[ The Morning Build ]
4 min read5 storiesAIBig TechDev Tools

The Morning Build for November 11, 2026: Formal Verification, Robot Vision-to-Act, and Business Profiles in ChatGPT

Today’s stories focus on how AI is moving from prototypes into engineering workflows: DeepMind used an agentic system to produce machine-checked proofs of Signal’s production cryptography; Google DeepMind and Nvidia both reported robotics advances that translate vision and language into real-world control; OpenAI launched a beta Business Hub to let companies publish profiles and agents inside ChatGPT; and browser vendors published the Modern Web Guidance repo and CLI for safer coding-agent outputs.

DeepMind introduces AlphaProof Code and publishes machine-checked proofs of Signal’s cryptographic implementation

  • What happened: DeepMind published AlphaProof Code, an agentic framework that, with human review, generated over six million lines of Lean to formally verify key security properties of Signal’s production cryptographic code and rechecked two previously fixed real bugs. The team says the end-to-end verification took four weeks and humans only needed to re-inspect about one percent of the codebase.
  • Why it matters: Engineers get a concrete example where multi-LLM agents plus human experts produced machine-checked proofs that connect cryptographic design to compiled implementation, reducing manual review scope and producing artifacts intended for community use. The output is Lean proofs, not heuristic test suites, so correctness is checked by the theorem prover rather than by examples.
  • Outlook: Open-source release of the technical paper and the formal verification artifacts (the Lean proofs and repository) that DeepMind says it is publishing.

Sources: deepmind.google

Nvidia publishes GR00T 2 research that converts human video demonstrations into robot control via world-action video prediction

  • What happened: Nvidia released GR00T 2, a world-action model trained on roughly 400,000 hours of head-mounted human video using a cluster of more than a thousand GB300 chips; the model predicts both next robot actions and short-horizon future video frames to map human motions to different robot morphologies. In lab tests, single-video transfers produced task success rates from 10 to 80 percent and, for a set of 50 tasks, under three minutes of conventional training raised median success from 40 to 95 percent.
  • Why it matters: The model’s video-prediction approach yields a transferable representation across form factors, allowing researchers to derive robot commands without teleoperating each robot for every task; success-rate ranges and remaining gaps, missing tactile inputs, false negatives like half-turned screws, and multi-second planning latency when paired with a higher-level language model, define current operational limits for production deployment.
  • Outlook: Published GR00T 2 research and lab metrics, including the Turbo variant running on Jetson Thor, serve as the immediate artifact to inspect for replication and latency/robustness numbers in follow-on papers or code releases.

Sources: the-decoder.com · techcrunch.com

OpenAI launches Business Hub beta to let verified companies create profiles and deploy specialized business agents inside ChatGPT

  • What happened: OpenAI rolled out a Business Hub beta where eligible companies can verify and update structured profiles, view limited interaction metrics (profile views, mentions, website clicks), and create business agents that answer customer queries within ChatGPT. OpenAI said verification uses multiple methods across markets but did not publish the specific verification criteria.
  • Why it matters: For engineers supporting customer-facing systems, ChatGPT now presents a platform where third-party data and agent endpoints can be surfaced directly in user conversations, changing where discovery and conversational integrations may need to expose structured APIs and verification hooks when companies opt into the Hub.
  • Outlook: Ongoing Business Hub beta availability and the unspecified company verification process during the beta period are the next public milestones to track for changes in profile verification signals and agent deployment capabilities.

Sources: techcrunch.com · the-decoder.com

DeepMind’s Gemini Robotics 2.5 scores 96 percent in a Mercedes-Benz parts-sorting pilot and ships the ASIMOV-Jailbreak safety benchmark

  • What happened: Google DeepMind introduced Gemini Robotics 2.5, reported 30–100 percent task improvement over its previous baselines, and says Apptronik’s Apollo 2 achieved a 96 percent success rate sorting metal parts on Mercedes-Benz production lines in Hungary and Germany. DeepMind also released the ASIMOV-Jailbreak benchmark on Hugging Face and used Gemini 4 Argon to generate simulation training data.
  • Why it matters: Engineers working on industrial automation get a concrete production-line pilot with measured success rates and an open benchmark for prompt-based safety testing; Argon-generated simulated runs and the ASIMOV-Jailbreak dataset provide artifacts for reproducing training data generation and stress-testing model jailbreak resistance.
  • Outlook: Rollout of Gemini Robotics 2.5 to declared partners Agile, Apptronik, and Mercedes-Benz as stated by DeepMind is the next deployment milestone for observing real-world transfer and performance retention.

Sources: the-decoder.com · arstechnica.com

Browser vendors publish Modern Web Guidance repo and CLI to steer coding agents toward modern web APIs and patterns

  • What happened: The Modern Web Guidance project, stewarded by Google Chrome, Microsoft Edge, Mozilla Firefox, and MDN teams, published a GitHub repo and a CLI (npx modern-web-guidance@latest) that injects curated, token-efficient web platform guides into coding-agent contexts and provides a legacy-scan and retrieval APIs covering 250 web features and 203 real-world use cases.
  • Why it matters: Teams embedding coding agents can use the published guides and CLI to reduce legacy-pattern outputs from LLMs and to supply up-to-date browser API usage, accessibility, and performance advice directly into an agent’s context window, making generated code more aligned with current platform primitives rather than older polyfilled patterns.
  • Outlook: Updates to the Modern Web Guidance GitHub repository and the NPM package (the ‘latest’ npx install) are the immediate artifacts to watch for new guides, eval dashboard results, and changes to the CLI interface.

Sources: github.com · arstechnica.com