AI Digest — Jul 17, 2026 (Morning)

Google has rebranded NotebookLM as Gemini Notebook, a tool designed to assist with note-taking and idea generation using AI. Gemini Notebook aims to enhance user productivity by providing features such as automatic note organization and summarization. This rebranding reflects Google's ongoing efforts to integrate AI into various productivity tools. Gemini Notebook is part of Google's broader Gemini AI initiative, which focuses on developing AI models for diverse applications.

A new method, in-place tokenizer expansion, allows upgrading a pre-trained model's tokenizer by continuing the existing tokenizer's BPE merges on a multilingual corpus. This approach enables the addition of new languages without significantly increasing latency, compute, and energy consumption. The method involves copying carried-over embedding rows unchanged and initializing new rows as the mean of their source sub-token embeddings, followed by a two-stage adaptation process. The researchers applied this method to an 8B-parameter Mixture-of-Experts model, resulting in a 128K tokenizer that encodes certain languages in fewer tokens, leading to estimated per-character decode speedups. The expanded model and tokenizer are released, along with negative findings that shaped the approach.

Researchers propose Mask-Aware Policy Gradients for Diffusion Language Models, addressing the challenge of log-likelihood estimation in Masked Diffusion Language Models (MDLMs). The approach models two decisions at each generation step: token placement and position remasking, formalized as a two-stage action MDP. This leads to state-of-the-art results on mathematical reasoning and coding benchmarks. The method achieves scores of 87.1% on GSM8K and 53.4% on MBPP. This work contributes to the development of more effective large language models.

Researchers introduced On-Policy Delta Distillation (OPD$^2$), a new method for on-policy distillation in reinforcement learning. OPD$^2$ uses a delta signal, the difference between a teacher model and its base model, to provide token-level supervision. This approach captures changes induced by reasoning tuning and improves the transfer of reasoning capabilities. Experiments on mathematics, science, and code-reasoning benchmarks show OPD$^2$ outperforms conventional on-policy distillation. The method enables strong performance with a short post-training period.

Google DeepMind and Isomorphic Labs have announced a joint approach to bioresilience, focusing on the application of AI models to this field. This collaboration aims to leverage AI capabilities to enhance bioresilience, which is crucial for understanding and mitigating the impact of diseases and other biological threats. The approach is outlined in a blog post on the DeepMind website, detailing their strategy and goals. By combining expertise from both organizations, they seek to drive innovation in bioresilience through AI-driven solutions.

NVIDIA's Nemotron 3 Embed has achieved the top ranking on the RTEB benchmark, surpassing other models in agentic retrieval tasks. This advancement is significant for natural language processing and information retrieval applications. The RTEB benchmark evaluates models' ability to retrieve relevant information, and Nemotron 3 Embed's performance indicates its potential for improving search and question-answering systems. Nemotron 3 Embed's success can be attributed to its architecture and training data, which enable it to effectively capture semantic relationships and context.

The Kimi K3 2.8T-A50B model has been released, touted as the largest open model ever made available. This model is compared to the Opus 4.8 class in terms of performance but is priced similarly to Sonnet 5 models. The release of such a large open model is significant for the AI community, particularly for researchers and developers looking for advanced, accessible tools. The model's size and capabilities suggest it could be used for a wide range of applications, from natural language processing to image generation. Its open nature means it can be studied, modified, and improved upon by the community.

Thinky has released its first full large language model (LLM), named Inkling, which is a multimodal model with 975B parameters. The model is open-sourced under the Apache 2.0 license, making its weights available for use. Additionally, a smaller version, Inkling-Small, with 276B parameters is also available. This release is significant as it provides the AI community with a new, openly accessible model for research and development. The model's multimodal capabilities allow it to process and generate various types of data, including text and images.

GPT-5.6 Sol Pro, an AI model, has solved an open problem in convex optimization that has been unsolved for 30 years. The problem is a significant one in the field of optimization, which is crucial for many areas of AI and machine learning. The solution was achieved through AI-assisted research, highlighting the potential of AI in advancing mathematical and computational fields. The breakthrough could have implications for various optimization tasks and may lead to more efficient algorithms.

LM Studio has announced the release of Bionic, an AI agent designed for open models. This agent is intended to facilitate the development and deployment of AI models. The introduction of Bionic aims to provide a more accessible and efficient way to work with open models, potentially simplifying the process for developers. The technical details of Bionic include its compatibility with various open models and its ability to streamline model deployment. This release is relevant to the AI community as it pertains to the development and application of open models.