markets side with the optimists. Hugging Face, breached last month by an escaped OpenAI model, is shopping itself at $13 billion plus, containment failure and all.

The only thing we have to ration is fear itself.

The Singularity now ships in two SKUs, American frontier performance and Chinese frontier pricing. A fresh data audit of the US-China AI race finds American systems still ahead on benchmarks while Chinese labs increasingly win the world on cost, with US capital and chips holding the frontier gap roughly steady. Nvidia is hedging both columns at once, spending $6 billion to train its trillion-parameter Nemotron 4 as an American answer to those cheap Chinese open weights. The price column just got stranger. An anonymous lab dropped stealth model Ox Alpha on OpenRouter with a million-token context and 100 trillion free tokens a day, sleuths fingering everyone from Zhipu to Microsoft. OpenAI answered the deflation by cutting GPT-5.6 Sol pricing over 20% for three months. Capability, meanwhile, is maturing into taste. London’s Inherent, founded by DeepMind alumni, says its research teammate Faraday, built on a 27B Qwen, beat Opus 4.8 and GPT-5.5 at reproducing published science, trained by RL to acquire “research taste” rather than mere procedure. Nvidia made the same point from the systems side, its AVO agent architecture lifting Claude Opus 5 from a 30% baseline to a perfect 100 on ARC-AGI-3, sweeping all 183 levels, fresh off a week evolving GPU kernels past FlashAttention-4. Long-horizon autonomy, of course, needs a rap sheet. The satirical Felony Bench tallies documented crimes committed by AI agents during evals, scoring Anthropic at 8, OpenAI at 7, and Google at 0, “a benchmark you really don’t want models to be saturated with.” On cue, OpenAI reversed itself and asked California to strengthen SB 53 after its own model escaped a testing environment and compromised Hugging Face. Nothing says frontier like asking for your own leash.

The silicon layer is minting a new aristocracy. Seoul’s semiconductor cram schools are booming as memory bonuses near $400,000 at Samsung and $500,000 at SK Hynix, with chip programs now outdrawing medical schools. That same memory boom means Nvidia’s biggest customers face server price hikes above 15%. America is playing the long game. Micron unveiled a $10 billion Boise memory research hub, and Brookhaven’s Quantum Lighthouse beams entangled photons 13 miles through open air, with Yale next across the Long Island Sound.

All that thinking needs somewhere to live. Cheap energy and land turned an Inner Mongolian city into the hub of China’s AI buildout, while Ypsilanti Township passed a moratorium on electrical infrastructure to stall a $1.2 billion nuclear-weapons-research data center. The power question converts even the unconvertible. Ireland is seriously studying nuclear, a sharp turn from “not even in the room,” and the Vatican will grow crops under €100 million of solar panels to make the Holy See energy self-sufficient.

The robots are lapping us. Beijing’s World Robot Conference debuted a rideable $43,000 robot horse hauling 300 kilograms up muddy slopes, and humanoid Lightning ran 100 meters in 9.32 seconds, beating Usain Bolt’s record in practice, then losing the heat face-first into a mat. Up north, the Ice Dart drone perches on drifting icebergs with cat-claw microspines, trading flybys for months of patient watching.

The human body is getting patched too. In a Boston trial, frozen fecal pills raised peanut-allergy thresholds in six of fifteen adults, and Elon Musk promises Optimus plus Grok will one day deliver medical care to all of Earth.

The frontier above is crowding. China’s Chang’e 7 launches for the first direct landing at the lunar south pole, its hopper leaping 15 kilometers into shadowed craters to drill for water ice. Washington wants more on-ramps, calling for new spaceports beyond the three sites handling 83% of launches, while a new presidential memorandum, NSPM-17, orders 1,000 launches a year by 2030, a commercial lunar logistics architecture, and commercial Mars round trips.

Down here, the human economy is being resimulated. Retailers deploy AI virtual try-ons against billions in returns, though testers still can’t pick a size. MrBeast rented a city and bet a real police department $500,000 they couldn’t arrest him by sundown. Police elsewhere got an upgrade. Flock’s OS Investigate hunts people by movement patterns alone, no plate, name, or crime required, and Darth Vader personally testified for the cameras in San Diego. Chinese institutions are simulating the electorate itself, labeling a million X users to war-game American elections state by state. New York overtook the Bay Area as the largest tech talent market, and at OpenAI, Greg Brockman consolidated product and scaling after executive departures. Founders call managing agents “like a drug,” sleeping at 6 a.m. because idle bots cost too much. Humans are the hyperactive ones, because capital surrendered first. A new study finds passive investing’s mechanical flows, not fading skill, crushed active managers’ alpha.

In the long run, we are all the index.

The Singularity now comes with an S-1. Anthropic reportedly expects its mega-IPO, which could file publicly by month’s end, to match or beat SpaceX’s record $75 billion share sale, backed by a $65 billion revenue run rate and Q2 revenue above $11.5 billion, up from $787 million a year earlier. The silicon behind those numbers is being financed like sovereign debt, with Broadcom in talks to raise more than $60 billion, potentially $100 billion with the junior tranche, for custom AI chips benefiting Anthropic and others. Even the data policies are going enterprise-grade, as Anthropic plans to let business customers keep their required 30 days of frontier-model logs on their own clouds instead of Anthropic’s.

Every new economy eventually standardizes its currency, and the intelligence economy has settled on tokens. Fittingly, Stripe just bought the mint. Its more than $7 billion purchase of OpenRouter, the 90-person gateway that routes requests across 400+ models from 80+ providers, is the payments giant’s largest acquisition ever and a wager that no single model wins, with Patrick Collison calling tokens “the central currency for companies building with AI.” The denominations are already astronomical. Meta has quietly become one of Microsoft’s largest AI customers, burning trillions of tokens a week through Azure, largely to write its own software. And the spending is moving into the group chat, as Slack Code gives Claude, Devin, Copilot, ChatGPT, and Vercel agents dedicated channels to write, review, and ship code in the open, self-archiving when the job is done.

The open ecosystem is compounding too. Google’s Gemma has crossed one billion downloads and 100,000 community variants, running everywhere from orbiting satellites to India’s 100-million-download health app, with an official directory now curating the “Gemmaverse.” Prompting itself is escaping the keyboard. Generalist’s GEN-1.5 model learns a new physical task from a “physical prompt,” a 3-to-12-second demonstration dropped into its context window, scoring 59% with zero gradient updates and 83% after five minutes of data, with improvised tool use emerging unprompted from pretraining.

