Google DeepMind says AlphaProof Code used Gemini-powered agents and Lean to check properties of Signal's deployed cryptographic implementation in four weeks.
Gemini Robotics 2.5 rolls out to robotics partners; Google reports a 96% success rate in an Apollo 2 automotive-kitting proof of concept.
Businesses can claim profiles and manage product information; eligible US ecommerce and online-service firms can offer separate Business Agent conversations.
Meituan releases a 1.6-trillion-parameter mixture-of-experts model, with around 48 billion active parameters per token and an MIT license.
A peer-reviewed SimPoly study reports polymer-density and glass-transition predictions trained on quantum calculations, and releases a 130-polymer benchmark.
AI NEWS ANALYSIS
The most consequential AI advances now connect models with checkable outcomes. DeepMind's AlphaProof Code tests whether production cryptographic code satisfies formally reviewed security properties; Gemini Robotics 2.5 takes physical-action models into an industrial pilot. OpenAI's Businesses in ChatGPT adds a governed channel for commercial discovery, while Meituan's LongCat-2.5 exposes unusually large-scale agent-model weights under an open license. The peer-reviewed SimPoly work applies quantum-trained learning to measured polymer properties. These are different claims with different standards of evidence: machine-checked proofs, company-run robot tests, product availability, open-weight artifacts and scientific measurements cannot be graded on the same marketing scoreboard.
Read previous source-backed editorials and dated briefings.
Send an AI model release, research finding or industry announcement for independent editorial review.
Pitch a Story →IN-DEPTH AI NEWS / NOVEMBER 11, 2026
Source-backed original reporting on AlphaProof Code, Gemini Robotics 2.5, Businesses in ChatGPT, Meituan LongCat-2.5 and SimPoly scientific research.
Google DeepMind says AlphaProof Code used Gemini-powered agents and Lean to check properties of Signal's deployed cryptographic implementation in four weeks.
Google DeepMind reported on November 10 that AlphaProof Code, its new multi-agent formal-verification framework, was used to prove key security properties of the implementation of Signal's cryptographic protocol. This differs from asking a language model whether code looks safe: specialists review a mathematical statement of what must hold, while an independent proof checker validates each formal proof. The work targets an especially difficult boundary between a protocol's design and the compiled software users actually run. Because billions of conversations depend on end-to-end encryption, a check that reaches implementation behavior is a material research milestone, but not a certificate that every part of Signal is invulnerable.
Formal verification starts with a precise threat model and specification. The verifier checks that a formal representation of the compiled implementation satisfies those rules, rather than sampling example inputs. That can reveal classes of bugs conventional testing misses, but the guarantee is conditional: a correct proof of an incomplete requirement is still incomplete security. Teams adopting this approach need trusted compiler and machine semantics, explicit environmental assumptions, reviewed specifications and an audit trail linking formal artifacts to shipped builds. Keep fuzzing, penetration tests, dependency checks and operational monitoring because cryptography can fail outside the proved boundary.
Previous AI code stories often measure how quickly agents write tests or patches. AlphaProof Code tackles a more consequential question: can agents accelerate rigorous verification of widely deployed code while a deterministic tool—not the model—decides whether a proof is valid? Its structured human-review stage is as important as the agent generation step. This workflow could become especially useful in payment systems, safety-critical firmware and communications, where failures are extremely costly. Still, the research does not make AI-generated proofs equivalent to independent security reviews of every surrounding component.
Gemini Robotics 2.5 rolls out to robotics partners; Google reports a 96% success rate in an Apollo 2 automotive-kitting proof of concept.
Google DeepMind introduced Gemini Robotics 2.5 on November 10, updating its vision-language-action system for robot manipulation. The important change is not simply a higher simulated benchmark score: Google describes a limited real-world automotive pilot with Apptronik's Apollo 2 humanoid robots at Mercedes-Benz facilities. Humanoid systems can demonstrate dexterity in controlled videos yet struggle with unfamiliar parts, obstacles and recovery after a failed grasp. This release places more emphasis on repeated physical tasks, safety reasoning and variation across hardware platforms, while still restricting the model's rollout to robotics partners.
A vision-language-action policy maps sensed scenes and instructions to physical movement, while an embodied-reasoning layer tracks task goals and state. For production evaluation, count successful complete work cycles over many shifts, not just grasp accuracy or one polished trial. Measure mean interventions per hour, cycle time at the 95th percentile, unexpected-contact events, recovery after a dropped part and total operating cost. Validate that the same task runs safely across different lighting, tooling and inventory variation. Use independent emergency-stop and robot safety systems even when the AI model claims strong hazard avoidance.
The factory pilot improves the relevance of the evidence relative to tabletop demonstrations. But one successful proof-of-concept does not settle the costs of maintenance, certification, supervision or long-duration reliability. The link with Gemini 4 Argon as a simulation and failure-analysis tool also illustrates how foundation models may help train physical models indirectly. This is a distinct November 10 release from NVIDIA GR00T 2 and Isaac Lab 3.0, already covered in Issue 010: both address robot learning, but today's story is Google's new model and named industrial field test.
