Today in AI — November 4, 2026

Washington changed hands overnight and the AI trade had to reprice it — while Tencent shipped a flagship, MIT showed a model editing another model's weights, and Uber admitted how few rides its agents have actually booked.

Models & Research

Tencent shipped the finished Hy4, and it can see. Ten weeks after the August preview, Tencent released the full model and says it will open the weights, adding native vision for the first time: read a photo to guess where it was taken, turn a textbook page into a 3D explainer, point at a screenshot to request a UI change. By Tencent's own benchmark table, it clears DeepSeek-V4.1-Flash on agentic coding, search and office tasks while staying behind GPT-6.1 Sol and Claude Opus 5.5 — the gap to the closed frontier has narrowed, not closed.

Two details matter more than the scores. Pricing sits between DeepSeek V4 Pro's off-peak and peak input/output rates, with cache hits at DeepSeek's peak price, and a Hy4 Flash is promised — Tencent is competing on the bill, not the leaderboard. And the model spent part of its own training run fixing its training run: working with Tencent's Hyra 2.0 research agent, Hy4 found bottlenecks in its pretraining stack and lifted end-to-end throughput 20.1 percent, then trained a Jev model that beat Tencent's own version on several evaluations. That is a vendor claim with no external replication, but it is the kind of release note most labs have not had to write yet.

MIT showed GPT-6 Astra reading and rewriting another model's weights. In a research blog, the researchers fed Astra the raw weight matrices of trained networks and asked what they had learned: it identified which of four games a network played, matched the training language of small models, and picked the sycophantic model out of three pairs — but only when it could also see attention patterns, not weights alone. Given game rules and source code, it wrote weights from scratch that beat a network trained over 100 million rounds at Breakout, 517 points to 25 — though it fell behind the longest-trained baselines on two other games.

The practical result is narrower and more useful. On Qwen3.5-0.8B, Astra diagnosed the thinking loops Qwen's own documentation warns about, changed 0.016 percent of the weights, and moved MATH-500 accuracy from 7.9 to 48.1 percent and instruction-following from 21 to 62 percent; on Gemma 3 270M, which did not loop, it managed nothing. So the technique unblocks capability that is already present — it does not install new skills. Jürgen Schmidhuber's verdict on X was that the idea is "old hat," which is fair as history and beside the point as engineering: nobody had a model that could do the audit.

Two papers took the air out of autonomous research agents. A pre-registered test ran 100 matched LLM-versus-random proposals per domain across three verifier tiers and found the obvious bet fails: scores rose with verifier strength, but the LLM's advantage over random controls did not scale with it, and at the strictest tier it shrank. What did help was widening the proposer pool — mixing pools added 1.52 holdout-value units, p=0.0002. A separate field study ran a 100-minute sprint with twelve researchers and logged 33 AI interactions: one was accepted outright, 27 had no recorded disposition. Generation outpaced evaluation, and nobody noticed in the room.

Industry

Oracle scrapped a $7 billion chip lease to Tencent under White House pressure. The Financial Times reported that Oracle walked away from the five-year deal to lease as many as 100,000 advanced AI chips to Tencent after intervention from the Trump administration, citing people familiar with the matter. Oracle has not commented publicly.

This is how export control now works in practice: not a rule published in the Federal Register, but a call that kills a contract. That is faster than regulation and considerably less predictable — a cloud provider cannot price geopolitical risk it cannot read, and the same ambiguity now hangs over every US infrastructure deal with a Chinese counterparty.

Chinese buyers ordered 6 million Nvidia RTX Pro 5500s and expect about 1 percent of them. Bloomberg reported that Alibaba, Tencent, ByteDance and others placed orders worth roughly $70 billion, with delivery of about one percent expected inside four months. The card is a workstation part Chinese buyers have been racking eight to a server to run models.

Put the two stories together and the shape of the constraint is clear: demand on the Chinese side is measured in years of backlog, supply is being throttled from the American side by both formal controls and informal pressure, and the gap is the strongest argument Beijing has for its domestic chip stack.

Meta paid up for AI debt, and the bond market made it. Meta raised €9 billion in its debut euro bond sale after a record €53 billion of orders, plus $15 billion in dollars that drew as much as $58 billion, the largest slice of a cross-border offering above $20 billion. We covered the plan on Tuesday; today's news is the price. Meta is paying roughly 1.4 times the risk premium Alphabet offered for similar paper this week.

Demand for the debt is not the issue — oversubscription says the opposite. The spread is the issue: investors are no longer pricing "big tech" as one credit, they are pricing capex intensity, and Meta's is the highest in the group.

Apple put its engineers on a token budget and is cutting it. Bloomberg's Mark Gurman reported that Apple cut some engineers' daily AI spending cap to $200 from $300 and plans cuts as low as $100, with the cost discipline feeding staff fears of layoffs. Apple has not confirmed the figures.