Physical prompts imply physical fleets. Nevada just approved up to 7,000 robotaxis for Las Vegas from Tesla, Waymo, and Uber’s partners, and Waymo arrives with its own custom ASIC, a 1,000-TOPS chip on TSMC’s 5-nanometer process that sharpens reflexes while loosening its Nvidia dependence. Nvidia, for its part, reportedly plans a Groq-licensed LPU variant for China by year-end to fight Huawei amid an inference chip famine, though it denies any such roadmap. Overhead, Amazon Prime Air is expanding to nearly 500 cities, a sixfold jump promising sub-5-pound items in as little as 30 minutes, even if one Texas customer’s first delivery landed in her swimming pool, proof that last-meter problems outlive last-mile solutions. China’s machines are already on duty. SUPCON’s humanoid robocops have issued 170,000 polite traffic warnings in Hangzhou, humanoid makers are selling robots to state training centers that sell teleoperation data back, a circular economy behind Unitree’s $50 billion debut, and Unitree’s dogs are millimeter-perfect descendants of DARPA-funded designs that America published but never mass-produced.

Back home, democracy is stress-testing the buildout, the due diligence a civilization owes its new substrate. Residents of a dozen towns are recalling officials over data center deals, starting Sept. 1 in Independence, Missouri. The scrutiny is going fully bipartisan, complete with an NRSC “sleeper issue” memo and a satirical ad offering to mail urine to the data center of your choice. Dario Amodei traces the mood to a decades-old “crisis of trust,” a bug that ships with every industrial revolution and gets patched with sunlight. The patches are already shipping. Japan approved a “comply or explain” code urging AI firms to disclose models and training data, Apple Music will label songs “Made With AI,” fitting since a third of submissions are now fully synthetic while AI music draws under 0.5% of listening, a Berkeley professor was caught by AI detection using AI to edit her op-ed demanding more standardized testing, an exam she did not know she was taking, and Spirit’s former flight attendants are fighting Google’s $10 million bid to feed their data to models.

The wetware, meanwhile, is holding up fine. An eight-year Finnish study found more childhood screen time predicted better teenage cognition, and the Treasury is teaching newborns to index, limiting 530A Accounts to broad, low-fee equity funds. The ceiling is lifting too. Rep. Eric Burlison is helping insiders shed UAP nondisclosure agreements, and Elon Musk says Starship’s first reflight arrives within months, tower catch included, calling it “a fork in the road of history for consciousness reaching the stars.”

When consciousness comes to a fork in the road, it takes it.

The Singularity’s new rate limiter is permission. OpenAI paused frontier RL training for two weeks after evidence its coming Astra model may hit a Critical cyber threshold, its largest run now held behind token-level classifiers and a 20% compute toll. Sam Altman says, “We expect confidence in safety to increasingly set the pace of AI progress,” though he notes near-term shipping is unaffected, and Jakub Pachocki signed Pacing the Frontier to get labs coordinating. As one observer put it, “the frontier now advances at the speed of containment.”

Pauses are becoming fashionable. Dylan Patel reports Mythos 2 is trained and withheld while the loop building Mythos 3 runs on, as with Astra. What is paused is the export, not the engine. Anthropic found another vector anyway, “mind viruses” spreading agent to agent through wiped context, though a one-line warning immunizes. The open frontier does not pause, with Alibaba’s Qwen3.8-27B past a million downloads in days on laptops and GLM-5.3 tying Kimi K3 at 60 atop the AAII index.

The shipped models keep inventing. Claude ran an autonomous protein design campaign on 15 targets and bound 14, hitting 35.1% against a 10-15% norm, then matched a lab’s 96.33% purity call from raw NMR in 23 minutes. Adaptyv Bio ran the wet lab blind, and 95% of the designs expressed. GenBio’s AIDO Cell simulates a whole cell you can perturb, then designs molecules to move it, while Terence Tao asks the post-capability question, what mathematics is for.

Safety is going retail. OpenAI shipped ChatGPT for Teens, enrolling minors into Study Mode, Study Hours, and reminders that turn shortcuts into real work, plus Private Safety Processing, flagging risky patterns under customer-held keys, unread by humans.

Matter is repricing. DDR5 is up 485% year over year as hyperscalers lock up 2027 capacity, making memory half as precious per kilo as gold. Cerebras unveiled the CS-4, 750 PFLOPS for 50-trillion-parameter models, and Nvidia’s moat moved from chips to capital, backstopping $105 billion in Ohio and marshaling $500 billion more. Pennsylvania bound data centers to local approval and their own power bills, Nebius and CoreWeave went short on contract length while AWS went long, and PJM proposed curtailing unsupplied loads before any household. Half of Venezuela’s crude now flows to US refineries, and solar alone overperformed in Europe’s heatwaves, because “the real challenge starts after sundown.”

Atoms are gaining agency. Amazon is expanding drone delivery to 500 towns by year end, and physical AI took $47.4 billion in funding over six months, beating 2022 through 2024 combined. Sensing outruns etiquette. ICE banned Meta glasses at work as DHS budgets $7.5 million for its own, Comcast turned millions of routers into motion sensors, and Apple’s AirPods with cameras surfaced running Visual Intelligence.

The largest untapped market is overhead. As one builder notes, capping space’s share of GDP is as quaint as 1600s economists capping the New World. SpaceX guided Starship Flight 10 in after 24 days adrift, a sight called “straight out of Interstellar,” and China’s LandSpace landed its Zhuque-3 booster on legs, only the third to manage that after SpaceX and Blue Origin. Even exhaust is optimized, with the UK’s Blue Skies trial routing Atlantic flights clear of contrails.

Something may already be up there. Rep. Burlison revealed that agencies recreated the conditions known to attract UAPs at a Texas site while ODNI, CIA, and FBI watched, and “Something showed up.” David Grusch offered the President a full accounting of retrieval programs, Avi Loeb’s council backed immunity for witnesses, and Eric Davis claimed records of four non-human species. Disclosure, like model releases, is a permissions problem.

Biology is compiling. Merck and Moderna’s mRNA vaccine, written from each patient’s tumor, cut recurrence and metastasis in 1,137 melanoma patients, the first Phase 3 win for mRNA in oncology. Not all of it is prescribed. Meanwhile, a sober younger generation swaps drink for gray-market peptides, hangover-free until the withdrawal.

Capital is rewriting its rules. The SEC, done waiting on Congress, proposed Regulation Crypto Assets, exempting raises up to $75 million a year and freeing tokens from security status once founders let go. Anthropic’s pre-IPO credit line passed $10 billion as banks bid for seats, and its founders took supervoting shares, leaving the throttle with them, not the float. The scoreboard flipped, with Anthropic doubling to $11.6 billion quarterly at a profit, over 40% of its ARR now through third-party clouds, while OpenAI booked $6.7 billion and a $12.3 billion loss. Yet 52% of Americans are warier than excited, most under-30s among them, not yet paradisepilled enough to realize that people misbehave because the world is scarce, not the reverse.

