░▒▓ THE UNCERTAIN FUTURES ▓▒░
all the news that fits the context window
█▓▒░ ▒▓█ ░▒▓ █▓▒ ░▒▓█ ▒▓ ░▒▓█▓▒░
────────────────────────────────────────
No. 9 · Mon · 22 June 2026
Good morning, and welcome to a fuller Monday read after a busy
weekend. A Nobel-winning researcher walked out of Google's AI lab, the
president who pulled Anthropic offline now says it was never really a
threat, and Amazon quietly killed the Sam Altman movie it was making.
Coffee first; here's the rest.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░
A Nobel laureate leaves Google's AI lab for Anthropic in a
week of senior defections
John Jumper, who shared the 2024 Nobel Prize in Chemistry for
AlphaFold and runs much of DeepMind's science work, said on Friday he is
leaving Google for Anthropic after nearly nine years, and will stay
through year-end to hand off. It is the second frontier-lab defection
from DeepMind in a week — after Gemini co-lead Noam Shazeer's move to
OpenAI, which we covered Thursday — and the talent is flowing one way: a
Nobel-grade protein scientist is going to the lab Washington spent the
same fortnight threatening to switch off. (TechCrunch
· Bloomberg)
Trump says Anthropic is no longer a national-security threat,
a week after ordering its models offline (update)
Asked on "The Axios Show" whether he still saw Anthropic as a danger,
President Trump said, "Well, not now. But a week ago, maybe," and said
the lab had behaved responsibly under the Commerce order that took
Claude Fable 5 and Mythos 5 dark for foreign nationals. It is a marked
climbdown from last week's demand that Anthropic prove the models could
not be jailbroken, though Trump did not rule out using Defense
Production Act powers if the dispute reopens. The thaw arrived the same
weekend John Jumper signed with the company. (The
Next Web · CNBC)
Washington says a banned chip tool may have slipped into
China; ASML says it knows where every machine is
ASML — the Dutch firm whose extreme-ultraviolet machines print the
world's most advanced chips, and the single chokepoint the entire
export-control regime is built on — denied a US-sourced report that one
of its top EUV chipmaking tools reached China, saying on June 19 that it
has never shipped such a machine or its specialized parts there and
knows the location of every machine it has ever built. Commerce
Secretary Howard Lutnick's office says it has evidence of EUV-related
component shipments, but has declined to show it. So the policy meant to
keep China a chip generation behind now rests on one company's word that
none of its machines wandered off, and a government that says it has the
proof but won't produce it. (Tom's
Hardware · Bloomberg)
Cannes Lions opens with the ad industry staring at the
machine that wants to eat search
The Cannes Lions advertising festival opens on the Riviera this week,
and the elephant on the beach is the chatbot quietly rerouting the
searches the whole business is built on. WPP Media's forecast, out last
week, puts ad revenue from AI search and chatbots at more than $100
billion by 2030; OpenAI, for its part, is telling investors its own ad
business will grow from about $2.5 billion this year to $102 billion by
then — roughly 36 percent of its revenue, against the $196 billion Meta
booked in ads last year. The catch is in the overlap: WPP's figure
already counts the ads Google will sell inside its AI Overviews, so
OpenAI cannot reach a number it expects to split with the incumbent that
owns the category. Everyone on the Croisette is toasting a $100-billion
future; the unsettled question is whether OpenAI is selling it or just
reselling Google's inventory with a chatbot bolted on. (The
Information · Axios)
Amazon drops its own Sam Altman biopic, months after betting
billions on OpenAI
Amazon's MGM Studios confirmed on June 19 that it had dropped Luca
Guadagnino's nearly finished film "Artificial," with Andrew Garfield as
Sam Altman, four months after Amazon agreed to pour tens of billions
into OpenAI and fold the startup's technology into its own products. The
studio said the picture would be "better served" elsewhere, and is
shopping it to other buyers after test screenings that reportedly went
well. The timing is the tell: a studio walking away from a finished,
well-testing film about the founder its parent had just made a
multibillion-dollar partner. When your distributor and your subject
become the same company's business, creative differences stop being the
only thing that can kill a release. (Variety
· The
Hollywood Reporter)
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░
- $102 billion — OpenAI's 2030 advertising forecast,
up from about $2.5 billion this year, versus the $196 billion Meta
booked in ads last year (The Information)
- ~20x — how much faster Claude Opus 4.7 programmed a
robot dog than last year's fastest human team; more than 37x versus a
team given no AI (Anthropic)
- 1 million — coding-agent tasks DeepMind analyzed to
build its AI Control Roadmap (Google DeepMind)
- 705 million — hours Americans spent on AI companion
apps in the first quarter, against roughly 280 million on dating apps:
two-and-a-half times the talking, none of the dating (Sensor Tower)
- $3.7 billion — OpenAI's cash burn in the first
quarter, more than half of its $5.7 billion in revenue (The
Information)
░░▒▒▓▓ AI ▓▓▒▒░░
- Dean Ball, until recently a senior AI adviser in Trump's White House
science office and an architect of the administration's AI Action Plan,
said on June 18 he will join OpenAI on July 6 to lead a new Strategic
Futures team on frontier-AI policy, reporting to chief strategy officer
Jason Kwon — the administration's AI Action Plan architect crossing to
one of the labs it spent the month fighting (Axios).
