📁 LLM

A new study evaluates the effectiveness of different prompting strategies (Zero-Shot, Few-Shot, Chain-of-Thought) on large language models (LLMs) such as GPT-3.5, GPT-4, and GPT-4o, applied to question answering on charts. The analysis focuses on the accuracy and format adherence of the responses, using the ChartQA dataset.

A new study explores the use of Large Language Models (LLM) for synthetic data generation, aiming to improve the performance of smaller models through fine-tuning. The research focuses on analyzing the diversity of generated data in the embedding space and proposes a targeted sampling method to optimize deliveries.

Research explores whether thought requires a language-like format, as posited by the Language of Thought (LoT) hypothesis. Simulations with AI agents reveal that emergent communication protocols outperform predefined languages in efficiency, suggesting that optimal collaborative cognition may not depend on symbolic structures.

GigaChat has released the weights of its GigaChat-3.1-Ultra (702B) and Lightning (10B) models under the MIT license. Ultra is designed for high-resource environments, while Lightning targets local inference. Both are pre-trained from scratch and optimized for English and Russian, with good multilingual performance and tool calling. GigaChat-3.1-Lightning achieves 0.76 on the BFCLv3 benchmark.

A user seeks LLM models capable of competing with Claude Opus, but with an extreme constraint: only 32MB of VRAM. The discussion explores hardware limitations and potential alternatives, considering local execution with Ollama on obsolete hardware.

The Reka AI team, a research lab focused on models useful for real-world applications, participated in an AMA (Ask Me Anything) session on Reddit, in the LocalLLaMA subreddit. The focus was on their Reka Edge model and the direction of their research, with a particular emphasis on inference and APIs.

HP introduces "HP IQ", a new local AI application integrated into its business laptops. This application allows users to chat, share files, record and summarize meetings, aiming to differentiate HP products in the market.

OpenAI has discontinued Sora, its app based on the Sora 2 video and audio generation model. Despite the impressive capabilities of the underlying model, interest in an AI-only social feed was not enough to sustain its growth.

OpenAI is preparing to shut down Sora, the video generation app that drew widespread attention when it launched in late 2024. The company thanked users for their contributions and promised more details on the timelines for the shutdown and the preservation of works created with the platform.

Anthropic's new auto mode for Claude Code lets AI execute tasks with fewer approvals. This reflects a broader shift toward more autonomous tools that balance speed with safety through built-in safeguards. The goal is to accelerate processes while maintaining control over the outcomes.

Google TV is enhanced with three new Gemini-powered features: visual responses, deep dives, and sports briefs. The integration aims to provide a more interactive and informative user experience directly on the TV.

OpenAI releases prompt-based safety policies to protect younger users when interacting with AI systems. The policies are designed for developers using gpt-oss-safeguard, aiming to moderate age-specific risks.

A Reddit user expresses enthusiasm for the performance of the Qwen 27B model, highlighting its successful use even with GPUs equipped with 24GB or 48GB of VRAM. The discussion focuses on the accessibility of large language models (LLMs) for users with less expensive hardware, opening new possibilities for local inference.

A novel fine-tuning approach, named Embedding Space Separation (ES2), aims to enhance the safety of large language models (LLMs) by increasing the distance between harmful and safe query representations in the embedding space. KL divergence regularization prevents degradation of general capabilities.

JointFM, a foundation model that directly predicts future joint probability distributions, has been introduced. Trained on infinite streams of synthetic stochastic differential equations, JointFM reduces energy loss by 14.2% relative to the strongest baselines in zero-shot scenarios.

AgenticGEO is a framework that uses self-evolving agents to optimize generative search engines. It overcomes the limitations of static approaches by dynamically adapting to changes in engines and content, outperforming traditional methods.

Researchers have found that instructing an AI model to impersonate an expert in a specific field can be counterproductive. While this technique may improve safety, it does not guarantee better results in terms of accuracy and quality of the deliveries.

A researcher has trained Qwen3.5 27B large language models (LLM) with repeated layers, suggesting that models might process information in an internal "universal language." Results indicate that repeating blocks in the middle of the transformer stack appears to be the most effective strategy. Several pre-trained models are available on Hugging Face.

Sen. Bernie Sanders tried to trick Claude into revealing the AI industry's secrets, but he really just exposed how agreeable chatbots can become. The attempt backfired, but it spawned numerous online memes.

A Stanford research analyzes interactions between people and chatbots, revealing how AI models can unintentionally fuel delusions and dangerous obsessions. The study raises crucial questions about the responsibility of AI companies and the need for greater regulation.