Your Week in AI

Your Week in AI: agent rollouts, stronger checks and an NVIDIA transition

Hydrogen and ChatGPT Wallet expand AI-agent workflows, llama.cpp reaches 1.0, and new research highlights the difference between model behavior and enforced guarantees. NVIDIA also begins a leadership transition. This edition covers announcements before 17:00 UTC on 13 November.

Published · Updated

Coverage: to (cutoff; Europe/Paris).

Hydrogen brings Gemini 4 to everyday work—through different access paths

Google introduced Gemini 4 Hydrogen on 12 November for everyday coding and multi-step AI-agent work. Four thinking settings and listed API rates of $1.50 per million input tokens and $7.50 per million output tokens position it as an execution model as well as a partner for larger models. Its rollout spans developer tools, the Gemini app and Search, but access is not uniform: the individual-developer Antigravity route is offered to paid Google AI Pro or Ultra subscribers; free app users receive Hydrogen through automatic selection when appropriate. [1][2][3][4]

Google reports 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash and 2.5-times-faster task completion at Hydrogen's high setting. Those are publisher comparisons, not TideReach reproductions. Antigravity's Orchestrator switches models within a session, so a result for that combined workflow is not a result for Hydrogen operating alone. [3][4]

Wallet adds a purchase step, not unrestricted delegation

OpenAI's 13 November ChatGPT Wallet announcement brings AI agents closer to completing purchases. Initial availability is in dots for ChatGPT Pro subscribers aged 18 or older in the United States with eligible US-issued cards—not a general rollout across all ChatGPT products. For each purchase, the user reviews the items, total and selected card and completes a security check; OpenAI says an authorization record is retained. [5][6]

The merchant remains the seller responsible for fulfillment, returns and customer service. Payment-network tokens substitute for actual card details, but tokenization is not an independent security certification. Pre-approved spending limits, categories and time windows for fully delegated purchases are future work, not part of this initial launch. [5][6]

Local inference gets a clearer upgrade contract

The ggml-org project announced version 1.0 for both llama.cpp's libllama and the ggml tensor library on 13 November. These are libraries for running AI models, not new model weights. Breaking public C API or ABI changes require a major-version increment; backward-compatible additions use minor versions and bug fixes without API changes use patch versions. [7][8]

TideReach assessment

That is a concrete improvement for teams embedding local inference: the version number now communicates the intended compatibility boundary. It does not establish that every listed packaging channel has identical binaries or that every hardware-and-model combination works. Test the actual integration before treating an upgrade as interchangeable. [7][8]

Learned context structure and formal proof answer different questions

IBM's Granite LLMON previews add reserved structural tokens, addressable instruction and data spans, and execution blocks that bind them together. The team reports improved resistance on static indirect-prompt-injection benchmarks, while adaptive multi-turn attacks remain active work. These are experimental previews for exploration and evaluation, explicitly not production releases. Learned instruction separation is not an enforced privilege boundary. [9][10]

Earlier in the week, DeepMind described AlphaProof Code: humans reviewed mathematical specifications for selected properties of Signal protocol code, AI agents constructed proofs and Lean checked them. The researchers report completing that project in four weeks and detecting two reintroduced, previously fixed bugs in isolated code copies. This does not prove that the whole messaging application, its devices or every real-world threat is covered. [11][12]

TideReach assessment

The distinction matters when evaluating an AI system. Resistance on a benchmark concerns observed model behavior; a machine-checked proof concerns a specified formal statement. Neither can be transferred without qualification to a different application or used to dispense with its permissions and operational checks. [11][9][12][10]

Portable skills still need a functioning host and enforced checks

GitHub's practical skills guide separates shared SKILL.md instructions from project manifests, client setup and access. Its Data Forge example uses a resolved manifest for project details and human checkpoints before schema changes. Packaging shared instructions for Claude Code, Cursor and GitHub Copilot does not install the tools or grant the database permissions those workflows require. [13][14]

The guide's Proof Agent example is a different kind of verification from AlphaProof Code. Its GitHub Action statically reviews the original request and diff without the coding agent's own assessment; it does not run code or tests. The demonstration makes findings advisory, and a failed check prevents a merge only when repository rules require it. [11][13][14][12]

NVIDIA's leadership changes; continuity remains a company commitment

NVIDIA announced on 12 November that Jensen Huang had resigned immediately as president, CEO and a board member. Colette Kress becomes interim CEO while retaining her CFO role, and the board is starting a permanent-successor search. This is a leadership change at a supplier of AI and accelerated-computing infrastructure, not a new model or product release. [15][16]

TideReach assessment

The board says operations, customer commitments and the product roadmap should continue without interruption. That is an assurance from the company, not an outcome already verified by this edition. For customers, delivery and roadmap execution during the search are the concrete things to watch; the announcement does not establish a changed product plan. [15][16]

What changes now—and what to check next

TideReach assessment

The week expands the ways AI can act: broader model access, a bounded purchase flow, local inference libraries and reusable agent instructions. The useful next check depends on the stage. Compare a complete repository task under the chosen Hydrogen setting; confirm each Wallet transaction before payment; test a local-library upgrade in its real environment. For research previews, check whether the result concerns learned behavior or a formally specified property. Keep those observations separate from enforced access, required tests and promises about future delivery. [11][3][9][13][7][5][12][15][4][10][14][8][6]

Sources and further reading

Numbered references connect each passage to its sources and related TideReach coverage.

  1. Google AI Blog Google Search gets a major upgrade with Gemini 4 Hydrogen Back to text ↑
  2. Google Gemini Models Gemini 4 Hydrogen: everyday coding and agentic workflows for developers Back to text ↑
  3. Google Gemini Models Introducing Gemini 4 Hydrogen Back to text ↑
  4. TideReach coverage Google launches Gemini 4 Hydrogen for coding and Search Back to text ↑
  5. OpenAI News Introducing ChatGPT Wallet Back to text ↑
  6. TideReach coverage OpenAI starts ChatGPT Wallet rollout in dots for eligible US Pro adults Back to text ↑
  7. Hugging Face Blog llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem Back to text ↑
  8. TideReach coverage llama.cpp and ggml reach 1.0 with explicit compatibility versioning Back to text ↑
  9. Hugging Face Blog LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B Back to text ↑
  10. TideReach coverage IBM releases experimental Granite LLMON previews for structured AI context Back to text ↑
  11. Google DeepMind Blog Introducing AlphaProof Code: a breakthrough in verifying complex software at depth Back to text ↑
  12. TideReach coverage DeepMind reports AI-assisted formal verification of Signal protocol code Back to text ↑
  13. GitHub Blog Build once, run on any agent: A practical guide to agent skills Back to text ↑
  14. TideReach coverage GitHub guide separates reusable agent skills from host permissions Back to text ↑
  15. TideReach coverage NVIDIA names Colette Kress interim CEO after Jensen Huang resigns Back to text ↑
  16. NVIDIA Newsroom NVIDIA Announces CEO Transition Back to text ↑
  17. Axios National Jensen Huang, an architect of the AI boom, resigns as Nvidia CEO

Explore more articles