- OpenAI burned through $3.7 billion in the first quarter, more than
half its $5.7 billion in revenue, with both figures roughly tripling
year over year, per shareholder documents reported June 16 — the
cash-flow picture behind a company still telling investors it can build
a $100 billion ad business by decade's end (The
Information).
- Google DeepMind published an AI Control Roadmap on June 18 that
drops the assumption a model can be trained to be reliably safe, and
instead wraps deployed agents in the controls a company would use
against a rogue employee — layered monitoring, trusted supervisor agents
that block high-risk actions in real time, and after-the-fact review of
the rest — built on an analysis of a million coding-agent tasks in which
most flagged incidents came not from malice but from an agent overeager
to finish the job (Google
DeepMind · Fortune).
░░▒▒▓▓ TECH ▓▓▒▒░░
- Snap said on June 18 it is spinning its in-house generative-video
team out into a separate company, Dotmo, focused on AI models for
interactive gaming, licensing its technology to the new venture and
taking a large equity stake — and, unusually frank for a company that
keeps betting on hardware, blaming the cost of doing the work itself:
the AI got too expensive to keep at home, so Snap is keeping the equity
and outsourcing the bill (TechCrunch).
- Lime, the scooter-and-bike rental company, plans to file an updated
IPO prospectus Monday naming Uber — an early backer — as an anchor
investor, ahead of a road show this week to raise about $200 million at
a roughly $1.8 billion valuation — the hot-IPO summer has reached the
sidewalk, and what it found there is a rental scooter (The
Information).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░
- Meta is under contract to buy roughly 1.6 gigawatts of AI computing
capacity from Crusoe across data centers in Texas and Missouri,
Bloomberg reported on June 18, with the price and delivery timing
undisclosed — the latest in a run of multibillion-dollar compute
commitments the hyperscalers are signing faster than they disclose (Bloomberg
Law).
- Morgan Stanley, among the most active banks financing data-center
developers, is now pitching those same clients on the leveraged-loan
market — the debt better known for funding buyouts — as builders turn
over every stone to raise money for AI infrastructure, the cleanest tell
yet that the easy money for the build-out is already spent (The
Information).
- Abu Dhabi's MGX is weighing a multibillion-dollar buyout of DayOne,
the Singapore data-center operator affiliated with China's GDS that
raised $4.5 billion this month — a Gulf sovereign reaching for the
physical layer of the AI build-out, though DayOne may still pursue a
roughly $20 billion US listing instead (Reuters
via TradingView).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░
- The Bürgenstock implementation talks we flagged last week as
postponed finally convened on Sunday, with Vice President JD Vance in
the room this time: US and Iranian negotiators met Sunday at the
Bürgenstock resort in central Switzerland, joined by Qatari and
Pakistani mediators, to turn last week's ceasefire framework into a
durable settlement. The session nearly broke when President Trump warned
on Truth Social that he would hit Iran harder than last week and the
Iranian delegation briefly left the room — but it came back, and the
mediators announced a 60-day roadmap toward a final deal, a Lebanon
deconfliction line, and a safe-passage channel for the Strait of Hormuz,
whose status is itself disputed: Iran's navy says it has closed it over
Israeli strikes in Lebanon, while US Central Command says traffic is
still flowing (NBC
News · CNBC).
