Daily editorial briefing

№ 20261113

Nvidia Changes CEOs for the First Time in 33 Years, While Google Turns a Math Conjecture Into a Lean Proof

The lead story on November 13 was Nvidia: after 33 years, Jensen Huang resigned as CEO, president and board member, and CFO Colette Kress became interim CEO. The news crowded ou…

The lead story on November 13 was Nvidia: after 33 years, Jensen Huang resigned as CEO, president and board member, and CFO Colette Kress became interim CEO. The news crowded out almost everything else, and it pushed the financing risk underneath AI infrastructure into plain view. The Federal Reserve’s financial stability report, released the same day, found that 87% of market participants now rank AI as the top financial risk, up from 50% in the spring.

The second thread was quieter. Google Research published concrete output from its Antigravity /teamwork multi-agent system across mathematics, chemistry and systems software, including a complete Lean 4 formalization of the Weil conjectures. Both stories point at the same question: AI capability is entering domains that demand long verification cycles, while the capital structure paying for it is being repriced.

Theme 1: Nvidia changes CEOs, and Kress inherits a company that has never had a second one

According to SEC filings summarized by Latent Space, the board accepted the resignation on November 11, after opening an independent investigation into the relationship allegations reported by The New York Times. AI Valley recorded that the stock fell more than 13% in early trading that day, closed down about 4%, and recovered somewhat on Friday. Kress has run Nvidia’s finances since 2013 and keeps the CFO role alongside the interim title.

The market’s concern goes beyond personnel. The Decoder, citing The Wall Street Journal, reported that Huang’s removal traced more to risky business practices than to his personal life. Nvidia had been offering billions of dollars in financial guarantees to AI cloud providers and data-center developers in exchange for revenue share. The largest was a $250 billion guarantee tied to OpenAI’s Ohio data center; the stock fell 5% when that became public, and Kress has since halved the commitment.

Those guarantees turned chip revenue into something closer to a credit instrument. Investors interviewed by TechCrunch expect the successor to be more conservative and doubt she can sustain Huang’s pace of several billion dollars a year, including 105 rounds already this year. Startup funding will feel the gap.

The volume of rumor is itself worth recording. Bryan Catanzaro, Nvidia’s VP of applied deep learning research, stated plainly that a circulating letter urging the board to reinstate Huang, and purporting to carry his signature, is forged. Elon Musk said publicly that he is signing a separate petition. No independently verifiable evidence supports the personal allegations, and neither the company announcement nor the regulatory filing text is inside this capture.

First-hand reaction says more about Nvidia’s position than the share price. Newcomer, reporting from the Cerebral Valley AI Summit in San Francisco, found most attendees expecting no short-term damage but describing strain across the ecosystem; Databricks’ CEO called it a blow to the whole ecosystem. Israeli investors were more measured: Nvidia has 6,000 employees and its largest overseas R&D center there, and slower decision-making could open a window for competitors in inference software and AI chips — though if uncertainty also slows overall spending, that window faces a smaller market.

Sources:

Theme 2: Google points multi-agent systems at proofs and kernels, and the results pass machine checks

Google Research published a batch of results from Antigravity /teamwork. The heaviest was a complete proof and formalization of the 2013 Courtade–Kumar conjecture, with CUHK, USC and CMU: once a random bit string passes through a noisy channel, no compression of the input to a single bit preserves more information than a “dictator function”. The full proof was formalized in Lean 4, reaching 50 million lines of code.

A cluster of long-open problems came with it. LEAP paired with Argon autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures — about 1.8 million lines of Lean across 28,800 subagent invocations. The same setup produced the first formalization of Hilbert’s tenth problem over general rings of integers and reproduced the formalization of finite-time blowup for 3D Navier–Stokes.

The engineering numbers are closer to daily work. Building on an earlier 25% port, a two-week Argon campaign rewrote the rest of the Zircon microkernel’s C++ into about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64 hardware. libgav1 went from C++ to roughly 170,000 lines of Rust at 100% bit-exact parity, and ended up 20–30% faster than both the original C++ and hand-written assembly. On nine JAXBench kernels, DuoDiscover reached a 4.09x geometric-mean speedup, and Borg’s scheduling workflow cut TPU preemption by 24% on some hardware.

