Agents Cross Apps. Models Shrink. Robots Go Industrial.

What happened
OpenAI and Oracle expanded their partnership to connect Oracle's business applications and enterprise data with ChatGPT. Oracle is developing AI Workspace to let employees access information and complete approved workflows across business systems, while new offerings include AI Security Mesh for vulnerability detection, App Composer for building applications with Codex, and an AI bundle for small businesses.

Why it matters
AI assistants are moving from standalone productivity tools into the systems where companies manage transactions, employees, and operations. Integrating ChatGPT with Oracle's applications could allow agents to coordinate work across enterprise functions while respecting existing permissions and approval processes.

What's next
Oracle is developing AI Workspace and its new agent-powered products, with rollout details still to come. U.S. customers can already purchase OpenAI API access, Codex, and ChatGPT Work through Oracle Marketplace using eligible Oracle Universal Credits.

What happened
ServiceNow and Google Cloud expanded their partnership to connect ServiceNow's Action Fabric with Gemini Enterprise, allowing AI agents to execute multi-step business workflows from Google's assistant. Integrations for retail operations, customer relationship management, and IT services are generally available to mutual customers, with telecommunications workflows planned for early 2027.

Why it matters
Enterprise AI is shifting from answering questions to completing work across existing software systems. By combining Gemini's reasoning with ServiceNow's operational context and governance, employees can investigate incidents, update customer records, or initiate service requests without manually navigating multiple applications.

What's next
The companies plan to expand industry-specific integrations through Google Cloud Marketplace. Telecommunications automation is expected in early 2027, while customers can begin testing whether the available integrations reduce operational handoffs and improve resolution times.

What happened
Liquid AI introduced LFM3, a family of six open-weight models covering text, vision, audio, multimodal interaction, and specialized decision making. Built around a shared architecture trained on 22 trillion tokens, the models contain approximately 2 billion parameters each, with Liquid reporting that its 1.7-billion-parameter text model outperforms its larger predecessor while fitting into roughly 1GB of memory at 4-bit precision.

Why it matters
AI doesn't always need a massive model to deliver useful intelligence. Liquid is targeting specialized applications that can run on phones, laptops, and industrial devices, potentially reducing cloud dependence, inference costs, and latency while keeping sensitive information closer to where it is generated.

What's next
LFM3 models are available through Hugging Face, Liquid's API, and its online playground, with support for local inference and fine-tuning on a single GPU. Developers can now evaluate whether smaller, specialized models outperform larger general-purpose alternatives on specific business tasks.

What happened
Google DeepMind unveiled Genie 4, a new world model capable of generating realistic, interactive environments with native audio. Users can steer simulations in real time using natural-language instructions, changing weather, camera movement, objects, and characters while the model attempts to preserve environmental consistency.

Why it matters
Generative AI is moving beyond creating static images and videos toward producing environments that respond to actions. Interactive world models could become valuable for robotics training, autonomous-agent testing, filmmaking, and game development by allowing developers to explore scenarios before deploying systems in the physical world.

What's next
Genie 4 is available to a limited group of trusted research testers, with broader access planned for Google AI Ultra subscribers through Project Genie and selected developers through Gemini API and AI Studio. Google acknowledges that object permanence, physical interactions, and long-term consistency remain significant research challenges.

What happened
Oracle and Google announced plans to bring Google's Tensor Processing Units to Oracle Cloud Infrastructure, making OCI the first hyperscale cloud platform outside Google to offer TPUs. Beginning in 2027, Oracle customers will be able to select Google's purpose built AI accelerators alongside existing infrastructure options for model training, inference, and agentic workloads.

Why it matters
AI infrastructure competition is expanding beyond individual cloud providers and GPU suppliers. Allowing Google's custom chips to run inside Oracle's cloud gives customers more flexibility to optimize performance, energy consumption, and cost without moving their entire AI operation to a different platform.

What's next
TPU capacity on OCI is scheduled to become available in 2027 and expand over multiple years. The deployment will include Google's eighth-generation TPU architectures, with TPU 8t targeting large scale training and TPU 8i designed for inference intensive workloads.

What happened
ANYbotics began customer deliveries of ANYmal X, which the company describes as the world's first legged inspection robot built for potentially explosive industrial environments. The autonomous quadruped is shipping to customers including Equinor and DSM Firmenich, using industrial AI, autonomous navigation, and the company's Shift platform to perform inspections and connect findings with existing maintenance systems.

Why it matters
Physical AI is expanding into environments where human inspections are dangerous, expensive, or operationally disruptive. ANYmal X demonstrates how autonomous robots can deliver value by collecting operational data in hazardous facilities while keeping employees out of high-risk areas.

What's next
ANYbotics plans to demonstrate the complete system publicly at ADIPEC 2026 with industrial partners. The company is expanding commercial deployment after years of early-access testing, with customers evaluating how automated inspections improve safety, maintenance efficiency, and equipment reliability.

What happened
BMW announced that its Automated Driving In-Plant technology has moved one million vehicles across its manufacturing network. Developed with Swiss technology company Embotech, the system autonomously moves newly manufactured vehicles through defined factory and logistics areas, completing approximately 3,500 movements daily across four plants in Germany and the United Kingdom.

Why it matters
Autonomous driving is finding commercial value inside factories before becoming universal on public roads. BMW's milestone shows how controlled environments, external sensors, and intelligent motion planning can eliminate repetitive vehicle-handling tasks without requiring fully autonomous consumer vehicles.

What's next
BMW plans to expand the technology to its Spartanburg, South Carolina, facility in 2027, making it the fifth plant to adopt the system. Further deployment will test how consistently autonomous vehicle logistics can improve efficiency across different factories, vehicle models, and production processes.

AI is moving from isolated intelligence to integrated execution. OpenAI and ServiceNow are connecting agents to enterprise workflows, smaller models and new infrastructure options are expanding where intelligence can run, and industrial robots are demonstrating measurable operational scale. The next competitive advantage will come from making AI accessible, governed, and dependable wherever work happens.

Put a 1.7-Billion-Parameter Model to Work

Liquid AI's new LFM3 family demonstrates how smaller models can handle specialized tasks without requiring massive computing infrastructure. You can test LFM3-1.7B in Liquid's Playground, which requires sign-in.

Try this: Select LFM3-1.7B and paste this fictional IT-support challenge:

******
You are an IT support triage agent. Classify these tickets as ACCESS, OUTAGE, or BILLING, assign HIGH or NORMAL priority, and return only JSON:

Company-wide outages must be HIGH priority. Do not claim to have fixed anything.
******

Look for: Does the model produce correctly structured JSON, classify the company wide outage as high priority, and avoid claiming actions it hasn't taken?

Insight: Smaller models can handle specialized decision making tasks, but reliable enterprise automation still requires validation, permissions, and execution controls.