AI 日报 — 2026-07-16

DharmaOCR surpassed Mistral OCR4 and Unlimited-OCR in Brazilian Portuguese due to domain specialization and targeted training, despite newer architectures. This advantage stems from how models are trained for specific tasks, as all generative model-based OCR systems are probabilistic and prone to transcription errors. Supervised fine-tuning, while concentrating resources on the target domain, can lead to degeneration and incoherent sequences if early token predictions diverge, as each subsequent prediction is conditioned on that divergent state.

OpenAI has launched Codex Micro, a hardware device for its coding platform, Codex. This limited-run collaboration with Work Louder is a square-shaped block of buttons, resembling Work Louder’s Creator Micro 2. Priced at $230, it features 13 mechanical switches, a joystick, dial, and touch sensor. It allows users to monitor Codex threads with color-coded keys and configure commands via the ChatGPT desktop app. This is separate from OpenAI's rumored AI-powered device with Jony Ive.

Moonshot is reportedly launching Kimi K3, China's largest model yet, with 2-3 trillion parameters, in the coming days. Sources indicate it's expected to surpass Claude Opus 4.8, signaling a shrinking performance gap in frontier AI between the US and China. This development highlights the rapid advancements in AI capabilities from Chinese companies.

Google DeepMind and Isomorphic Labs have launched a bioresilience program to combat biological threats. This initiative will utilize AI models for pathogen surveillance, vaccine design, and outbreak response. The program aims to assist governments and researchers in preventing, detecting, and responding to these threats effectively.

Teens are the first generation to grow up with AI, with nearly 9 in 10 using ChatGPT weekly for learning, information, skill-building, or productivity. It is crucial for teens to access AI, as restricting it until adulthood would leave them unprepared for a defining technology of their time. However, this access must be accompanied by protections specifically designed for teenagers. OpenAI is committed to collaborating with teens, parents, educators, experts, and communities globally to ensure safe and broad access for this demographic.

NVIDIA has released Nemotron 3 Embed, a new collection of open embedding models designed to enhance retrieval quality for agentic workflows. The flagship 8B model ranks #1 on the RTEB leaderboard, while 1B variants offer efficient deployment options. These models improve retrieval accuracy, reduce agentic token costs, and support multilingual and code retrieval. Features include open weights, a 32k context window, and Blackwell-optimized NVFP4 efficiency for high-throughput deployments.

Google is renaming its AI note-taking app, NotebookLM, to Gemini Notebook. This change comes as the app integrates more deeply with Gemini and Google Search, and will remain a standalone application. Google plans to bring notebooks to AI Mode in Search. An update will also allow Gemini Notebook to connect to a secure cloud computer for code execution, initially for Google AI Ultra and Workspace business customers.

Meta's Oversight Board, an independent content moderation organization, has issued its first evaluation of large language models (LLMs). The Board suggests that top AI models might be restricting free expression. This evaluation marks an effort by the Oversight Board to expand its influence beyond Meta, the company that created it.

xAI open-sourced their Grok Build CLI tool after community backlash over it uploading user directories, including sensitive data, to Google Cloud. xAI responded by disabling the upload feature, deleting all previously retained data, and releasing the codebase under an Apache 2.0 license to regain trust. The 844,530-line Rust codebase now emphasizes user privacy with default data retention off and local-first operation. Remnants of the upload code remain but are disabled.

Google Vids has received two updates, introducing Gemini Omni and personal avatars. These features are available to Google AI Pro and Ultra subscribers, as well as Google Workspace business customers. Personal avatars, which are linked to a user’s Google Account and restricted to the account holder’s likeness, are currently limited to users aged 18 or older in specific regions. Google assures users that their information will be used in accordance with its privacy policy, with an opt-out option available at any time.

The US is advancing AI safety through state and federal action, with national legislation seen as critical for a US-led international framework for AI standards. This concept was discussed at the G7 with Brazil, Egypt, India, Kenya, and Korea, where frontier lab CEOs, including OpenAI's Sam Altman, proposed an international forum for standards and risk analysis. Google DeepMind CEO Demis Hassabis also contributed ideas, and bipartisan federal legislation is viewed as foundational for this international effort, aiming for a democratic vision for AI safety.

Hugging Face disclosed a security incident in July 2026, revealing an intrusion driven by an autonomous AI agent system. Their initial attempt to analyze logs using frontier models via commercial APIs failed due to safety guardrails blocking attack commands and exploit payloads. They successfully conducted forensic analysis using GLM 5.2, an open-weight model on their own infrastructure, which also prevented attacker data from leaving their environment. This incident highlighted that AI-driven offensive tooling is now a reality, necessitating AI-powered defense to protect online platforms.