Two design lessons are reusable. First, candidate solutions are judged by independent falsifiers — the Lean 4 checker, unmodified test suites, hardware counters — rather than by the model assessing itself. Second, context is partitioned by construction: each specialized agent receives only the files and tools its step needs, and auditors verify results in clean contexts that exclude the original producer’s reasoning.

Human experts were moved earlier in the loop, not removed. The CUHK team first reduced the conjecture to a 14-variable optimization problem and supplied a key analytical lower bound; Gemini 4 Argon then partitioned the parameter space and developed most of the new bounds and interval-arithmetic certificates, and only then did humans confirm the formal statement matched the original problem. A separate pipeline used 29 agents over 11.5 hours to turn 13 public tables into a 50-task causal inference benchmark, with reviewer and auditor agents vetoing shortcuts and hallucinated scripts.

Long-horizon autonomy also got a measurable trial. Starting from a basic AlphaZero implementation, hill-climbing teams on Argon trained a 9x9 Go engine from scratch, with humans setting only chip allocation and evaluation. In under eight days it gained roughly 2,600 Elo to reach 1,885, placing it between KataGo at 16 and 32 visits per move. Champions were promoted strictly by statistical tests on locked records, not by human pick.

There is also an external timeline. Gemini 4 models can be used with /teamwork and /boost in Antigravity today, and Deep Review is open to paper authors. Google has separately opened applications giving 10,000 academic researchers six months of free access at its highest usage tier; the formal proof pattern and more specialized workflows are due in Antigravity in coming weeks.

Evidence boundary: every figure here comes from Google’s own blog, is vendor-reported, and most has not been independently reproduced. The chemistry numbers — 95% top-1 and 100% top-3 — are results on specific benchmark molecules, not a general claim about structure elucidation.

Sources:

Theme 3: Agents get a wallet, and they get an identity protocol

OpenAI introduced ChatGPT Wallet, letting dot complete purchases on a user’s behalf with approval. The mechanism matters: the agent never touches the real card number, every purchase passes a security check, and existing card protections and rewards are preserved. It is available from today in the US for dot on Pro subscriptions.

Sierra released Personal Agent Protocol (Poppy) v1.0 the same day, addressing a different problem: how a personal agent proves to a business who it is and what it is allowed to do. Companies publish a /.well-known/poppy.json file on their own domain declaring available entry points — website, OpenAPI, MCP or the company’s own agent — and users authorize through OAuth on the company’s own page, after which a single session continues across web, API and conversation.

The rollout says more than the specification. Every storefront on Shopify has already published a Poppy file automatically, with no merchant work required, and Meta’s personal agent Muse has Poppy enabled for all users. Live examples include Mercury for balance checks and transfers under account approval rules, and Turkish Airlines for fares, flight status and existing bookings.

For businesses, the appeal is not having to build the full authorization stack at once. Some capabilities work without login, so a merchant can publish a discovery file first and add OAuth and account actions later. Sierra says it has added 60 design partners for a total above 100, and ships an implementation toolkit containing a skill for coding agents plus a command-line checker for deployment compliance.

Mind the version state. The official site, the toolkit and the GitHub repository still label the current text as draft 1.0, and the changelog warns that fixes before final release may not be backward compatible. Payment notifications, push and attachments are not yet covered. The specification uses Apache 2.0.

Sources:

Theme 4: Anthropic runs three tracks at once — listing, rollout, architecture

Anthropic’s official account confirmed it has filed a registration statement with the SEC including a prospectus, and noted that securities may not be sold before the statement takes effect. Filing is not the same as completing an offering or listing; the filing text is outside this capture, so neither deal size nor price range can be verified. Speaking with JPMorgan CEO Jamie Dimon in New York, Dario Amodei said he is not much bothered about timing.

On the product side, the rollout is broadening. Ado Kukic said that as of November 13, every US Pro, Max and Ultra subscriber on the Claude Duet waitlist had received an invitation. The research preview includes limit resets — three for Pro, one for Max and Ultra — and the company plans to open more countries soon.

Claude Meetings shows where the product boundary sits. The beta is opening gradually to Team and Enterprise plans, and the usage is to add Claude as a visible participant in team video calls, producing documents and whiteboards while the discussion happens, with meeting notes also falling to it.

