Author: Ziyu
Editor: Shen Xiao
OpenAI released ChatGPT Enterprise to provide greater security and privacy
On August 28, OpenAI announced that it will launch ChatGPT Enterprise Edition, which specifically addresses the needs of enterprise users, provides a higher level of security and privacy, and the conversation data will not be used for model training, and it will perform twice as fast as the regular version of GPT-4. OpenAI says that ChatGPT Enterprise not only provides unlimited quick access to powerful GPT-4 models, but also enables deeper data analysis to help businesses quickly understand information and can ask more complex questions to ChatGPT.
https://www.ithome.com/0/715/406.htm
GitHub Copilot updates three major features
On August 28, GitHub Copilot welcomed three major updates, some of which are only available in Visual Studio and some of which only support VScode. New interactive experience, ask questions directly in the line of code, no need to switch chat windows, users can directly use Copilot Chat in the code editor to optimize the code; The official has officially launched 8k context windows for all code completion requests; Slash command improvements, context extension to terminals, new conversation history, and support for quick slash commands.
https://www.qbitai.com/2023/08/80132.html
NVIDIA introduces DLSS 3.5, which leverages new AI models to create higher quality ray-traced images
On August 28, NVIDIA launched NVIDIA DLSS 3.5, which once again promoted the development of rendering technology. The innovation lies in the use of an enhanced AI-powered neural renderer. This is the use of a new AI model that utilizes Ray Reconstruction technology to create higher-quality ray-tracing images for intensive ray-tracing games and applications. Moreover, because the light reconstruction technology itself does not rely on hardware, the previous RTX series graphics cards can also enjoy this technology upgrade.
https://36kr.com/p/2408488309203971
NVIDIA and Meta access SK hynix seek AI servers
News on August 29, Meta executives recently visited SK hynix headquarters in Icheon, South Korea, inspected DDR5, HBM production facilities and EUV equipment in Icheon Park, and conducted quality audits of DDR5. Meta is one of SK Hunix's main customers, and has invested heavily in AI servers, requiring SK Hynix to also supply high-performance and high-efficiency DDR server DRAM. In addition, NVIDIA is expected to visit SK Hynix's Lichuan campus this week and is expected to review SK Hynix's fifth-generation HBM (HBM3E) production line. The report pointed out that in view of the fact that NVIDIA has considered significantly increasing the production of AI chip H100, the two sides may discuss the supply of HBM.
https://finance.ifeng.com/c/8Sdy5LrLHN2
Google trains the next generation of large model Gemini on TPUv5
News on August 29, it is said that the training scale of Google's large model Gemini has reached 5 times that of GPT-4. The original version of Gemini may have been trained on TPUv4, using 14 pods, training FLOPS up to 1e26, while GPT-4's FLOPS was only about 2e25. The next-generation version of Gemini has begun training on TPUv5 with up to 1e26FLOPS, which is 5 times that of GPT-4. Gemini has been iterating quickly on TPUs, first TPUv4, now upgraded to TPUv5, and its performance is also better than GPT-4.
https://www.chinaz.com/2023/0829/1554861.shtml
Google unveils a new supercomputer powered by an Nvidia H100 "Hopper" GPU
News on August 29, Google Cloud announced at the Cloud Next Conference that it will launch A3 virtual machine instances next month. The biggest highlight of this instance is that it is equipped with NVIDIA H100 Tensor Core GPU to meet the needs of generative AI and large language models. The A3 instance uses 4th generation Intel Xeon Scalable processors, 2TB DDR5-4800 memory, and 8 NVIDIA H100 "Hopper" GPUs, achieving 3.6TB of bisecting bandwidth via NVLink 4.0 and NVSwitch. The supercomputer "is specifically designed to train and service the most demanding AI models that power today's generative AI and large-scale language model innovations."
