The 2023 Chinese Gonggong Intelligence Conference opened in Fuzhou
The 2023 Chinese Gonggong Intelligence Conference opened at the Fuzhou Digital China Convention and Exhibition Center. With the theme of "Digital Intelligence Leading the Way and Building the Future Together", this conference aims to focus on the research progress and academic frontiers in the field of artificial intelligence at home and abroad, and provide new inspiration and ideas for the breakthrough of basic theory of artificial intelligence, key technological innovation and industrial application.
Biden on artificial intelligence: emphasizing the three principles of safety, security and trust
Biden met with the heads of 7 leading technology companies in the field of artificial intelligence in the United States. These 7 companies include the four Internet technology giants of Microsoft, Google, Amazon Web Services, and Meta, as well as OpenAI (the parent company of ChatGPT), Anthropic, and Inflection AI, three fast-growing AI startups. Sign the Voluntary Commitment to AI and commit to developing AI responsibly.
The UN Security Council held its first meeting on artificial intelligence, and China put forward five principles of governance
The UN Security Council held a high-level open meeting on the theme "Opportunities and risks brought by artificial intelligence to international peace and security". Security Council Secretary-General António Guterres has called for the creation of a global regulatory body to oversee the new technology. This is the first time that the Security Council has held a meeting on artificial intelligence, and Zhang Jun, China's permanent representative to the United Nations, proposed five principles on artificial intelligence governance: first, adhere to ethics first, second, adhere to safety and controllability, third, adhere to fairness and inclusiveness, fourth, adhere to openness and inclusiveness, and fifth, adhere to peaceful use. A legally binding resolution is still far away.
Google's founders returned to build a more powerful model
One of Google's founders and one of the pioneers of AI technology, Sergey Brin, is about to return to Google with the goal of building AI models that are more powerful than ever before. In an interview, Sergey Brin expressed his motivation for returning: "My passion for artificial intelligence has never waned. As technology continues to evolve, I believe we can create models that are more powerful and intelligent than they are today. This moment is the perfect time for me to return to Google. ”
Apple was exposed to quietly develop a large model, creating a chatbot "Apple GPT"
Apple has built its own framework to create large-scale language models, called Ajax, and on top of that it has also created a chatbot service that some engineers call "Apple GPT." People familiar with the matter also said that in recent months, the advancement of artificial intelligence has become an important work for Apple, and multiple teams have collaborated on the project, mainly to solve potential privacy issues related to artificial intelligence technology.
Meta started the turnaround battle with the "free version of ChatGPT" and pulled Microsoft on board
Following LLaMA open source, Meta today teamed up with Microsoft to open source Llama 2, with three versions: 7B, 13B, and 70B. Llama 2 was trained on 2 trillion tokens with a context length of 4k, which is twice as long as Llama 1. Fine-tuning models have been trained on more than 1 million human labels. Llama 2's performance is a second-to-second beat to many open source language models, achieving SOTA in reasoning, coding, ability, and knowledge testing. Most importantly, this time Llama 2 can not only be studied, but even commercially available for free!
Musk founded a new company to return to the artificial intelligence track
American entrepreneur Elon Musk announced the establishment of an artificial intelligence startup xAI, officially entering the artificial intelligence track and challenging explosive AI applications such as ChatGPT. According to the official website of xAI, the company's technical team is composed of engineers and university researchers from large American technology companies, spanning many fields in the industry and academia. The team includes 12 people, including Musk, and the remaining 11 have worked for large US technology companies such as DeepMind, OpenAI, Microsoft and Apple. Musk will operate independently of the new company but will work closely with them and develop technology that will benefit other companies' businesses.
Tsinghua's ChatGLM2-6B model allows free commercial use, opening a new era of AI research
The NLP team at Tsinghua University, along with Zhipu AI and Tsinghua KEG Lab, announced that their ChatGLM-6B and ChatGLM2-6B models are now fully weighted to allow academic research and commercial use. These two models are of great significance in the field of artificial intelligence research. The ChatGLM-6B model is a Chinese-English bilingual dialogue model based on the General Language Model (GLM) architecture with 6.2 billion parameters. Through model quantization technology, the model can be deployed locally on consumer-grade graphics cards with as little as 6GB of video memory. The model has Q&A and dialogue functions, which initially shows its application potential in the field of artificial intelligence.
Google DeepMind releases CoDoC to improve the reliability and accuracy of medical AI diagnostics
Google DeepMind has partnered with several organizations to release an AI system called CoDoC (Complementarity-Driven Deferral to Clinical Workflow) to improve the reliability and accuracy of medical AI diagnoses. The system learns when to predict when the AI will provide the right information and follow the clinician's recommendations when needed. CoDoC explores ways to deliver the best outcomes in healthcare settings through AI collaboration with humans. In screening for breast cancer and tuberculosis, CoDoC improves accuracy compared to relying on clinician or AI baselines alone. In terms of breast cancer screening, CoDoC reduced false positive rates by 25% and reduced clinician workload by 66% compared to the "double-reading arbitration" in the UK screening programme.
Google's touting of its AI-powered news-writing tools to the news giant has raised concerns in the press
Google's press spokesperson issued a statement saying that Google is exploring the possibility of using artificial intelligence tools to assist news writing, and has been in talks with some giant news organizations to seek to apply relevant tools in the writing process of journalists in these media. The spokesperson did not say which news organizations had discussed the issue, but The New York Times reported that Google had preliminary discussions with The Washington Post, The Wall Street Journal, The New York Times and other media outlets.
ChatGPT introduces custom directives that remember conversation context and user preferences
Shortly after releasing the blockbuster "Code Interpreter", Open AI announced the launch of custom instructions for ChatGPT, which is in response to feedback that users need to reset each time they start using Chat GPT, and also give users more control over ChatGPT. The feature is first available in beta for Plus users and is not currently available outside the EU and UK. But in the coming weeks, the feature will be available to all users. This new operation of ChatGPT can better meet the individual needs of users. Custom directives allow users to control how ChatGPT responds. By setting preferences, ChatGPT will keep those preferences in mind in all future conversations, without repeating personal preferences or information in each conversation.
The Android version of ChatGPT went live weekly
OpenAI announced on Twitter that ChatGPT for Android will launch this week, but did not list a specific date. And the tweet links directly to the Google Play Store booking page, where users can sign up and install it once the app is available. OpenAI has made the Android version of ChatGPT a "big move". If the Android version is as popular as the iOS version, it is bound to become the strongest GPT released by all platforms (Web+iOS+Android).
Interactive Chinese pathway language models in the medical field
Researchers from Macau Polytechnic University, Shanghai Jiao Tong University and Opera have proposed a large language model based on LLaMA, IvyGPT, which is trained and fine-tuned through high-quality medical question answering and reinforcement learning based on human feedback (RLHF). After supervised fine-tuning, IvyGPT performed well at multiple rounds of dialogue, but failed to reach the level of human doctors in other areas, such as comprehensive diagnosis. In addition, through RLHF, IvyGPT can output richer, more human-like diagnostic and treatment options. Experimental results show that IvyGPT outperforms other medical GPT models.
MediaGPT: Large language models for Chinese media
Recently, Xinhua News Agency Chinese launched a large-scale language model trained on various media data - MediaGPT, which focuses on solving the actual needs of Chinese media. To further validate the effectiveness of the model, they built a unique dataset specifically applicable to the media domain and developed validation methods specifically for generative tasks.
Today's AI morning report is here, if you like it, please pay attention to the public number: AI has fan