Sunday's haul has a shape, and the shape is a squeeze at the bottom and an escape at the top. At the bottom, the measurement stack closed around Google's AI answers this week: three trackers now put AI Overviews on more than 80% of searches, a randomized experiment finally put a causal number on the missing clicks—39.8% fewer outbound clicks when an Overview appears, with no gain in how satisfied searchers say they are—and Google answered the arithmetic the way Google always answers arithmetic, by switching on ads inside AI Mode in ten more markets. The answer is no longer a feature on the page. The answer is the page, and the page is now inventory. At the top, computing started leaving the building. Google's strongest model went to cyber defenders only, one point clear on the leaderboard and unavailable to almost everyone else. IBM wired two 120-qubit processors together across a meter of cold cable and made them behave as one machine. A Google TPU in orbit answered its first prompt with a haiku. And in Rome, Pope Leo XIV gave the week's most precise sentence about all of it—machines are tools, not persons—and aimed it at the one delegation he says must never happen. Fewer doors at the bottom, wider sky at the top, and a pope insisting the hand stays human.
Over 80% of searches carry an AI Overview, and the click loss is now causal
Search Engine Roundtable's October 26 roundup of Semrush, SE Ranking, and seoClarity data puts AI Overviews on more than 80% of Google searches, up from about 50% in June 2025 and roughly 15% the year before. The components disagree, which is what honest measurement looks like: seoClarity measured over 62% with an industry spread of 55% to 84%; SE Ranking found Overviews on 82.5% of US branded searches; Semrush found them on 88% of 400 tracked commercial keywords in August 2026, and on 96% of “best X for Y” queries—the searches people run right before they choose.
The causal number comes from academia, the way this paper prefers. Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon ran a pre-registered field experiment in which a browser extension randomly hid AI Overviews for 1,065 US desktop Chrome users in January and February 2026. On queries where an Overview appeared, showing it cut outbound organic clicks by 39.8% and raised zero-click searches by 34.5%; hiding it lifted clicks from roughly 0.37 to 0.62 per search. The clicks that came back were not junk—engagement on destination sites did not differ—and a closing survey found no difference in satisfaction, perceived quality, or ease of finding information. Label the study exactly as its authors do: a preliminary draft, US desktop users, a short window. It is still the first design that answers Google's standard reply. The reply has been that correlation is not causation. Now causation has a number, and satisfaction does not move to excuse it. Yesterday's halo experiment said absence from the answer costs a site twice; today's experiment says presence in the answer costs the open web two-fifths of its clicks. Both can be true, and for Chris' clients both point the same way: the citation is the remaining door, so the facts behind it had better be accurate everywhere they can be read. The experiment itself: Agarwal and Sen, working paper (PDF).
AI Mode ads switch on in ten markets
Google is switching on ads inside AI Mode in ten markets—Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines, and Singapore—with Brazil, Mexico, and the UK following later in the fourth quarter, per reports from Thurrott and Search Engine Land in late October. The sponsored placements sit inside the generated answer, labelled sponsored, and multiturn ads—placements that can surface on follow-up questions inside the same conversation—launch in the United States first. Advertisers do not buy any of this separately: existing Search, Shopping, and Performance Max campaigns become eligible automatically, cannot target AI Mode alone, cannot opt out of it alone, and get only aggregate reporting back.
The timing is in Alphabet's own numbers. Search revenue growth stepped down three quarters running—19% in the March quarter, 17% in June, 12% in September—while total revenue rose 25% to $127.6 billion and Cloud rose 103% to $30.7 billion. Chief business officer Philipp Schindler told the earnings call the company is scaling ads within AI Overviews and AI Mode to more users and more queries. Google says the ads do not influence the organic answers, and that promise is now the load-bearing wall of the whole product: the moment a reader suspects the answer bends toward the sponsor, the answer stops being an answer. For a small business the practical shape is already visible. Your existing campaigns will spend inside conversations you cannot inspect line by line, your product feed becomes your storefront inside the answer, and the work that earns the unpaid citation next to the paid one is the same work as ever—accurate facts, complete profiles, pages that say plainly what things cost and where they are.