https://techcrunch.com/2023/08/29/googles-new-a3-gpu-supercomputer-with-nvidia-h100-gpus-will-be-generally-available-next-month/
HUAWEI CLOUD and CNKI will jointly build a large AI model for the knowledge service industry
On August 30, at the HUAWEI CLOUD Pangu Model Forum, Tongfang CNKI Digital Publishing Technology Co., Ltd. (hereinafter referred to as "Tongfang CNKI") and Huawei Cloud Computing Technology Co., Ltd. (hereinafter referred to as "HUAWEI CLOUD") officially signed a cooperation agreement between the China Knowledge Model (hereinafter referred to as the "Huazhi Model") and the Artificial Intelligence Joint Innovation Lab. Based on the HUAWEI CLOUD Pangu model, the two parties will build an AI model for the knowledge service industry.
https://news.sina.com.cn/sx/2023-08-30/detail-imziyppc7320069.shtml
360 released the 2023 semi-annual report, and the "360 Intelligent Brain" large model has generated nearly 20 million yuan
News on August 30, 360 released its 2023 half-year report, with a total revenue of 4.503 billion yuan, of which the security business achieved revenue of 908 million yuan. The 360 security model has been applied in 360's internal and own products, and has created nearly 20 million yuan of related business revenue. 360 Security Big Model is 360's self-developed cognitive general artificial intelligence big model, with a number of core capabilities, has launched enterprise-level AI large model solutions, and launched industry large models with partners to promote the development of the large model industry.
https://finance.sina.com.cn/jjxw/2023-08-30/doc-imziytuz7224758.shtml
The former CEO of Google launched the AI+Science moonshot program, modeled on OpenAI
News on August 30, former Google CEO Eric Schmidt is building a large AI+Science non-profit startup, hoping to solve challenges in scientific research through AI. Schmidt hired Samuel Rodriques, founder of the Francis Crick Institute's Applied Biotechnology Laboratory, and Andrew White, a professor at the University of Rochester and a pioneer in the use of AI in chemistry. According to people familiar with the matter, the work is modeled on OpenAI. Rodriques said companies will need teams of core AI researchers and core scientists to work together and adopt rapid iteration cycles to build tools that can leverage cutting-edge technologies and bring real value to scientists.
https://36kr.com/p/2410198303040258
Google has partnered with General Motors to bring conversational AI to its vehicle segment
News on August 30, General Motors announced on its official website that the company has cooperated with technology giant Google to explore opportunities to use AI in cars. It is understood that GM will have a "broader cooperation" with Google Cloud around generative AI, after the two sides have already targeted GM's OnStar IVA GM said that the IVA system will be powered by Google-provided "intent-recognition algorithms" to provide OnStar users with responses to common queries, as well as route and navigation assistance.
https://www.ithome.com/0/715/870.htm
Google introduced SynthID, which embeds invisible watermarks in AI-generated images
News on August 30, in response to the growing problem of AI-generated images, Google's DeepMind team today launched a tool called SynthID, which can embed a watermark invisible to the human eye in AI-generated images, but can be detected with specialized artificial intelligence detection tools to distinguish between real and false content. The tool is already available to select customers on the Google Cloud Platform.
https://www.ithome.com/0/715/671.htm
Google plans to promote AI search in India and Japan
On August 31, Google plans to promote its artificial intelligence-driven search generation experience (SGE) program in India and Japan, and the feature is still in the stage of invitation. SGE is designed to provide users with curated answers, rather than traditional web page lists. Google's system differs from Microsoft in that it integrates artificial intelligence directly into the existing search bar. The feature will provide multilingual versions and voice input capabilities, and continue to display ads on the search results page.
https://www.chinaz.com/2023/0831/1555322.shtml
Google launches AlloyDB AI to help developers create generative AI applications
On August 31, Google issued a press release announcing the launch of the AlloyDB AI service, which allows developers to combine the functions of large language models and real-time manipulation data to provide full support for vector embedding. AlloyDB AI uses simple SQL functions to convert data into vector embeddings, and it can execute vector queries ten times faster than standard PostgreSQL.
