Caiyun Technology recently held the 'From Paper to App' themed communication conference in Beijing, officially releasing the general large model Yunjin Tianzhang based on the DCFormer architecture and announcing the upgrade of its AI RPG platform Caiyun Xiaomeng to version V3.5 based on DCFormer. This marks a significant breakthrough in model architecture efficiency in the field of artificial intelligence. In the AI field, the Transformer architecture has always been the core technical support for mainstream large models like ChatGPT and Gemini.

Generating high-quality, temporally continuous videos requires a significant amount of computational resources, especially for longer time spans. The latest Diffusion Transformer model (DiTs) has made significant progress in video generation, but its reliance on larger models and more complex attention mechanisms has resulted in slower inference speeds, exacerbating the challenge. To address this issue, researchers at Meta AI have proposed a training-free method called AdaCache to accelerate video DiTs.

Recent research from Alibaba's Tongyi Laboratory indicates that existing text-to-image Diffusion Transformer models already possess the ability to generate multiple images with specific relational contexts. With just a little 'guidance', they can achieve a synthesis of understanding and produce high-quality multi-image collections. Traditional Diffusion models are more like 'rote learning' students, requiring massive datasets to generate high-quality images. However, with the support of IC-LoRA, they become 'high-achieving students', capable of learning through analogy.

This week, MIT showcased a new model for training robots, aimed at addressing the challenges faced by imitation learning when introduced to small obstacles. Researchers pointed out that imitation learning can fail in conditions such as varying lighting, different environments, or new obstacles because robots simply do not have enough data to adapt. The team sought powerful data methods, like those used in models such as GPT-4, to tackle the problem. They introduced a new architecture called the Heterogeneous Pretrained Transformer (HPT), which aggregates information from different sensors and environments.

Chinese Academy of Engineering academician Wang Jian shared his in-depth thoughts on AI, AI+, and AI infrastructure at the opening ceremony of the 2024 Bund Conference. He emphasized that AI+ is not merely a simple integration of AI and industry, but rather the integration of data, models, and computing power, with cloud computing serving as the infrastructure of the AI era. Wang Jian pointed out that although artificial intelligence has a long history, it has only truly impacted industries for the past 7 years, with the critical turning point being Google’s introduction of the Transformer in 2017. Under the logic of AI+, ChatGPT is not just an application, but a broader application platform.

xAI has released two intelligent models, Grok-2 and Grok-2mini, showcasing significant performance improvements with powerful reasoning, programming, dialogue, and image generation capabilities, surpassing Claude 3.5 and GPT-4 Turbo. The new models excel in solving mathematical problems, news querying, writing cover letters, and game recommendations, and are integrated with FLUX.1 for multitasking. The interface and features on the 𝕏 platform have been updated, offering richer services for Premium and Premium+ users. Enterprise A

The Transformer architecture, a star in the field of artificial intelligence, has led a revolution in natural language processing with its self-attention mechanism at its core. However, when handling long contexts, the resource consumption of self-attention calculations becomes a bottleneck. To address this issue, researchers have proposed the Tree Attention method, which decomposes the calculation tasks through tree reduction to improve efficiency. This method not only reduces communication overhead and memory usage but is also 8 times faster than existing methods in a multi-GPU environment. The introduction of Tree Attention not only enhances performance.

This is an interactive visualization tool specifically designed for non-experts. With this tool, we can directly run the GPT-2 model in the browser and observe in real time how the model 'understands' our text step by step and predicts the next word.

In the field of artificial intelligence, the Transformer model has gained widespread attention for its outstanding performance in language tasks. The recent research paper 'Transformer Layers as Painters' explores the hierarchical structure of the Transformer model from an innovative perspective, comparing each layer to a painter that creates complex and rich texts on the canvas of language. The study reveals the working mechanism of Transformer layers through experiments, particularly how they collaborate with each other.

Zhipu AI announces the open source of its video generation model CogVideoX, aiming to promote the rapid development and application of video generation technology. The CogVideoX model is based on advanced large model technology and is capable of meeting commercial-grade application requirements. The currently open sourced CogVideoX-2B version requires only 18GB of GPU memory for inference at FP-16 precision and 40GB for fine-tuning, allowing inference with a single 4090 GPU and fine-tuning with a single A6000 GPU.

Scientists at ShanghaiTech University have developed CLAY, an AI model that can generate detailed 3D objects from text or 2D images, outperforming earlier technologies in quality, diversity, and generation speed. At its core, CLAY includes a multi-resolution variational autoencoder and a diffusion transformer, allowing direct processing of 3D content without conversion. By training on over 500,000 3D models, CLAY can produce objects ranging from everyday items to complex creatures and achieve precise control through additional inputs, generating urban scenes and reconstructing 3D models from sketches.

MusiConGen模型是在预训练的MusicGen-melody框架基础上进行微调的，主要用于生成各种风格的音乐片段。研究团队通过设置和弦和节奏的控制参数，展示了模型生成的音乐样本，涵盖了五种不同风格:休闲布鲁斯、平滑酸爵士、经典摇滚、高能放克以及重金属。

Recently, Apple's research team and Meta AI researchers have jointly introduced a new technology called LazyLLM, which aims to enhance the efficiency of large language models (LLMs) in long-text reasoning. As everyone knows, current LLMs often face slow processing speeds when dealing with long prompts, especially during the pre-filling phase. This is primarily due to the modern transformer architecture's computational complexity, which grows quadratically with the number of tokens in the prompt

In the realm of artificial intelligence, the competition between open-source and closed-source models has never ceased. The recent release of Meta AI's Llama 3.1 model, however, seems to have drawn a watershed line in this contest. This is not merely the launch of a new model; it is a sign of the maturation of open-source AI, heralding the arrival of a new era.Llama 3.1, developed by Meta AI's team, is a new generation of large-scale language models. In over 150 benchmark tests, its 405B paramet

DeepSeek has open-sourced the RWKV-CLIP model, a visual-language representation learner that combines the strengths of both Transformer and RNN architectures. The model has significantly enhanced performance on vision and language tasks by pre-training on image-text pairs sourced from web data, expanding its dataset with information obtained from various websites.To address the issue of noisy data and improve data quality, the research team introduced a diverse description generation framework.

The groundbreaking Transformer acceleration technology FlashAttention-3 has been launched! This is not just an upgrade; it heralds a direct increase in the inference speed of our Large Language Models (LLMs) and a direct decrease in costs! Let's talk about FlashAttention-3 first, which is a significant improvement over its predecessors: Significant increase in GPU utilization: Training and running large language models with FlashAttention-3 doubles the speed, up to 1.5 to 2 times faster, and thi

The Million Experts Mixture model proposed by Google DeepMind, a revolutionary study that has taken a significant step forward in the Transformer architecture. Imagine a model capable of sparse retrieval from a million mini-experts - doesn't that sound a bit like a science fiction novel plot? Yet, this is the latest research achievement from DeepMind. The core of this research is a highly parameter-efficient expert retrieval mechanism, which separates the computation cost from the parameter coun