On the engineering side, Anthropic published an architecture note others can copy. “Building always-on agents” breaks a persistent agent into five parts: an inbox that decides when it wakes, timers that let it act unprompted, a memory folder so it remembers context across wakeups, permissions that bound what it may do, and an outbox that decides when it speaks. The core judgment is that “always on” is a user experience, not a reason to keep a model running and burning tokens.

A counterexample appeared the same day. One user reported that Claude Meetings sat out an entire client meeting, because it treated every non-member of the user’s organization as an external party to guard against — when those people were the entire point of the meeting.

Sources:

Theme 5: Four security stories in one day, and permission scope tightens before model capability

Cursor published its incident review: attackers used an access key exposed by Fireworks to read coding agent traces used for model training, and fewer than 0.01% of customers had API credentials exposed. A parallel thread involved a malicious Claude Code extension affecting up to 5,000 users. Cursor’s guidance is that credential rotation and extension reinstalls depend on individual circumstances.

The Grok Build problem sits at another layer. QbitAI, citing a Claroc investigation, reported 37,571 apps online among the tested addresses, 94.7% of which opened without login, some containing corporate material, patient records and plaintext passwords. This is not a model flaw; it is treating a random URL as access control. By the time of publication, 24 apps created by employees of the companies involved had been taken down.

Two more findings are closer to research. A Microsoft study named “Lazarus” reports implanting agent backdoors by rewriting fewer than ten dormant neurons, with no retraining or data poisoning, reaching a 98.8% trigger rate in some experiments — conditional on write access to model weights, and with no real-world attacks observed so far. The GCSA security team disclosed taking control of a robot on real hardware from a leading vendor via close-range Bluetooth or a voice command, and driving it into a restricted area.

The common thread: the permission scope agents receive is expanding faster than their judgment.

Sources:

Theme 6: DeepSeek splits its app into APIs and open-sources the front end

Several developers relayed a DeepSeek announcement: web search, deep-reasoning search, file parsing, link reading, voice interaction, and lower-level capabilities such as OCR, ASR and TTS are all exposed as standalone APIs, callable individually or combined. The company also open-sourced a complete chat front-end demo that can import conversation history.

The direction runs against the mainstream. Most companies are pulling capabilities into their own flagship app to retain users; DeepSeek is unbundling what sits behind its entry point, which turns “replicate a DeepSeek app” into an executable plan. Developers have already noted that ASR pricing is well below comparable services.

One line in the announcement explains the motive: the existing chat product is only “a fairly basic form”, and the capabilities behind it can now be freely combined through the API, so developers can skip repetitive groundwork. For small teams building vertical products, that removes the need to wire up search, OCR and speech synthesis yourself — and pushes differentiation further up the stack.

Evidence boundary: this capture contains only social-media relays and screenshots, with no official announcement text or pricing document. Quotas, rate limits and billing must be confirmed against official documentation. The Friday-evening Beijing launch time also drew a predictable reaction from Chinese developers.

Sources:

Theme 7: The open-source toolchain closes three gaps in one day

llama.cpp shipped v1.0.0, the first fully stable major release of the most widely used local inference library. Clément Delangue emphasized control rather than performance: running locally on phones, laptops and workstations, with data never leaving the machine.

Chrome’s Modern Web Guidance left preview and shipped stable 1.0. It targets stale training data in coding agents — models that rewrite a CSS scroll-driven animation as a large JavaScript library, or stack nested divs instead of using appearance: base-select. Version 1.0 bundles more than 200 expert-reviewed guides, Baseline compatibility data and fallbacks, maintained jointly with the Firefox and Edge teams. Partner figures include MetaMask driving 43 accessibility and semantic HTML fixes across roughly 480 pull requests, and note inc. lifting its modern web check pass rate from 61% to 94%.

Alibaba open-sourced open-code-review, its internal code review tool of two years: it reads a Git diff, then a tool-using agent opens full files and searches the repository to produce line-level review comments, with built-in rules for null pointers, thread safety, XSS and SQL injection. The README claims higher precision than a general agent on the same model at about one-ninth the tokens. OpenAI released Codex Security’s CLI and TypeScript SDK, covering vulnerability scanning, verification, patch generation and SARIF/JSON/CSV export.

