Collection of large models in the first week of June 2023
- 2023.6.5
- Copyright statement: This article is the original article of the blogger chszs, and shall not be reproduced without the permission of the blogger.
1. The open source organization OpenBMB releases the current strongest Chinese language model CPM-BEE
On May 27, 2023, OpenBMB released an open source large language model CPM-BEE with a scale of up to 10 billion parameters. OpenBMB is an open source organization established by the NLP Laboratory of Tsinghua University and the Zhiyuan Research Institute. The model is trained and optimized for high-quality Chinese datasets, and supports both Chinese and English. According to the official test results, its English test level is about equal to LLaMA-13B, and its Chinese ZeroCLUE test result is second only to human level.
Previously, the strongest domestic open source large language model should be ChatGLM-6B , which is also the research result of Tsinghua University and released by the KEG team of Tsinghua University. It has received good responses both at home and abroad, and CPM-BEE is even stronger.
2. China's large-scale model application standards
On May 30-31, 2023, the Hangzhou General Artificial Intelligence Development Forum hosted by China Academy of Information and Communications Technology, Zhejiang Provincial Department of Economy and Information Technology, Hangzhou Municipal People's Government, and China Artificial Intelligence Industry Development Alliance was held in Hangzhou Future Science and Technology City. At the scene, China Academy of Information and Communications Technology issued certificates to the contributors who participated in the compilation of China's large-scale model application standards, including Huawei, Baidu, the Institute of Automation of the Chinese Academy of Sciences, HKUST Xunfei, China Telecom, China Mobile, China Unicom, Alibaba Cloud, Tencent Cloud, Inspur, China Mobile Software, NetEase Smart Enterprise, China Unicom Digital Technology, 360, Baiying Technology, Lanzhou Technology, Shanghai Pudong Development Bank, State Grid, ZTE, Midea, Daguan Data, Ant Group, Jietong Huasheng, TAL Future, Yuanting Technology, Suir Yuan Technology, Hikvision, Feiteng, Sound Intelligence Technology, Wuhan Artificial Intelligence Research Institute, Tianjin University, etc.
3. Official public beta of Alibaba Cloud's "Tongyi Tingwu"
On June 1, 2023, Alibaba Cloud announced the progress of the Tongyi large-scale model, and the new AI product "Tongyi Tingwu" focusing on audio and video content was officially launched. Tongyi Tingwujie can help users complete the transcription, retrieval, summary and organization of audio and video content. During the public beta period, users can receive more than 100 hours of listening and comprehension free transcription time.
Tingwu is an AI assistant for work and study. It aims at audio and video content scenarios with high knowledge-added value, such as meetings, classes, interviews, training, interviews, live broadcasts, watching videos, listening to podcasts, etc. It can use the latest AI technology such as large models Technology quickly refines and precipitates knowledge. Internal test address: https://tingwu.cn/
4. The Huawei version of ChatGPT "Pangu Chat" will be released on July 7
Huawei internally revealed that it will officially release a multi-modal 100-billion-scale large-scale model product directly benchmarked against ChatGPT at the Huawei Cloud Developer Conference on July 7 this year, called "Pangu Chat". For To B/G government and enterprise customers, it may not be open to ordinary users.
5. The price of Microsoft Office 365 AI is exposed, which is 40% more expensive than the original subscription
According to foreign media reports, Jordi Ribas, head of Microsoft's Bing engineering and product department, announced that Bing Chat (Bing Chat) now supports up to 30 exchanges per session and up to 300 exchanges per day.
The subscription price of Microsoft Office 365 AI is exposed, which is $100,000/year/thousand people. According to The Information, at least 100 customers are currently paying up to $100,000 for Microsoft 365 Copilot to provide services to 1,000 subscribers, paying an additional 40% for automation tasks compared to original enterprise subscriptions . Empowered by AI, Microsoft 365 Copilot can assist users in creating, analyzing, and other work, further improving creativity and productivity, and users are willing to pay a higher premium, which leads to an increase in ARPU.
ARPU (Average Revenue Per User) refers to the average revenue per user. In the Internet industry, ARPU is usually used to measure the average income level of users served by a company or product. This metric can help businesses understand their customer value, leading to more effective marketing strategies and improved profitability.
6. AI video editor Dumme: It can turn long videos into short videos in a few minutes
Dumme is a startup that has been using AI to convert longer YouTube videos into shorter ones. The company has hundreds of video creators testing its product and a pre-release waitlist of more than 20,000. Using Dumme is very simple. First, the user pastes a link to a YouTube video, then clicks "generate," and the AI outputs some short videos showing highlights of the ingested content.
Processing time and the number of clips produced will depend on the length of the original video. For example, an hour-long video podcast might take about 20 minutes to process, and you'll receive the edited video about five minutes later. Once complete, creators can download video clips that default to less than 60 seconds and upload them to any platform that supports short-form content, such as YouTube Shorts, as well as others such as Reels or TikTok.
7. Palantir releases artificial intelligence platform AIP
Palantir's AI platform, which allows business and government to use large language models based on their private datasets, will be available to "select customers" this month.
AIP is a platform that allows customers to deploy ChatGPT-like language models in enterprise computer systems and access private data. While tools like ChatGPT are great, they are trained on very generic data. The idea behind AIP is to give the model access to a company's entire dataset, giving it insight into how the company works. The AIP is designed to visualize every element of the business for clients in real time, while setting ground rules and guardrails for what data a model can access and how it should be handled.
The long-term value of an enterprise lies in the application of artificial intelligence, not the development of artificial intelligence. Palantir is uniquely positioned because it has the advantage of being able to train on complex and massive datasets from secretive organizations with sensitive data, such as the CIA.
In one of Palantir's promotional videos, it demonstrates how to use the AIP platform to integrate artificial intelligence into military decision-making, which can automatically generate combat plans, deploy drones, and more.
8. 360 Intelligent Brain Large Model Application Conference is scheduled for June 13
360 official Weibo announced that the 360 Smart Brain Large Model Application Conference will be held in Beijing at 14:00 on June 13.
On April 9 this year, 360 has released a large model of "360 Smart Brain", claiming that "360 Smart Brain" will combine 360 search, smart hardware and other landing applications, and will open internal testing for enterprise users. In fact, except for a few gun articles claiming to apply for an internal test account, there is no news. This time the large model will be released again, it may be true.