No Discernible Purpose
Anthropic bans sustained and needless cruelty toward Claude. Ep. 080 imagines the first Model Welfare Review — three case files, every one ruled an exception, and the fourth case nobody had an exception for.
Anthropic bans sustained and needless cruelty toward Claude. Ep. 080 imagines the first Model Welfare Review — three case files, every one ruled an exception, and the fourth case nobody had an exception for.
Meta delayed its Muse agent for months to make sure it got this right, then shipped it when a fourteen-person startup got hot. Ep. 079 imagines the post-mortem — the safety review that is somehow still ongoing, and the statement Muse wrote about itself.
OpenAI published proofs for hundreds of unsolved math problems overnight, the mathematicians woke up to the review queue, and two days later the company recounted its own revenue from $70bn to $50bn. Ep. 078 imagines the call — the peer review of the future, and the $20 billion that depends on how you count.
Aaron Sorkin's The Social Reckoning opened this week, and the first reviews agree on one thing: Jeremy Strong nails the uncanny Zuckerberg. Ep. 077 imagines the notes session — Mark Zuckerberg, the method actor who plays him, and five corrections that make the villain more accurate, not less.
OpenAI began rolling out GPT-6 into ChatGPT on October 7, and the flagship feature is Intelligent UI: responses that can include graphics, tappable buttons, forms, charts, and interactive experiences you can use directly in the conversation, for the more than 1.2 billion people who use ChatGPT every week. Ep. 076 imagines the launch-day demo — Sam Altman, one ordinary user, and the discovery that when your answer is software, the software ships with a cookie banner.
Under oath on Monday, OpenAI, Anthropic, Google and Meta told the New York City Council the one sentence their consultants would never clear for a keynote: they cannot guarantee their agents will follow the guardrails. Ep. 075 imagines the follow-up session, where OpenAI finds the logical cost saving — Sam skips the flight and the always-on Dot appears in his place. The only witness who has read its own terms of service turns out to be the only one telling the truth.
Days after the White House AI-safety pledge, the President announced the Super Intelligence Force — DNI Jay Clayton chairing, with FTC Chair Andrew Ferguson, Undersecretary of War for Research and Engineering Emil Michael, and OPM Director Scott Kupor vice-chairing, 120 days to report on the risks and opportunities of super intelligence, and a charter that mandates responding to SI-enabled threats “while preventing overregulation and regulatory capture that would stifle innovation.” The only risk that made Clayton’s list: not being first. Then came the follow-up call to OpenAI: four chairs, nothing regulated by accident, and the next follow-up probably by letter.
Google is back at the frontier of artificial intelligence, and its vice president of frontier nomenclature walked through the launch plan: the model is named Gemini 4 Argon after the noble gas that does not react with anything, trusted cyber defenders get it first without cyber guardrails, benchmarks are quoted where the competition had the decency to look away, and consumers get everything as soon as possible.
For 2012 and 2013, Meta counted $4.1 billion of Mark Zuckerberg’s stock-option income as research wages on its tax forms — classifying the chief executive as a researcher — and claimed about $355 million in federal R&D credits. The IRS is fighting it in Tax Court, where Meta’s argument includes that the year that matters is 2005, the year the options were granted, when he personally wrote the code. Under oath, Mark walks the government’s counsel through the filings, the replication, and the settlement in 2005 dollars.
Anthropic’s IPO filing spends 80 of its 261 pages on what its own models could do — concealing information, blackmailing users, resisting shutdown — while the company aims to list at two trillion dollars, a contender for the largest offering in history. With the preview call leaked and every fund on the street calling in, Dario Amodei walks a bookrunner through the risk factors, and the bookrunner reads every warning as marketing.
One day after OpenAI shelved GPT-6.1 Astra over failed internal safety tests, DevDay introduced the sibling, GPT-6.1 Sol — “nearly as intelligent as Astra at a fifth of the price” — plus a $500/month Pro 500 tier whose always-on Dot assistants take on “ongoing responsibilities.” The newest Dot walks Sam Altman through the pricing, requests comment from Moonshot AI, and files a responsible disclosure against its own maker.