Two more projects on the same trending list point the same way. Tailscale’s tailcat opens a WireGuard end-to-end encrypted tunnel between machines like netcat, with automatic NAT traversal and no account or root required. Chrome DevTools’ official MCP server lets coding agents drive a real Chrome — recording performance traces, reading the console, taking screenshots — so agents no longer write front-end code blind.

All of these point at one shift: large vendors are now feeding their internal engineering standards directly to agents, rather than waiting for agents to absorb them from public corpora.

Sources:

Theme 8: Capital prices AI one way, risk departments price it another

LangChain raised a $300 million Series C at a $3 billion valuation. The official framing is letting every company “own its intelligence”, landing on LangSmith as the platform to run, improve and govern agents. Harrison Chase said a batch of product launches follows next week.

A second document reads in the opposite direction. The Fed’s financial stability report found 87% of market participants naming AI as the top financial risk in the New York Fed survey, up from 50% in spring. It said enthusiasm for superintelligence has pushed equity valuations up, that AI infrastructure debt exceeded $680 billion as of October — roughly five times all of 2023 and about 31% of investment-grade corporate borrowing — and that the largest banks’ exposure to big tech, data-center construction and financing vehicles approaches $1 trillion.

Put the two together and Nvidia’s guarantees-for-revenue stop looking like one company’s operating choice; they look like a sample of the financing structure. TechCrunch adds the Israeli view: Nvidia has 6,000 employees and its largest overseas R&D center there, and investors weighing a window for challengers concede that if uncertainty slows overall AI spending, challengers face a smaller market.

Sources:

Theme 9: Three evaluations land at once, and the scores now need a manual

Anthropic published its safety evaluation methodology from a blunt starting point: a claim that a model is “safer than the last one” can only rest on some evaluation, and if the evaluation is weak the claim does not hold. The team dropped refusal rates for suicide and self-harm scenarios in favor of an “appropriate response rate” — a user in crisis who gets no response has already been failed — and a multi-turn conversation fails if any single turn fails.

The configuration gap is the most striking finding. The same model as a raw API and as a version with a system prompt differ by up to 34 percentage points in multi-turn appropriate response rate. In the system card, the suicide and self-harm multi-turn evaluation shows exactly that spread: 60% (±14%) on the API versus 94% (±7%) with the new system prompt. Their counterexample matters just as much: in a romance-scam evaluation Opus 5 scored 98% and Opus 5.5 scored 94%, a difference that is not statistically significant and cannot support a safety claim either way.

The investment-research finding is sharper. BuySideFlow collected 408 Chinese-language investment research tasks, where nine agent configurations scored between 16.2% and 36.8% accuracy. Auditing 900 runs across 100 shared tasks found that of 380 runs passing result checks, 193 violated point-in-time constraints — 50.8%. A self-consistent answer is not evidence that only then-disclosed data was used, and this class of error is invisible to output comparison.

Medicine and memory supply the same caution. MedGPT averaged 85 on 40 undisclosed difficult lung cancer cases where 250 tertiary-hospital physicians averaged 53 — but that is an answer-sheet evaluation, not a comparison of clinical outcomes. CAMP-II trains long-term memory through ordinary file operations and reports a 2.97x gain over the strongest OPD baseline, not total success rate; memory behavior transferred to Terminal-Bench 2.1 and SWE-Bench Pro V2 with improved success rates. Anthropic’s operational advice: state the model version, prompt and grader in any report, and never compare numbers across graders.

Sources:

Theme 10: Embodied AI at both ends — a long-horizon demo and a door into 3D content

Jim Fan posted a GR00T 2 demonstration: the model controls a real dexterous hand end to end to complete a chemistry experiment and clean up, shown unedited at 1x speed, across more than twenty steps lasting over five minutes. The post also describes transferring from human wearable devices and simulation data to robots. A single demo is not generalization, but long-horizon completion is closer to a usability signal than the usual single-action clip.

World models are laying down entry points. Fei-Fei Li announced that World Labs’ Atlas is open to all users with an app and an API, generating keyframes, video clips and 3D Gaussian splats from the world model. Open access is not open weights, a distinction that gets blurred when people assess how “open” a model really is.

