📁 LLM

OpenAI has announced that on February 13, 2026, it will retire the GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini models from ChatGPT. The decision does not currently impact the APIs. This announcement follows the previous communication regarding the retirement of GPT-5 (Instant, Thinking, and Pro).

The emergence of "distilled" models like Qwen 8B DeepSeek R1 has demonstrated reasoning capabilities exceeding their size. The article questions why there aren't more models of this kind, capable of operating on hardware with limited resources.

While major companies pour billions into large language models, San Francisco-based startup Logical Intelligence is taking a different approach to achieving AGI, aiming to emulate the human brain. The company seeks to develop artificial intelligence that more closely resembles human reasoning.

The Kimi AI team sent an appreciation email to a user who reviewed Kimi K2.5 on their YouTube channel, offering premium access to "agent swarm". The news was shared on Reddit.

OpenAI built an in-house AI data agent that uses GPT-5, Codex, and memory to reason over massive datasets and deliver reliable insights in minutes, enhancing data processing and analysis efficiency.

OpenAI has released Prism, a free AI-powered workspace for scientists. This tool, integrated with GPT-5.2, aims to facilitate the writing of scientific papers and collaboration. However, some researchers fear that Prism could contribute to an increase in low-quality publications, an existing problem in the sector.

Google has announced Project Genie, a new tool for generating virtual worlds powered by advanced AI models like Genie 3, Nano Banana Pro, and Gemini. Initially available to AI Ultra subscribers in the U.S., it offers new creative possibilities.

Google has initiated a testing phase for Project Genie, offering AI Ultra subscribers in the U.S. the opportunity to experiment with interactive worlds. The project represents a step forward in exploring the potential of generative artificial intelligence in creating virtual environments.

Anthropic's secret to building a better AI assistant might be treating Claude like it has a soul—whether or not anyone actually believes that's true. Anthropic released Claude's Constitution, outlining the company's vision for how its AI assistant should behave, notable for the highly anthropomorphic tone it takes toward Claude. It remains unclear whether this is a development strategy or a genuine belief about the nature of AI.

The Qwen3-ASR family includes 1.7B and 0.6B parameter models, capable of identifying the language and transcribing audio in 52 languages and dialects. The larger model achieves performance comparable to proprietary commercial APIs, offering a valid open-source alternative for speech recognition applications.

Google Maps now allows users to interact with Gemini while walking or cycling. You can ask contextual questions like "What neighborhood am I in?" or "What are the top-rated restaurants nearby?".

An engineer has developed Mini-LLM, an 80 million parameter transformer language model from scratch, based on the Llama 3 architecture. The project includes tokenization, memory-mapped data loading, mixed precision training, and inference with KV caching. Suitable for students wanting to understand modern LLM architecture.

OpenMOSS has released MOVA (MOSS-Video-and-Audio), a fully open-source model with 18 billion active parameters (MoE architecture, 32 billion total). MOVA offers day-0 support for SGLang-Diffusion and aims at scalable and synchronized video and audio generation.

A developer has created a system where an LLM generates procedural spells for a virtual reality prototype. The system uses a pool of spell components and converts words into instructions to create unique effects. The soundtrack was made with Suno.

A user discovered that Devstral 2 123B and 24B models can be forced into more consistent logical reasoning through the use of Jinja templates. Adding a specific Jinja statement appears to significantly enhance the reasoning capabilities of the models, although the smaller version may have difficulty exiting the thinking process in some configurations.

A new study shows that, with proper training, human experts can outperform automated systems in identifying Korean texts generated by LLMs. The approach relies on a detailed rubric that analyzes the peculiarities of the language.

A new study introduces Gap-K%, a novel technique for identifying data used in the pre-training of large language models (LLMs). The method analyzes discrepancies between the model's top-1 prediction and the target token, leveraging the optimization dynamics of pre-training to improve detection accuracy.

A novel approach, Self-Querying Bidirectional Categorical Planning (SQ-BCP), addresses the challenges of large language models (LLMs) in reasoning with incomplete information. SQ-BCP uses targeted queries and hypotheses to resolve unknowns, significantly reducing constraint violations in complex tasks such as WikiHow and RecipeNLG.

A 2025 workshop explores synergies between neuroscience and artificial intelligence, identifying promising areas such as embodiment, language, robotics, learning, and neuromorphic engineering. The goal is to develop NeuroAI to improve algorithms and the understanding of biological neural computations, analyzing benefits and risks through SWOT analyses.

Assistant_Pepe_8B, an 8 billion parameter LLM, has been released, designed to combine top-tier shitposting capabilities with actual helpfulness. The model boasts a 1 million token context window and aims to provide useful and irreverent responses, while avoiding excessive pandering. No system prompt is needed.