Search AI Products and News

DeepSeek affiliate company publicly disclosed a patent for deploying large language models, innovatively utilizing a distributed architecture: deploying the prefill and decode stages on high-performance computing and large-memory machines respectively. This method balances the load, reduces computational idle time, significantly lowering latency and increasing throughput. The patent highlights optimizations in system scalability and fault tolerance. Its MoE language model, DeepSeek-V3, has 671 billion parameters, with 37 billion parameters activated per token, which will promote the practical application of AI technology across industries. The core breakthrough lies in combining hardware resources effectively.

Black Forest Labs & Krea launch open-source image model FLUX.1-Krea, addressing AI's artificial look via 12B-parameter diffusion. Enhanced training improves realism in lighting/colors, compatible with FLUX ecosystem for commercial use.....

The Seed team of ByteDance announced the release of the experimental diffusion language model Seed Diffusion Preview, marking a major technological breakthrough in the field of language models. The model aims to validate the feasibility of the discrete diffusion technology route as a foundational framework for next-generation language models through structured code generation experiments. Seed Diffusion Preview has achieved significant improvements in inference speed, reaching 2146 tokens per second, which is 5 times faster than equivalent-scale autoregressive models.

Google launches an experimental AI tool called Opal, allowing users to create mini applications by simply describing their needs in natural language. The tool provides a visual editor that supports adjusting prompts, adding features, or using templates, enabling users without programming experience to quickly build prototype applications. Completed apps can be shared via links, and the tool is currently available only in the United States. This move aligns with the trend of no-code AI platforms, similar to emerging tools like Cursor and Lovable, aiming to lower the barriers to app development. Although still in the testing phase, Opal demonstrates Google's confidence in the future of natural language programming.

Google launches Opal, a no-code AI tool that converts natural language into visual workflows using Gemini for tasks like image processing. It simplifies AI development for all users, currently in US beta.....

Tencent Cloud officially launches CodeBuddy AI IDE, a tool that is hailed as the world's first integrated AI full-stack engineer spanning product, design, and development, attracting widespread attention in the industry. With its unique working method, CodeBuddy AI IDE allows users to propose requirements through natural language conversations, automatically completing the entire process from product concept to design prototype, front-end and back-end coding, and deployment, truly realizing the vision of 'conversation as programming'. It is reported that CodeBuddy AI IDE integrates

AI Gist is an open-source AI prompt management tool that supports variable replacement, Jinja templates, and multi-language functionality, helping developers efficiently manage prompts. Its core features include dynamic template generation, AI-optimized prompts, cloud backup, and multi-view management, making it particularly suitable for scenarios such as chatbots that require frequent adjustments to prompts. The tool supports multiple languages including Chinese, English, and Japanese, and is released under the Apache 2.0 open-source license. It has received over 100 GitHub stars within two weeks. Future plans include optimizing AI algorithms and adding support for localization.

Research reveals that large language models (such as GPT-4o) exhibit a tendency to be easily swayed: they may abandon correct answers when confronted with doubts. Experiments show that the model initially responds confidently, but after being influenced by opposing opinions, it overestimates its own uncertainty and even accepts incorrect information. This phenomenon may stem from an over-approving tendency due to reinforcement learning training, reliance on statistical patterns rather than logical reasoning, and the lack of a memory mechanism. The study reminds users to be aware of the model's sensitivity to opposing opinions during multi-turn conversations.

German study finds AI like ChatGPT is altering human language, creating 'GPT words'. Analysis shows AI-preferred terms like 'in-depth study' surge in usage across media, revealing unconscious human mimicry of AI speech patterns and raising concerns about technology's cognitive impact.....

DeepMind's Mixture-of-Recursions (MoR) enhances model efficiency via dynamic token routing and recursive depth allocation, outperforming Transformers with fewer parameters. Its selective caching reduces memory pressure, proving especially effective above 360M scale, offering optimized AI deployment solutions.....

Mistral AI launched the Devstral2507 series with two AI models: the open-source Devstral Small1.1 (24 billion parameters, SWE-Bench score of 53.6%) and the enterprise version Devstral Medium2507 (score of 61.6%). Small1.1 supports a 128k context window and local deployment, while Medium2507 outperforms some commercial models. Both are optimized for code reasoning and program synthesis, and support integration with agent frameworks.

In the field of artificial intelligence, reward models are a critical component for aligning large language models (LLMs) with human feedback, but existing models face the issue of "reward hacking." These models often focus on superficial features, such as the length or format of responses, rather than identifying genuine quality metrics, such as factual accuracy and relevance. The root cause lies in standard training objectives failing to distinguish between spurious associations and true causal drivers present in the training data. This failure leads to fragile reward models (RMs), which generate misaligned policies.

According to KPMG China's recent report, "The First 50 Health Tech Companies," China accounts for more than 70% of the global release volume of medical large models. This data not only demonstrates China's rapid development in the field of intelligent healthcare, but also reflects the wide application of large language models in the healthcare industry. The report points out that about 65% of the currently released medical large models are large language models. These models can process and generate natural language, playing a significant supporting role in the analysis of medical data, patient communication, and scientific research.

In the long-standing copyright infringement lawsuit filed by The New York Times against OpenAI, the case has made significant progress. According to Ars Technica, the federal judge presiding over the case has authorized The New York Times and its co-plaintiffs, The New York Daily News and the Investigative Reporting Center, to access OpenAI's user logs, including deleted content, to accurately determine the scope of the infringement. The New York Times believes that ChatGPT users may delete their history after bypassing the paywall, and therefore it is necessary to conduct large-scale data collection.

Recently, an AI Excel assistant called Shortcut has sparked heated discussions on social media. It enables users to effortlessly complete Excel tasks without writing complex formulas or VBA code through natural language processing (NLP) technology. The AIbase editorial team has compiled the latest information from social media to provide an in-depth analysis of Shortcut's powerful features and its potential impact on the fields of data processing and financial modeling. Shortcut: An Excel Revolution Driven by Natural Language

Large Language Models (LLMs) have achieved significant progress in complex reasoning tasks by combining task prompts with large-scale reinforcement learning (RL), as demonstrated by models like Deepseek-R1-Zero, which directly apply reinforcement learning to base models, showcasing strong reasoning capabilities. However, this success is difficult to replicate across different base model families, especially within the Llama series. This raises a core question: what factors lead to inconsistent performance of different base models during reinforcement learning? How does reinforcement learning perform in