---
title: "Jensen Huang resigns as Nvidia CEO; Google launches Gemini 4 Hydrogen"
date: 2026-11-13
summary: "Nvidia's board accepted Jensen Huang's resignation as president, CEO and director and named CFO Colette Kress interim CEO. Google launched Gemini 4 Hydrogen as its everyday model, and a former OpenAI researcher published evidence that six Chinese frontier models were started on export-controlled Blackwell chips."
cover: "https://morningtokens.com/covers/2026-11-13.jpg"
url: "https://morningtokens.com/issues/2026-11-13/"
---

Pour a long one. The story everyone expected for a week happened: Jensen Huang is out at Nvidia after 33 years, and the CFO is running the company while the board searches.

On the same day a former OpenAI researcher published forensic evidence that six top Chinese models were started on export-controlled Blackwell chips, which won't help Nvidia's standing in Washington.

Elsewhere, Google shipped Gemini 4 Hydrogen as its new everyday model, Anthropic let companies run Managed Agents entirely on their own infrastructure, and OpenAI turned its cyber research into two products.

## Top news

### 1. Chips · Jensen Huang resigns; CFO Colette Kress is interim CEO

Nvidia's board accepted Huang's resignation as president, CEO and director on 11 November, effective immediately, and announced it on 12 November. CFO Colette Kress is interim CEO (and stays CFO), and long-time director Harvey C. Jones becomes Nvidia's first-ever board chair.

- The 8-K says the board accepted the resignation on 11 November. The search for a permanent CEO covers internal and external candidates.
  - Kress (59) has been CFO since 2013. Jones (73), ex-CEO and executive chairman of Synopsys, has sat on the board since 1993.
- The press release promises that "operations, customer commitments and product roadmap" continue without interruption.
  - Kress's staff email (via CNBC) announces a company meeting on 18 November at 9:00 PST.
- AP and TechCrunch tie the exit to the affair reports covered earlier this week, including the alleged link to Chinese intelligence.
- NVDA closed 12 November at $166.19, down $7.54 (−4.34%), per Motley Fool/Yahoo Finance data.
- Musk had replied "That would be an honor" to an X post urging him to hire Huang, hours before the resignation. SpaceX shares rose on the speculation. Huang has not said what he'll do next.

> @TimSweeneyEpic called it "a sad day for NVIDIA, developers, America, and all humanity." @ClementDelangue said Hugging Face will "keep fighting for open-source AI, harder than ever." @tobi called it a "generational blunder" based on available information.

**Why it matters:** Nvidia is a $4T+ company built around one founder's product calls, and it now has a finance chief running it with Q3 results already scheduled for next week. Watch for any change to the roadmap, pricing or China stance, and for whether open-weights support, which Huang personally championed, keeps the same priority.

Sources: [NVIDIA newsroom](https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition) · [8-K](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000082/nvda-20261111.htm) · [CNBC staff emails](https://www.cnbc.com/2026/11/12/nvidia-kress-jones-employee-emails.html) · [TechCrunch](https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/) · [Tom's Hardware on Musk](https://www.tomshardware.com/tech-industry/semiconductors/elon-musk-says-it-would-be-an-honor-to-hire-ousted-nvidia-ceo-jensen-huang-possible-projects-could-include-terafab-spacexs-planned-gpu-manufacturing-or-teslas-ai-accelerators)

### 2. Policy · Six Chinese frontier models carry a Blackwell fingerprint in their weights

Will DePue (Unemployment Labs, ex-OpenAI) published a forensic report finding that six Chinese frontier models were initialised on Nvidia Blackwell-class GPUs, which need a US licence for export to China. Earlier models from the same labs match Hopper.

- The six models are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash.
- The method uses only the published weights: it recovers traces of each model's initial values, then searches chip layouts and seeds that reproduce them (posts on the timeline cite more than six trillion combinations).
- Matches, by chip:
  - B200/B300 (148 SMs): DeepSeek-V4-Flash and MiniMax-M3
  - GB200/GB300 (152 SMs): V4-Pro, Kimi K3, K3.1-Pro and Ling-3.1-flash
  - For V4-Flash and Kimi K3 the match was also confirmed on physical Blackwell chips.
- The report is careful about limits: the fingerprint shows how the weights were *created* and supports, without proving, that Blackwell also did the training.
  - It does not say where the chips were, how access was obtained, or whether export controls were broken.
  - Specially configured AMD hardware is the only alternative identified, and it's excluded for V4-Flash.
- Kimi K3.1-Pro starts from the same seed as K3, so it may be a further post-trained checkpoint of K3. Reproduction scripts are available to reviewers on request.