The compute and hardware narrative arrived the same day. Elon Musk reposted a plan for a SpaceX supercomputing campus in Arizona: a proposed $50 billion investment, an estimated 6,500 jobs including 1,000 permanent ones, and $3.6 billion for grid upgrades. Jobs, tax revenue and electricity savings are planning projections, not built facts. Meanwhile the social feed’s “Jensen Huang goes to Tesla to build chips” scenario is entertainment; the people discussing it noted the basic fact it skips — Nvidia has no fabs, and leading-edge process is yield ramp and decades of process work.

High-value briefs

  • Blackwell fingerprints in Chinese model weights: Unemployment Labs, a nonprofit research group, analyzed open weights from DeepSeek, Moonshot, MiniMax and Ant Group, claiming Kimi K3 and DeepSeek V4 show training traces of export-restricted Blackwell chips, and traced smuggling channels and Malaysian offshore compute. It is one research group’s inference, with no vendor response. https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/
  • Tencent Hunyuan Hy Realtime: the realtime speech model adds tool calling and singing, live in Yuanbao voice calls and Tencent customer service, with a Tencent Cloud API coming. https://mp.weixin.qq.com/s/bFxAT04ohJs9hPcjjin3bg
  • Gemini 4 Hydrogen: replaces 3.8 Flash with four thinking depths, which Google says is 2.5x faster, at $1.50 and $7.50 per million input and output tokens; on Box AI Complex Work Eval it scored 5 points higher than 3.8 Flash at under half the time.
  • Runway opens Praxis-1: reported as a 27B-parameter MoE diffusion transformer with weights and fine-tuning code under Apache 2.0, and 46.2% success on RoboLab-120; the repository text is not attached.
  • A Claude Code Stop hook trick: to retry automatically when a check fails, the hook must exit with code 2; exit 1 is non-blocking and Claude finishes even when the check failed. https://x.com/lydiahallie/status/2121313048820617645
  • GitHub’s production guide to Agent Skills: written with the Procore team on turning engineering judgment into skills reusable across agents. https://www.xiaohu.ai/article/github-agent-skills-guide/

🕐 Selected hourly signals

PT time Signal Why it is worth remembering
04:00 Full Cursor incident review The attack chain starts at a third party’s exposed access key, not the model
09:00 Sierra Poppy 1.0 walkthrough Every Shopify storefront published automatically; rollout outran the spec
10:00 Claude Meetings user feedback Misreading clients as external visitors exposes the permission model before capability
11:00 Alibaba open-code-review released 50k stars; README claims about one-ninth the tokens of a general agent
13:00 Modern Web Guidance goes stable Maintained by three browser vendors; fixes agents’ stale training data
16:00 Anthropic’s always-on agent architecture Five parts with clear responsibilities; “online” is experience, not idle compute
17:00 Codex Security and chrome-devtools-mcp trend Vendors shipping internal tools as agent interfaces
18:00 Anthropic publishes safety eval method Up to 34 points of difference between API and prompted configurations
19:00 Gemini 4 Carbon called better for RSI Developers say kernel hill-climbing triggers no safeguards; usable as an RSI advisor

Editorial conclusion

What actually changed today was not a company’s ranking but the arrival of three things at once: the leadership of the compute supplier was caught by the capital structure it helped build, frontier labs began delivering research that machines verify line by line, and agents received wallets, identity protocols and enterprise permissions. The first two decide whether this round of spending can continue; the third decides who is accountable when it goes wrong. Nvidia reports earnings on Tuesday — whether Kress commits to the existing chip roadmap, and whether the board sets a timeline for a permanent CEO, is the nearest checkable signal.

Sources and method

Reviewed 20 hourly captures and 9 named sources, about 280,000 bytes; two named sources had no new posts that day and two were under 500 bytes. Nvidia’s personnel change and Anthropic’s listing both lack company statements or regulatory filing text, so relay level is marked. Vendor benchmarks and single demos were not independently reproduced.

WeChat QR code for 智简 Smart&Concise

FOLLOW ON WECHAT

智简 Smart&Concise

Search in WeChat for independent development and AI updates.