Today in AI — November 9, 2026

Nvidia's Washington problem turned from a scandal into a criminal referral, the data-center fight got a physical explanation, and Baidu put hard numbers on an AI that improves its own inference stack.

Models & Research

Baidu published RSI results instead of RSI promises. Its ERNIE team detailed ERNIE Evolve, an automated research system that proposes directions, edits code, runs experiments, and keeps only changes that survive an independent evaluator and a reviewer checking for goal drift and eval holes. Applied to inference work on Kunlun P800 chips, Baidu says the loop took generation speed to 6.2x baseline over about 168 hours, deterministic throughput to 133.7% with 26% lower per-token time, and extreme-decoding performance to 114.9% across seven research directions, according to QbitAI's writeup of the technical blog. ERNIE 5.5 Flash Base went open-weight on November 5; a new Flash model is expected shortly. This is the first Chinese lab to publish an RSI ledger with numbers attached rather than a manifesto.

Someone finally built a benchmark for agents that rewrite themselves. Flex-RSI scores recursive self-improvement under a fixed budget: an executor improves the answer, then an improver edits the harness — prompts, tools, memory, control loop — including its own role, with gains kept only if held-out validation improves, and final output scored on hidden tests across perception, reasoning, robot action, 3D reconstruction and infrastructure tasks. The paper's findings are more interesting than the leaderboard: self-improvement pays off most when the agent builds its own solution, feedback that records only success or failure holds open-weight models back, and open-weight models fall furthest behind wherever output must pass strict physical-stability or exact-correctness checks. Harness editing itself, not just retrying the task, drives the gains.

ByteDance's Seedream 5.5 sells stability, not spectacle. The new image model holds edits steady across up to 100 rounds with untouched regions preserved at the pixel level and no color drift, generates native RGBA with transparent backgrounds so output drops straight into a design file, and renders dense small text more accurately. ByteDance claims it leads GPT Image 2.5 Sunburst on text structure and realism, per its Seed blog. It is rolling out through Jimeng, Doubao, Doubao Work and Volcengine Ark, with a Pro tier planned.

The nastiest prompt injection hides in the model's own voice. A new study fed reasoning agents untrusted text dressed up as their own deliberation traces, and hit attack success rates up to 50% on operational goals like reverse-shell exfiltration and repository backdooring inside a containerized cyber range, per the paper. The same malicious objective phrased as ordinary user input never succeeded against the production-class open-weight models tested — GLM-5.3, MiniMax-M2.5 and Qwen3.8-Flash-Next. Payload style was model-specific, which means defenders cannot write one filter and ship it.

Retrying is what keeps a saturated inference server down. Overload experiments on vLLM found cancellation alone restored service in none of the runs, and only 12% to 38% of calls arriving after the overload were delivered: retries typically started prefill less than half a second before timing out, and executions vLLM later aborted had already consumed 79% to 85% of the prompt tokens it computed. The paper finds a retry budget restored service by shedding retries but still dropped calls, while request coalescing delivered every call after the overload unless cancellation had already killed the execution. Worth reading before the next capacity incident.

Industry

SpaceX wants to generate its own power and run its own supercomputer at Starbase. A public hearing notice shows the company planning a gas-fired power plant and a super-compute facility in Texas, with roughly $24 billion of investment in the initial phases, as Bloomberg reported. It is the same playbook xAI used in Memphis, now attached to a rocket company: bring the turbines, skip the interconnection queue.

The data-center backlash is about electricity, and the industry keeps calling it a culture war. Reporting from Virginia's Data Center Alley puts numbers on the fight: the average facility there draws 54 MW, more than Denver International Airport, the smallest site attractive for AI training is 300 MW, and the ideal is closer to 6,000 MW — the draw of Chicago, per MIT Technology Review. Virginia now makes data centers pay in full for new transmission lines, and developers are preparing for rules that cut them off the grid at peak. The same week brought 150 to 200 arrests at a Microsoft site near Amsterdam, a planning fight over Google's Thurrock campus east of London led by a wildlife trust and a motorcycle club, and a North Dakota town voting down an anti-data-center measure.

Disney is building its own models with Adobe, trained on its own characters. The multiyear deal, announced at Adobe's MAX conference, gives Disney early access to Adobe's technology across theme parks, cruise line and ESPN graphics, and puts its studios to work inside Firefly Foundry on proprietary models trained on Disney's character attributes, environments and visual styles to produce "commercially safe" output, The Verge reports. Disney's Imagineering unit was already using Firefly for park development. The pitch to every rights holder is the same: your IP, your model, no lawsuit.

