📁 LLM

Deep Agents simplifies building complex AI systems through specialized agents. It introduces subagents for context isolation and skills for progressive capability disclosure. The article illustrates how to implement multi-agent systems, preserving context, specializing functions, parallelizing processes, and minimizing toolsets.

A WIRED analysis of over 5,000 papers from NeurIPS, using OpenAI's Codex, reveals unexpected collaboration between the US and China in AI research. The findings challenge narratives of pure competition and suggest a more complex and nuanced landscape.

A researcher fine-tuned the Qwen3-14B language model using 10,000 DeepSeek traces, achieving a 20% performance increase on a custom security benchmark. This demonstrates how fine-tuning smaller models with specific datasets can be a viable and more cost-effective alternative to using large models, especially in contexts like code analysis.

Higgsfield transforms simple ideas into cinematic-quality videos for social media. The platform leverages the power of advanced models like OpenAI GPT-4.1, GPT-5, and Sora 2 to automate the creation of engaging and visually stunning video content, opening new possibilities for digital creators.

Microsoft has released VibeVoice-ASR, a new model for Automatic Speech Recognition (ASR). The model is accessible via Hugging Face, opening new possibilities for developers working on voice applications. The release includes a link to the Hugging Face page and discussions on Reddit.

Anthropic has introduced a new constitution for Claude, its flagship language model. This update aims to improve the model's alignment with human values and make it safer and more effective in its applications. The initiative represents a crucial step forward in the responsible development of artificial intelligence.

OpenAI is trying to alleviate concerns about its new Stargate datacenters. The company promises plans that take into account local needs, minimizing the environmental impact and the impact on electricity costs. The initiative comes at a time of increasing attention to the energy consumption linked to artificial intelligence.

A new model named GLM-OCR from Z.ai has been spotted on GitHub. The finding was reported on Reddit, in the LocalLLaMA subreddit, via a post including an image and links to the discussion and the original resource. Further details on the model's capabilities or technical specifications are currently unavailable.

YouTube is introducing a feature that will allow content creators to make Shorts using AI versions of themselves. Viewers might soon see AI avatars of their favorite YouTubers while scrolling through Shorts feeds.

A bug in GLM-4.7-Flash-GGUF causing looping and poor outputs has been fixed. Users are advised to redownload the model for significantly improved results. Z.ai has suggested optimal parameters for various use cases, including general use and tool-calling. The update is available on Hugging Face.

We compared the AI models from Google (Gemini 3.2 Fast) and OpenAI (ChatGPT 5.2) to evaluate their performance. The tests, based on complex prompts, aim to simulate the standard user experience, that is, those who do not pay for subscriptions. The analysis combines objective evaluations and subjective impressions, updating the comparative tests carried out in 2023.

Here's how to get GLM 4.7 working on llama.cpp using Flash Attention for improved performance. The guide includes configuration details and a link to a specific Git branch. Note that quantizations may need to be recreated to avoid nonsensical outputs.

Adobe is integrating artificial intelligence tools into Acrobat, offering new features such as automatic podcast summary generation, presentation creation, and file editing via text prompts. The goal is to simplify and speed up user workflows.

Microsoft CEO Satya Nadella warns that artificial intelligence must generate benefits for a broad segment of the population, otherwise it risks losing social permission and turning into a speculative bubble. A wider impact is needed to prevent the benefits from being concentrated in the hands of a few.

Large language models (LLMs) continue to be vulnerable to prompt injection attacks, a technique that tricks AI into performing unauthorized actions. The difficulty lies in their inability to understand context as a human would, making them susceptible to manipulations that bypass security measures. New approaches are needed to effectively protect these systems.

Microsoft CEO Satya Nadella says datacenter location is "the least important thing" for AI sovereignty. Ownership of models and embedded corporate knowledge matters more than server location, according to Nadella.

OpenAI is committed to ensuring that electricity prices do not increase in the communities where it builds its Stargate data centers. The company will fund grid upgrades and flexible load management systems to reduce stress on the energy supply. The goal is to ensure that the expansion of AI infrastructure does not burden consumers.

AI cost efficiency clashes with data sovereignty, forcing companies to rethink their risk frameworks. The case of DeepSeek, a Chinese AI lab, raises concerns about data sharing with state intelligence services. This requires stricter governance, especially in sectors like finance and healthcare, where transparency on data provenance is crucial to avoid violations and reputational damage.

OpenAI introduces "Edu for Countries", a new initiative designed to support governments in adopting artificial intelligence. The goal is to modernize education systems and prepare the workforce of the future, providing tools and resources to integrate AI into learning and professional development.

The Davos 2026 Forum will feature artificial intelligence as a key topic. Global leaders will discuss crucial issues such as the necessary computing power, the control of algorithms, and the ethical and social implications arising from its development. The event promises to be a turning point in defining the future of AI and its impact on the world.