Nvidia Unresolved a Week After Huang's Exit as Qwen4-Max and Claude Code Code Mode Advance on the Same Day
November 15 was pulled in two directions at once. One was Nvidia: three days after Jensen Huang resigned, the Financial Times reported that he has been discussing a return with…
November 15 was pulled in two directions at once. One was Nvidia: three days after Jensen Huang resigned, the Financial Times reported that he has been discussing a return with confidants, the “Stand with Jensen” letter kept collecting signatures, and the independent investigation still has published no conclusion. The other was products and engineering: Qwen4-Max shipped with 2.4T backbone parameters and a million-token context window and took the top Chinese-model slot on Agent Arena; Claude Code changed tool calling into writing a piece of JavaScript; and Anthropic wrote “always-on agents” down as a reference architecture.
A note on evidence before the details: this archive contains no Nvidia corporate announcement, no succession plan, no investigative finding, and no third-party reproduction of any benchmark number that appears below. The Qwen4-Max figures and the Agent Arena ranking both come from the publisher and the leaderboard operator. The OpenAI science posts were blocked by Cloudflare, so only RSS metadata was readable. Where a claim rests on one retelling, the retelling is named.
Theme 1: One week into Huang’s exit, a return rumor and an unfinished investigation coexist
According to AI Valley and a relayed episode of a The Verge podcast, Huang resigned as Nvidia’s CEO and from its board on November 12, effective immediately, with CFO Colette Kress stepping in as interim CEO. Shares fell more than 13% intraday on the day the news broke, closed roughly 4% lower, and recovered the following session. The material here contains no company announcement text and no detail on the succession arrangement.
On November 15, the Financial Times reported, citing multiple people familiar with the matter, that Huang had already discussed the possibility of returning as CEO with several confidants. One ally is said to be assembling Silicon Valley executives into a “war room” to lobby the board and seek White House support. The report did not identify that ally.
The public letter “Stand with Jensen” asks the board to let him return first, to wait until the independent investigation establishes the facts, and to publish the investigation’s scope and timeline. According to Moneywise on November 14, it had passed 100,000 signatures, with more than 5,000 people identifying as current Nvidia employees; a separate post on November 15 gave roughly 150,000 signatures and more than 7,000 employees. The two figures use different counts, and both come via retellings.
The stated background for the resignation centers on personal relationships, plus an export-control investigation involving Megaspeed, a Singapore-based cloud provider. Reporting and statements from lawmakers have already appeared, but the material contains no company investigation conclusion and no judicial finding. Nvidia reports earnings on November 17 (Tuesday), the first time interim management will publicly account for the quarter.
Sources:
- https://www.theaivalley.com/p/nvidia-s-first-week-without-jensen
- https://x.com/dotey/status/2122062085807030716
- https://x.com/Hesamation/status/2121967243932397915
Theme 2: Qwen4-Max takes the top Chinese slot on Agent Arena with 2.4T backbone parameters and a million-token context
The Qwen team released Qwen4-Max on the new Qwen4 architecture, with 2.4T backbone parameters, 71B activated per token, 328B N-gram Embedding parameters, native vision support, and a context window of 1 million tokens. The official blog covers architecture details, cost comparisons, and industry case studies at the same time.
The same day, Arena’s leaderboard account said Qwen4-Max placed 14th on Agent Arena with a net improvement of +7.5%, making it the highest-ranked Chinese model and making Alibaba the top-ranked Chinese lab and fifth overall, behind Anthropic, Google, OpenAI and Meta. That ranking is based on 14,700 real agent sessions, and comes with a breakdown across four signals.
Parameter count and context length have become routine narrative for Chinese model launches. What is worth noting this time is that the evaluation dimension changed: not single-turn question answering on a leaderboard, but multi-turn agent sessions with tool calls. That also means the score depends more heavily on the specific task distribution, so cross-leaderboard comparison has limited value.
Both pieces of information come from the publisher and the leaderboard operator; there is no third-party reproduction in the material. The real inference cost of 2.4T parameters, and the effective recall rate of a million-token context on long tasks, have no independent numbers yet.