Businesses can claim profiles and manage product information; eligible US ecommerce and online-service firms can offer separate Business Agent conversations.
OpenAI announced Businesses in ChatGPT on November 10, adding a beta Business Hub for claiming and managing business information and a Business Agent experience for eligible merchants. This is a significant change to AI-assisted discovery because companies will have a structured path to update their own business details and product feeds instead of depending solely on third-party crawls or conversational mentions. The launch also adds aggregated discovery analytics, giving businesses some visibility into how users encounter their information. Crucially, a Business Agent is not the same as ChatGPT's independent answer: it is a clearly separate conversation that users may choose to open.
A business should check the ownership-verification flow, confirm its public URLs and keep product identifiers, stock status, service regions and prices synchronized with its canonical commerce source. Measure clicks and qualified leads against existing analytics rather than treating aggregate mentions as sales. Maintain ordinary technical SEO: crawlable product pages, clear business identity, structured product information where appropriate and accurate canonical tags. Product-feed updates should be monitored for stale prices and discontinued items. Any agent handling customer questions needs transparent escalation and should not be allowed to promise unavailable inventory or unsupported guarantees.
The launch could change how organizations allocate effort between conventional search, answer engines and conversational commerce. It creates a place for verified business input, but that is not equivalent to paid control of ChatGPT's independent recommendations. The separation of original answers, optional business-led conversations and labeled sponsored destinations is editorially important. Small retailers may gain access to discovery tools previously associated with large advertising budgets, although eligibility and the quality of analytics will determine how useful the product becomes in practice.
Meituan releases a 1.6-trillion-parameter mixture-of-experts model, with around 48 billion active parameters per token and an MIT license.
Meituan open-sourced LongCat-2.5 on November 10, positioning a large native multimodal mixture-of-experts model for long-running software, visual coding and professional office workflows. The most concrete news is not a proprietary performance chart: Meituan publishes a model-weight destination and says the model is licensed under MIT for commercial inference and additional training. Its very large total parameter count is offset by sparse routing, activating only a small portion per token. Even so, sparse activation should not be mistaken for a small download, a single-GPU deployment or inexpensive operation at million-token context lengths.
Start with the exact model repository, model-card requirements and license file. A trillion-scale mixture-of-experts network typically needs distributed storage and expert parallelism despite its modest active-compute fraction. Benchmark actual model load time, GPU and host memory, communication overhead, prefill latency and tool-call reliability; a million-token advertised window can impose substantial key-value cache and ingestion costs. For enterprise use, confirm that deployment code and dependencies are under compatible licenses, use trusted serving images and run constrained agent tests against a repeatable environment. Compare generated artifacts rather than accepting curated screenshots.
A permissively licensed model of this scale broadens research and commercial experimentation without mandatory per-token vendor access. Its significance rests in the engineering trade-off between sparse compute, enormous model footprint and long tasks. Meituan reports training and serving improvements on domestic accelerators, which may matter to teams seeking heterogeneous hardware supply. But practical portability is a systems question: memory distribution, quantized kernels, communication networks and complete reproducibility determine whether open weights translate into usable capacity.
A peer-reviewed SimPoly study reports polymer-density and glass-transition predictions trained on quantum calculations, and releases a 130-polymer benchmark.
A peer-reviewed Nature Communications paper published November 9 introduces SimPoly, a machine-learning force-field approach to polymer simulation funded by Microsoft Research. The goal is to make quantum-informed material prediction practical without running prohibitively expensive quantum-chemical calculations at every step of a large polymer simulation. In materials science, small-scale electronic accuracy and macroscopic industrial behavior are separated by many orders of magnitude in time and length scale. SimPoly attempts to bridge part of that gap by learning interatomic forces from first-principles calculations, then simulating larger systems to estimate experimentally observed properties.
Classical force fields provide affordable approximations of atom interactions but can struggle to transfer between different polymer chemistries. First-principles quantum calculations are much more expensive for long polymer chains and long times. A machine-learned potential offers a possible middle layer: train on accurate atomic configurations, perform molecular dynamics, and derive density or temperature-dependent properties for comparison with measurements. Reproduction requires matched polymer composition, chain-length distribution, equilibration procedure, simulation time, thermostat and uncertainty estimates. Researchers should benchmark out-of-distribution chemistry and temperature ranges instead of reporting only the most representative materials.
The public benchmark is as important as the model because it can support independent comparison with classical molecular dynamics and alternative machine-learning potentials. Faster candidate screening could help teams prioritize laboratory tests for coatings, packaging and structural materials. Yet simulation remains one layer of discovery: polymer processing, additives, aging, impurities and manufacturing conditions can shift observed performance. No result in this paper proves a commercial material's regulatory suitability or lifetime reliability.