英特尔公布了旗下生成式AI大模型Aurora genAI，该模型参数量高达1万亿，是ChatGPT的近6倍（参数量1750 亿），依赖于Megatron和DeepSpeed框架，这些结构增强了模型的强度和容量。
Meta在GitHub上再次开源了一款全新的AI语言模型——Massively Multilingual Speech (MMS，大规模多语种语音)，这款新的语言模型可以识别4000多种口头语言并生成1100多种语音（文本到语音）。上线短短的几个小时，在GitHub库便收获了23kStar，Fork数量高达5.5k。
来自斯坦福&普林斯顿大学学者联合发布的一篇名为【Siamese Masked Autoencoders】的论文中，发表了一种用于视频学习的蒙面自动编码器，SiamMAE用于从视频中学习视觉对应关系，可以在没有显式标签或注释的情况下使得机器进行自主学习。学习到的表示可以用于视频分类、动作识别或对象跟踪等下游任务。
Source: Wall Street News
1. Wall Street has underestimated cross-industry demand for generative AI. Nvidia's Q2 performance guidance exceeded Wall Street's expectations by 53%, and the AI supercomputing system will bring more excess revenue to Nvidia;
2. Intel announced the Aurora GenAI big model. The parameters are nearly 6 times that of ChagPT. The multi-specification parameter model is the “optimal solution” to balance cost and efficiency;
3. Meta also open source a multi-lingual voice model MMS that can recognize 4,000 languages and generate 1,000 voices;
4. Scholars such as Stanford announced major breakthroughs in machine vision tracking, and SiamMae can save high costs;
An anecdotal perspective
1. Wall Street has underestimated the cross-industry demand for generative AI, and the AI supercomputing system will bring more excess revenue to Nvidia.
Nvidia's Q1 earnings report far exceeded expectations and became the focus of attention today. More importantlyThe Q2 performance guide will be the company's highest quarterly revenue ever, exceeding Wall Street's estimate of 53.2%.
Behind Inventec's performance exceeding expectations, the biggest contribution came from the data center business among the four major businesses. Revenue reached a record high of 4.28 billion US dollars, up 14% year on year and 18% month on month. Furthermore, although the automobile business accounts for a relatively small share, it has also reached a month-on-month growth rate. And demand for gaming and professional visualization has clearly not returned to the same level as last year.
Insightful research suggests that:The market has far underestimated the demand for generative AI, and the AI supercomputing system will bring more excess revenue to Nvidia.
Nvidia's data center revenue surge is mainly due to growing demand for generative AI and big language models, which has driven the company's demand for GPUs based on Hopper and Ampere architectures to exceed expectations. Currently, the visibility of demand for data center products has been extended for as many as several quarters,The supply of H100 will increase further in the second half of the year.
Judging from customer needs:Cloud computing service providers, consumer network companies, and enterprise customers all want to apply generative AI to existing businesses as soon as possible. However, judging from the order schedule situation, GPUs will be in short supply throughout the year.As a result, it is expected that Nvidia's revenue will continue to benefit from data center business growth driven by generative AI.
Generative AI is a disruptive entity for even more industries, and it is in the beginning stage from scratch, and the value space it can create is very imaginative.
According to Gartner's forecast, by 2025, the proportion of new drugs and materials developed using generative AI technology systems will rise from 0% now to 30% +, and this is just one of its many industry use cases. In addition, generative AI technology also brings new value to many fields such as chips, component design, and synthetic data.
What is worth paying attention to, generative AI is driving exponential growth in computing demand and a rapid transition toNvidia's Accelerated Computing. The company also said it will start selling AI supercomputing systems to technology companies that pay more premiums.
Currently, Nvidia has advantages in terms of high-performance networks, and will further optimize the computing structure, memory usage, and communication efficiency and speed, while improvingThe need for high performance switches, optical modules, and light.
2. Intel announced Aurora GenAI, a big AI model with parameters nearly 6 times that of ChatGPT. The multi-specification parameter model is the “optimal solution” to balance cost and efficiency.
Intel announced its large generative AI model, Aurora GenAI. The number of parameters in this model is as high as 1 trillion, nearly 6 times that of ChatGPT (175 billion parameters). It relies on Megatron and DeepSpeed frameworks. These structures enhance the strength and capacity of the model.
The Aurora GenAI model is a purely science-centered generative AI model mainly used for scientific research; running on the Aurora supercomputing developed by Intel for the Aragon National Laboratory, its performance reached 20 billion times, double that of Frontier, the current TOP500 supercomputing champion.
Insightful research suggests that:As a strong contender for ChatGPT, Aurora GenAI's announcement heralds the arrival of new big players on the AI big model circuit and is likely to have a major impact on various fields of science in the future.
Also worth paying attention to is,The development of LLM models will continue to expand training parameters, but the operation of larger models will inevitably incur higher costs. Currently, how to balance effective demand and cost is an issue worth focusing on for big model developers. To prevent models from generating unnecessary operating costs due to parameter redundancy when applied, it is necessary to prevent models from generating unnecessary operating costs due to parameter redundancy, so it will become an inevitable development path for large model developers to launch diversified parameter models for specific fields.
1. Meta is also open source a multi-lingual voice model MMS that can recognize 4,000 languages
Meta has once again open-sourced a new AI language model — Massively Multilingual Speech (MMS, Massive Multi-lingual Speech) on GitHub. This new language model can recognize more than 4,000 spoken languages and generate more than 1,100 kinds of speech (text-to-speech). In just a few hours of launch, the GitHub library received 23 kstars, and the number of forks reached 5.5k.
Meta is seen as a dark horse for big model development. It is the originator of the open source big model. LLama, which was released earlier, is the general model that has been fine-tuned the most. Previously, the company's SAM visual model dropped a heavy bomb in the CV field. The company's continued efforts in the open source multi-modal field further provided strong technical support to the open source community.
2. Scholars at Stanford & Princeton University discover new computer vision tracking technology that can save high costs
Establishing correspondence between images or scenes is an important challenge in computer vision, especially considering occlusion, changes in perspective, and changes in the appearance of objects.
In a paper called “Siamese Masked Autoencoders” jointly published by Stanford & Princeton University scholars, a masked autoencoder for video learning was published. SiamMae is used to learn visual correspondences from videos, which allows machines to learn independently without explicit tags or comments. The learned representation can be used for downstream tasks such as video classification, motion recognition, or object tracking.
The features learned through SiamMae performed excellently in self-supervised methods such as video object segmentation, posture keypoint propagation, and semantic partial propagation. This method is particularly useful when labeling data is scarce or data acquisition is expensive.