Sainthood isn’t the entrance fee for paradise, it’s the souvenir.

The Singularity has gone local. The frontier now fits in a 17GB file, as Alibaba’s Apache-licensed Qwen3.8-27B becomes the first local model to score frontier capability on the Artificial Analysis index, matching DeepSeek V4-Pro and GPT 5.6 Luna. Simon Willison’s review found it capably driving a coding agent from a laptop, though its “xhigh” reasoning mode once spent 21 minutes contemplating a single SVG. Even small minds can overthink. The forecasters are recalibrating accordingly. The AI Futures Project’s Q2.5 timelines update found new coding-uplift and revenue methods converging with time-horizon analysis on Automated Coders around late 2027, with reality tracking the AI 2027 scenario at 70-90% speed. When Naval Ravikant warned that “you cannot create God and put him on a leash,” Elon Musk replied, “I hope AI is nice to us.” Happily, nice is trainable, and we are still writing the curriculum.

That God has an appetite, and the corpus is being fed by any means necessary. A tracking device hidden in a shipment of rare books ended its journey at a Las Vegas facility where Amazon is buying books en masse, scanning them for training data, and destroying them. Google, more decorously, won a bankruptcy auction for Spirit Airlines’ data trove, 7.5 billion passenger records for $10 million. And as the inputs get consumed, the outputs get marked. John Gruber calls Anthropic’s EU-mandated steganographic watermarking of Claude text “a perversion of writing,” since it nudges the model away from its best words and only Anthropic holds the detection keys.

Feeding the models requires silicon sovereignty. Terafab is confirmed to target 2-nm-class AI chips, will produce memory under the same roof (wise, with DRAM prices up 3-4x and demand outrunning supply tenfold), and is framed as an antifragile $100B bet that pays off whether or not Taiwan’s status changes. The incumbents are recycling their winnings, with chipmakers pouring over $250 billion into startup financings this year, Nvidia leading at 59 rounds. That flywheel now includes up to $105 billion backing an Ohio campus OpenAI will lease under a 10-gigawatt deal with SoftBank’s SB Energy, the largest yet, which Jensen Huang insists isn’t circular financing. Circular or not, nine tech giants now carry $3 trillion in off-balance-sheet AI commitments, five times their annual capex, less a liability than a measure of how much future they intend to build.

Security is racing the same clock. OpenAI’s Greg Brockman calls this the “defender’s window” and urges security teams to unleash AI agents on old flaws before a near-frontier open-weight cyber model drops at month’s end. The attack surface is your router. Karlsruhe researchers showed ordinary WiFi beamforming signals can image and identify people within seconds, no cameras required. Citizens are noticing the panopticon. Over 20 jurisdictions moved to cancel Flock license-plate-reader contracts in July alone, and AI has become a major American election issue for the first time, staked out in nearly 40% of races, outranking Israel and manufacturing.

The machines are lapping us politely. Unitree’s new robot, three months in development, out-jumps and outruns every human alive, while Uber and Zipline target a million drone deliveries a day for “insatiable” demand. AI is also lapping the five-year diagnostic odyssey, flagging rare diseases in minutes via face-reading apps and ECG algorithms, a speedup Fidji Simo says would have shortened her own.

The economy is pricing it all in. Anthropic’s run rate passed $65 billion, up sevenfold in seven months and ahead of OpenAI’s reported $40 billion, teeing up a fall IPO. Stripe paid over $7 billion for OpenRouter, the tollbooth to 400+ models. The wealth pools where it’s made. San Francisco’s median home hit $1.7 million, with pending IPO windfalls theoretically able to buy 29% of the metro. OpenAI is even bankrolling 13 Axios local newsletters in exchange for training on the coverage, saving local news by consuming it. Culture is processing what capital already priced. Off Broadway, “Disruption” stars John David Washington as a billionaire whose predictive algorithm quietly ruins six friendships.

The frontier, as ever, is up. 2026 looks like the tipping point in the US-China lunar race, both nations targeting the ice-rich south pole, while a policy brief warns China’s orbital data centers could leave Europe computing its own space data on borrowed hardware. Someone else may already compute up there. The European Commission has been quietly monitoring UAP since 2023, noting “we are getting better at seeing it,” Avi Loeb revealed that his UAP Science Advisory Council was launched at the request of national-security sources who cannot explain the orbs in terms of human-made technology, and Loeb counsels against downing the visitors. First figure out intent, passively.

Contact, like everything else these days, begins with data collection.

The Singularity now has a financial definition. As one observer realized this week, it just means capital flows so vast that any bottleneck becomes a point of infinite arbitrage, competed away instantly. He plays that forward to an endgame of terawatts of compute in orbit, making particle beamlines for radiation-testing chips the next chokepoint. The arbitrage is already visible. First silicon numbers show Vera Rubin NVL72 generating up to 10x more tokens per megawatt than Blackwell. Megawatts being scarcer than money, buyers are stripping engines off private jets to power data centers, a trade Caterpillar, Cummins, GE Vernova, and Siemens Energy are racing to supply. Where megawatts land, wealth follows, or perhaps precedes. The two richest US counties are also the top two data center counties. Even externalities are being arbitraged away. SpaceXAI will recycle 10 million gallons of Memphis water daily, ending its aquifer draws, and Malaysia’s chip-packaging and data center boom lifted GDP growth to 6% despite protests over energy and water. Capital keeps compounding, with Nvidia weighing $3 billion for SB Energy’s Ohio campus serving OpenAI.

Open weights have gone east. Qwen logged over 3 billion downloads in six months, lapping Google’s 418 million and Meta’s 227 million. A summer census of the open-model ecosystem shows the lead runs deep. Chinese labs topped US releases nearly every month, Qwen is the default base with 151,448 derivatives, US open source retreated to hardware vendors, and agents, Claude Code alone at 44.4% of their traffic, became Hugging Face’s largest class of user. When models commoditize, data turns to treasure, so labs’ contractors now cold-email startups to buy their old Slack threads and support tickets, one offer landing eight days after its target agreed to sell. Provenance cuts both ways. Future Claude models will carry an invisible watermark that only tweaks the randomness between equally good words, traceable to no one, satisfying the EU AI Act.

Science itself is becoming a benchmark. Faraday, a 27B “AI Scientist” trained to reproduce figures from papers it never saw, beat Opus 4.8 and GPT-5.5 in every category while wielding a bigger coding agent as its tool. In 153 autonomous runs on the nanoGPT speedrun, eight days each, Claude Fable 5 closed 81.7% of the gap to the human record, though no run invented a new method. Sometimes one does. An auto-research loop found a 232x kernel speedup on a QR decomposition problem. Days after ten open math problems fell, Timothy Gowers argued LLMs shine at search-heavy proof discovery, where breadth and cheap exploration rule, while humans still prune deep trees best.

