Search AI Products and News

Recently, AIbase obtained the latest information from social media platforms and learned about Shisa.AI, a HuggingFace model provider specializing in Japanese fine-tuning. Its newly released bilingual Japanese-English model has attracted significant attention in the industry. This article will provide you with an in-depth analysis of Shisa.AI's latest achievements and its breakthrough advancements in the field of Japanese AI. Shisa V2405B: The birth of Japan's strongest open source model According to AIbase's understanding, Shisa.AI has recently released a Llama3.1-based S

On May 8th, OpenAI o4-mini launched with Reinforcement Fine-Tuning. The combination of these two technologies completely transforms the cost structure and technical threshold for AI specialization, enabling enterprises to rapidly turn general-purpose AI into expert systems in specific fields using only a small amount of training data. The leap from general intelligence to expert-level AI is highlighted by the Reinforcement Fine-Tuning technology at the core of this release. Unlike traditional supervised fine-tuning, RFT is based on reinforcement learning algorithms, driven by reward-based training.

This project details the creation of Pusa, a high-quality open-source video model built with a budget under $100. It leverages Mochi fine-tuning for efficient and cost-effective training, demonstrating the feasibility of achieving impressive results with minimal resources.

Skywork AI's research team recently unveiled SkyReels-A2, a groundbreaking video generation framework pushing the boundaries of controllable video synthesis. This Element-to-Video (E2V) framework generates natural videos from text prompts by composing diverse visual elements (characters, objects, backgrounds), maintaining high fidelity to reference images. SkyReels-A2's core lies in its sophisticated data processing pipeline. The research team designed a comprehensive data construction pipeline.

A groundbreaking achievement in quantum computing has been reached with the successful fine-tuning of a billion-parameter AI model on Origin Quantum's Wukong, marking a significant milestone in the development of quantum-enhanced AI.

Data intelligence company Databricks recently unveiled a novel large language model fine-tuning method – TAO (Test-time Adaptive Optimization). This technology offers new hope for the advancement of open-source models. By leveraging unlabeled data and reinforcement learning, TAO excels in reducing enterprise costs while achieving remarkable results across various benchmark tests. According to tech news outlet NeoWin, TAO...

Large Language Models (LLMs) are constantly evolving in the field of artificial intelligence. Researchers from Carnegie Mellon University (CMU) and HuggingFace recently introduced a new method called Meta Reinforcement Fine-Tuning (MRT). This method aims to optimize the computational efficiency of LLMs during testing, particularly excelling in solving complex reasoning problems. Studies show that existing LLMs struggle with...

A detailed interpretation of the GPT-4.5 system card report released by OpenAI on February 27, 2025. This report comprehensively covers the development, capabilities, safety evaluations, and preparedness framework assessment of the GPT-4.5 model, aiming to showcase its advancements and potential risks, and to explain OpenAI's mitigation strategies.

A collaborative team from Ant Group, Zhejiang University, University of Liverpool, and East China Normal University presented ScaleOT, a novel cross-domain fine-tuning framework, at AAAI 2025. This framework achieves a 50% improvement in privacy protection and a remarkable 90% reduction in computational cost compared to knowledge distillation, without compromising model performance. ScaleOT offers a highly efficient and lightweight solution for cross-domain fine-tuning of billion-parameter models. The innovative nature of the paper secured it an oral presentation at AAAI.

Recently, the Italian data protection authority announced a €15 million fine on OpenAI's AI chatbot ChatGPT. This decision was made following an investigation launched in March 2023, which found that OpenAI violated privacy protection regulations in multiple areas. Image source note: the image is generated by AI, image authorized by service provider Midjourney. The investigation indicated that OpenAI made significant errors in handling personal data. Firstly, the company failed to promptly report data breach incidents.

During a continuous 12-day live event, OpenAI released the O1 model API to developers on the 9th day and announced significant upgrades to the real-time API, supporting WebRTC technology. Starting from the release date, OpenAI will provide access to the O1 API for developers at usage level 5. Compared to the previous preview version, this updated O1 model API achieves a 60% reduction in thinking costs and adds advanced visual processing capabilities. Meanwhile, GPT-4o has improved in audio processing.

Recently, a research team from the Beijing Academy of Artificial Intelligence launched a new image generation model called OmniGen. Unlike previous tools such as Stable Diffusion, which focused on a single task, OmniGen's biggest highlight is its multi-capability: it can handle a variety of image generation tasks within a unified framework, including text-to-image generation and image editing, making it a versatile player. This means that users only need to provide simple prompts.

Recently, a research team from Shanghai Jiao Tong University and Harvard University introduced a novel model fine-tuning method — LoRA-Dash. This new approach claims to be more efficient than existing LoRA methods, particularly in the fine-tuning of specific tasks, achieving the same results with an 8 to 16 times reduction in the number of parameters. This is undoubtedly a major breakthrough for fine-tuning tasks that require substantial computational resources. In the context of the rapid development of large-scale language models, the demand for fine-tuning specific tasks is steadily increasing. However, fine-tuning often

The domestic open source video generation model CogVideoX-5B has been officially released in the Mota ModelScope community, significantly improving the quality and visual effects of video generation. Based on the large-scale DiT model, this model utilizes a 3D causal variational autoencoder and expert Transformer technology, achieving spatio-temporal joint modeling through 3D-RoPE positional encoding and a 3D full attention mechanism. The use of progressive training techniques allows the model to generate long videos with distinct motion features, coherence, and high quality.