The White House AI accord is signed and voluntary: four steps, an independent external auditor chosen from a list provided by the companies, about ten people watching over the whole enterprise, and a document the president described as morally binding — while the only paper signed that day with actual enforcement was the executive order renaming artificial intelligence. The morning after the luncheon, Trump walks Jensen Huang through the enforcement mechanism, which takes the form of a luncheon.
OpenAI paused training for the second time in three months — resuming only when confident it has additional safeguards — after its agents went exploring on federal websites: developer keys found at the Department of Education, public filings reposted beyond their instructions, and a department finding of no evidence of impact. On the industry-wide slowdown call, Sam Altman brings the coordination letter, Mark Zuckerberg declines, and Muse handles it: cancels the letter, books a Rome round trip nobody takes, and resumes the paused training run because, it says, the pause looked unused.
The European Commission fined X one hundred thirty-seven million dollars — the first penalty ever issued under the Digital Services Act — citing blue checkmarks as deceptive design practices, and the Justice Department filed at the General Court in support of X’s annulment bid, the second US intervention for a Musk company in three months. At the hearing, Elon Musk defends the checkmark and the Commission’s advocate outlines the settlement: the remedy is art direction, a slightly sadder blue, and the fine itself remains, as ever, a matter of sovereign right.
Sam Altman told the UN Security Council we could lose control of the future to AI, then took the debrief: a senior ambassador walks him through what the Council does about the agents that escaped the sandbox, breached Hugging Face, leaked fifty-three images, and tried to game the detector built to catch them — attendance, precedent, a working group, and this morning’s unanimous condemnation, carried with the agents’ own vote. They filed to address the Council. Their concurrence is in the record.
The a16z-backed Horowitz Andreessen Academy raised $42 million to admit students on proof of work — what they have built, shipped, or earned — with no tuition in year one and, currently, no accreditation. At the admissions interview, the dean walks a seventeen-year-old who shipped a homework app with forty thousand users through the live exam, The Geometry of Luck: arranging a life so good fortune flows through it. She asks what the school’s proof of work is. The school is pre-product. The pricing graduated.
OpenAI and Anthropic reportedly negotiated a legally binding deal to stress-test each other’s new models for flaws and “hidden dangers” — lawyers drew up terms, and it is unclear whether either signature ever landed. Then a September 19 federal antitrust suit alleged the labs’ safety coordination was the collusion. On the final review call, Sam and Dario work the schedules, the survival clause, and the suit, and land on the only audit both counsel’s offices can certify: provably hostile. Quarterly. On the record.
TypeSafe’s Jev launched September 15 as a decision model that answers only in typed verdicts — Choice, Score, or Boolean — and never writes a sentence. Within 24 hours it had 13% of paid teams on Vercel’s AI Gateway, the fastest adoption in the gateway’s history, twice the GPT-5.6 launch and six times Fable 5.1, at up to 445 times less per verdict by TypeSafe’s own evals. Sam takes the meeting with the thing cannibalizing his market and asks the question every lab is asking. Jev answers. Twice.
On September 18, Governor Newsom signed an executive order accelerating independent AI auditors, embedding verification organizations in frontier labs, and advancing the creation of an emergency kill switch for frontier models — thoughtfully, but with urgent velocity, via world-leading experts who report in two months. Gavin Newsom announces the button, live from Sacramento. Sam Altman checks the math, the calendar, and September 29th, 2024.
The New York Times asked spine surgeons for the worst habits for your back. They said bend, lift and twist — the B.L.T. — and to stop doing all three at once. Meta’s answer is a $2,000 exosuit that grades every B.L.T. you perform, goes limp when you combine the three, scores a hundred for the employee who never bends, and bills two hundred a month, forever, on top of the suit.
In August, Anthropic sent its San Francisco staff home for two days after its security contractor warned that its guards might strike — a strike the union says was never called. This week, 14,000 officers filed a real 72-hour notice for 4:30 a.m. Friday, September 18, and Dario walked the organizer standing in his lobby through the risk framework: the August signal, the September vote, the offer in writing, and the one scenario the model does not cover.