https://www.ithome.com/0/715/950.htm
Baidu, SenseTime and other 11 domestic large model products have passed the filing and will be opened online one after another
On August 31, 11 domestic AI large model products filed through the "Interim Measures for the Management of Generative Artificial Intelligence Services" will be launched one after another, open to the whole society. The first batch of lists include Baidu's "Wenxin Yiyan", Douyin's "Skylark", Baichuan Intelligence's "Baichuan Big Model", Zhipu Huazhang's "Zhipu Qingyan", the Chinese Academy of Sciences' "Zidong Taichu", SenseTime's "Discussion SenseChat", MiniMax's "ABAB Big Model" and Shanghai Artificial Intelligence Laboratory's "Scholar General Large Model". In addition, Huawei, Tencent, and iFLYTEK are also about to open their large-model products. Ali's "Tongyi Qianwen" has completed the filing and is waiting to go online.
https://36kr.com/p/2411653492204549
Huawei Pangu Large Model Edition Xiaoyi opened crowd testing recruitment
On August 31, Huawei began to recruit a new Xiaoyi crowd testing experience officer. The first models to support crowdtesting include the MATE 60/P60 series. The new Xiaoyi can better understand users and provide more accurate feedback, including native digital content search, dialogue recognition screen content, help users quickly summarize, copywriting content assistance, image secondary creation and other functions.
https://www.chinaz.com/2023/0831/1555383.shtml
"Discuss SenseChat" ranked first in 12 performances, suspected to be close to GPT-4
On August 31, SenseTime's natural language application "SenseChat" was officially opened to users. Its base model is InternLM-123B released by SenseTime and a number of top domestic scientific research institutions, ranking second in the world in terms of test results on 51 well-known evaluation collections (including MMLU, AGIEVAL, ARC, CEval, Race, GSM8K, etc.) with a total of 300,000 questions. Surpassing GPT-3.5-turbo and Meta's newly released LLaMA2-70B models, it surpassed GPT-4 in 12 major evaluations and ranked first.
https://www.sensetime.com/cn/news-detail/51167247?categoryId=72
Meta may be working on Llama 3, which is said to be more powerful than GPT-4 and free and open source
On August 31, Jason Wei, an OpenAI engineer and former Google Brain engineer, learned at the Generative AI Group social event organized by Meta that Meta may be developing a more powerful AI model Llama3 than GPT-4 and keeping it open source and free. Upgrading from Llama2 to Llama3 can be more challenging than simply scaling with more training, and may take longer than jumping from Llama1 to Llama2. Llama2 has reached GPT-3.5 levels in some applications and is being optimized by the open source community through fine-tuning and additional features.
https://tech.ifeng.com/c/8ShQFq6NXdN
Meta Releases AI Benchmarking Tool FACET to Assess the "Fairness" of AI Models
On August 31, Meta released a new AI benchmark, called FACET, which aims to assess the "fairness" of AI models that classify and detect people and objects in photos and videos. The FACET consists of 32,000 images containing 50,000 individual tags labeled by human annotators. In addition to occupation- and activity-related categories, such as "basketball player," "DJ," and "doctor," demographic and physical attributes are included, allowing Meta to make a "deep" assessment of biases against these categories.
https://techcrunch.com/2023/08/31/meta-releases-a-data-set-to-probe-computer-vision-models-for-biases/
Baijia Cloud released a one-stop AIGC scenario-based platform - Baiwen Big Model
On August 31, Baijia Cloud released the one-stop AIGC scenario-based platform "Baiwen Big Model" at the strategy conference. Based on the technology accumulation and R&D strength of Baijia Cloud, the platform ensures the security of enterprise data and can access various AIGC services at low cost, solving the difficulties of enterprise application model technology. The platform provides a variety of data security audit methods, supports zero-code rapid application generation, centralized management and control of enterprise knowledge base, meets the requirements of information creation, and supports independent and controllable private deployment. Baijia Cloud also showcased the application cases of the platform in scenarios such as intelligent large model AI customer service, official document writing, personal AI assistant, digital human interactive live broadcast and marketing copywriting assistance.
https://baijiahao.baidu.com/s?id=1775736573951503811
NVIDIA introduces FlexiCubes, which uses generative AI to create high-quality 3D meshes
On September 1, Nvidia launched a new way to use AI to generate 3D meshes, FlexiCubes, which can generate 3D meshes for a wide range of 3D applications. FlexiCubes can be integrated with physics engines and can easily create soft objects in 3D models. This innovative approach to grid generation promises to transform the AI pipeline, delivering high-quality meshes for a variety of applications.