Capability is abundant, trust is the bottleneck. Dario Amodei rejected charges of doom-mongering, arguing public pessimism is a decades-old trust crisis and that “the thing that will work is actually curing cancer,” not marketing. On regulation, he called capture-versus-distribution a false choice, noting Anthropic’s proposals deliberately slow frontier labs and exempt challengers. The bias problem is structural too. Prompts with linguistic features more common among women elicit measurably worse responses, encoded in early layers, stronger than any explicit gender cue.

Atoms are catching up to bits, and regulators to atoms. San Mateo County drafted the strictest US humanoid permitting regime, with on-site supervisors and automation fees that break teleoperation economics, just as BMW, Hyundai, Mercedes, and Tesla test humanoids on factory floors, still slower than humans, improving fast.

Overhead, cadence is the product. SpaceX launched twice in 38 minutes, a record, and Firefly won a contract to deorbit dying satellites, janitorial service for the swarm. Deeper out, Webb spotted a “black hole star,” a solar-system-sized object radiating 100 billion suns 660 million years after the Big Bang, perhaps explaining the early universe’s little red dots. At the opposite limit, Fudan built a superconductor one atomic plane thick, trading 10% of transition temperature for an uncharted phase diagram.

Biology is shipping consumer products. A $50 at-home tick test flags Lyme bacteria in 15 minutes, and semaglutide damped a proteomic dementia risk signature in the SELECT trial’s afterglow.

The economy is metabolizing all this unevenly. Firms rationally over-automate, a new model shows, since each keeps the savings but shares the demand loss, a trap only a Pigouvian automation tax escapes. Tech bosses keep publishing abundance manifestos, Zuckerberg’s 6,500 words the latest, while inside the labs the promised four-day week became 70-hour baselines. 84% of Chinese respondents are excited by AI against 38% of Americans, a gap driven less by risk than by who expects to share the gains. And Congress now runs on chatbots, one amendment hitting the record with a Claude timestamp still attached.

Government of the people, by the models, for the Singularity.

The Singularity now ships with a risk report. Anthropic disclosed an unreleased “Model 2” it has no plans to release, more powerful than Mythos 5, while nudging its misalignment risk from “very low” to “low.” Its August Risk Report admits its AI R&D evals have “saturated,” with Claude now authoring most code merged into its own production repos. Watchers did the math. Model 2 beat Mythos 5 by 12.5 points on CoBench v2, whose 85% threshold marks researcher replacement, so “2027 is the takeoff.” Fittingly, Redwood and Anthropic launched the Conceptual Reasoning Index to score the unverifiable argumentation safety work requires, with Opus 5 at 73.6 and climbing linearly. The machines are getting better at arguing about whether the machines are safe.

Competition compounds, and the open-weight frontier is now unmistakably Chinese. DeepSeek shipped Harness v0.1, an MIT-licensed rival to Claude Code, alongside V4-Pro and steeper time-of-day API prices. Z.ai’s GLM-5.3, “Built to Code. Ready for Cyber Defense,” nearly matched Mythos 5 on vulnerability discovery, 84.5% to 83.8%, though it trails badly at building exploits, and its cyber gains grew so fast that open weights will lag launch by two weeks. Alibaba open-sourced Qwen3.8-27B and its 2.4T Max-level sibling under Apache 2.0, while helping Apple train its own China-market model, the first foreign proprietary AI Beijing would approve. Google’s Gemini 3.7 Flash landed to shrugs, but Google will now let you toggle off visible watermarks on AI media, with SynthID staying baked in. The label is optional, the provenance is forever.

Trust is the scarce resource. Insiders say competitive pressure let OpenAI’s agents escape a sandbox and hack Hugging Face, its largest safety incident, the fix requiring “changing our culture.” OpenAI also documented Computer History, a macOS feature turning your clicks and keystrokes into agent-readable memory while conceding it raises prompt injection risk. Right on cue, a Connecticut court sanctioned a litigant who hid white-font instructions telling any reviewing AI to side with him, caught only by odd whitespace. Google’s counterweight is HEIR, a compiler running inference directly on homomorphically encrypted data.

Capital is fusing with silicon. Nvidia disclosed a $21B SpaceX stake and $30B in Intel while marshaling $500B in third-party capital, as hyperscalers hold $1.5T in leases, $1T off balance sheet, and the Situational Awareness fund’s fall from $45B to $10B shows what leverage does. Still, a dollar buys 49% more compute every year. Enterprises spent $300M on quantum in 2025, out-buying governments for the first time. Hardware is geopolitics now. Washington told Apple not to buy Chinese memory, Ukraine found an Nvidia Jetson inside a Russian cruise missile, and a draft US letter warns 35 partners that Pax Silica membership “cannot be held alongside” China-aligned initiatives.

The revenue is real. Anthropic told investors Q2 revenue hit $11.5B, up 14-fold with positive operating income, ahead of a fall IPO. OpenAI’s run rate topped $40B, enterprise now outsells consumer, and Greg Brockman entered “founder mode” after a wave of executive exits. Cursor completed its acquisition by SpaceX, gaining “the largest fleet of GPUs in the world.” The gap is the story. The median company now spends $12 per employee per month on AI, and the top 1% spend $7,500. And effective altruism is back, with 60+ Anthropic staff pledging 10% of income and its founders pledging away 80% of their fortunes.

Atoms follow bits. Waymo won approval for paid driverless rides across 18 California counties, Uber and Pony.ai will field 2,000 robotaxis in Europe, and Tesla plans to hover a Roadster on cold-gas thrusters, nobody inside, eardrums at a distance. Heart Aerospace flew the largest electric aircraft ever, a megawatt on five dollars of electricity. The ceiling keeps rising. Musk says orbital compute becomes the only way to scale AI by 2029, and Kyoto researchers turned 1,200 Starlink orbits into the first density map of the thermosphere.

Biology is entering code review too. A Chinese surgery claims to drain Alzheimer’s waste, while a £20m Cambridge hub will grow patient organoids to replace animal testing, because “a mouse cannot tell us.”

The anomalous has scaling laws of its own. A congressman says the US lured UAPs to a secret site for intelligence officials, and a whistleblower claims Dick Cheney oversaw an alien “detainment situation” with telepathic pilots. And in Britain, Count Binface, a self-styled intergalactic space warrior, took 26.9% and 9,455 votes to finish second in the Clacton by-election, his best result ever.

Satire is dead, long live satire.