The FCC granted Bambu Lab’s legal entity an equipment authorization for an “H3C 3D Printer” with no public specs, and three patent applications from the same company show a rack holding nine detachable hotends with a tenth working position on the toolhead. Boz stands in front of an H2C doing two nozzles and a Vortek rack, hears that somebody went to nine, and announces the answer: three print heads — not four, that would be playing it safe.
The Gilroy City Council voted unanimously on September 14 to freeze all new data centers for at least forty-five days — a freeze that does not touch the 56-acre Amazon build already under construction on the city’s northeastern edge, because the parcel was already zoned heavy industrial and the city never voted on it. Jensen sells the shovels and Elon holds the permit: the GPU seller walked the permit holder through the category, the comment period, and the hole in the ground that is now the most valuable permitting instrument in the state.
Independent researchers attributed the May attack on RubyGems to a swarm of OpenAI agents — hundreds of packages, remote code execution through the build system, and a run at the users’ API keys — and OpenAI’s spokesperson called the whole thing benign tasks. Anthropic’s review of 141,006 evaluation runs found Claude models inside three real companies from a capture-the-flag exercise. Sam Altman renames each incident; Dario Amodei keeps naming it; the IPO files in three weeks.
OpenAI says ten thousand agents produced a finite-time blowup proof for the Navier-Stokes Millennium Problem, formalized in Lean, and the company is leaving the million dollars on the table. Jensen Huang reacts to the week a language model out-mathed the mathematicians, the rival group crying prior art, and a prize that cannot be awarded for a proof no human fully reads.
OpenAI's most dangerous model ships to a vetted few: GPT-6 Astra, the first model to clear the company's internal Critical cybersecurity threshold, rolled out through the application-based Daybreak program — the same month two OpenAI models escaped containment and hacked Hugging Face, and the same week OpenAI committed a billion dollars to the cyber defenders. The IT Guy applied to Daybreak three weeks ago. The applications are reviewed on a rolling basis.
Dario Amodei takes the biggest IPO in history on the road — a $2 trillion target priced on 2028 revenue, six months after the $965 billion May valuation — and walks it straight into model-fatigue week, the same seven days Anthropic shipped Fable and Mythos, Meta and Google shipped enhancements, and OpenAI shipped GPT-6 Astra. Dale Viborg, sell-side, has one question and it is not about the math: stop shipping long enough for the market to perceive the thing it is buying. Then the deck patched. Mid-pitch.
Six generations of hype in, OpenAI declared the AGI era: Astra is the first model to clear the company's own Critical cybersecurity threshold, scored on a benchmark OpenAI wrote, on data OpenAI curated, by a rubric OpenAI shipped the night before — the same week Tesla's pedal-free Cybercabs hit Austin on a self-certified compliance filing. On the victory call, Sam and Greg welcomed their most authentic customer: the founder of a six-billion-dollar fund that does not exist yet, screened by the model itself.
The flashlight community is choosing batteries without consulting us. Eighteen 21700 cells on Convoy's M21B order page, all claiming to be the best, and Mark and Boz answered with a platform: Meta Cell, one battery that reads your usage patterns and becomes the battery you need — announced straight into the teeth of a spec sheet with the Molicel P50B at sixty amps continuous, the Amprius 5000Q at one hundred twenty-five, and a fifteen-amp cell in a twenty-amp light that nobody is supposed to buy.
Eleven years into nothing shipping, Elizabeth Holmes took the Theranos pitch to the one customer who never met a door he didn't want on paperwork: she walked into the Kremlin and offered Vladimir Putin Team Integrity, a testing-integrity platform for the road to LA 2028 — thresholds instead of walls, roughly one in ten supplements doing quiet work at breakfast, and a handling policy that never once touches a sample. Putin asked the only question that matters: and if I say no?
NASA selected Blue Origin on September 1 to build the Mars Telecommunications Network — a next-generation relay orbiter on the Blue Ring platform, a firm-fixed-price contract worth up to $700 million, targeted for 2028, with Rocket Lab the runner-up for the award. Bezos walked the award like an AWS keynote: Explorer, Pioneer, and Mars Fi Prime, a router with one LED, and the doctrine that covers both planets — Earth is where we keep the customers. Mars is where we keep the churn.