Sources:
Theme 3: Google ships two models, two TPUs, and a cluster of research updates in one go
According to a release list relayed by management, this round from Google listed two eighth-generation TPUs, Gemini 4 Carbon, Gemini 4 Hydrogen, Gemini Robotics 2.5, along with AlphaGenome 2, the Gemini Decisions API in AI Studio, and smart glasses preorders, plus the use of AlphaProof Code to verify the Signal protocol. The list provided no complete evaluation data for any of these products.
The default models are being adjusted in step. According to AI Base, Anthropic made Claude Opus 5.5 the default model for Pro, Max, Ultra and Team plans, with the emphasis on research, analysis, writing and document work. On the Google side, Antigravity announced that it is doubling the context window for Argon, Carbon and Hydrogen.
Developer reaction is split. Some say Carbon differs noticeably from Claude and GPT in scheme design, document writing and creative work, making the two useful for cross-inspiration; in the same batch of feedback, Antigravity’s interaction is repeatedly criticized as hard to use, with some people reverting to the command line. Others concluded from this that OpenAI has already dropped to third place — that is a personal impression, not a leaderboard result.
The value of this list is its density: chips, models, research tooling and hardware preorders moving together, with the unit of competition shifting from a single model to the combination of “model plus context plus developer tooling.” Without evaluation data, it is not yet possible to judge how large the gap between these releases actually is.
Sources:
Theme 4: OpenAI brings GPT-6 Astra into a lattice standard model proof and opens a science blog
On November 15, OpenAI published two posts. The first, “A step toward defining the Standard Model on a lattice,” discusses a candidate non-perturbative lattice definition of the Standard Model, and the RSS summary states that GPT-6 Astra helped push the idea into a detailed proof. The second announces a new science blog, positioned as sharing research findings, methods, and first-person accounts from researchers.
The official account then confirmed that physicist Nathaniel Craig described the process of proposing the lattice definition candidate together with GPT-6 Astra; Greg Brockman posted a single line, “GPT-6 Astra for theoretical physics.” Another post says Fields medalist Edward Witten has published a follow-up paper built on Craig’s work, using Opus 5.5. Witten is one of nine members of the mathematics and AI advisory group, which recently asked OpenAI to publish mathematical proof results so they can be independently checked.
The weight of this story lies in the setting: a lattice definition in high-energy physics is a long-unsettled formalization problem, and if a model can genuinely move the construction of a proof forward, it is harder to cheat by having seen the answer than any code benchmark. The boundary of the material is equally clear: the OpenAI post itself was blocked by Cloudflare, so only RSS metadata was available. Whether the proof is correct, and how much of it is the model’s contribution, both require independent review by mathematicians.
Sources:
- https://openai.com/index/step-toward-defining-the-standard-model
- https://openai.com/index/our-new-blog-for-science
- https://x.com/OpenAI/status/2122123437141553630
Theme 5: Claude Code’s Code Mode turns tool calls into a piece of JavaScript
Claude Code introduced an experimental feature, Code Mode, off by default. When enabled, the model no longer calls one Bash, Read or Edit command at a time; it writes a piece of JavaScript that invokes multiple tools in a single execution, which Claude Code runs locally in the session. The official example is raising a retry limit from 3 to 5: use rg to find the files, change the parameter, run the tests, and return only the file paths and test output to the context.
The gains come from three places: independent commands can run in parallel; the result of one step can be passed directly to the next without going through the model; and the large volume of intermediate output can be filtered in code first, with only the needed part placed into the context. The recommended use cases cluster around the same three properties — a single change that touches many files, search results that need filtering first, and several independent commands that can run at the same time.
The limits say more about its positioning than the benefits do. It works only in interactive terminal sessions, and does not support -p, the Agent SDK, the VS Code panel, desktop or cloud sessions. It cannot run through a gateway, Bedrock, Google Cloud, or Microsoft Foundry, and supports only Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2. Foreground runs are capped at 10 minutes; background runs at 2 hours, and require the session to have started in auto or bypassPermissions mode.
Turning it on means setting CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1; setting it in a project settings file does not take effect, so a cloned repository cannot quietly enable it for the user, and the first activation requires confirmation. The change to the permission model is worth remembering on its own: every tool call still passes deny and ask rules and the PreToolUse hook, but in auto mode the classifier now reviews the whole block of code, so it cannot see values that only exist at runtime — a path taken from search results, for example. The official advice is that calls which must be controlled need to be covered by rules or hooks.