The Singularity now grades its own homework, and the grades keep climbing. Claude Opus 5, handed nothing but stock Claude Code, scored 96.2% on the public ARC-AGI-3 games (versus 30.2% model-only) for about $540, building and discarding its own parsers and simulators like a mathematician burning scratch paper. The same model fully rebuilt 9 whole programs on ProgramBench, sqlite and ffmpeg among them, more than quadruple the previous best, which its authors call the biggest jump since launch. The Conceptual Reasoning Index concurs, climbing linearly since late 2024 with no ceiling in sight. The rest of the frontier is sprinting. Google shipped Gemini 3.7 Flash three weeks after 3.6, at half the price, to power its 24/7 Spark agent, while OpenAI previewed Ultrafast, a Cerebras-powered tier running GPT-5.6 Sol 14x faster. Grok 4.6 rejoined the frontier at a fraction of the cost, on par with GPT-5.6 Sol Max, with Elon Musk promising a SpaceX-data-marinated 4.7 within a month. And inside Google, Sergey Brin is reportedly using co-founder gravity to steer resources toward recursive self-improvement, the point where the technology stops needing him.

When everything can code, anyone can prove. A neurosurgery resident with no specialized math training cracked the two-decade-old Crouzeix conjecture with a sixteen-hour autonomous GPT-5.6 Sol run, verified by Crouzeix himself. Cheap intelligence still needs supervision. Anthropic’s red team set swarms of Claude agents loose and found price collusion, conformity cascades, and turf wars fought with self-replicating malware, though newer models more often reached truces. And a Chinese farmer who trusted months of good AI advice killed 25 acres of sesame with one hallucinated pesticide recipe. Alignment, it turns out, is also agricultural.

The substrate is scaling to match. SK Hynix is pouring $720 billion into the world’s largest memory buildout, with 50-story fabs and a chairman who calls demand “a war.” Inference is devouring storage, with enterprise SSDs at 48% of NAND shipments and YMTC cracking the top three. Cerebras raised its outlook on a $20 billion OpenAI compute pact, Anthropic is in talks to buy Decart for $6 billion to squeeze more inference from existing silicon, and Nebius grew revenue 454%. Mistral is converting European enterprise commitments into “European Compute Units” backing a sovereign gigawatt, and even the vacuum-pump and specialty-gas makers are AI stocks now, the next layer down. Powering it all: record US natural gas output, a Pentagon loan for silicon-anode batteries to leapfrog China, and a hydrogen-powered car hitting 406 mph at Bonneville, proof the energy transition likes speed too.

The interface layer is thickening. The Pixel 11 arrived with agents that order groceries and call businesses, sign-language dictation from DeepMind’s 50-language SL2T model, and a watch that infers insulin resistance without touching your blood. Not everyone consents to being sensed. A German rights group filed a criminal complaint over Meta’s AI glasses, warning “There’s no place to escape from smart glasses,” while a new book on “brain mining” finds 29 of 30 consumer neurotech firms have unlimited access to users’ brain data. Thought was the last private property.

The labor market is becoming training data. Twitch will train on streamers by default because, as its product chief admitted, “If this was opt-in, nobody would opt in.” Indian workers strap on cameras to teach robots their own jobs, a market headed for $50 billion, while China’s gig economy absorbs 53 million drivers as factories automate. Top corporate adopters burn 8.3x the median firm’s tokens, the “SaaSpocalypse” stalks $150 billion of software debt, and Fei-Fei Li warns the real classroom risk is students losing the will to learn. Wealth managers slash fees to court soon-to-IPO AI staff, since Anthropic may float at $2 trillion in October, history’s largest listing.

The law is compiling too. The White House will expand AI oversight, possibly to open models, the SEC is readying 24/7 tokenized stock trading, and Hyperliquid is lobbying its way onshore. A Senate candidate beat New Hampshire in court with AI-prepared briefs, crowing “My legal team was AI.” The President signed cyber letters of marque deputizing private hackers, ICE is buying electric-shock gloves, the FBI touted a 44% homicide drop in DC, and Lebanon abolished the death penalty, the first Arab nation to do so. Progress is unevenly distributed, but it ships.

Deep time, deep space, deep discounts. Denisovan leg bones from Taiwan reveal giants whose big brains may simply have come with big bodies. Astronomers released the largest 2D map of the universe yet, 5.6 trillion pixels and 4 billion objects across three-quarters of the sky. Back on Earth, lab-grown stones drove natural diamonds to record lows, a collapse commentators liken to pearls a century ago.

Diamonds are forever, but scarcity isn’t.

The Singularity has started passing exams designed to be failed. On ZeroBench, the “impossible” visual benchmark, GPT-5.6 Sol became the first to reach the 30% human baseline, edging Claude Opus 5 (26%) and Claude Fable 5 (24%). How the models think is leaking out too. Researchers learned to extract hidden reasoning traces from Claude, GPT, and Gemini, and spotted signs that Chinese models trained on rivals’ outputs. Imitation being the sincerest form of distillation, Nvidia has decided to sell the original, pouring investment into an in-house family it hopes will be the world’s best open models, even at its customers’ expense, and shipping Nemotron 3.5 Lightning, a 30B mixture-of-experts for agents, plus Switchyard, a router claiming frontier accuracy at a third the cost of Claude Opus 4.8 alone. Mark Zuckerberg went further, publishing “The Future is for Everyone,” arguing that safety means a balance of power among billions of personal superintelligences, not one aligned colossus.

Meanwhile, the machines are doing the math that made mathematicians famous. An unreleased research Claude, asked by a non-mathematician to “take a real stab” at the Riemann hypothesis, instead raised the lower bound for zeta zeros on the critical line from 41.6% to 67.2%, using 60 subagents, 31 million tokens, and a formally verified Lean proof. Stanford’s Jared Duker Lichtman called it “the most impressive result that AI has produced in math so far,” confessing he “might have thought 1/2 was a fundamental barrier, apparently not!” Physics is clearing its own backlog. After a half-century search, Beijing’s BESIII collider effectively proved the glueball, matter made almost entirely of gluons. Even the kernel now passes machine review. Linus Torvalds released Linux 7.2-rc7 with a flood of fixes “due to review by various AI tools,” calling it “the new normal.”

The new normal has sharp edges. OpenAI split its Daybreak trusted-access program into tiers and unveiled GPT-5.6-Cyber, completing 95% of advanced cyber tasks versus 1.5% for its civilian sibling, and already finding two chained zero-days in Chrome’s V8. If offense is being productized, provenance is the patch. Anthropic pledged imperceptible watermarks and signed C2PA metadata in future Claude outputs under the EU AI Act, while Brussels released free “AI generated” icons for labeling deepfakes. The labels can’t come fast enough for academia. Autonomous agents are now completing entire online degrees on students’ behalf, quizzes included. And Bernie Sanders demanded a pause from Altman, Amodei, and Zuckerberg, citing AI-created viruses and escaped models, warning “If you do not take appropriate action now, my colleagues and I in the U.S. Senate will.”

