AI Today: NVIDIA Cosmos and OpenAI Decisions API

State of knowledge as of 2026-10-23. Source material gathered on 2026-10-23.

Today's technology review covers a suite of AI tools, ranging from environmental monitoring systems to computing infrastructure. OpenAI introduces the Decisions API, Google DeepMind releases the Forest Typology 2020 map, and NVIDIA expands the Cosmos platform while introducing the Minerva system. The roundup concludes with the IBM Granite Timeseries Ensemble solution, which integrates four time-series model architectures to streamline data forecasting in analytical applications.

Key Takeaways (TL;DR)

• Google DeepMind Forest Typology: The ForTy 2020 map has been released at a 10-meter resolution, tracking changes in forest structure from 2017 to 2025.
• OpenAI Decisions API: A new interface powered by the GPT-6 Luna model enables rapid decision-making for AI agents with prompt caching support.
• NVIDIA Cosmos: The platform has been expanded with Data Factory tools and architecture blueprints supporting the creation of world models for physical AI systems.
• NVIDIA Minerva: This AIOps-class platform integrates infrastructure telemetry data to detect anomalies and predict failures in compute clusters.
• IBM Granite Timeseries Ensemble: A new mechanism allows combining four open time-series model architectures into a single unified forecasting interface.

Google DeepMind Forest Typology

Google DeepMind, alongside the World Resources Institute and the International Institute for Applied Systems Analysis, released the Forest Typology (ForTy) 2020 map on October 22, 2026. This tool presents the structure of forests and agricultural land worldwide at a 10-meter resolution. In addition to static data, the project includes a time series tracking changes between 2017 and 2025, enabling the identification of natural forest conversions into plantations or farms.

The map categorizes each 10-by-10-meter area into one of 6 forest types and land-use categories, in accordance with the definitions of the Food and Agriculture Organization (FAO) of the United Nations. This data is intended to support companies and regulatory bodies in implementing the European Union Deforestation Regulation (EUDR), which requires verifying the origin of commodities such as coffee, cocoa, and palm oil.

The model creation process was based on a dataset of 1.7 million labeled locations obtained from expert surveys, national and regional datasets, and analyses of terrain where the model encountered difficulties. The system architecture utilizes AlphaEarth foundations. Unlike previous maps that focused solely on tree presence, ForTy makes it possible to distinguish primary forests from secondary forests, tree plantations, or agricultural crops, which is crucial for assessing carbon stocks and biodiversity. Users can access the map via an application in Earth Engine.

For AI builders: individuals designing tools for monitoring environmental changes can test the API and model architecture based on the Earth Engine documentation. The press release does not specify the costs of commercial data usage, bandwidth requirements for large-scale analysis, or the exact license for utilizing model results in external reporting systems.

OpenAI Releases Decisions API

OpenAI announced the general availability of the Decisions API on October 22, an interface designed for rapid decision-making by applications and AI agents. The tool, powered by the GPT-6 Luna model, enables the processing of inputs in text, image, and JSON formats, offering latency improvements and advanced caching features.

The Decisions API focuses on questions with a limited scope of answers, operating up to 10 times faster than the Responses API. Developers can submit up to 400 questions within a single request, which is double the previous limit. Among the new features are integration with GPT-Live, the ability to apply prompt caching for repeating contexts, and ensured inference residency in the US and EU. Additionally, throughput in Asian regions has been increased.

The system supports three main types of responses:

• probability: determination of the probability (from 0 to 1) of whether a given state is true;
• choice: selection of one of predefined options along with a confidence score;
• score: an evaluation of contribution based on ordered scale levels.

Integration with GPT-Live allows voice applications to make real-time decisions, for example, to manage avatar reactions or steer the flow of a customer conversation. Use cases include categorizing transaction descriptions, assessing communication sentiment, or verifying product damage.

For AI builders: Teams deploying automated customer support can test the API in conversation path selection or input classification scenarios. The announcement does not provide exact pricing per thousand requests or specify whether the system supports languages other than English — this should be verified in the technical documentation before production rollout.

NVIDIA Cosmos Expands Tools for World Models

NVIDIA announced on October 22 the expansion of the Cosmos platform with new training tools, simulation checkpoints, and data architecture blueprints. These solutions are intended to support creators of physical artificial intelligence systems in building world models that simulate reality.

The Cosmos platform now includes Cosmos Data Factory, a set of tools designed for labeling, curation, and augmentation of multimodal data. Developers have access to five cloud-hosted architecture blueprints available via GitHub, Microsoft Azure, Nebius, and CoreWeave. These functionalities include generating data for autonomous vehicles, image attribute augmentation, creating synthetic videos of rare events, generating defect images, and video data augmentation considering changing weather conditions.