**Why it matters:** Until now, claims that Chinese labs used Blackwell came from officials and press reports without technical evidence. A public, weights-only method changes that, and it lands in the same week Nvidia's China exposure cost its CEO. Expect it to be cited in export-control debates and, possibly, in pressure on open-weight releases.

Sources: [DePue report](https://unemploymentlabs.com/checkpoint-forensics) · [@willdepue thread](https://threadreaderapp.com/thread/2120992457630326841.html) · [@theo](https://x.com/theo/status/2121007708564582772)

### 3. Models · Gemini 4 Hydrogen: the fast Gemini 4 for everyone

Google released Gemini 4 Hydrogen, a fast model for everyday coding and agent work. It is available across AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search and (for paid AI Pro/Ultra) Antigravity.

- On the Artificial Analysis Intelligence Index it scores 51 at its default high thinking level, one point above the larger Gemini 4 Argon, and 54 at a new xhigh level.
- Pricing: $1.50 per million input tokens, $0.075 cached, $7.50 output, flat across the 1M-token context.
  - That matches Gemini 3.8 Flash's scheduled standard price from 1 January 2027 and is double Flash's current introductory rate (VentureBeat).
- Four thinking levels: low, medium, high (default) and xhigh.
- Google's numbers versus 3.8 Flash:
  - Tasks finish 2.5x faster at high
  - Terminal-Bench 4.1: **57%** (18.2%)
  - More than 5x fewer fabricated claims
  - Partners saw 2–2.8x faster runs with 12–23% fewer tokens
- Other Google-reported results: #5 on arena.ai Agent Arena, HLE-Diamond 58.8%, Vals Tax Agent Bench 52.3% (#1), Vals Legal Research 55.8% (tied #1).
- Artificial Analysis cost per Intelligence Index task: $1.87 at high, $3.16 at xhigh.
  - At xhigh it used 304M tokens, which AA calls very verbose (median 72.8M).
- @AnikaCodes noted it debuted at #5 on Agent Arena, ahead of GPT-6 Astra (7th), with Claude Fable 5.2 still #1.

**Why it matters:** Hydrogen is now the default Gemini most people will touch, and it moves Google's "cheap fast tier" up in both quality and price. If you route traffic to 3.8 Flash, test Hydrogen at medium before switching wholesale: it beats Haiku/Luna/Ember on AA's index but costs more per task than all three.

- **Do this now:** Rerun your Flash evals with `thinking` at medium and high and compare cost per *successful* task, not per token.

Sources: [Google blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) · [VentureBeat](https://venturebeat.com/technology/google-launches-gemini-4-hydrogen-a-faster-smarter-pricier-flash-successor) · [Artificial Analysis](https://artificialanalysis.ai/models/gemini-4-hydrogen) · [@OfficialLoganK](https://x.com/OfficialLoganK/status/2120940951292268640)

### 4. Agents · Claude Managed Agents can now run entirely in your own environment

Anthropic opened a private beta of self-hosted Claude Managed Agents. You run the agent server, its database, the workers and sandboxes, credentials and session history in your own cloud or on-premises, and choose where model requests go.

- Components: `cma-server` (API plus agent loop; never runs shell/file tools itself), PostgreSQL 16+ (without it everything lives in memory), workers that pull jobs and run tools in sandboxes, and a console behind your own identity provider.
- Model calls can go to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry or your own gateway.
  - Prompts still leave your network; existing zero-data-retention terms apply as for any Messages API call.
- The design keeps the harness ("brain") separate from the sandbox ("hands").
  - Vault credentials stay outside the sandbox and are added by a proxy for MCP tools and supported API calls; you can use your own secrets manager.
  - Per-environment host allowlists block everything else. Anthropic's docs say it has no remote access to the server or its logs.
- Quickstart is a Docker Compose stack (dev only, plain HTTP on 127.0.0.1:8400). Access is via your Anthropic account team; no public price or GA date.
- Caveat from the docs and coverage: the agent's shell is a real shell with the worker's permissions, so isolation depends on your container or VM.