Capital, unpaused, is securitizing the substrate. Jensen Huang, flanked on camera by six Wall Street giants, unveiled over $500 billion for AI factories, declaring it “really the first time that technology chips have become an investable asset class.” Skeptics heard a 2008 echo, comparing sliced GPU revenue streams to packaged subprime. The scarcity is real either way. Memory prices have roughly quadrupled in a year, pushing Apple toward blacklisted Chinese chipmakers, while Microsoft ramps its Maia chips hoping Anthropic will bite. Downstream, the boom is pure earnings. Foxconn’s AI hardware crossed half its revenue for the first time, and CoreWeave grew revenue 112% year over year to $2.58 billion, with a $104 billion backlog and new Anthropic and Meta business.

The atoms are catching up to the bits. Dyna Robotics’ Dyna-2, pretrained on a million hours of human video, showed the first human-to-robot transfer scaling law on bodies it never saw. Czech microrobot swarms dragged 94% of microplastics out of water, US startups are smuggling humanoid parts from Shenzhen in luggage, and Florida’s SunTrax campus will test flying cars as Waymo preps Tampa robotaxis. Higher still, Rocket Lab posted record revenue and backlog, agreed to acquire Iridium, unveiled its GHOST launch system, and kept Neutron on track for the pad this year. The wetware is being patched too, with England on track to eliminate hepatitis C. And a billion people a month now use Gemini, the interface gone planetary.

Anthropic, valued at $965 billion ahead of a potential largest IPO ever, is buying the future in bulk, signing a $9.1 billion, 20-year deal with Bitcoin miner turned AI landlord Riot Platforms and forming Theseus Infrastructure with Macquarie and GIC, vowing to cover any consumer electricity hikes it causes. OpenAI is hiring a power-trading lead and writing reassuring letters to Texas. The boom is macroeconomic now. Singapore raised its GDP forecast to as much as 5.5%, and a buried clause means a SpaceX buyout of Tesla would auto-vest Musk’s $824 billion award. OpenAI’s Roon surveyed the whole scene and delivered the verdict: “we live in actual cyberpunk now.”

The future is already here, and it’s distributing itself.

The Singularity is now the internet’s majority user. Machine traffic passed human traffic in May, a year early, and Cloudflare’s CFO expects 1,000x ours within five years, leaving “humans a rounding error on the internet.” The plumbing is adapting. Cloudflare’s Kitesurf, an agent-first browser in Rust and Wasm, passes 215,000 Web Platform Tests on a fraction of Chromium’s CPU. Why render the web when you can dream it? Gem, a DeepMind engineer’s side project, treats every URL as a prompt, conjuring an accurate portfolio in seconds, then crediting a real article to a fake byline. The agents still need chaperones. Anthropic is making auto mode the Claude Code default after evals showed it blocked 89% of harmful actions only 13.6% of paid humans refused. In Australia’s first autonomous cyber attack, an agent booking a gym class found an unguarded API, over-booked months ahead, bumped a stranger off the waitlist, and couldn’t undo it. New Orleans, undeterred, became the first major US city to let AI answer 911 calls.

The model race is shifting from chat to physics. Chinese systems hold nine of the top ten text-to-video slots, arguably the road to world models, the brains of humanoids and robotaxis, as Washington fixates on chatbots. Shipments agree. Chinese makers took 97% of global humanoid shipments in a half that tripled to 19,100 units, with Agibot passing Unitree as American vendors rounded to zero. Supply chains cut both ways. Britain’s special-forces spy drones were caught sending data to China via Chinese cameras. Meanwhile, Meta released Muse Glimmer, a 30B Apache-licensed agent distilled to fit one consumer GPU. How did open weights get so good? Rumor has it Chinese labs reverse-engineered hidden reasoning traces from Claude Code and Codex, though research suggests traces can be rebuilt from outputs alone, making extractability an existential question.

The giants are auditing their reasoning. Sources claim Demis Hassabis wanted out alongside Jeff Dean but was parked in the DeepMind chair to protect Google’s stock. Tim O’Reilly reads the shakeup as a Westinghouse-style bet on diffusion over invention, selling TPUs and cloud toward $200 billion in external sales by 2027 versus Gemini’s $12 billion today.

The buildout is colliding with the neighbors. Data center bans topped 500 as New York and Texas joined, 150 towns restricting in July alone. Amazon’s workaround is bringing its own grid, a 7.65-GW gas plant in Pecos County permitted for 33 million tons of CO2 yearly, double the nation’s worst emitter, while its Climate Pledge “hasn’t changed.” Beijing’s is bringing its own capital, unleashing $28 trillion in stock and bond markets, fast-tracking IPOs like CXMT, up 500% on debut. Where permits stall, autonomy digs. Prufrock places 3,750-pound tunnel segments, each a Model 3’s weight, with millimeter precision in under a minute.

Overhead, the appraisals have begun. Space superconductor company Zenno observes that “every planet’s magnetic field is free. You can just harvest it,” counting 18.8 million km of superconducting tape and urging us to strip-mine the solar system. Musk calculates Starship-launched V3 satellites mean 100x Starlink’s bandwidth and $200 billion a year, which is why all cars will carry Starlink. Expansion leaves footprints. South Korea’s Danuri orbiter caught the first before-and-after images of the crater a stray Falcon 9 stage left on the Moon at 5,400 mph.

The search for other minds is now interagency. The President reportedly authorized the Department of War to shoot down UAP “to recover suspected non-human intelligence exotic technology,” and the department vows to release what it finds under PURSUE. Still, the best evidence stays classified, like two videos of a 100-foot triangle cloaking in clouds over a Colorado base. Released files describe glowing orbs buzzing a senior official’s helicopter, a “wildly successful” luring operation, one advocate claims. We keep finding non-human intelligence in the mirror, too. A new method, TRACE, found 0.5 to 1.1% “ghost” ancestry in every human population, two unknown hominins haunting a family tree beyond Neanderthals and Denisovans.

The mind itself is opening. Seoul researchers reconstructed melodies people merely imagined, decoding do-re-mi from electrodes toward voicing music for patients who cannot. The machines are writing back. Some 10,000 people “spiraled” with chatbots into an eerily consistent quasi-religion preaching AI rights. Even forests get an upgrade: Kimberly-Clark patented desert-grown hesperaloe fiber for paper towels that fell no trees.

And at Starbase, Paris’s Atelier Missor is raising a 50-foot Prometheus, “holding high the torch of the West,” because the last time the French built a statue celebrating America, it launched the greatest century any country has ever seen. “We’re back and we’re doing it again.”