Commercial entities are already utilizing these tools. Porsche uses Cosmos Data Factory to build a shared dataset, which is expected to shorten development cycles by 85%. ANYbotics generated over 1,500 synthetic video clips to train inspection robots, and Ultralytics used synthetic defects to fine-tune the YOLO model.

Within the Cosmos Framework, creators can train their own architectures or use pre-trained models such as GR00T, Alpamayo, QWEN, or Pi0.5. Updated NVIDIA Cosmos 3 NIM microservices offer up to 2x faster inference while maintaining quality. New variants of the Cosmos 3 Nano model enable testing autonomous policies in multi-polygon simulations. Additionally, the NVIDIA Metropolis VSS 3.3 blueprint allows monitoring 40% more video streams using the same GPU.

For AI builders: Teams working on robotics or computer vision systems can test Cosmos Data Factory blueprints in cloud environments. The announcement does not specify the costs of using NIM microservices or licensing details regarding the commercial use of individual base models, which must be verified in the technical documentation.

NVIDIA Introduces Minerva Platform for Managing AI Factories

According to an announcement on October 22, NVIDIA released the Minerva platform, designed for end-to-end observability of infrastructure supporting artificial intelligence computing. This AIOps-class tool integrates telemetry data coming from GPUs, networks, storage systems, job schedulers, and the workloads themselves.

Minerva builds a dynamic knowledge graph that maps physical infrastructure, including SuperNICs, switches, and cabling, connecting them with currently running processes. The system analyzes billions of data points per hour using supervised and unsupervised machine learning methods. This allows for anomaly detection, component failure prediction, and grouping related events into incidents with an indicated root cause. Operators can use a built-in AI assistant and connect the tool with external agents via a natural language interface, aiming to extend uptime and shorten mean time to repair.

The platform focuses on the AI factory performance index, encompassing token production, cluster availability, and GPU utilization. Entities planning to use Minerva solutions include IREN, Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI, and Tesla. Furthermore, Dell Technologies, Lenovo, and Supermicro will integrate the platform into their AI infrastructure operations offerings.

For AI builders: Teams responsible for maintaining large computing clusters can verify the compatibility of their existing network and server infrastructure with the requirements of the Minerva platform. The source does not specify deployment costs, licensing requirements, or the full list of supported operating systems, which requires clarification in direct contact with the vendor.

IBM Granite Timeseries Ensemble

According to an announcement on October 22, 2026, IBM released the Granite Timeseries Ensemble mechanism, which allows combining four different time-series model architectures into a single forecasting interface. This solution aims to reduce the difficulty of choosing the right model for a specific task, offering a default method of combining pre-trained models without the need to retrain them.

The system utilizes four open architectures: PatchTST-FM-r1, PatchTST-FM-r2, FlowState, and Tiny Time Mixers. Each of them analyzes trends, seasonality, and local changes differently, allowing for the generation of point and quantile forecasts (defaulting to a range from 0.1 to 0.9).

The package offers two main ways of combining results:

• Probabilistic Ensemble: applies uniform probability weighting, giving each model equal influence on the final result.
• IQR Ensemble: adjusts weights based on the model's interquartile range (IQR), preferring narrower forecast intervals.

The configuration does not combine model weights, but instead loads four separate checkpoints at runtime, which requires accounting for additional time on the first run. The interface and included models are released under licenses allowing for free use.

For AI builders: a data analyst or forecasting systems engineer can install the Granite TSFM library and test the ready-made recipe on their own time-series datasets. The announcement does not specify the hardware specifications required to simultaneously load four models, nor does it indicate potential performance limitations when working with very long time horizons — these issues require independent load testing.

Frequently Asked Questions (FAQ)

What distinguishes the Forest Typology (ForTy) 2020 map from previous solutions?

Unlike previous maps that focused solely on tree presence, ForTy makes it possible to distinguish primary forests from secondary ones, tree plantations, and agricultural crops. The tool utilizes a 10-meter resolution and includes data from 2017–2025, supporting the verification of commodity origins under the EUDR regulation.

How does the IBM Granite Timeseries Ensemble forecasting system work?

The system combines four time-series model architectures — PatchTST-FM-r1, PatchTST-FM-r2, FlowState, and Tiny Time Mixers — into a single interface. Users can utilize Probabilistic Ensemble or IQR Ensemble result-combining methods by loading four separate checkpoints at runtime without needing to retrain the models.

Sources

• Beyond tree cover: mapping the world’s forest types — 2026-10-22
• Build more responsive apps with the Decisions API — 2026-10-22
• NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models — 2026-10-22
• NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories — 2026-10-22
• Ensemble forecasting with IBM Granite time-series models — 2026-10-22

--- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.