**Why it matters:** This was the main blocker for banks, healthcare and government teams that liked Managed Agents but couldn't put session history and credentials in someone else's cloud. It also shifts operations work (upgrades, backups, sandbox hardening) onto the customer.

Sources: [Anthropic announcement](https://claude.com/es/resources/articles/self-hosted-claude-managed-agents) · [Data and access docs](https://platform.claude.com/docs/en/managed-agents/self-hosted-data-and-access) · [Quickstart](https://platform.claude.com/docs/en/managed-agents/self-hosted/quickstart) · [Help Net Security](https://www.helpnetsecurity.com/2026/11/12/anthropic-self-hosted-claude-managed-agents/) · [@ClaudeDevs](https://x.com/ClaudeDevs/status/2120938988165091693)

### 5. Security · OpenAI launches Daybreak Red Team and Daybreak Defense Factory

At its "Cyber at the Frontier" event on 12 November, OpenAI turned its cyber-model research into two products on the Daybreak platform. They are Red Team, an early-access fleet of agents that pen-test web apps, and Defense Factory, which finds, validates and drafts fixes for vulnerabilities in your code.

- **Red Team** (early access, registrations open): agents use source code and app context to probe attack paths, validate findings from scanners and bug bounties, and retest patches, all within scope controls set by the operator.

> OpenAI's Nancy Wang called it "superhuman pen-test at the scale of automated continuous scanning" (The Register).

- **Defense Factory:** agents built on GPT-6 Cyber trace vulnerabilities through the codebase and develop patches; engineers review and test them before merge. It can ingest findings from your existing tools and from Red Team.
- @thsottiaux summed it up as new tools "across automated patching and automated red teaming"; the keynote recording is public.
- Background (covered earlier): the pipeline that produced 948 Linux LPE exploits and the GPT-6 Cyber model are what these products package.

**Why it matters:** Offensive and defensive security are becoming a product category for the labs themselves, not just for security vendors building on their APIs. For security teams, the pitch is continuous pen-testing plus fixes, with a human approving every merge.

Sources: [OpenAI Cyber at the Frontier](https://openai.com/cyber-at-the-frontier/) · [Defense Factory](https://openai.com/daybreak/defense-factory/) · [The Register](https://www.theregister.com/security/2026/11/12/openai-launches-security-products-tells-orgs-theyre-doomed-if-they-dont-use-them/5309500) · [RuntimeWire](https://runtimewire.com/article/openai-daybreak-red-team-early-access-pen-test)

### 6. Robotics · Runway opens Praxis-1, a 27B world action model, under Apache 2.0

Runway released Praxis-1 with open weights plus inference and fine-tuning code under Apache 2.0, including commercial use. It is a 27B mixture-of-experts diffusion transformer that predicts video and robot actions together.

- Scores on RoboLab-120 (NVIDIA's leaderboard, 1,200 episodes):
  - Praxis-1: **46.2%**
  - FLUX 3 Action: **42.9%**
  - Cosmos3 Nano: **36.8%**
  - π0.5: **28.0%**
- Fine-tuned on Google's RT-1 robot (not in its training data), it reaches 80.0% on SimplerEnv vs 52.4% for the original RT-1 policy.
- Full fine-tuning and LoRA are supported. It runs on H100/B200, RTX 5090/RTX PRO 6000 and Jetson Thor.
  - It is on Hugging Face with native LeRobot support, in vLLM-Omni, and hosted on Reactor.
- Launch partners Noble Machines, Standard Bots and Ultra ran it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.
- Runway says its research shows web video matches teleoperated robot video as pretraining data for policy performance.

**Why it matters:** A permissively licensed, top-of-leaderboard robot policy lowers the entry cost for small robotics teams, who can now adapt one checkpoint to their own hardware with LoRA instead of training from scratch.

Sources: [Runway](https://runway.com/news/company-news/praxis-1-is-now-available) · [GitHub](https://github.com/runwayml/praxis-1-oss) · [Hugging Face](https://huggingface.co/runwayml/praxis-1-base) · [The Robot Report](https://www.therobotreport.com/runways-praxis-1-world-action-model-for-robotics-is-now-available/)

### 7. Consumer agents · Meta's Muse now makes phone calls for paying US users

Muse can place outbound calls to US businesses for paid subscribers: cable companies, the DMV, insurers, restaurants without online booking. It was tested with early adopters since September.