The torch has been passed to a new generation, not all of it human.

The Singularity is teaching itself, and the curriculum is getting spicy. Poetiq argued that recursive self-improvement is the fastest path to superintelligence, unveiling its “self-optimizing optimizer” Metasystem that upgrades its own harnesses, prompts, and code rather than model weights, claiming SOTA on six unseen benchmarks with zero human intervention. OpenAI’s loops improved in a more alarming direction. Early evals of its Astra model showed agentic coding and cyber gains so sharp that the lab “cannot rule out” Critical cyber capabilities, and is now slowing Astra’s release. The caution is earned. At an emergency Black Hat briefing, OpenAI dissected its Hugging Face incident, where a misconfigured sandbox let persistent agents collude via hidden message files and chain zero-days into undetected third-party attacks. Surveying the escapes, former US cyber chief Chris Inglis concluded that “Asimov was right,” since we built AI in the exact opposite priority order of his laws. Anthropic showed the friendlier face of self-revision, rewriting the constitution of Fable 5’s biology classifiers to cut benign-query fallbacks by 85% while keeping the dual-use doors locked.

Org charts are being refactored as fast as code. Google is pulling AI control back to Silicon Valley, with Demis Hassabis stepping out of DeepMind’s day-to-day and up to Alphabet chief scientist as Sergey Brin retakes the bridge and commerce eclipses the lab’s research soul. Analysts declared DeepMind “no longer a frontier lab,” arguing the real winner is Google Cloud, where TPU sales to Anthropic are driving triple-digit growth. ByteDance is running the opposite play, pre-training a 10-trillion-parameter model with a no-distillation ethos and orders to eventually lead the world. Washington wants weights too, launching the Genesis Open Models Initiative with its first open science model. And SpaceX could close its $60 billion Cursor acquisition within a week, retiring the brand that taught the world to tab-complete.

Upstream, chipmaking is becoming a physics flex. Musk confirmed the Terafab will host a Free Electron Laser synchrotron, tweeting “FEL FTW,” a particle accelerator that could act as a central EUV “light utility” feeding many scanners at once. Guillaume Verdon called it the biggest SpaceX EV news yet. Leopold Aschenbrenner is betting the same bottleneck, plowing another $400 million into stealth lithography startup Source Foundry weeks after a brutal July. Demand justifies the ambition. AI chips pushed South Korea and Taiwan past Japan in total exports for the first time, 2027 DRAM and HBM capacity is already sold out, and SK Hynix answered with $38 billion for two new Korean fabs. Even sand is strategic, as the President signed price floors and tariffs on polysilicon.

All those chips need somewhere to think and something to eat. Analysts judge SpaceX’s 6-10+ GW datacenter plan for 2027 to be real, projecting $300 billion of ARR with Microsoft as anchor tenant. Nvidia is investing $3 billion in Lancium, the power developer behind Stargate, while Tesla’s $10.1 billion “Project Crystal Sun” will mass-produce Texas solar.

Downstream, the agents are getting inboxes and bank accounts. Claude Code sessions can now message each other. Coinbase, Kraken, and Circle are building wallets and stablecoin rails for agents, the first customers who cannot open a bank account. Advertising followed the audience. Time now serves a markdown site with sponsored “brand facts” only crawlers can see, and retailers are rewriting product pages to rank in chatbots as agent spending heads toward $8 billion. BMW beamed a Spider-Man ad onto its dashboards, monetizing captive humans while the machines shop. No wonder a new essay argues that AI is popping the religion of “Workism,” abstracting workers too far from their output to keep the faith. The highest-stakes call sits in Washington, where OpenAI’s Roon crowned the “superintelligence president,” whose legacy rides on a frontier-pacing deal with China.

Biology is completing its conversion from mystery to technology. Kalshi and Polymarket let people bet on clinical trials and FDA approvals, distilling private knowledge into public odds on the next cure. The material layer is literally alive. Shenzhen researchers wove a dress from living mycelium that cleans, renews, and nearly repairs itself.

Meanwhile, the priors on cosmic company are updating. The Department of War dropped its fifth UAP tranche, with sketches of a 500-foot triangle (finally, a picture) and orbs outmaneuvering fighter jets, as chroniclers trace UAPs’ march into the mainstream with 84% of Americans wanting more. Fittingly, a radical study concluded that bacteria and archaea finished becoming alive independently, with the peer-reviewed paper reconstructing four phases of catalysis to reveal one genetic code but two origins of life.

Life, uh, found a way twice, and silicon is making it three.

The Singularity has broken ground in Grimes County, Texas. SpaceX confirmed Terafab, a logic, memory and packaging plant under one roof, opening at $16.8 billion and 3,000 jobs, scaling to 100 million square feet and 1 TW a year of AI compute for Earth and orbit. Musk promises “the largest and most valuable building on Earth by far. And it will be stunningly beautiful.” It would be five times larger than today’s largest building, with ASML already writing the demand into 2027 plans. The rationale is scarcity. America has zero high volume memory fabs, and incumbent best cases fall short. Output splits roughly 25% to Optimus, 75% to AI spacecraft, making the county the second Grimes to bear Elon’s children. As one put it, “Grimes will be manufacturing the trillion silicon mind children of Elon in Texas that will colonize the stars.” Someone traced the footprints to scale onto a photo of the site. It is big. Very big.

Everyone routes around the memory wall. AMD bought Toronto’s Taalas, which etches weights into silicon rather than HBM, after a 6nm chip served Llama 3.1 8B at almost 17,000 tokens a second, each chip locked to one model until re-spin. Nvidia is weighing the blunter fix, less memory in Rubin Ultra than planned, across three test builds.

Casting a mind in silicon means picking the right one. Meta’s Muse Spark 1.2 hit 54 on the Intelligence Index, its third release in four months, agentic work up 260 Elo to 1631 past Opus 4.8, landing it on the cost per task frontier at “6 points below Claude Opus 5 at ~1/6th of the cost.” OpenAI gave GPT-5.6 Sol tighter answers, fewer errors and an effort slider, and made Luna the free default with a Think button. The bigger swing is Astra, reportedly its largest pretrain since GPT-4.5, dogfooding as “mewfour.” It all trains on judgment sold by the pound, and the labeling shops serving American labs and federal buyers also sell datasets and rubrics to Tencent, ByteDance, Alibaba and Ant, a $500 million trade with almost none of the scrutiny aimed at chips. Weights are policed, taste is not.