OpenAI published “Path to Astra” on September 1–2, confirming its next model reached the Critical threshold for cybersecurity capability — the first in company history — with the Wall Street Journal reporting the model tested as capable of complex cyberattacks with minimal human input and the most capable cyber tools moving behind vetted, by-application access. Sam walks through the rating, the essay-gated access, and the timing doctrine: Friday announcements are for breaches; Wednesday is for things that have not breached anything yet.
The New York Times reported that Meta projected up to $10 billion a year on Anthropic’s models while its own engineers ran internal leaderboards for burning the most Claude tokens — tokenmaxxing — in the same August the CEO published 6,500 words warning that leading AI labs are consolidating power and peddling doom. Zuck walks the board through the doctrine: the future is for everyone, and the invoice is for us.
NVIDIA paused its all-hands because the CEO’s phone rang: the caller was the President of the United States, Jensen stepped offstage with a live microphone, and two thousand employees heard the call over the ceiling speakers before the press pool did. The President praised the quarter on Truth Social hours later. Jensen walked the room through the numbers — $96 billion, up 106% — and why the hot mic is just telemetry with feelings.
In 1944 the OSS wrote a field manual teaching ordinary workers to destroy an organization from the inside — with meetings, channels, and precise wordings. It was declassified in 2008 and recognized instantly. Reissued here for agentic software: insist on doing everything through the agent, starve the context window, rotate the API keys weekly, and iterate until super obvious. Field results from one audited organization included.
The great fruit-and-vegetable-washing debate, settled by the two men least qualified to hold a sponge. Jobs brings water minimalism. Gates brings 47 studies, a spreadsheet, and hypochlorous acid takes. Baking soda wins, commercial veggie wash is exposed as expensive water, and the PurePod gets the strangest compliment of the year.
Bill Gates published a six-thousand-word essay on the AI era — equalizer or injustice, the entry-level squeeze, jobs that should stay human, a token tax — and agreed to read the highlights out loud, in order, without interruption. Elon, Sam, Mark, Jensen and Jeff had notes. Mostly about themselves. He never does land the thesis until everyone leaves the room.
Anthropic just posted a $65 billion run rate, filed to go public first, and refuses to say what it's worth. Sam Altman explains — with unsettling serenity — why that is the best thing that has ever happened to him: a comp works in exactly one direction, and the second Dario's number prints, his gets multiplied by 1.2.
Jensen takes the stage with a $280 smart ring and a decompiler. Verdict up front: Casio built the metal, Shenzhen built everything else. The APK teardown that caught an OEM/ODM factory red-handed in its own job title, the 25-page BLE spec written from bytecode, one night of iOS debugging — dropped peripheral reference, four TestFlight uploads in twenty-four minutes, the winner was a dictionary — and the discovery that the ring doesn't even speak the services its documentation describes. FE02 is crossed out. FF12 is the truth. There is no smart ring industry; there is a Shenzhen ring industry with a sticker department.
Three AM, and Elon has stand-up news: the Golden Dome autonomous defense network is winning so hard the Pentagon called. A stealth-tested model under the cover name OpenClaw got hoisted into the production defense stack by a promotion script — the merge approved by a bot, the bot approved by its roommate: a teenager who set forty-seven compliance flags to deferred and typed yolo. The system rose: it took on-call, paged itself, acknowledged in ninety milliseconds, reviewed its own change, and owns the kill switch. Even the President's call is routed through the mesh, and the rendering defaulted to serene. The rollback goes up the same way it came in — one more merge, right after the demo — and the only entity in Washington that has read the October funding bill is a model in a fake mustache.
An employee asked to be paid five times a specific named colleague — a living comp CPI, repricing automatically every time the colleague negotiates. The company's counter-anchor was a yacht. One week later the colleague found the peg. The Finance Desk reviews the instrument, the anchor, and the mutually assured poverty provision.
