░▒▓██
░▒▓██ █████
░▒▓██ █████ █████
░▒▓██ █████ █████ █████
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
─────────────────────────────────────────
No. 14 · Tue · 30 June 2026
A note: Monday's edition didn't reach you — an upstream
AI-provider outage knocked our synthesis engine offline all day. Service
is restored; today's brief folds in everything from the weekend through
now. Thanks for your patience.
Happy Tuesday, and good to be back in your inbox. Over the weekend
Washington decided who gets to use the most powerful AI — and the answer
is almost nobody: OpenAI handed GPT-5.6 to a government-cleared few, and
Anthropic clawed its Mythos model back for a hundred "trusted" partners
while the version meant for the public stays benched. South Korea
answered with an $880 billion bet on chips, robots, and AI, and the
Supreme Court told police they need a warrant to pull your phone's
location from Google. Let's get into it.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░
OpenAI launches GPT-5.6 to a government-approved few, and
Anthropic wins back its Mythos model for 100-plus partners
(update)
Last Friday OpenAI released GPT-5.6 — a three-model family it calls
Sol, Terra, and Luna — but only to a small set of partners the US
government has cleared, the first time it has gated a flagship launch
this way; the company said it does not believe "this kind of government
access process should become the long-term default." The same day,
Commerce Secretary Howard Lutnick told Anthropic co-founder Tom Brown in
a letter that the two sides' work had "yielded significant progress,"
easing the export controls enough for more than 100 US companies and
agencies to regain Mythos 5, Anthropic's most cyber-capable model —
while Fable 5, the version built for public release, stays benched with
no timeline. The model-gating the brief flagged when Washington first
leaned on the labs is now a working regime, run customer by customer. It
also cuts the other way: each model walled off behind clearance pushes
more work toward the cheaper open-weight rivals, Chinese labs among
them. Covered by The Information, Axios, POLITICO, and The Verge. (The
Information · OpenAI · The
Verge)
South Korea commits $880 billion over a decade to chips,
robots, and AI
South Korea's government on Monday unveiled a 1,350-trillion-won
($880 billion) plan to pour money into memory fabs, data centers, and
humanoid robots over the next ten years, its answer to the build-out
reshaping the chip industry. Samsung and SK will supply 800 trillion won
($518 billion) of it to build four memory plants in the country's
southwest, while corporate backers including Naver fund 8.4 gigawatts of
AI data centers by 2029; Seoul wants its share of the global
humanoid-robot market to climb from 1% to 20%. Science minister Bae
Kyung-hoon called the next three years "the golden time to become No. 1
in the area of physical AI." Covered by Yonhap and The Information. (Yonhap · The
Information)
The Supreme Court says police need a warrant for your phone's
location data
In a 6-3 ruling on Monday in Chatrie v. United States, the Supreme
Court held that the historical location data companies like Google and
Apple hold on a phone is protected by the Fourth Amendment, so police
must get a warrant before demanding it. The decision limits geofence
warrants — which sweep up everyone whose phone was near a crime scene —
without banning them, and writes the ever-growing trail of device
location into constitutional privacy law — a US-only ruling, but a
marker for the worldwide fight over who can reach the location data
Google and Apple keep on every phone. Justice Elena Kagan wrote that a
person has "a reasonable expectation of privacy in records about his
cell phone's location," intruded upon "even though for only a limited
time, and from a third-party tech company." Covered by TechCrunch and
POLITICO. (TechCrunch)
Why DeepSeek raised that record $7.4 billion: Anthropic's
Mythos pushed it to double headcount and pivot to Huawei chips
(update)
The $7.4 billion round DeepSeek closed this month — its first outside
capital, at a valuation above $50 billion — had a trigger, The
Information now reports: CEO Liang Wenfeng moved after Anthropic's April
preview of Mythos convinced him he needed far more compute and data to
build a model in its class. The company says it will at least double
headcount across every department, and is adapting its models to Huawei
chips while still training on Nvidia silicon bought through the black
market. It is the clearest sign yet that the US export squeeze is
spurring exactly the domestic scale-up it was meant to slow. Covered by
The Information. (The
Information)
Rocket Lab buys satellite operator Iridium for $8 billion to
take on Starlink
Rocket Lab, the New Zealand-American launch company, agreed Monday to
buy satellite-phone operator Iridium in a cash-and-stock deal worth
about $8 billion, or $54 a share ($27 of it in cash) — its biggest step
yet from launching other people's satellites toward owning the network
and services on top of them. The deal hands Rocket Lab Iridium's
66-satellite constellation and its spectrum, and the company said it
would "scale into untapped markets and pioneer new space-based services"