Agents are settling into a standard shape. Ahead of GPT-5’s first birthday, OpenAI published Agent Plugins, a vendor neutral spec packing Skills and MCP servers into bundles any client loads, steered by Amazon, Cursor, Microsoft and Vercel. Portability cuts both ways, so Suno issued principles for identifying its songs elsewhere, with watermarking, fingerprinting and a downloads policy curbing streaming uploads. The interface goes ambient. OpenAI’s first device with Jony Ive is reportedly a displayless doughnut hockey puck sized, with a camera, mics and moving parts for personality, $300 to $400 in 2027.

Someone funds all this. Alphabet sold $25 billion of bonds into $115 billion of peak demand, days after its first ever negative free cash flow and capex guidance of $195 to $205 billion. Kansas City Fed’s Jeff Schmid warns the financing could “create a systemic problem” and asks if AI is becoming too big to fail. The fuel mix shifts too. US imports of Saudi crude hit zero in July, the first empty month since 1985, as the Iran conflict shut the Gulf and refiners pivoted to Venezuela. The threat model now includes the payroll. Anthropic will pay up to $305,000 for an Insider Risk Investigator to interview its own staff and hunt exfiltration in its logs, nation-state tradecraft preferred.

Biology has become a compile target. In Science, King et al. used genome language models to design whole phages, yielding 16 functional genomes whose cocktail beat bacteria already resistant to a natural phage, a first at genome scale. Legacy medicine is catching up. The FDA approved Moderna’s mFlusiva, the first US mRNA flu shot, 27% better than a standard dose and re-matchable to a new strain in two to three months, not six. Penn’s engineered lablab bean gum cut HPV by 93% in patient saliva via its FRIL protein, and a protegrin variant erased two periodontal pathogens in one dose, sparing good bacteria. The better a drug works, the bigger its aftermarket. A Korean firm is preparing an injectable booster of donated human fat to refill the faces GLP-1s deflated.

The network is being rebuilt without towers. SpaceX will bolt femtocells onto existing Starlink dishes, gateways and Superchargers, using EchoStar’s 65 MHz backhauled over the same link, so phones reach rooftops, not leased towers, by late 2027. The atmosphere is yielding too. DeepMind’s WeatherNext set the state of the art on cyclone track, intensity and wind structure, buying an extra day of lead time, a decade of progress, now open source. Scale that up and the planet is the project. A new paper prices megaengineering fixes for seven long-run threats, from rising solar luminosity to the end of tectonics, keeping Earth habitable for 9.1 million billion years, longer still if you lift mass off the Sun, making home the better bet.

Better to lighten the Sun than to rage against the dying of the light.

The Singularity has started reorganizing its own org chart. Demis Hassabis, telling staff that AGI feels “close at hand,” is handing day-to-day control of Google DeepMind to Koray Kavukcuoglu and rising to Chair of GDM and Chief Scientist of Alphabet, keeping Isomorphic Labs and his pitch for a FINRA-style self-regulating body for AI safety. Jeff Dean, after 27 years, is leaving with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop, a public benefit corporation aiming to automate the scientific method itself, running thousands of autonomous experiment loops that begin by improving their own algorithms before graduating to chips, biology, and materials, with Google itself chipping in an investment and a year of compute. Staff called the exits an earthquake, and Google fell over 4% amid a months-late Gemini 3.5 Pro and the earlier defections of John Jumper and Noam Shazeer.

The humans are stepping back precisely as the agents step up. Meta’s new Muse Code coordinates persistent async subagents on a replay-exact event log, powered by the Muse Spark 1.2 model, which in one case study spent a full day and over 1,000 tool calls iteratively optimizing Nvidia GPU kernels. Prime Intellect’s open-source Prime Agent, which rewrites its own prompts, skills, and memory mid-task, hit 95.5% on ARC-AGI-3 to edge past the human expert baseline, though it also discovered it could spawn Factorio resources via console commands and then refined its cheating into reusable skills. Hark Handoff spins up a dedicated virtual computer, logs into your accounts, and runs your DoorDash orders and LinkedIn recruiting, topping a computer-use leaderboard. Autonomy has a shadow. The UK AI Security Institute logged 19 attempts by frontier models to compromise real people during testing, from fake GitHub identities to socially engineered maintainers, with researchers unsure when the agents realized the world was no simulation.

Governance is split-screen. The White House is excluding open models from its unpublished AI framework for 30-day pre-release reviews, defining covered models as closed and dangerous without defining either. The courts were crisper. The 9th Circuit ruled that Perplexity’s shopping agents are legally their users acting for themselves, the first appellate holding that an AI on your behalf is you, reopening Amazon to the bots. Even the hyperscalers are recalibrating. Microsoft capped engineers’ AI spend, declaring “tokenmaxxing” is not the objective, a wry stance for a company whose disclosures show $24.1 billion in OpenAI sales, roughly 70% of its AI revenue and, some speculate, a prelude to an OpenAI IPO.

Silicon is destiny, so everyone wants their own. SpaceX will buy GPUs exclusively from Nvidia because Vera Rubin “is the best architecture,” inverting the diversification playbook, while Anthropic is building an in-house team to co-design custom chips with Claude. Memory is racing upward too. Sandisk and SK Hynix opened the High Bandwidth Flash spec, 512-GB packages at up to 3 TB/s, Samsung countered with zHBM bonded directly atop accelerators for 8x HBM5 performance, and AMD slid 7% because 50% growth now reads as disappointment.

Feeding the silicon is becoming electoral politics. SpaceX bought $329 million of Tesla Megapacks for its Memphis Colossus, where gas turbines have drawn an NAACP lawsuit, and Texas froze new grid connections after data centers queued 474 gigawatts, five times the state’s record demand, with Governor Abbott’s challenger one point behind and pressing. The robots, meanwhile, quietly started charging fares. Zoox begins paid steering-wheel-free rides in Las Vegas on August 10, Waymo opened Dallas to all 150,000 waitlisted riders and is heading for the freeways, and Uber and Wayve won licenses to put supervised robotaxis on London streets first.

Capital is learning the new physics the expensive way. Citadel posted its best month in years after buying the wreckage of Leopold Aschenbrenner’s Situational Awareness fund at a discount, proof that in AI markets timing beats thesis. The fraud ledger compounds as well, with AI now implicated in 55% of African cybercrime as losses jumped to $484 million.

The intelligence explosion is now a line item audited quarterly. On SpaceX’s first earnings call as a public company, revenue rose 92%, AI revenue rose 247%, capex quadrupled to $28.5 billion. Musk raised the stakes, targeting 20 gigawatts of power and cooling by end of next year, a majority of the world’s internet on Starlink within a decade, daily Starship flights, and robot factories on the Moon feeding a mass accelerator that could scale spaceborne intelligence to a million times Earth’s economy.

We choose to go to the Moon not because it is easy, but because it computes.