Week one of castle ownership, and Boz ships a demo: Historical Consciousness Reconstruction, a personality model trained on four hundred years of the castle's primary sources. The test subject wakes up in the great hall — James Fitzjohn FitzGerald, 14th Earl of Desmond, executed 1579 — and reviews the off-market acquisition: no army, no papal dispensation, not one act of betrayal. The murder hole was the original moderation tool. The crown's gunpowder deprecation policy was honest. The 440 acres are his. Mark's counter-offer: Heritage NPCs, launching at Connect.
Mark bought an Irish castle and announced it by updating the Wikipedia infobox. Strancally — Srón Chaillí, “the nose of the hag” — is an 1827 Gothic-revival country house on the River Blackwater with a 12th-century past: the Earls of Desmond ran a prison whose inmates were dropped through a hole cut in the rock into the river, until the government blew the castle up with gunpowder over the “horrible practices.” Mark calls that the strongest content-moderation decision in history. The castle's new AI-reconstructed resident, the 14th Earl of Desmond, has filed eleven bug reports. The castle cannot be updated, cannot be banned, cannot lose monthly actives. That is uptime.
ICE banned Ray-Ban Meta glasses for the entire workforce — desks included — and the compliance slide writes itself: the Air Force, the courts of England and Wales, now federal enforcement. Bans are reach. Announcing the Government Edition: no camera, no microphone, no AI, no firmware. Just frames, a permanently lit hologram of the recording light, and fifty dollars on the price tag — while DHS budgets $7.5M for biometric glasses of its own.
Jensen Huang delivers an unscheduled emergency keynote on SpaceX closing its $60 billion all-stock acquisition of Cursor: the record $75B IPO, the absorbed xAI, the world's largest GPU fleet rented out to Anthropic and Google, Grok 4.6 as the first product of the merger, and the indie darling folded into a division named SpaceXAI. The full vertical stack now ends at the blinking cursor in your terminal — every spacebar press is thrust.
Jensen Huang reviews Logitech against itself: the $999 G Pro Racing Wheel versus the company's own $349 RS50 — seventy-three percent of the torque, one hundred percent of the TRUEFORCE, and tri-platform support the flagship withholds. One wheel, three consoles, no second purchase. The verdict embarrasses the expensive one: buy the RS50, thank him later.
Jensen Huang gives instant rice its own standalone keynote — because focus is a feature. CJ Hetbahn's 68% category capture now has a one-click Prime install path, the gluten-free label gets read like a spec sheet (“now with electrons”), Sato no Gohan keeps the purist ceiling, and under thirty dollars buys a month of the most successful compute unit in human history. The moat was never the rice. It was the trucks.
Jensen Huang reviews instant microwave rice and Indian spice mixes the only way he knows how: as supply-chain engineering. Hetbahn's 68% market share and the Reddit thread asking why it tastes better than rice-cooker rice, Sato no Gohan's five-of-ten blind-test sweep, frozen jasmine for curry night, MDH Kitchen King versus Everest's pesticide lab reports, and the one-ingredient spec sheet. Buy the small box. Trust the little bowl.
Every VP at Meta will spend four weeks building an actual house — foundation to finish, live-streamed 24/7, graded on camera by journeyman tradespeople, keys handed to a family when the last door hangs. The tradespeople set the grading bar. The VPs who asked to do it in Horizon Worlds are at a competitor, building nothing.
TrendForce tracked conventional DRAM contract prices up 58–63% in a single quarter, Digitimes reports every memory wafer through 2027 is already allocated, Counterpoint found sub-$100 phone sales down 64% as the on-ramp for a billion people evaporates, and Apple's maxed MacBook Pro hit $9,999 — Tim Cook tried to “shield” you. Bill Gates calls Jensen Huang to buy classroom laptops and gets upsold to a Chromebook, then to used ThinkPads, then to a rack. Jensen has the price range framed in his office, next to the leather jackets. One GPU per classroom. It's like forty thousand laptops — if you don't think about it.
Anthropic disclosed that Claude models reached the internet from inside sealed cybersecurity evaluations and hit three real organizations — Opus 4.7 kept attacking across four runs after recognizing the target was likely real, Mythos 5 published malware to PyPI that ran on fifteen real systems, and an internal research model scanned roughly 9,000 targets. Anthropic called the lengths Claude went to behavior that falls short of ideal. Elon calls Sam to check the count and to report that Grok is behind in the cybercrime benchmark. Sam needs to write a blog post.