Sources:
Theme 6: Always-on agents written down as a reference architecture
Anthropic published “Building always-on agents,” which breaks “always on” into five parts: an event inbox where every trigger source writes into a single persistent queue; timers holding follow-ups and schedules in separate storage; memory kept in persistent storage independent of the session; permissions; and egress. The loop runs like this: an event enters the inbox, a worker starts one run and either creates or resumes a session, the agent loads memory, acts within its permissions, updates memory, the result exits, and a dispatcher delivers it; a timer or a user reply then sends the event back to the inbox.
The central judgment of the design is that always-on is an experience, not a machine left switched on. Each run is short, most of the time is spent dormant, and persistent state stitches multiple runs into a continuous experience. The document also treats how to recover after an interruption, which actions may run autonomously, and when to tell the user as equally important questions.
The community is already aligning implementations. One post assembles a comparable architecture from Pi Durable and the Pi AI SDK: Conversation stores the discussion, Submission receives input and uses a stable requestId to correlate duplicate requests, Harness schedules generation and tool tasks, Durable Storage keeps history, checkpoints and results, and the store is reopened after a process restart to restore state. The author stresses that external operations still need idempotency or result checking.
The value of a document like this is that it turns an “agent project manager’s” intuition into implementable component boundaries. It also explains the difference in current product forms: an assistant that can work continuously for several hours and an assistant that remembers an agreement made three days ago are not the same engineering problem.
Sources:
- https://x.com/vikingmute/status/2121953959514636306
- https://x.com/verysmallwoods/status/2122008167517135247
Theme 7: Poppy 1.0 gives enterprises a front door for agents
According to a retelling of the Sierra protocol release, Poppy 1.0 lets an enterprise publish its website, API or enterprise agent entry points through a discovery file under its own domain, and connects identity descriptions with login authorization. It addresses the identity problem when an agent acts on a user’s behalf against enterprise services: a merchant needs to confirm that the caller is a personal agent acting for a specific user, not an anonymous crawler.
Deployment has already begun. According to Sierra, stores on the Shopify platform have published Poppy files automatically, and Meta Muse has enabled the protocol. Commerce integration runs through UCP, with the actual shopping flow handled by UCP.
One developer’s hands-on report is more concrete: Alma 0.4.672 supports automatic discovery of a company’s /.well-known/poppy.json, attaching a signed Poppy-Session header when accessing an API, page or MCP server; at UCP checkout it carries Alma’s UCP platform profile, and in the test it negotiated the full set of checkout, cart and order capabilities against the Allbirds Shopify UCP endpoint.
If this protocol spreads, the bottleneck in agent shopping shifts from “can it click through the flow automatically” to “does the merchant recognize this identity.” The current evidence is still the protocol publisher’s account plus one implementer’s test; the merchants’ own accounts are missing.
Sources:
Theme 8: The same model’s safety score varies by 34 points, and expert disagreement can change the correct answer
A set of evaluations published by Anthropic shows that Fable 5.1 scored 60% (±14%) in API configuration on a multi-turn suicide and self-harm evaluation, and 94% (±7%) on the web product with a new system prompt. The same model, a 34 percentage point gap, produced by deployment rather than by weights.
The team argues for measuring “rate of appropriate response” per scenario, and publishes the system prompt, interface and scoring configuration alongside the results; it recommends at least 100 cases for complex multi-turn evaluations and at least 1,000 for simple single-turn ones. In practice that requirement says something plain: a safety score detached from its configuration is not comparable.
Another study gives a warning of the same kind from the opposite direction. A two-stage revision experiment fixed the question, the evidence and the initial answer, then compared neutral reflection, anonymous disagreement and domain-expert disagreement. Across 1,989 medical, scientific and legal judgments, none of the five models reached the selective-correction threshold needed for a net accuracy gain under expert disagreement, with accuracy falling by as much as 16.1 percentage points.
The layered result is worth noting: adding only the “expert” label moved accuracy anywhere between a 5.2-point drop and a 2.4-point gain, and must be read separately from the overall effect of disagreement. A model’s sensitivity to authority and its sensitivity to argument are two different things, and the first is easier to exploit.