Every large company is about to learn the same thing: inference is a per-seat operating cost, not a capex line, and it shows up in exactly the budget where managers have the least room. The interesting question is not whether Apple trims — it is whether capping tokens changes which models its engineers choose.

Uber fulfilled four rides through third-party AI agents last quarter. On its earnings call, Uber disclosed for the first time that four rides were arranged by AI agents in the third quarter, and about 500 Uber Eats orders in the week of Halloween — against 320 million rides and deliveries, The Verge reported. "We hope that there will be more traction," Dara Khosrowshahi said. "It is just very, very early."

The disintermediation thesis has a number now, and the number is four. Agents still cannot reliably complete the boring middle of a booking flow — payment, address confirmation, a driver's question — which is precisely the part platforms control.

SpaceX raised its 2027 data center target to about 13 gigawatts. SemiAnalysis estimates the new target is roughly equal to all of Google's data center capacity from 1999 through 2026, built in about eighteen months. Hours after the Texas result, Elon Musk posted that "data centers in space looking pretty appealing right now."

Whether or not orbital compute is serious, the domestic number is: 13 gigawatts in a year and a half is a schedule that assumes interconnection queues, turbines and local consent all cooperate — the three things that just became less reliable.

Policy

A Democratic Congress makes a real frontier AI bill likely in early 2027. Transformer's Shakeel Hashim expects a bill stronger than the Obernolte–Trahan FRONTIER Act — tough safety standards, mandatory testing and approval for the most powerful models, narrower preemption — to pass in the first months of next year, with strict liability for developers or data center moratoria possibly on the table. His reasoning: with as many as 56 Senate seats, Democrats need only four Republicans, and industry PACs just lost the races they played.

That last point is the one to watch. AI policy has been set by threat of electoral consequence for two years; if that threat did not pay, the lobbying position weakens from both directions — the industry has less to spend credibility on, and members have less to fear.

The money lost. Crypto and AI super PACs spent about $250 million on the midterms and lost nine of the twelve competitive races called so far, putting $74.7 million behind losing candidates against $28.3 million behind winners, per Molly White's tally. Elon Musk spent a record $446 million as a non-candidate and his preferred candidates lost almost everywhere, Politico reported.

Two readings are available and both are probably true: voters were not buying AI policy at the federal level, and they were voting on data centers at the local level, where the money never reached.

Anthropic's text watermark is a default, not a control. A second retrofit wave lands November 16 for Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, after Opus 4.6 slipped from its November 3 date; the watermark applies only when temperature, top_p and top_k are all set to their defaults, so API users can dodge it by changing one sampling parameter.

Read that next to the cryptographic proof-of-inference work covered earlier today and the pattern is consistent: provenance schemes keep arriving with an off switch, and the off switch is wherever a paying customer finds it inconvenient. A mark that vanishes on a temperature change will not survive contact with a regulator's evidence standard.

Apple will scan public iCloud links for "harmful content" in iOS 27.3. Items shared to "Anyone with the link" will be checked in the 27.3 OS versions, and Apple says it may delete the content or disable accounts, 9to5Mac reported. Apple has not published the classifier, its thresholds or an appeal path.

This is the client-side scanning fight rerouted through the sharing link. The narrow scope — only public links, not private libraries — is a real limit and worth saying plainly; so is the fact that the enforcement decision sits with an automated system whose error rate nobody outside Apple can measure.

Tools

DuckDuckGo started courting schools the day after Google's ultimatum. CEO Gabriel Weinberg pointed educators to deployment documentation for switching a default search engine and offered help directly; chief public affairs officer Kamyl Bazbaz said Google had shown "how a monopolist behaves when they think people are powerless," TechCrunch reported. Google's own figures put roughly 50 million Chromebooks and 150 million Classroom users in schools worldwide.

The switching cost is not the search engine, it is the admin console, the SSO and the lesson plans. DuckDuckGo is making a bet that a procurement fight inside districts is winnable on principle — a bet that has failed before, but rarely with the incumbent's CEO handing out the argument.

Pangram shipped image detection and pushed it into video. Pangram Image 1.0 claims state-of-the-art accuracy on AI-generated image detection, and Image Flash 1.0 is the model that powers its AI video detection, per Pangram. Both are vendor claims pending independent evaluation.

Detection is the least glamorous part of the AI stack and one of the few with a defensible business model: every platform that has to make a provenance call needs a second opinion, and the labs selling generators are not credible suppliers of it.

ChatGPT generates animated GIFs now. OpenAI's Images 2.6 model, rolling out to ChatGPT and Codex users on web, desktop and mobile, turns a prompt or an uploaded photo into a short looping animation with adjustable speed and frame-level edits, The Verge reported.

The feature is trivial; the capability under it is not. Frame-to-frame character consistency is the same problem video generation has been stuck on, and doing it in a flipbook is a cheap way to ship the fix.

Today's through-line: the price of everything AI got repriced — the debt, the chips, the token budgets, the political spending, the agent bookings. Which of those repricings do you think is the real one? Write to us via our contact page.