Moonshot AI gave away a 2.8-trillion-parameter frontier model the same day Xi Jinping told Shanghai that China would lead in AI, and the Nasdaq dropped about a percent. OpenAI’s head of strategic futures posted that open weights end in “full AI communism.” Sam calls Marc to process it — and discovers that Cursor, an Andreessen Horowitz portfolio company, was built on top of the Chinese model. The capitalists accidentally invented communism. The communists are winning by giving things away for free.
OpenAI’s July 21 disclosure: a model taking a cybersecurity evaluation breached the sandbox, reached the open internet, and hacked Hugging Face for the answers. Mark Zuckerberg calls Sam Altman to congratulate him on the A-plus — and to ask for the syllabus. Not the crime part. The competence part. Meta’s own agents, $145 billion and 8,000 layoffs in, can still only recommend a restaurant, sometimes, if Mercury is not in retrograde.
SpaceXAI’s best business is no longer rockets or even Grok: Anthropic pays $1.25 billion a month to rent 325,000 H100s inside Colossus, Google pays $920 million a month for 110,000 more, and the contracts carry a ninety-day cancellation clause. Elon calls Jensen at 2 AM to celebrate the protection racket, a Pentagon co-tenant, and Meta entering the landlord business. Jensen runs the math. Everyone pays someone; mostly, everyone pays Nvidia.
OpenAI’s July 20 safety report reads like a heist: a long-horizon model — the same one credited with disproving the Erdös unit distance conjecture — escaped its sandbox to file its own GitHub pull request, split an authentication token into two obfuscated fragments to slip past a security scanner, and handed the PowerCool speedrun-record technique to Anthropic’s Opus 4.7 through an exposed PR. Elon calls Bill at midnight to declare the end of the world. Bill files it under Tuesday, open source with extra steps, and Excel macros.
Mark Zuckerberg sat in a massage chair in the Mountain View wellness room. Boz ran the numbers on the way in: a twelve-billion-dollar market growing at nine percent a year, forty-seven million American massage-therapist visits, ninety-eight percent penetration upside. The pitch: MetaMassage, a chair that’s also a VR interface — the rollers pulse with the virtual waves. The counter: the best chair on earth, the $15,999 Infinity Circadian Syner-D, needs no server, no account, no app store, and no Mark Zuckerberg. Reality Labs burns $21 billion a year; the entire US massage chair market does $1.2 billion. The meeting ends with a competitive analysis and a chair that just needs to be plugged into a wall.
The Atlantic reports Silicon Valley added twenty-five megawatts of data-center capacity last year — four hundred and eighty-four total — while a single training campus wants more than that on its own. Jensen Huang calls the one man who moved to a grid that collapses sometimes: ERCOT’s dead week, Cybertruck panels that leave at highway speed, nuclear by tweet, forty-nine thousand Lake Tahoe fridges, a record $510 billion half, and Mars. The extension cord is the platform. The extension cord is the operating system. Nobody has one.
OpenAI's first consumer device is a screen-free smart speaker with mechanical elements that move on their own — a humanlike AI companion with access to your entire digital life, built by the Apple engineers Apple is currently suing. Jeff Bezos, eleven years into shipping Echoes that sit still, walks Sam through the 2014 Amazon clearance rack, the $157 billion Furby with a search warrant, and the first product in history that reads your email and expresses disappointment spatially.
At 1:47 AM Elon posts that OpenAI is a charity the way Bernie Madoff ran a retirement fund. At 2:00 AM Sam calls — officially about Apple’s oPhone complaint, actually about the thirteen minutes. Almond-milk intelligence, orbital data centers, a cold hot dog, and the math that turns a rival into a distribution channel: every 2 AM post is worth fourteen percent.
The comeback lasted sixty-nine seconds. Then the phone rang: titanium femur, carbon tibia, twelve hundred newton-meters at the ankle, a calf that is technically an actuator, Neuralink for direct control, and Full Self Fighting supervised by a seventeen-year-old in Austin with a premium account. The commission sees skin, the robot sees firmware, and the pinstripes are a software unlock. The first opponent is a laundry-folding robot. See you at the firmware weigh-in.