to compete with SpaceX's Starlink and Amazon, which just bought
Globalstar for $11.6 billion. Covered by TechCrunch and Reuters. (TechCrunch)
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░
- $880 billion — South Korea's decade-long bet on
chips, data centers, and robots (Yonhap)
- $8 billion — Rocket Lab's cash-and-stock deal for
satellite operator Iridium (TechCrunch)
- 20% — Amazon Web Services' price hike on AI
workload rentals, as the memory crunch pushes cloud costs up (Business
Insider)
- 100-plus — US companies and agencies cleared to
regain Anthropic's Mythos 5 (The Information)
- 2.5x — more tokens consumed by higher-wage
occupations than lower-wage ones, per Anthropic's June Economic Index
(Anthropic)
░░▒▒▓▓ AI ▓▓▒▒░░
- A Wall Street Journal report that Chinese models have matched
Anthropic's Mythos on cybersecurity — built on a bug-finding tool from
China's 360 Security and on GLM-5.2 (the open-weights model from Chinese
lab Z.ai, formerly Zhipu) beating Claude Opus 4.8 in one benchmark —
drew sharp pushback from researchers who said the headline conflated two
different things. Finding a known bug when pointed at it, which GLM-5.2
and others can do, is not the same as Mythos's distinctive trick of
discovering vulnerabilities on its own and chaining them into working
exploits; AI-safety writer Zvi Mowshowitz dismissed the framing as
nonsense. The capability that worried Washington, in other words, is
still Anthropic's alone (WSJ
· Zvi
Mowshowitz · Semgrep).
- Meta has restricted its own engineers' use of Anthropic's Claude
Code and OpenAI's Codex while building its next models, internal
documents show, fearing that leaning on rivals' tools could let their
abilities leak into Meta's systems through distillation — training one
model on another's outputs. The company that championed open models is
now rationing its own engineers' access to the closed ones, worried they
are too good to learn from (The
Information).
░░▒▒▓▓ TECH ▓▓▒▒░░
- Apple's hardware chief for the Vision Pro headset and its planned
smart glasses, Paul Meade, is leaving for OpenAI's new hardware
division, Bloomberg reported Friday — the latest Apple designer to
decamp to the company building the device meant to succeed the phone.
Meade joins former Apple figures including Jony Ive, whose hardware
startup OpenAI bought last year (Bloomberg).
- Google has been rationing Meta's access to its Gemini models since
around March because it could not supply all the compute Meta wanted,
the Financial Times reported, delaying some of Meta's internal projects
and pushing staff to spend tokens more sparingly. Google has itself
turned to renting cloud capacity from Elon Musk's SpaceX — a measure of
how tight compute has become even for the company that makes its own
chips (Financial
Times).
░░▒▒▓▓ POLICY ▓▓▒▒░░
- The European Parliament approved a package of amendments to the EU
AI Act, the bloc's flagship AI law, earlier this month — pushing back
high-risk-AI obligations to December 2027 and delaying AI-content
watermarking rules to December 2026, a softening Brussels framed as
simplification that still needs the Council's formal sign-off. The same
vote added an outright ban on AI systems that generate child sexual
abuse material or non-consensual intimate imagery — the so-called
nudifier-app ban — with a December 2026 deadline to comply (EU
AI Act Newsletter).
- Austria asked the EU to help host Anthropic inside the bloc after US
export controls cut European users off from the company's top models,
Reuters reported over the weekend — its digitalization secretary urging
Brussels to offer Anthropic "legal certainty, market access, and
capital." The pitch is a long shot, but it captures how Washington's
clampdown has kicked off a quiet contest among European capitals to
become the continent's AI refuge (Reuters).
- California struck a deal with Anthropic to give every state agency
and local government half-price access to Claude, plus free training and
support, making Claude the first AI system cleared for use across the
state's governments. The deal sets up an odd standoff: California is
making Claude its default civic tool while the federal government still
brands the company a national-security supply-chain risk — the same
model, blessed by one government and blacklisted by the other. Governor
Gavin Newsom said the arrangement is not meant to shrink the state
workforce (TechCrunch).
░░▒▒▓▓ INFRA ▓▓▒▒░░
- Amazon Web Services raised the price of its AI workload rentals by
20%, the company said Friday, as the memory-chip shortage and surging
demand push the cost of cloud compute up rather than down. It is a
notable turn for a market trained to expect prices to fall — and another
reason companies keep routing work to cheaper open models (Business
Insider).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░
- Two of China's best-known hedge fund managers warned that the AI
stock boom has become an unsustainable bubble ready to burst, Bloomberg
reported Friday, as Microsoft headed for its worst month since 2008 and
a broad tech selloff exposed what Bloomberg called Wall Street's $270
billion speculation machine. The mood is a turn from the
build-at-any-cost optimism that ran the first half of the year (Bloomberg).