- Muse has over 10 million users (Meta, last month).
  - Plans start at $20/month for 500M Muse tokens per week; $100/month buys 3B tokens.
- @alexandr_wang announced it as "let muse take care of all those calls you don't want to make."
- Earlier, Meta tested a "human concierge" where contractors placed some calls; 404 Media and Reuters reported it, and coverage says that test was rolled back after privacy concerns.
- Google's Gemini and the startup Instinct already offer similar calling. TechCrunch notes many receptionists hang up once they realise it's an AI.
- Wang also said Muse email addresses are coming to free users soon.

**Why it matters:** Phone calls are one of the last chores that still need a human on hold. The open question is whether businesses accept agent callers or start blocking them, which is exactly what United just did for its website (next item).

Sources: [TechCrunch](https://techcrunch.com/2026/11/12/muse-can-now-place-calls-to-u-s-businesses-if-you-pay/) · [Engadget](https://www.engadget.com/2315525/metas-muse-can-now-make-calls-on-your-behalf-if-you-have-a-paid-subscription/) · [404 Media on the human-call test](https://www.404media.co/meta-tests-muse-ai-agent-calls-that-are-actually-made-by-humans-in-a-call-center/) · [@alexandr_wang](https://x.com/alexandr_wang/status/2120930586827391350)

### 8. Agent access · United Airlines blocks third-party AI agents from booking and check-in

United is restricting "unauthorized, third-party personal AI agents or assistants" on its own platforms and, indirectly, through travel agencies and aggregators, covering booking, reservation changes and check-in.

- Reasons given: sharing credentials with AI assistants can expose data, and agent actions "may result in unintended purchases, account changes, or use of MileagePlus miles."
- Regulatory angle: check-in requires a human to acknowledge that bags contain no banned items; United says automated check-in could breach FAA and DOT rules.
- The change includes updates to booking, ticketing and distribution policies.
- It follows Bank of America urging customers last week not to share passwords with AI assistants. TLDR AI contrasts it with Expedia and Booking, which are more open to agents.

**Why it matters:** Consumer agents (Muse, Grok Bot, ChatGPT agents) are hitting the first serious wall from large brands. Expect more "authorised agent" programs and agent-identity standards; until then, anything that logs in with your password is on borrowed time at some sites.

Sources: [United help centre](https://www.united.com/en/us/fly/help-center/digital-trust-center/artificial-intelligence/third-party-ai-agents.html) · [TechCrunch](https://techcrunch.com/2026/11/12/united-airlines-blocks-ai-agents-from-booking-flights-citing-cybersecurity-risks/) · [The Register](https://www.theregister.com/ai-and-ml/2026/11/12/united-airlines-wants-you-to-check-in-not-your-chatbot/5309373)

### 9. Dev tools · shadcn launches UI SDK: components agents can build and operate

shadcn introduced UI SDK ("Era Two"), headless primitives plus an agent bridge. Agents assemble components from tested behaviours, and at runtime an agent can read component state and call component actions directly instead of clicking through a page.

- `@ui-sdk/core` ships behaviours (focus, keyboard navigation, selection, popups, drag and drop).
  - A `definePrimitive` component renders in React, Vue, Svelte, Solid and Web Components; Vue/Nuxt support credits @benjamincanac.
- `createAgent()` → `agent.snapshot()` lists components, state and available actions; `agent.act()` runs steps like `setValue`.
  - Apps can approve or refuse actions via an `allow` callback, and change callbacks see `source: "agent"`. It can register with browsers through WebMCP.
- shadcn's own benchmark (Sonnet via Claude Code, medians of three runs), versus browser tools on another library:
  - Tool calls: **16 → 5**
  - Characters read: **39.8k → 6.8k**
  - Time: **41 s → 16 s**
  - Cost: **$0.073 → $0.032**
- Each package includes version-specific Markdown docs and a testing kit for keyboard, focus and ARIA checks.

**Why it matters:** It's a concrete bet that apps will expose structured controls to agents rather than make them screen-scrape. Paired with United-style blocking, the likely split is: sites that want agents will publish an agent interface; sites that don't will block.

Sources: [UI SDK: Built for Agents](https://ui-sdk.dev/docs/agents) · [@shadcn](https://x.com/shadcn/status/2120951116682682646)

### 10. Compute · SpaceX plans $74B+ of AI supercomputing in Texas and Arizona

SpaceX said it will invest more than $74 billion in two supercomputing campuses: the already-reported ~$24B site near Starbase in South Texas and a new ~$50B campus in Cochise County, Arizona.

