Tuesday's haul runs from the inside of a neural network out to the borders of search. At MIT CSAIL, a frontier model edited another model's weights directly—0.016% of them—and the smaller model's math score rose as much as sixfold, with no training run at all. In Google search, the brand panel a company used to claim and correct started carrying a generated AI Overview instead, on the same day 18 media buyers told Digiday that search budgets no longer get waved through. In Turkey, the competition board ordered interim terms for news publishers inside AI Overviews and AI Mode before its investigation has even finished. At WorkOS, an AI offensive framework called ARTEX reached an internal sales application and the customer conversations inside it. In a RIKEN lab, five silicon qubits ran together above 99.9% control accuracy. And a live GEO benchmark put rewritten product pages on the real web and found the engines disagreeing about the winner almost half the time. The thread is control: who edits the model, who edits the answer, who sets the terms, and who can prove any of it.
MIT: one model rewrote another model's weights
Computer scientists at MIT's Computer Science and Artificial Intelligence Laboratory asked OpenAI's GPT-6 Astra to edit the weights of Qwen3.5-0.8B to make it better at math. Astra changed only 0.016% of the weights, reduced the smaller model's reasoning loops, and improved its MATH-500 performance by as much as sixfold, while its instruction-following score on IFEval rose roughly threefold. There was no gradient descent and no training loop. Astra reasoned about what parts of the network compute and revised the numbers directly. The team—Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab—calls the technique Weight-Space Programming, and reports that Astra could also write a policy network for Breakout from scratch, zero-shot. An independent overnight write-up that reran the edited model put the MATH-500 move at 7.9% to 48.1% and noted a caveat worth keeping: a free setting matched the gain in that tester's hands, so the size of the effect is still being argued about in public, with the logs open.
“We find early signs that frontier LLMs are surprisingly good at natively reading, writing, and reasoning about neural network weights.”—Rikiya Takehi, MIT CSAIL, November 2, 2026
“Instead of training, models may treat raw weights like readable, editable code: diagnosing failures and patching them directly.”—Rikiya Takehi, MIT CSAIL, November 2, 2026
A model treating another model's weights as code is a concrete step toward recursive self-improvement, and the authors name that door themselves, along with uses in interpretability and reducing sycophancy. The honest limits sit in the same report: one small open model, benchmark scores, and a technique the authors themselves describe as early signs and black magic on the way to a benchmark and a harness. Watch for that harness. The day weight editing gets a standardized test is the day it stops being a demonstration and becomes an engineering discipline, and labs will compete on it the way they compete on everything else that can be scored.
Google moves brand AI Overviews into the panel brands used to own
On Monday, November 2, Search Engine Roundtable reported that Google has moved AI Overviews for many brand keywords into the right column of desktop results, into the spot where a Knowledge Panel used to sit. Chris Long of Nectiv spotted the layout the same day; David McSweeney had flagged the pattern on October 31. The examples included Apple, Zillow, Coke, and Google itself. Ahrefs had already measured the spread: AI Overviews on 63 of 100 brand names by September 30, 2026, with 48 in position one, against five of 98 usable results on July 1, and a separate measurement put Overviews on 82.91% of branded keyword results. The Zillow panel also carried a factual wobble, giving 2006 as the founding year in the AI text while a fact row in the same panel gave 2004. The same Monday, Digiday published conversations with 18 media buyers who said clients no longer treat Google search spend as automatic, with contested slices put at 10% to 20% of the Google budget and one buyer seeing 10% to 30% reallocated. Google search ad revenue grew 12% in the third quarter, Alphabet's slowest pace of the year, and Google's own help documentation now says Performance Max may hand brand queries back to dedicated Search campaigns.
The Knowledge Panel was the one box on the page a brand could claim, correct, and largely control. A generated summary in that box answers in the brand's place and cannot be edited by the brand. For Chris' clients the work changes address: the facts that feed that panel—name, founding, locations, services, the GBP record underneath it—now have to be right everywhere the model reads, because the correction route is the source data, not the panel. Yesterday this page covered agents getting the keys. Today the brand's own nameplate is one of the doors.
Turkey orders interim fair terms for publishers in AI Overviews
Turkey's competition authority, Rekabet Kurumu, announced on November 2 that it is investigating Google for possible abuse of a dominant position with AI Overviews and AI Mode, under Article 6 of Act No. 4054. The Competition Board had already imposed an interim measure on October 8, and the two were announced together. Under the measure, Google must give news publishers non-discriminatory terms that reflect the economic value of using their content in the two features, and let each publisher decide how its content is used there without any effect on its visibility or indexing in Google's other services. The Board cited strong suspicion of discriminatory and exploitative conduct. The announcement sets no compliance deadline, names no fine, and gives no traffic figure, and the measure covers news publishers only. It lands on top of the European Commission's investigation opened December 9, 2025, the French publisher complaint filed October 29, and the UK CMA's publisher-controls requirement, whose metrics arrive in Search Console on December 3.
London mandated an opt-out. Brussels is still investigating. Ankara has already ordered terms that reflect economic value, interim or not, and that ordering is the news. Lessig's chair and Sohn's chair share this item: the architecture question—may a publisher refuse the answer engine without losing the search engine—now has a regulator answering it in writing, and the value question rides with it. Every market that watches this template, and yesterday's China item showed how fast templates travel, will study whether Google pays, licenses, or redesigns its way out.