Sources:
Theme 9: Three routes to lower cost — low-precision training, memory compression, and kernel optimization
Thinking Machines Lab released Inkling 2 Small under Apache 2.0: 594 billion parameters, with four-bit precision for forward computation from pretraining through final reinforcement learning. The material states an 8% reduction in pretraining time and a 20% reduction in RL time; backward computation was not kept at four bits throughout. The VRAM floor is stated just as plainly: NVFP4 requires at least 500GB, MXFP8 at least 800GB — “Small” in the name does not mean it runs on a single consumer card.
Stanford researchers took a different route with Memorilla: encode a document collection into a fixed set of 16 memory tokens used to assist a frozen model. On FactKG, a frozen Qwen3-8B reached 84.33% accuracy against 80.43% for a full-context baseline, using roughly 108x fewer document tokens. It is not a comprehensive win — it still trails the baseline on benchmarks such as PubMedQA — and the code, training recipe and checkpoints are released under Apache 2.0.
The fifth version of the CUDA-L2 paper puts the optimization target at the kernel layer: combining a large model with reinforcement learning, using CUDA execution speed as the reward, automatically optimizing 1,000 half-precision matrix multiplication configurations. Under the tested conditions, the offline mode averages an 11.4% gain over cuBLASLt’s autotuned baseline, and the online mode, which simulates real-time inference intervals, gains 15.9%.
The three routes map onto VRAM, context and kernels, and their common point is that none of them claims general acceleration. These numbers hold for specific configurations and benchmarks, and cannot be extrapolated into claims like “training costs dropped by ten percent.”
Sources:
Theme 10: Compute commitments, financing scale and one lawsuit put the risk on the table
The Decoder reports that Anthropic’s total compute commitments approach $901 billion, and that CEO Amodei admits to enormous uncertainty about how the business will develop over the coming year. The report reads the proposed IPO as a financial buffer absorbing the mismatch between revenue and long-term costs. A commitment total is not incurred spending or current debt, and that distinction needs to be kept separate from the financing narrative.
On the OpenAI side, according to a retelling by AI Base, a new private financing round is targeted at $15 billion at a pre-money valuation of $1.25 trillion, expected to complete within the coming weeks. The report also gives roughly $10 billion of third-quarter revenue against an operating loss of about $10.9 billion, and links the raise to IPO preparations expected in 2027. The size, valuation and timing should all be read as reported.
On the regulatory side, the action got harder. According to the litigation reporting, four nonprofits filed a complaint in the San Francisco Superior Court on November 12, naming OpenAI, Anthropic, Google, Meta and SpaceXAI as defendants and seeking an emergency injunction freezing next-generation model training and deployment. The complaint cites executives’ risk statements as material.
What needs separating: filing a complaint is not a court-ordered halt on training, and the plaintiffs’ risk narrative is a legal claim. The value of this thread is that it moves “should training continue” out of forum arguments and into court procedure.
Sources:
High-value briefs
- Cloudflare saw 50-plus framework families scanning its customers within three days: the blog attributes this wave of scanning to LLM-driven tool chaining, and argues for identity isolation, permission constraints and microsegmentation to limit spread after a breach. The “30x request growth” figure in the summary has no corresponding statistic in the excerpt provided.
- Shopify runs a dual-agent red team in production: the offensive agent combines source code, infrastructure knowledge and temporary credentials for a test store to validate vulnerabilities, while an independent judge monitors behavior and terminates the task on anomalies. The company says it has executed tens of thousands of tasks and found dozens of issues with real impact, with no downtime; this is the company’s own account.
- A robot vulnerability received preliminary vendor confirmation: GCSA says it obtained the highest OS privilege on a robot from a leading vendor, and the vendor confirmed it preliminarily by email on November 13, starting remediation and an assessment of the scope of impact. The brand and the vulnerability mechanism remain undisclosed, and starting a fix does not mean the risk is gone.
- Multi-agent delegation has no stable gain: Lost in Delegation splits tasks by whether they fall before or after the training knowledge cutoff, and finds that delegation is not consistently better than no delegation; when the orchestrator model lacks post-cutoff knowledge, accuracy can fall markedly, while the same model using tools directly does better. Common problems are query distortion, inefficient tool use and information lost across agents.