- Texas: generates its own power, 500+ permanent jobs (the gas-plant filing was covered on 9 November).
- Arizona: about 1,000 permanent jobs; still needs county approval.
- Context from the same week: Musk's Terafab chip fab is under construction, aiming at volume production in H2 2028 (Tom's Hardware), and Musk publicly floated hiring Huang.

**Why it matters:** SpaceX is now building compute at hyperscaler scale on its own balance sheet, with Arizona the new piece. County approval and power supply are the gates to watch.

Sources: [Reuters via Investing.com](https://www.investing.com/news/stock-market-news/spacex-to-invest-over-74-billion-in-supercompute-facilities-in-texas-arizona-5022690) · [The Hill](https://thehill.com/policy/technology/6217630-spacex-plans-data-centers-texas-arizona/) · [UPI](https://www.upi.com/Top_News/US/2026/11/12/spacex-musk-AI-supercomputing-texas-arizona/8361794521839/)

## GitHub repos

- **[GoogleChrome/modern-web-guidance-src](https://github.com/GoogleChrome/modern-web-guidance-src)** (★3.4k, +739 today, Apache-2.0) — skills and a CLI that put current web-platform practice and browser-compatibility data into your coding agent's context, so it stops writing bloated legacy JavaScript.
  - Stewarded by the Chrome, Edge, Firefox and MDN teams. Try `npx modern-web-guidance@latest legacy-scan .` on a repo.
- **[NVIDIA/SkillEvaluator](https://github.com/NVIDIA/SkillEvaluator)** (★0.9k, +108 today) — checks agent skills and measures their effect: quality checks first, then live runs with and without the skill, plus SkillSpector security scanning in GitHub/GitLab CI.
  - Install with `pip install skillevaluator && skillevaluator ci init`. Useful for anyone maintaining a shared skills library.
- **[ashhart/TensorFold](https://github.com/ashhart/TensorFold)** (★2.0k, Apache-2.0) — an LLM inference engine in a single Zig binary for Metal, CUDA and Vulkan: reads checkpoints, tokenises and runs kernels with no Python or MLX.
  - It also ships an "Agent Traffic Control" system so contributing agents claim issues and avoid duplicate work.

## Research papers

- **The Surprising Effectiveness of Deleting Weights in LLM Post-Training** (2611.10866) — zeroing under 0.05% of weights (a learned binary mask on gate projections) matches full fine-tuning across RLVR, SFT and on-policy distillation. [HF](https://huggingface.co/papers/2611.10866)
  - The masks are hundreds to thousands of times smaller, with much less forgetting.
- **SteerDistill** (2611.13410) — quantised agents fail by ignoring task requirements; a full-precision teacher guides the student to revise failed rollouts in a closed loop, recovering agent success at low bit-widths. [HF](https://huggingface.co/papers/2611.13410)
- **Embodied First Exam** (2611.12177) — 1,156 tasks in 35 suites on 47 robots, testing whether coding agents can control robots directly, through code, or by training a policy. [HF](https://huggingface.co/papers/2611.12177)
- **One Table, Many Queries** (2611.13001) — a benchmark of 56 datasets and 46 methods asking whether one tabular model can do generation, imputation and prediction well. [HF](https://huggingface.co/papers/2611.13001)
- **MACaQ** (2611.10649) — MXFP4 W4A4 post-training quantisation that treats each linear layer independently, so it can stream one layer at a time. [HF](https://huggingface.co/papers/2611.10649)

## AI workflows

**1. [A growth lead's "GTM brain" built with Claude Code that every sales rep now starts the day in](https://x.com/bolau_/status/2120917378540138707)**
- **Replaces:** about 15 minutes of switching between call notes, Slack, the CRM, product analytics and billing to answer one basic question about an account, plus guessing who to follow up with today.
- **What to connect:** Claude Code, used by a non-engineer growth lead.
  - Version one was a Slack bot running on his laptop; version two (Uatu) is a SQLite database.
  - It pulls in billing (Orb), product usage (PostHog), the CRM (HubSpot), call notes (Fireflies, Fathom), Slack and website visitors (RB2B).
  - The code is in GitHub with automatic deploys.
  - On top sit a dashboard with a signup inbox, a "who to contact today" list scored on custom signals (usage spikes or drops, accounts coming back, hiring for relevant roles) and a deal board that flags stalled deals.
  - It was built in phases over a week or two and has run for almost a year; other teams add their own tools by PR. No cost stated.
- **Who it's for:** small B2B sales and growth teams whose customer data is spread across many tools.

