Abstract: Alibaba's Intelligent Information Business Group officially released the self-developed Quark large model with a trillion-level parameters, which will be applied in general search, medical health, education learning, workplace office, and other scenarios. Nvidia recently revealed that the Blackwell architecture B100 GPU, to be launched in 2024, outperforms A100, H100, and H200 in GPT-3 175B inference performance, with its AI performance being more than twice that of the Hopper architecture H200 GPU.
Image Source: Generated by Wujie AI
Hot News:
Nvidia: B100 to be Launched Next Year, AI Performance More Than Twice That of H200
According to "Science and Technology Innovation Board Daily" on November 14, Nvidia recently revealed that the Blackwell architecture B100 GPU, to be launched in 2024, outperforms A100, H100, and H200 in GPT-3 175B inference performance, with its AI performance being more than twice that of the Hopper architecture H200 GPU. It is expected that Nvidia will commission TSMC to manufacture the Blackwell GPU using 3nm process technology, and Nvidia plans to advance the production timeline to Q2 2024. The Blackwell GPU will be Nvidia's first HPC/AI accelerator using chiplet design.
Alibaba's Intelligent Information Business Group Releases Trillion-Parameter Quark Large Model
According to Caixin on November 14, Alibaba's Intelligent Information Business Group officially released the self-developed Quark large model with a trillion-level parameters, which will be applied in general search, medical health, education learning, workplace office, and other scenarios. The Quark app will also be comprehensively upgraded with this large model.
OpenAI CEO: Next-Generation AI Model GPT-5 Under Training, Requires More Data
According to AI New Intelligence on November 14, OpenAI CEO Sam Altman recently revealed in an interview with FT the company's further plans. Altman stated that the company is developing the next-generation AI model GPT-5. However, he did not commit to a release schedule. Altman mentioned that this will require more data for training, which will come from publicly available datasets on the internet and the company's proprietary data. OpenAI recently issued a call for large-scale datasets, especially those that are "not easily accessible on the internet today," particularly long-form writing or any format of dialogue.
Altman also mentioned that, like most other large AI companies, OpenAI uses Nvidia's advanced H100 chips to train its models. He noted that due to the shortage of Nvidia's chip supply, there has been a "serious tension" throughout this year. However, with Google, Microsoft, AMD, and Intel preparing to release competing AI chips, the reliance on Nvidia may not continue for too long.
Furthermore, despite OpenAI's consumer success, Altman stated that the company is seeking progress in building artificial general intelligence. He believes that large language models (LLMs), which support ChatGPT, are "one of the core parts of building AGI, but there will be many other parts on top of it." He also emphasized the importance of language as a means of information compression, which he believes is a factor that companies like Google DeepMind have overlooked.
Google and UC Berkeley Jointly Introduce New Generative AI Method "Idempotent Generative Network," Capable of Generating Realistic Images in a Single Step
According to IT Home on November 14, Google recently collaborated with the University of California, Berkeley to develop a new generative AI method called "Idempotent Generative Network (IGN)," which can replace diffusion models.
Current mainstream generative AI models, including Generative Adversarial Networks (GAN), diffusion models, and the consistency models released by OpenAI in March, all map inputs such as random noise, sketches, low-resolution, or other corrupted images to outputs corresponding to the given target data distribution (usually natural images). For example, in the case of diffusion models, they learn the target data distribution during training and then perform "denoising" through multiple steps.
The Google research team proposed a new generative model called the Idempotent Generative Network (IGN), which can generate suitable images from any form of input, ideally in just one step. This model can be imagined as a "global projector" that projects any input data onto the target data distribution, unlike other existing model algorithms, which are not limited to specific inputs.
Lee Kai-Fu's AI Company "Zero One Everything" Open Sources Yi Large Model, Accused of Plagiarizing LLaMA
According to IT Home on November 14, Lee Kai-Fu, Chairman and CEO of Innovation Works, founded the AI large model startup "Zero One Everything" this year. The company has released two open-source large models, Yi-34 B and Yi-6 B, claiming to be completely open for academic research and simultaneously open for free commercial applications. However, on the Hugging Face open source page for Yi-34 B, developer ehartford questioned that the model used the architecture of Meta LLaMA, only modifying the names of two tensors, input_layernorm and post_attention_layernorm.
In addition, there is a circulating post from a friend of Jia Yangqing, former Chief AI Scientist of Alibaba, stating that "a certain domestic company's new model is exactly the architecture of LLaMA, but to show that it is different, they changed the names in the code from LLaMA to their own name, and changed a few variable names."