Gemini 4 Carbon tops the index, and almost nobody can use it
Google DeepMind released Gemini 4 Carbon on October 28 and scored 59 on version 5.0 of the Artificial Analysis Intelligence Index—one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57. The security numbers headline the launch: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, 71.0% on CWE-bench for patching, all run without guardrails. On the separate Coding Agent Index, Carbon in Google's Antigravity CLI scored 59, second to Claude Code paired with Opus 5.5 at 62. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20, with no end date given for the lower rate.
“Carbon delivers improved performance across domains—ranking #1 on v5.0 of the @ArtificialAnlys Intelligence Index and sets a new state-of-the-art in cybersecurity and long-horizon capabilities.”—Google, October 28, 2026
Then the door: Carbon is available this week only to a subset of cyber defenders, expanding in stages through the Fairwind Program, which counts over 650 partners—governments, national cyber authorities, critical-infrastructure operators—and Google gives no date for developers, businesses, or consumers. It is also in the US government's voluntary pre-release testing process. Hold the caveats in the same hand: a one-point lead is close to a tie, and except for the independent rankings, the scores are Google's own internal testing, which outsiders cannot yet reproduce because outsiders cannot yet run the model. The interesting fact is the distribution, not the point. A model whose best party trick is turning flaws into working attacks is being handed first to the people who defend networks, under vetting. That is Lessig's architecture-as-policy in a release schedule: who gets the capability, in what order, is the regulation, written before any legislature convenes.
IBM links two 120-qubit processors into one machine
IBM Research linked two of its 120-qubit Nighthawk R1 processors with four superconducting cables, each about 80 centimetres of aluminum running inside equipment chilled to near absolute zero, and ran the pair as one machine. Quantum states crossed the links at an average accuracy of 99%, reaching 99.29% on the best link; two-qubit operations spanning both processors averaged 94.75%. The team created entangled states shared across the two chips, including 4- and 6-qubit states, and generated two entangled pairs at once through neighbouring links. All four connections survived three warming-and-cooling cycles, and—this is the engineering detail that matters—the cables were installed after both chips were already mounted, plugged in from above so they can pack densely.
Earlier links mostly joined small experimental processors through a single connection. This one joins two large processors through four, installed the way a technician would have to install them in a real machine room. The paper is an arXiv preprint, Sundaresan and colleagues, DOI 10.48550/arxiv.2610.29986, and the team is plain that quantum information still does not live long enough across the link for commercial use. The frontier beat does not need the caveat softened or the claim inflated: modularity is how classical computing scaled, from boards to racks to data centers, and this is the first meter of that road for superconducting quantum machines—two big processors, four cables, one entangled system.
A TPU in orbit answers its first prompt with a haiku
Google's Project Suncatcher prototype satellite, launched in October on a SpaceX Falcon 9 alongside more than 100 other spacecraft, has returned what Google calls the first-ever AI response from a data-center-class TPU in orbit. After about three weeks of commissioning, engineers uplinked a prompt during a ground-station pass to the Gemma 4 model loaded on board, asking for a haiku about being in space. The satellite passed over Svalbard, Norway, and downlinked the poem.
“Blue marble below / Data streams in silent grace / Starlight feeds my core.”—Gemma 4, running on a TPU in orbit, via Google
Price the milestone honestly. The satellite carries four TPUs and generates up to 1 kW of power, which means it can run one TPU at a time for about 20 minutes; the test that matters—how the chips hold up under real thermal stress and radiation, where a bit flip can corrupt data without tripping any monitor—has barely begun. If the chips prove capable, Google plans two more satellites in 2027 with higher power and inter-satellite optical links, and a peer-reviewed analysis from the Suncatcher team estimates SpaceX would need to fly Starship some 1,800 times to achieve the learning-driven cost reductions its earlier vehicles saw. Nobody is moving a data center to orbit this decade on those numbers. What happened this week is smaller and stranger: the inference stack—Google silicon, Google model, a prompt, an answer—ran end to end off the planet, and the answer was a poem about the view. Negroponte's chair claims the item without argument. Bits over atoms, and now bits above them.