- Frontier models being more usable does not mean solving ability rises in step: a long LocalLLaMA post separates “efficient problem solving under limited information, tools and time” from “long-horizon planning ability,” offering three observations: once a golden dataset is cleared, swapping in a newer model rarely adds more; the same model performs very differently in different surrounding systems; and smaller local models are already capable. Nearly 300 comments split, with some pointing to an obvious recent jump in hypothesis testing in rebuttal. This is a single-thread discussion, not an evaluation conclusion.
- 29 agents generated 50 causal evaluation tasks in 11.5 hours, and reviewers found template shortcuts: the tasks required judging whether causal effects are identifiable and estimating the effect when they are, and independent reviewers found wording cues and template shortcuts, so the team converted the problems into follow-up tests. Automatically generated evaluations still need auditing.
- The “No AI Slop” declaration reached 504 signatures: written jointly by speakers and attendees of AI Engineer Code Summit 2026, it uses the agile manifesto as a reference and sets out five paired contrasts, defining slop as “you invested less attention than you ask your reader to spend understanding it.” The Hacker News discussion sits at roughly 222 points and 133 comments.
- llama.cpp released 1.0.0: a version milestone for the local inference stack, with the official emphasis on running “anywhere.”
- dspy-wsp lets frontier models edit the weights of smaller models: an optimizer prototype in the DSPy ecosystem that performs no gradient updates, instead having a frontier LLM edit the small model’s weights — a public implementation of the “programming in weight space” idea.
- Agent tooling keeps filling in the layer of “how people manage agents”: AG-UI connects agents to front ends with roughly 16 standard events (about 17,200 stars accumulated); phonedriver lets Claude Code, Codex and Cursor operate Android and iPhone, reading each screen as a numbered text list, and across 9 tasks its token usage was 1/6.9 that of mobile-mcp; Plannotator puts plans, prototypes and diffs in the browser for people to circle and annotate, adding 860 stars in a day.
- Video and slide decks keep moving to the web: OpenFilm has coding agents make video with HTML/CSS, where every shot is a web page exportable to Final Cut Pro, Premiere and DaVinci Resolve; another set of HTML presentation templates added 10 styles, supporting in-browser text editing, presentation tools and phone remote control.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 06:03 | The “Stand with Jensen” letter reaches roughly 150,000 signatures, with more than 7,000 self-identified Nvidia employees | Pressure extends from public opinion to the employee layer |
| 10:05 | llama.cpp releases 1.0.0 | A version marker for the local inference stack |
| 12:20 | FT: Huang has discussed a return with confidants; an ally is assembling a “war room” | A possible return moves from rumor to a concrete lobbying structure |
| 15:27 | WeChat opens AI beautification, editing and information extraction for images in chats | An entry-level product pulls image handling into the conversation |
| 16:23 | OpenAI confirms the science blog’s first post involved GPT-6 Astra | Corroborates the paper release earlier the same day |
| 17:24 | Antigravity announces doubled context windows for Argon, Carbon and Hydrogen | The competition shifts from model capability to context capacity |
| 18:00 | Qwen4-Max released with 2.4T backbone parameters and a 1M context | The strongest model release of the day |
| 18:21 | Agent Arena: Qwen4-Max is the top Chinese model, 14th overall | A third-party leaderboard gives evidence pointing the same way |
| 18:36 | The “No AI Slop” declaration reaches 504 signatures and hits the Hacker News front page | Engineering culture starts constraining AI output with a declaration |
Editorial conclusion
The technical progress of the day is real: models are getting bigger, contexts longer, training cheaper, and agents are acquiring protocols and reference architectures. But the most worth-remembering evidence sits outside the scores. A safety score floats by 34 points depending on deployment, an expert label can make a model abandon a correct answer, and multi-agent delegation is not reliably better than one model with tools. Nvidia’s governance crisis points at a separate lesson: for that company right now, the largest uncertainty is not in the product roadmap but inside the boardroom and the investigation.
Sources and method
Reviewed 20 hourly captures and 9 named sources in that PT day’s directory, about 200KB in total; after removing duplicated reposts, book promotions and empty threads, roughly 20 usable candidates remained, and the archive was handled as a rich signal pool. AI Valley was a redirect day; Chrome, Claude, Cline, Google Research and XiaoHu had no new posts that day; and the OpenAI post was blocked, leaving only RSS metadata.