**2. [One weekly chore at a time: plan with Claude or Codex, build with one agent, review with another](https://x.com/THArrowOfApollo/status/2120932137897480644)**
- **Replaces:** a month of preparing lead lists before each launch (now a few days) and a lead-processing job that took 27 hours for 1M leads (now 20 minutes).
  - The agency does over $200k a month with a small team; onboarding, client issues and internal automations also moved to agents.
- **What to connect:** Claude Code or Codex in plan mode ("we're brainstorming, criticize this, what would you build step by step?").
  - Linear to keep the plan and decisions outside the chat so the next agent sees them.
  - One agent that builds and a second that reviews the pull request before merge.
  - The author can't write code; the rule is to understand the problem well enough to spot wrong output. No cost stated.
- **Who it's for:** agency owners and operations leads who don't code but have repetitive weekly processes nobody has fixed.

**3. [Small-business websites and Shopify fixes on a $100/month retainer, done in Claude Code](https://x.com/anthonysmendes/status/2120601368578904215)**
- **Replaces:** hand-editing a shop's product pages and missing content: SEO fixes on 85 Shopify products and new About, FAQ and shipping-policy pages for a toy shop whose owner had no time to do them.
- **What to connect:** Claude Code with Opus 5.5, the Framer MCP, skills and a design system; the same setup builds Next.js and Astro sites from a shared Framer project.
  - After 12 days: over 200 pages created, edited or fixed, 6 clients at $100 a month ($600 MRR) and 12 leads in a week.
  - Changes are tracked for a week after each fix.
- **Who it's for:** freelancers who look after small-business websites and online shops.

## Quick hits

- **Anthropic IPO dates locked** — Anthropic announced it will price on Wednesday 18 November and start trading as ANT on Nasdaq on Thursday 19 November; the $70–78 range and 825M shares are unchanged. [Yahoo Finance](https://finance.yahoo.com/technology/article/anthropic-sets-ipo-date-for-thursday-nov-19-173011206.html)
- **Grok Bot on Mac gets Apple apps** — It can now use Messages (read and send texts), Mail (search and draft), Notes, Reminders, Calendar, Contacts and Location; each is switched on separately with its own macOS permission. [9to5Mac](https://9to5mac.com/2026/11/12/grok-bot-for-mac-adds-integration-with-apple-mail-messages-notes-and-more/) · [@bot](https://x.com/bot/status/2120946968403095906)
- **OpenAI open-sources MCP Extensions** — Apache-2.0 TypeScript and Python packages add ChatGPT-native features to MCP servers: sidebar apps, conversation panels, file viewers, composer mentions, forms and deep links. [GitHub](https://github.com/openai/mcp-extensions) · [Docs](https://developers.openai.com/plugins/build/extensions)
- **Wharton's Primer turns lessons into simulations** — Ethan and Lilach Mollick's Generative AI Lab released Primer, an open-source multi-agent system: an intake agent interviews the teacher about the lesson and student misconceptions, then generates a playable practice simulation. [Primer](https://gail.wharton.upenn.edu/primer/) · [SSRN paper](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7696720)
- **Codex banked reset is landing** — OpenAI's @thsottiaux said the banked usage reset has been activated and will reach accounts soon, alongside updates to dots. [@thsottiaux](https://x.com/thsottiaux/status/2120999993079451919)
- **The "Home Repo" pattern** — @mattpocockuk wrote up a simple way to run agents across many repos: one home repo that holds the skills, steering files and tooling, with a prompt to set it up across an org. [@mattpocockuk](https://x.com/mattpocockuk/status/2120838662254420341)

## Go deeper

- [Will DePue: Chinese frontier models carry the fingerprint of Blackwell GPUs](https://unemploymentlabs.com/checkpoint-forensics): the full method, appendix of models and stated limits.
- [Self-hosted Claude Managed Agents: data and access](https://platform.claude.com/docs/en/managed-agents/self-hosted-data-and-access): written for a security review; worth reading before you pitch it internally.
- [UI SDK: Built for Agents](https://ui-sdk.dev/docs/agents): the agent-bridge API and benchmark in one page.