WorkOS: ARTEX agents reached the sales application
WorkOS confirmed unauthorized access to an internal sales application, detected on October 30, 2026, and published its security bulletin on November 2. Founder and CEO Michael Grinich said the attack involved ARTEX, the AI-powered offensive framework that automates reconnaissance and exploitation attempts. The application held names, email addresses, phone numbers, job titles, and company affiliations from CRM systems, plus customer account histories and summaries and excerpts from sales calls, support discussions, and customer Slack conversations. WorkOS says there is no evidence that application data or production services were accessed, disabled external access to the affected application, began reviewing access controls across internal tools, and retained an independent forensics firm. Google Threat Intelligence Group and Mandiant describe ARTEX as a multi-agent platform that turns broad attacker goals into commands, testing permissions and adjusting failed commands through several attack stages with limited human input.
Read this beside yesterday's VAPOR PANDA item and the pattern sharpens. The framework is the same; the target class is different. A sales application full of conversation summaries is a phishing library: real names, real threads, real context, ready to make the next message sound true. The defense is boring and known—least access on internal tools, short-lived credentials, human approval outside the agent's reach—and the incident list is now long enough that no board can call it novel. Chris' clients run smaller shops on the same SaaS stack. The audit fits on an index card: which internal apps can be reached from outside, what conversations live inside them, and whose agent can read both.
RIKEN: five silicon qubits, one shared line, above 99.9%
Researchers at the RIKEN Center for Emergent Matter Science and collaborators reported in Nature Physics that they can control several silicon spin qubits at once through a shared microwave control line while keeping control accuracy high. Individually, the qubits achieved single-qubit operation accuracies above 99.99%, some approaching 99.999%. With up to three operating simultaneously, accuracy stayed above 99.99%; with all five active, it stayed above 99.9%. The team traced a major error source to microwave-induced shifts in energy levels, AC Stark shifts, measured how pairs of qubits affect one another, and used those pairwise measurements to adjust the control signals. That pairwise method matters for scale: calibrating every possible combination separately grows combinatorially, while a method built on pairs stays manageable as arrays grow.
Five qubits is a laboratory array, and the figures describe control operations, not a finished machine's answer rate; readout, coupling, and error correction across many operations still stand between this result and a useful computer. The board's lab doctrine applies anyway. Wiring is one of the unglamorous walls in front of larger silicon processors, and a shared control line that holds precision while its neighbors run removes part of that wall, in a material the chip industry already knows how to manufacture. Yesterday's IBM link and today's RIKEN control result are the same story told at two layers: the machine gets built out of plumbing that survives contact with scale.
LIVE-GEO: rewritten pages, live engines, no single winner
A new arXiv benchmark called LIVE-GEO published generative engine optimization rewrites of product pages as live webpages, then checked how real AI engines ranked them. The authors rewrote product descriptions and reviews for 80 fictional brands across 18 product families with five strategies—CORE, C-SEO, E-GEO, AgenticGEO, and Statistics Addition—and ran them through ChatGPT (GPT-6), Gemini 3.6 Thinking, and Grok in Expert mode, with prices frozen within ±0.2%. CORE won the top spot 66.3% to 76.3% of the time across the three engines. The engines picked the same top product for only 48.8% to 66.3% of brands. ChatGPT also reacted to candidate order: CORE's top rate was 73.8% under one fixed link order and 93.8% under the opposite order, against 82.5% randomized. A numerical-faithfulness check flagged 86 of the 400 rewritten descriptions, including all 80 Statistics Addition descriptions. The preprint is not peer reviewed, and the brands are fictional and the pages are product pages only.
Most GEO testing hands the model text directly. This one made the rewrites live pages and queried the engines the way a customer would, which is why its disagreements are the finding. If three engines agree on the winner only about half to two-thirds of the time, and link order alone can swing a result by twenty points, then AI visibility is a panel to be measured engine by engine, not a ranking to be won once. For Chris' practice the Statistics Addition flag is the practical warning: padding a page with numbers the source cannot support got caught by a faithfulness check in a student-grade benchmark, and the engines' own checks will only get stricter. Write the true numbers, in the question's own words, and let retrieval do its work—the answer is the page now rule from Sunday, tested live on Tuesday.
What the day adds up to
Control is the day's unit of account. A frontier model edited 0.016% of another model's weights and the field is now arguing, in public logs, about how much of the sixfold gain survives scrutiny. Google moved the brand answer into the panel the brand used to control, and buyers started pricing that loss at 10% to 30% of search budgets. Turkey ordered value-reflecting terms before its case is decided. ARTEX reached the conversations a company keeps about its customers. RIKEN held five qubits above 99.9% on one shared line. LIVE-GEO showed three engines crowning different winners on the same live pages. None of these stories is about a model getting bigger. Every one of them is about who holds the editing hand—over weights, over panels, over terms, over internal tools, over qubits, over citations—and what evidence they have to show for it. That is the desk Chris' clients are paying for: the facts placed where the machines read them, verified at the source, with the open questions left labeled open.
Sources: Weight-Space Programming, MIT CSAIL via The Register, November 3, 2026, with the independent rerun and caveat on Medium, November 3, 2026; brand AI Overviews move, Digiday buyer reporting, Performance Max documentation change, and Ahrefs brand measurements via PPC Land, November 2, 2026; Turkey Rekabet Kurumu investigation and October 8 interim measure via MACCUS, November 2, 2026; WorkOS security bulletin, November 2, 2026, and ARTEX description from Google Threat Intelligence Group and Mandiant, via Cyber Security News, November 3, 2026; RIKEN silicon qubit simultaneous control, Nature Physics via Knowridge, November 3, 2026; LIVE-GEO arXiv preprint via Anything Engine Optimization, November 3, 2026, not peer reviewed.