This week's top AI updates

GLM has released its 5.1 open-source LLM, which has reportedly surpassed leading models like Opus 4.6 and GPT-5.4 on the SWE-Bench Pro benchmark. This significant advancement suggests AI is now capable of handling complex software engineering tasks, potentially enabling it to perform work equivalent to a human's full workday.

A new AI memory system, 'mempalace,' has been released and is freely available. This system has achieved the highest benchmark score ever recorded for an AI memory system, indicating a significant advancement in AI's ability to store and retrieve information. Its open availability could accelerate research and development in more capable AI models.

This research extends an existing framework to analyze the occupational displacement potential of "agentic AI," systems capable of completing entire workflows end-to-end, not just single tasks. The study applies this new model to 236 occupations across five major US tech metros, including SF Bay and Seattle. It aims to provide a more accurate assessment of AI's labor market impact by addressing limitations of previous exposure measures that assumed task independence.

Anthropic, a prominent AI research company, is launching a new artificial intelligence model specifically designed for cybersecurity applications. This initiative marks their entry or expansion into the critical domain of digital security, aiming to leverage advanced AI capabilities to enhance defense mechanisms.

Large vision-language models frequently produce object hallucinations in image descriptions, highlighting a need for improved detection. This paper reveals that common detection strategies relying on coarse-grained attention weights on visual tokens are unreliable. It identifies hidden confounders like token position and object repetition that distort attention trends, leading to phenomena like Simpson's paradox where expected patterns reverse or vanish.

Frontend-VisualQA addresses a critical limitation where AI coding agents generate valid HTML/CSS but fail to visually verify their UI output, leading to broken layouts or off-screen elements. This tool provides "eyes" for agents like Claude Code and Codex, enabling visual testing and QA by allowing them to verify website UIs against natural language claims given a URL. It significantly enhances the reliability and quality of AI-generated front-end code.

Anthropic has launched a preview of its powerful new AI model, Mythos, specifically tailored for defensive cybersecurity work. This model is currently being utilized by a small, select group of high-profile companies to bolster their security measures. Its introduction signifies a notable advancement in applying sophisticated AI to critical cybersecurity challenges.

Current Large Language Models are limited by their static "train then deploy" paradigm, hindering dynamic adaptation to new information. Test-Time Training (TTT) offers a promising alternative by updating model parameters during inference, but faces significant challenges like architectural incompatibility and computational inefficiency in existing LLM ecosystems. This paper, "In-Place Test-Time Training," aims to address these critical barriers, potentially enabling more adaptive and continuously learning LLMs.

Firmus, an Nvidia-backed AI data center builder, has achieved a $5.5 billion valuation after raising $1.35 billion in just six months. Specializing in AI infrastructure in Asia, Firmus is known for projects like the 'Southgate' data center. This significant investment highlights the accelerating demand and financial commitment towards foundational AI computing capabilities.

OneManCompany introduces an AI company featuring a full corporate organizational structure, where independent AI agents serve as HR, COO, engineers, and designers. Unlike simple chatbot wrappers, these are distinct models and backends, each performing specific roles, with the user acting as the sole human CEO. This project aims to demonstrate a new paradigm for multi-agent systems by mimicking a real company's operational structure.

Google's Gemini AI is being updated to expedite access to mental health resources for users who express distress. This initiative aims to make it faster and easier for individuals in need to connect with appropriate support and services. It represents a significant step towards integrating user well-being and safety into AI platforms.

Google has launched a new, free AI dictation app that operates entirely offline. This application goes beyond basic transcription by automatically polishing and refining spoken words, aiming to improve accuracy and readability without requiring an internet connection.

Developers in the OpenAI community are reporting significant disruptions to their agent workflows due to hard usage limits that lack any visibility. This issue is particularly affecting users with Codex and ChatGPT subscriptions, causing applications and integrations to break without clear reasons or ways to anticipate failures.

ACP (Agentic Control Plane) is a new governance layer specifically designed for AI coding agents like Claude Code and OpenClaw. It addresses the critical challenge of authenticating LLM users and their agents, a problem identified as surprisingly difficult. ACP provides essential governance features, including permissions, usage limits, and audit logs, to ensure better control and accountability for AI-driven development.

This paper reveals that pre-trained diffusion models inherently possess restoration capabilities, which can be activated by directly learning prompt embeddings. Unlike current methods that rely on fine-tuning or Control-Net modules, this work demonstrates that these models secretly know how to perform All-in-One Restoration (AiOR). This discovery could lead to more efficient and direct ways to leverage diffusion models for image restoration without extensive modifications.

Spain's Xoople has successfully raised $130 million in Series B funding to further its ambitious goal of mapping the entire Earth specifically for AI applications. This significant investment will enable the company to build a robust geospatial data infrastructure. Xoople has also forged a strategic partnership with L3Harris, who will be responsible for developing the advanced sensors for Xoople's dedicated spacecraft.

A new reference-free method has been developed to detect hidden behaviors and biases in LLMs, eliminating the need for a base model comparison. This technique, which uses approximately 100 chat calls, effectively matches or surpasses existing auditing methods on benchmarks. Crucially, it can also uncover RLHF-induced opinion biases on socially sensitive topics in models like Llama 70B.

The rapid expansion of the Internet of Medical Things (IoMT) is transforming healthcare but also introduces severe cybersecurity and patient safety concerns due to increasing cyberattacks. This paper proposes a novel Tsetlin Machine (TM)-based Intrusion Detection System (IDS) designed to detect a wide range of cyberattacks targeting IoMT networks. The aim is to enhance the security and resilience of these critical medical infrastructures.

A new GitHub repository, 'caveman,' introduces a novel Claude Code skill designed to drastically reduce token usage. This technique achieves an impressive 75% token cut by employing a highly concise, 'caveman-like' communication style. The method aims to improve efficiency and lower costs for AI interactions by simplifying language.