Ada and Dialpad agreed to merge in an all-share deal valuing Ada at about $430 million, according to people familiar with the matter cited by the Globe and Mail. Both sell AI customer service software, and the combination is consolidation rather than expansion — a crowded category where the buyer's main asset is the installed base.

Alibaba's office suite now treats an agent as a team member. Qwen Office 2.0 shipped a collaboration space where several people and one agent work in the same document context, with the agent reading the space's files and tracking how the human discussion shifts, QbitAI reports. Alibaba is framing it as the first Team Agent for office work, and the comparison it invites is Anthropic's Claude in Slack.

The Megaspeed chip-smuggling inquiry is now a Justice Department matter. A Commerce Department investigation into Megaspeed — the Nvidia buyer at the center of the Huang affair — was transferred to the DOJ after Nvidia intervened, according to New York Times sourcing. Trump spent the weekend defending Huang as an "honorable guy," after being briefed on the suspected intelligence operation in October 2025 and taking no action, while CNBC counts at least $26 million steered by Huang and other Nvidia executives toward Trump-backed projects and a record $2.15 million in federal lobbying last quarter, reported Monday. Nvidia fell about 5% and Huang withdrew from the ElevenLabs Summit. The parallel story today, in our Take, is that Nvidia retained a crisis-communications firm known for handling sex scandals as early as October 2025 — months before its board opened a review.

Policy

Europe's draft Digital Fairness Act takes aim at addictive design and synthetic relationships. The draft would require companies to let adults switch off addictive features, and would target AI systems that simulate relationships, Bloomberg reported. It is a draft, not law — but it is the first EU text written specifically for companion bots.

Trump said he does not mind AI guardrails, then described threatening an AI executive with jail. Asked on 60 Minutes about safety warnings from lab leaders, he said the risk "is not nearly as real" as they claim and that the DOJ is managing it: "We had a situation where somebody was reported to us to be doing something bad with AI. We called them very strongly and said, 'Get ready to go to jail. Please give your wife a kiss good night.'" He refused to name the person, saying only that it was the head of "one of the big" AI companies, The Verge reported. He also said he would work with Democrats on safety, with a condition: "I don't want to put us in a deep second position."

Voice-cloning fraud has cost businesses at least $129 million in 18 months. Federal prosecutors say scammers mimicking executives' voices pulled that total from companies between July 2024 and now, the New York Times reported. The loss figure is the useful part: it is the first official accounting of what a cloned voice is worth to a criminal.

Tencent's AI chief says the bioweapon problem is coming and neither open nor closed models solve it. Yao Shunyu, who joined from OpenAI, argues for a mix — Anthropic's case for closed weights is "strong," but defenders need open models too, and he points to the Hugging Face breach where Claude's safety filters refused to parse attack logs and the platform used open-weight GLM-5.2 for forensics instead, per The Decoder's account of his remarks. Tencent plans to release Hy4 with a safety report, saying its biological capability sits below the high-risk threshold and its hacking ability was filtered out of the training data — a claim the abliteration market will test.

Tools

Claude Code settings files are leaking secrets into public repos. Ars Technica reports that thousands of credentials and tokens have been exposed through Claude Code's committed settings files on GitHub, which is a configuration problem rather than a model one — and the kind that spreads quietly because nobody reviews a dotfile. If your team checks in agent config, audit it today.

Medium built a notebook around the blank spot in a draft. Its new app, TK — named for the newsroom placeholder for "to come" — puts an assistant inside the editor that you invoke by typing @TK and a request, with skills for editorial and headline feedback and for matching a publication's style guide, TechCrunch reports. It ships with provenance notes that explain how AI was used, and can connect to your own agent over MCP. Web first, no encryption — it is built for collaboration, not privacy.

Apple started selling AI usage by the pack. Logic Pro now offers a $49.99 one-time Intelligence Usage Pack when you exhaust the limits on cloud-backed Apple Intelligence features, as The Verge spotted. Packs are tied to the purchasing Apple Account, do not work with Family Sharing, roll over monthly and expire after 12 months. More Creator Studio apps are expected this month, which tells you Apple's cloud AI has a marginal cost and intends to bill for it.

ChatGPT Enterprise is now sold through AWS Marketplace. Organizations can buy and manage it on their existing AWS billing, applying eligible spend toward AWS commitments, with ChatGPT for Healthcare and Financial Services available the same way, OpenAI announced. Procurement friction, not model quality, is what stalls most enterprise rollouts — and this removes a purchase order from the path.

Is a model that rewrites its own harness a research result or a governance problem? Write to us via our contact page.