Zero One Everything Responds to Controversy: The developed large model is based on the mature structure of GPT, and a lot of work has been done to understand the model and training
According to "Science and Technology Innovation Board Daily" on November 14, Zero One Everything, a company under Lee Kai-Fu, was accused of completely using the architecture of LLaMA for the open-source large model, only modifying the names of two tensors. In response, Zero One Everything stated: GPT is a recognized mature architecture in the industry, and Llama summarized it on GPT. The structural design of the large model developed by Zero One Everything is based on the mature structure of GPT, drawing on the industry's top-level public achievements. Since large model technology is still in its very early stages, maintaining a consistent structure with the industry mainstream is more conducive to overall adaptation and future iterations. At the same time, the Zero One Everything team has done a lot of work to understand the model and training, and is also continuously exploring breakthroughs at the model structure level.
Zhang Yueguang, Product Manager of Miaoyaduck Camera, Resigns, and the Popularity of the Red-Hot AI Application Cools Down
According to Tech Planet on November 13, Zhang Yueguang, the product manager of Miaoyaduck Camera under Alibaba's Great Entertainment, has resigned. He was involved in the planning of Alipay's Spring Festival "Collect Five Blessings" and "Flick One Flick" projects, and has worked at ByteDance, Alibaba, and other companies.
Miaoyaduck Camera is an AIGC product that has gained popularity at the consumer level. By uploading more than 20 photos containing faces, paying 9.9 yuan, and then choosing favorite templates and styles, users can create their "digital avatars" and obtain portrait works. Miaoyaduck Camera dominated the application product rankings for a period of time, but its ranking subsequently declined. As of November 13, the latest data from Qimai shows that Miaoyaduck Camera is ranked 64th in the iOS "Social" list. Currently, the market's long-term demand for AIGC products has become a difficult issue at the market level.
AI Startup Silo AI Launches Open-Source Language Model "Poro" for Europe, Covering 24 EU Languages
According to VentureBeat, Silo AI, an artificial intelligence startup headquartered in Helsinki, Finland, released a new open-source large language model "Poro" this week, aiming to enhance the multilingual artificial intelligence capabilities of European languages. Poro is the planned first open-source model designed to eventually cover all 24 official languages of the EU. These models were developed jointly by Silo AI's SiloGen Generation AI Department and the TurkuNLP research group at the University of Turku.
The Poro 34B model has 342 billion parameters and is named after the Finnish word for "reindeer." It was trained on a partition of a 210 trillion token multilingual dataset covering English, Finnish, as well as programming languages such as Python and Java.
Peter Sarlin, CEO of Silo AI, stated that the design purpose of Poro is to address the core challenge of training excellent performance for low-resource languages in Europe, such as Finnish. By using cross-lingual training methods, the model can leverage data from high-resource languages, such as English.
Rakuten Group Collaborates with OpenAI to Launch Rakuten AI for Business Platform
AI New Intelligence, November 14th - Rakuten Group announced a strategic partnership with OpenAI and launched the new artificial intelligence platform Rakuten AI for Business. The platform supports various core business functions including marketing, sales, customer support, operations, strategic planning, and engineering. It is currently available by invitation only, with plans to expand its services in 2024 and beyond. Research: AI companies are facing a crisis of training data depletion, with high-quality data expected to be exhausted by 2026. According to an article by Rita Matulionyte, a professor of information technology law at Macquarie University in Australia, AI researchers have been sounding the alarm about the shortage of data supply for the past year. A study by the Epoch AI Artificial Intelligence Forecasting Organization last year suggested that AI companies may deplete high-quality text training data before 2026, while the exhaustion of low-quality text and image data may occur between 2030 and 2060. For data-thirsty AI companies, using synthetic data generated by AI models for training may not be a feasible solution. Research indicates that training AI models using AI-generated content may lead to inherent distortions in the model, resulting in chaotic and bizarre outputs. Faced with this potential problem, the solution may lie in establishing data partnerships, where companies or institutions with rich, high-quality data reach agreements with AI companies to exchange data for funding. Vietnam's technology unicorn company VNG plans to launch an AI service similar to ChatGPT tailored for Vietnamese language users. Supported by Tencent and Alibaba's Ant Financial Group, the technology unicorn company, which already has a chat app more popular than Facebook in its local market and recently added translation functionality, aims to add AI-generated capabilities for tasks ranging from writing emails to finding query answers. Reports suggest that OpenAI is attempting to lure AI talent from Google with compensation packages of up to $10 million. The company is engaging in a talent war with Google, offering high salaries and top-notch technical resources such as AI accelerator chips for running tests to attract some of Google's finest researchers. AIGC Recommended Reading: - "The Most Powerful Model Training Chip H200 Released! 141GB Large Memory, AI Inference Up to 90% Increase, Compatible with H100" - "Musk's ChatGPT 'Grok,' How Does It Work?" For more information, visit: - https://www.aixinzhijie.com/article/6838346 - https://www.aixinzhijie.com/article/6838336
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