Pope Leo XIV: machines are tools, not persons
On October 30, in the Consistory Hall of the Apostolic Palace, Pope Leo XIV addressed participants in “The Future of AI,” a dialogue between the Vatican and the scientific world organized by the Pontifical Academy for Life and the Future of Life Institute, and laid out five considerations drawn from his May 15 encyclical Magnifica Humanitas, On Safeguarding the Human Person in the Time of Artificial Intelligence. The future of AI is not predetermined, he said; it depends on human choices. Machines are tools, not persons—they do not experience, do not feel joy or pain, do not mature through relationships—and granting them rights or legal personhood risks hiding the humans who should answer for what they do. Certain decisions must never be delegated: a person on a target list or in a border database is at risk of being reduced to a data point, and decisions affecting human life, freedom, and the use of force must remain under effective, informed, and responsible human oversight. For that reason AI must be “disarmed,” and he encouraged the states meeting in Geneva in November to take concrete steps toward prohibiting lethal autonomous weapons—recalling Pope Francis' June 14, 2024 words to the G7 that no machine should choose whether to take a human life.
This paper keeps the question of machine consciousness open, and Leo's argument is not actually a claim about what stirs inside the model; it is a claim about responsibility. Personhood for a machine is, in his framing, a place for a human to hide. Read beside this morning's other items, the address is the week's sharpest piece of systems analysis from the least expected desk: Google is deciding by release schedule who may hold a capability, IBM is deciding by cable what counts as one machine, and Rome is asking who answers when the machine acts. The faith beat and the engineering beat turn out to be the same beat—delegation, and where it must stop. Geneva will decide whether anyone with the power to sign a ban was listening.
What the day adds up to
Add the week up and a single negotiation has replaced the old search economy. The answer surface now covers four searches in five, takes two clicks in five when it appears, charges advertisers for placement inside itself with aggregate-only receipts, and offers publishers a voluntary robots.txt line as the consolation architecture—which is why yesterday's permission slip and today's causal experiment belong in the same frame. Against that squeeze, the frontier did what the frontier does: it went up and it went modular. A model too sharp to hand out, two quantum processors behaving as one across a meter of cold cable, a TPU answering poetry prompts from orbit, and a pope drawing the one line he says no delegation may cross. The practical residue for a Sunday morning is short. Assume the Overview is on your customer's search whether you planned for it or not, and make the cited facts—the profile, the prices, the plain pages—accurate enough to survive being quoted. Watch the December metrics fight in Britain, where a regulator has ordered Google to show publishers their AI numbers separately. And keep an eye on Geneva: the states meeting there this month are the first audience that could turn “machines are tools, not persons” from a sentence into a rule.
Sources: AI Overview prevalence, Search Engine Roundtable, October 26, 2026, from Semrush, SE Ranking, and seoClarity data; Agarwal and Sen, “Google AI Overviews and Publisher Traffic: Evidence from a Field Experiment,” preliminary draft, April 2026, revised 2026; AI Mode ads expansion, Tech Insider, November 1, 2026, from Thurrott and Search Engine Land reporting, late October 2026; Gemini 4 Carbon, BetaNews, October 29, 2026, from Google DeepMind's October 28 announcement and Artificial Analysis Index v5.0; IBM meter-scale quantum interconnects, Phys.org, October 31, 2026, and Sundaresan et al., arXiv, 2026, DOI 10.48550/arxiv.2610.29986; Project Suncatcher first orbital response, TechCrunch, October 30, 2026, and Google, October 30, 2026; Pope Leo XIV address to the Vatican AI dialogue, October 30, 2026, via Unite.AI and the Holy See transcript.