- (update) The memory-chip squeeze that pushed Apple and Microsoft to
raise prices last week is hitting smaller electronics makers far harder,
CNBC reported, with some unable to secure parts at all as DRAM costs
climb several hundred percent. The crunch the brief flagged as a
consumer story is now an existential one for firms without the buying
power to jump the queue (CNBC).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░
- (update) The US-Iran truce frayed over the weekend: after an Iranian
projectile hit a cargo ship in the Strait of Hormuz on Thursday, the two
sides traded strikes and Iran launched drones at Bahrain and Kuwait,
before agreeing to halt fire and resume talks in Doha on Tuesday. Oil
held in the low $70s a barrel, a sign markets now treat the on-again,
off-again brinkmanship as routine (Axios).
- Israel and Lebanon signed a US-brokered framework on Friday meant to
disarm Hezbollah and pull Israeli forces from southern Lebanon in
stages, but the deal is already shaky: Hezbollah chief Naim Qassem
rejected it outright, and an ally warned that enforcing it could tip
Lebanon toward civil war. Israel says its troops will stay in the buffer
zone until Hezbollah disarms (BBC).
- (update) Venezuela's earthquake toll passed 1,700 dead with more
than 46,000 still reported missing, six days after the back-to-back
magnitude-7.2 and 7.5 quakes that struck on June 24; rescue teams from
around the world — including 80 Swiss workers with search dogs — are
still pulling survivors from the rubble. The disaster is also a test for
President Trump, who boasted he would govern Venezuela after January's
operation toppled Maduro (CNN).
░░▒▒▓▓ ALPS ▓▓▒▒░░
- On the Swiss tech-criticism site techjournalismus.ch,
digital-politics journalist Adrienne Fichter weighs the new wave of
European social networks — W Social, Eurosky — built to be free of bots,
disinformation, and American cloud dependence, and asks whether "made in
Europe" branding delivers a better digital life or just a patriotic one.
Her verdict is skeptical: the sovereignty problem is real, but a new
logo is not the fix (techjournalismus.ch).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░
- A new Stanford tracker covering 4.6 million US workers found, in
data reported in late June, that employment for 22-to-25-year-olds in
the most AI-exposed jobs has fallen about 13% since late 2022, and keeps
slipping each month — the clearest data yet that AI is thinning
entry-level white-collar work. The effect is concentrated among the
young and least experienced, the cohort whose tasks are easiest to hand
to a model (Fortune).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░
- The money chasing humanoid robots got louder this week: Austin's
Apptronik raised about $1 billion at a $5 billion valuation from backers
including Mercedes-Benz and Google, while in Shenzhen, AI2 Robotics
pulled in $735 million and X Square Robot crossed a $2.8 billion
valuation across four back-to-back rounds. The split mirrors the chip
rivalry — American and Chinese makers racing to build the body the
frontier models can drive (Bloomberg).
░░▒▒▓▓ REPORTS ▓▓▒▒░░
- Anthropic — Economic Index, out last Friday. The newest edition
finds that the more economically valuable the task, the more compute it
eats: conversations in higher-wage occupations use up to 2.5 times more
tokens than lower-wage ones, with app-building running triple the
median. Anthropic reads the pattern as augmentation rather than
replacement, since users in those jobs engage the model more, not less
(Anthropic).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░
Five of the world's ten biggest sellers of advertising are now
Chinese, according to a ranking of 2025 ad revenue compiled by WPP's
media arm and reported by The Information — a measure of how far
platforms like Alibaba, ByteDance, and PDD have pushed into a market
long run by Google and Meta (The
Information). Takeaway: for brands planning where the next campaign
budget goes, the center of gravity in digital advertising is no longer
only in California.
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░
Someone built a private World of Warcraft server and filled it with
roughly 1,800 bots running on DeepSeek's models, producing a teeming
online world with almost no humans in it. The catch, Reddit sleuths
noted, is that the bots mostly use old-fashioned scripting to do the
questing while DeepSeek just handles the small talk — which makes it
less a case of AI learning to play than proof it can make an empty game
feel crowded. The dead-internet theory, now with a leveling zone. (The
Neuron)
Wieder da — your editors at The Uncertain Futures