AI Digest — Jun 10, 2026 (Morning)

DeepMind has announced Gemma 4 12B, a multimodal model that unifies various tasks into a single architecture. This encoder-free model is designed to handle multiple modalities, such as text, images, and audio. Gemma 4 12B aims to simplify the development process by eliminating the need for separate encoders for different modalities. The model's unified architecture could lead to more efficient and effective multimodal processing. The introduction of Gemma 4 12B is significant for the field of multimodal learning and could have implications for various applications.

A new approach to supervised fine-tuning (SFT) has been introduced, focusing on target distribution design. This method, called Target-SFT, allows for more flexibility in how the model learns from demonstrated trajectories by explicitly considering the reliability of observed tokens and allocating probability to alternative tokens. The Q-target framework provides a unified perspective on various SFT variants, highlighting the importance of target distribution design in SFT. Target-SFT has been shown to outperform existing methods across multiple dataset-model settings. This work contributes to a deeper understanding of SFT and opens up new avenues for improving SFT objectives.

Notion is utilizing Codex, an AI model developed by OpenAI, to enhance its capabilities. This integration allows Notion to implement one-shot specs and build AI Voice Input for web applications. The goal is to amplify engineering power, especially in small teams, by automating certain tasks and improving user interaction. This collaboration demonstrates the potential of Codex in real-world applications, particularly in enhancing productivity and user experience. The integration is expected to streamline Notion's development process and provide more intuitive user interfaces.

Google DeepMind has released Gemini 3.5 Live Translate, a technology that allows for near real-time, natural speech translation. This feature is integrated into Google AI Studio, Google Translate, and Google Meet. The technology aims to facilitate more fluid and natural voice translation, enhancing communication across languages. Gemini 3.5 Live Translate is a significant update to Google's translation capabilities, leveraging advancements in AI and machine learning.

Google DeepMind is investing in the future of robotics in Europe through various initiatives. This includes partnerships with European research institutions to advance robotics research. The goal is to develop more advanced and capable robots that can interact with and understand their environment. This investment is expected to drive innovation in the field of robotics and have significant technical implications.

Hugging Face has published a benchmarking study on the Frontier Automatic Speech Recognition (ASR) system's ability to handle bilingual customers who engage in code-switched speech. The study aims to assess the system's performance in understanding and responding to customers who switch between languages during conversations. The benchmarking process involves evaluating the ASR system on a dataset of code-switched speech to identify its strengths and weaknesses. This research is significant for AI researchers and developers working on voice agents and conversational AI systems. The study's findings can help improve the design and development of more effective and inclusive voice agents.

Cohere Labs has introduced North Mini Code, its first model designed specifically for developers. This model is intended to provide developers with an efficient tool for coding tasks. The release of North Mini Code marks an expansion of Cohere's offerings into the developer space. The model's capabilities and performance will be of interest to those working on coding and developer-focused AI applications. North Mini Code is available through the Hugging Face platform.

A project utilized Hugging Face Spaces to create a 3D Paris gallery by chaining two spaces together. This involved an agent that could generate and manipulate 3D models. The project demonstrates the potential of combining different AI models and tools to achieve complex tasks. The use of Hugging Face Spaces allows for the integration of various AI capabilities, such as text-to-image and 3D modeling, to create interactive and immersive experiences.

A German court has ruled that Google is liable for false information provided in its AI-generated overviews. The decision states that the AI-generated content is considered Google's own words, making the company responsible for its accuracy. This ruling could have significant implications for tech companies and their use of AI in generating content. The case highlights the growing need for accountability in AI-generated information. The ruling may influence how companies develop and deploy AI models in the future.

Nucleus is a new, security-hardened, Nix-native container runtime. It has been released on GitHub by the sig-id team. Nucleus aims to provide a more secure environment for running containers by leveraging the Nix package manager's reproducibility and isolation features. This project could be of interest to those working on container security and infrastructure. The code is available on GitHub for review and contribution.

A man was wrongfully arrested due to AI misidentification. The incident highlights the potential flaws in AI-powered facial recognition systems. The man is now seeking justice, raising concerns about the reliability of such systems in law enforcement. This case underscores the need for more accurate and transparent AI technologies. The incident has sparked a discussion about the limitations and potential biases of AI in critical applications.