January 2026 Summaries

A team of Google Developer Experts (GDEs) conducted a groundbreaking test at Thunderhill Raceway, utilizing Google's new Antigravity (AGY) framework to develop an AI system that offers real-time, physics-based guidance to race car drivers. This AI system, leveraging a "Split-Brain" architecture, separates reflexive actions from strategic decisions using Gemini Nano for immediate responses and Gemini 3.0 for higher-level analysis. The development, compressed from three months to two weeks, involved rapid prototyping in Google AI Studio and transitioning to Vertex AI for production-grade deployment. The framework's neuro-symbolic training method ensures that AI guidance, such as braking instructions, is mathematically verified for safety, while persona-based routing enhances human understanding by delivering context-aware advice. The success of this test, praised by Thunderhill CEO Matt Busby, highlights the potential of AGY to revolutionize the racing industry by merging instinct with objective data, setting a new standard for accuracy and repeatability in racing telemetry.

Gemini CLI introduces hooks as a way to customize and control its operations without altering the source code, allowing users to inject custom logic at specific lifecycle points. These hooks function like middleware, enabling tasks such as adding context, validating actions, enforcing security policies, and optimizing tool usage. A practical application involves using a BeforeTool hook to automatically scan for and block sensitive data like API keys from being written into code. Hooks are executed synchronously, ensuring that custom logic is respected before the CLI proceeds, and are supported in both local configurations and extensions, such as the "Ralph loop" technique, which enhances Gemini CLI’s capabilities by enabling continuous, iterative processing. Hooks are enabled by default from version 0.26.0 onwards, and users are encouraged to explore their potential to tailor Gemini CLI to their specific workflows.

LiteRT, introduced in 2024, has evolved from its TensorFlow Lite foundation into a modern on-device AI framework, offering advanced hardware acceleration for developers. This evolution includes significant performance improvements, with GPU acceleration delivering 1.4x faster performance than TensorFlow Lite and new NPU capabilities providing up to 100x speed over CPU. LiteRT simplifies the deployment process with a unified workflow for GPU and NPU across platforms like Android, iOS, and Windows, supporting cross-platform Generative AI deployments for models such as Gemma. It enables seamless model conversion for PyTorch and JAX, ensuring compatibility with popular frameworks. LiteRT also offers a robust solution for reducing latency in real-time AI applications through asynchronous execution and zero-copy buffer interoperability. Additionally, the framework maintains long-term reliability and compatibility with the existing .tflite model format, catering to both existing and next-generation AI needs.

Agentic AI's FunctionGemma model, a version of the Gemma 3 270M model, is specifically fine-tuned for function calling, enabling developers to create agents that efficiently translate natural language into executable API actions. Despite its initial capabilities, fine-tuning is necessary to tailor the model to specific business rules, resolve tool selection ambiguities, and specialize in niche tasks. A case study highlights how FunctionGemma can be trained to distinguish between internal and external information sources, using the Hugging Face TRL library and the bebechien/SimpleToolCalling dataset. The process involves careful dataset preparation to ensure a balanced representation, which is crucial for the model to learn effective routing logic. The FunctionGemma Tuning Lab offers a no-code interface to simplify the fine-tuning process, allowing users to define function schemas, import custom data, and visualize training progress, ultimately transforming the model into a specialized agent that adheres strictly to enterprise policies.

The Universal Commerce Protocol (UCP) is an open-source standard developed by Google and industry leaders like Shopify and Walmart to streamline agentic commerce by creating a common language for consumer interfaces, businesses, and payment providers. UCP aims to integrate seamlessly with existing retail infrastructure and supports secure payments through the Agent Payments Protocol. It allows businesses to showcase their products across various consumer platforms while maintaining business logic and a customized checkout experience. For AI platforms, UCP simplifies onboarding with standardized APIs and supports agentic shopping. Developers are encouraged to contribute to the evolving community-driven standard, while payment providers benefit from a modular design that ensures interoperability and cryptographic proof of user consent. UCP addresses the challenges of traditional tech infrastructure by reducing integration complexity and enabling real-time inventory checks, dynamic pricing, and instant transactions. It offers a unified integration point, shared language, and extensible architecture to support the full commerce lifecycle securely. UCP's integration with Google allows for a seamless buying experience in Google's conversational platforms, supporting the use of Google Pay for transactions. The protocol invites collaboration from developers and businesses to expand its capabilities and support the future of commerce.

JAX on Cloud TPUs offers robust acceleration for machine learning workflows by providing specialized tools for debugging and profiling in distributed cloud environments. Central to these tools is the libtpu, a critical software component housing the TPU Runtime, which facilitates interactions with hardware and configuration for various debugging tools, including logging and monitoring. JAX and its C++ backend, jaxlib, act as the bridge to libtpu, enabling efficient model code execution and diagnostics. Essential debugging practices include enabling verbose logging to capture extensive runtime data and utilizing tools like the TPU Monitoring Library and tpu-info for real-time performance insights. These tools help users gain a comprehensive understanding of TPU utilization, capacity, and memory usage, allowing for more effective issue resolution. The post further explores leveraging these components and tools to optimize machine learning workflows and sets the stage for advanced debugging techniques like generating HLO dumps and using XProf for code profiling.