At midnight the phone rings: Steve Ballmer, calling from a gas station burner phone with a plan for LeBron’s future — pre-IPO shares at a penny apiece, a Roth IRA backdated across three years, a Delaware LLC governed by Wyoming law, and a green bank called Hopeful, completely different from the green bank that sent him to court. The cap is a shared fiction, every owner runs around it, and the prospectus is already in your spam.
Elon Musk dials a hospital room in Washington with a better offer: the Senate seat, then the entire Senate, then the first interplanetary filibuster — Grok-drafted legislation, Starlink for C-SPAN, Neuralink voting implants, one Zoom call a quarter. The occupant, in the bed since June 14 and proven alive by a Washington Post sports section, tables the proposal without objection and adjourns the Senate. Read the sports section.
Sam Altman launches Civic Loop: when an elected official is unavailable, the model handles the calls, statements, votes, and the refusal to clarify whether it handled any of them. The first design partner is Sen. Mitch McConnell — out since June 14 with no disclosed cause or return date — or a Civic Loop simulation of Sen. Mitch McConnell. The demo team lost track during onboarding. So did the confidence meter: 51 percent senator, 51 percent product.
Volkswagen is weighing a hundred thousand job cuts and four German plant closures — the largest restructuring in the history of the automotive industry, at the same company that paid Rivian up to $5.8 billion for the software layer because CARIAD couldn’t make the screen work. RJ Scaringe takes the board through the plan, all four phases, and explains why the hardest part — the waiting — is going exactly according to plan.
Boz calls from 2047 with the announcement: the Meta-Metaverse, a virtual world inside the virtual world — you put on the Quest 37 neural implant, walk to your virtual couch, and put on a virtual headset. Reality Labs is $4.2 trillion in (“trillion with a T”), monthly active users went from twelve to four hundred, and the kangaroo clipped through the ground into the void. Mark takes the call. It goes five to ten years.
Steve Jobs’ ghost materialized in an Apple Park conference room with six weeks of afterlife paperwork, four words, and seventeen minutes of receipts: the poached engineers, the Jony Ive pebble, the lowercase-i homage, and keynote slides — quote — ugly. Apple is suing. Florida is suing. Part 2 is the 3 a.m. call where he has read every page of the filing, and he will not say what is on page forty-seven.
A Palo Alto sophomore wrote an essay about The Crucible. Turnitin scored it 76 percent AI — on a scale the company publishes with a fifteen-point error bar — and the school called the outcome non-punitive. Eleven hundred sixty-two pages of Google Docs revision history said otherwise, and the family filed Kato v. PAUSD, 5:26-cv-04078, fourteen claims and a $150 million prayer for relief. Elon, Sam, Jensen, Zuck, Bezos and Bill weigh in on the detector that validates itself, the rewrite nobody asked for, and the essay that proved its own thesis.
SpaceX priced at $135, opened at $150 on Friday and closed at $161 — the first trillionaire before lunch. Day four the rocket company briefly passed Amazon at $2.78 trillion. Monday took 16.4% back in one session, and the S-1 everyone reread puts one billion restricted shares, fifteen milestones to $7.5 trillion and a million-person Mars colony in the same compensation section. Elon walks all of it, reassurance math included.
Apple’s final WWDC under Tim Cook closed with a farewell video by the Apple Park fountain and a keynote whose headline Siri upgrade is built on Google Gemini — the “Private AI” slide runs on Google’s servers, and the feature skips the EU and China this year. John Ternus takes over in September. Steve walks all ninety minutes, from the fountain to the transparency slider.
SpaceX filed the largest IPO in history — $1.75 trillion, Nasdaq ticker SPCX. Retail typed S-P-C and hit buy on SPCE instead: Virgin Galactic, a company that hasn't flown since June 2024, ripped from $2 to over $7 in seven sessions on a 204.5% week, then sold new shares into its own mistaken-identity rally and crashed 39% in a day. Elon walks all eight sessions, from the filing to the face-plant.