https://www.chinaz.com/2023/0901/1555781.shtml
X (Twitter) adjusts its privacy policy to train AI models on information posted by users
News on September 1, Elon Musk's social platform X (Twitter) recently adjusted its privacy policy, allowing X to use the information posted by users to train its artificial intelligence (AI) model. The new privacy policy will take effect on September 29. The new policy states that X may use the platform information collected and publicly available information to help train X's machine learning or artificial intelligence models.
https://www.ithome.com/0/716/397.htm
OpenAI publishes AI teaching guides
On September 1, OpenAI released an official guide called "Teaching with AI" to help teachers use ChatGPT more effectively in the classroom. OpenAI's "Teaching Artificial Intelligence" guide explains how ChatGPT works, its limitations, the role of detectors and how to avoid bias, and provides a practical set of instructions for teachers to interact with students.
https://www.chinaz.com/2023/0901/1555747.shtml
Seven top universities including Hong Kong Chinese and Tsinghua proposed a new CSV method, and the accuracy rate of GPT-4 MATH rose to 84.3%
Recently, models such as GPT-4 and PaLM-2 have demonstrated superior performance in mathematical reasoning. In order to deeply study the impact of code generation on language model reasoning, researchers from seven top universities including Hong Kong Chinese and Tsinghua University jointly published a paper to explore the impact of code usage frequency on model performance. The study found that the success of GPT-4 Code Interpreter is due in part to its ability to generate, evaluate, and correct code. Based on this, the researchers proposed a code-based self-validation (CSV) method, which further improved the mathematical reasoning performance of GPT-4. By combining Code Interpreter and CSV methods, GPT-4 improved its accuracy on the MATH dataset from 53.9% to 84.3%.
https://36kr.com/p/2413040960136198
Tencent will release the hybrid model in early September, and Tencent Documents and Tencent Games have been put into testing
News on September 2, a few days ago, China Securities News learned from Tencent that Tencent will officially release the mixed element model at the 2023 Tencent Global Digital Ecology Conference held in early September, but did not mention whether it is open to the public. Tencent previously said that the hybrid model has entered the company's internal application testing stage, and a number of Tencent internal businesses and products such as Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Fintech, Tencent Meeting, and Tencent Documents have been connected to Tencent Mixed Element Model for testing and achieved preliminary results.
https://www.ithome.com/0/716/485.htm
The large model of domestic Chinese-English bilingual voice dialogue was exposed, and Kai-Fu Lee participated in the research
On September 2, a paper on speech-text multimodal large models appeared in arXiv, which proposed a Chinese-English bilingual commercial dialogue model LLaSM that supports both recording and text input. The study, from LinkSoul.AI, Peking University and 010,000 Things, has been open-sourced. LinkSoul.AI is an AI startup that previously launched the first open-source Llama 2 Chinese language big model. According to the researchers, LLaSM is the first open-source commercially available dialogue model that supports bilingual speech-text multimodal dialogue in Chinese and English.
https://www.qbitai.com/2023/09/80972.html
Baidu Input Method launched the "Super Write" AI assistant, which is based on the Wen Xin One Word model
News on September 3, Baidu Input Method recently released the 11.7.19.9 Android version, which is an important update after Baidu announced that the first batch of "Wen Xin Yiyan" was approved to open to the whole society. The highlight of this update is that Baidu Input Method has added a full-scene AI creation assistant "Super Write", which is based on the Wenxin large model, which can help users create text with efficiency, high quality and high emotional intelligence in various situations.
https://k.sina.com.cn/article_6678407719_18e106a27001012ld8.html
The co-founder of DeepMind says OpenAI is secretly training GPT-5
On September 3, DeepMind co-founder Mustafa Suleyman, now the CEO of Inflection AI, revealed in an interview that OpenAI is secretly training GPT-5, which is 100 times larger than the current model. His company, Inflection AI, is building one of the world's largest supercomputers, and he thinks in the next 18 months, they could run a training run that is 10 or 100 times larger than the language model training run that made GPT-4.
https://www.ithome.com/0/716/586.htm
![Strong ranking evaluation, 360 Wisdom Brain, which is firmly in the first echelon, has begun to make money[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)