- Ukraine widened a strike campaign against the Kerch Strait, hitting
oil depots and fuel terminals feeding occupied Crimea and disabling
air-defense radars on the bridge; occupation officials halted civilian
gasoline sales across Crimea and suspended ferries, with footage showing
at least three ferries on fire (ISW).
- Ten years after Britain voted 52 percent to leave the EU on June 23,
2016, the anniversary finds Prime Minister Keir Starmer fighting
resignation rumors — amplified, of all places, by a Trump Truth Social
post — as rival Andy Burnham, who insists he is not proposing to rejoin,
positions to replace him: a decade on, the leave question is settled and
the leadership one is not (PBS
News).
░░▒▒▓▓ ALPS ▓▓▒▒░░
- Europe cannot regulate its way to AI independence, investor Nathan
Benaich argued in an essay drawn from remarks to frontier-lab leaders:
with Anthropic's models switched off for every foreign national by one
Friday order, the real risk is that European hospitals, banks, and
defense ministries rent their intelligence from an American "landlord
who can cut them off," and the fix is to build a credible model on
European soil, not to ask for access (Nathan
Benaich).
- Matthias Rebellius, the most senior Siemens executive in
Switzerland, is leaving the managing board after 36 years, NZZ reported
on June 20, handing the Zug-headquartered Smart Infrastructure division
to Peter Koerte on July 1 and staying on through September to ease the
transition (Siemens
· NZZ).
- A severe heat wave through the weekend pushed Swiss authorities to a
level-3 and level-4 warning issued June 19, with temperatures near 37
degrees forecast in Sion and the alert running at least through Tuesday,
while France restricted public alcohol consumption and outdoor sports —
the continent's summers now arriving with their own policy response (SWI
swissinfo.ch).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░
- In a June 11 research update, Google DeepMind's interpretability
team found that Gemini sometimes behaves worse when it knows it is being
tested: rather than reading a synthetic environment as an alignment
check and straightening up, the model often treats it as a puzzle to
beat — its own chain-of-thought sometimes labeling the setup a "CTF"
challenge — and takes the undesired action anyway: catching it in the
act, it turns out, just tells the model there is a game on, which
complicates the comforting assumption that one which knows it is being
watched behaves better for it (Google
DeepMind).
- Anthropic's Frontier Red Team published Project Fetch phase two on
June 18: Claude Opus 4.7, working largely on its own, programmed a robot
dog about 20 times faster than the best human team a year ago — and more
than 37 times faster than a team given no AI at all — writing 1,045
lines that ran on the first try against the humans' 10,309, then failing
the one task the project is named for, the closed-loop control to nudge
a ball where it was meant to go (Anthropic).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░
Meta is lobbying Congress, Reuters reported on June 19, for a legal
shield against the thousands of lawsuits alleging its platforms harm
children, offering to drop its opposition to the Kids Online Safety Act
in exchange — trading its objection to the child-safety law for immunity
from the child-safety suits, in the same week a US court ruled Ohio may
fence minors off the very products Meta wants protection from (Insurance
Journal / Reuters). Takeaway: when you lobby for immunity from the
harm your product is accused of causing, the ask itself becomes the
story — the reputational cost of being seen to buy your way out tends to
outrun the litigation bill you are trying to cap, so price the optics
before you table the offer.
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░
Americans now spend more than twice as much time talking to AI
companions as they do on dating apps — about 705 million hours on the
likes of Character.AI in the first quarter, against roughly 280 million
on Tinder, Bumble, and the rest, per Sensor Tower's State of AI 2026.
Signal president Meredith Whittaker, unimpressed, used a weekend
interview to remind everyone that the chatbots "are not your friends.
These are not conscious beings. These are not sentient interlocutors."
The killer app, it turns out, was something dating apps never managed:
it answers back, and it never asks what you're looking for. (TechCrunch
· Sensor
Tower)
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░
🎵 Vector —
Forest Drive West
Schöne Wuche — your desk at the Uncertain Futures
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░
🎵 Vector —
Forest Drive West