[Big Model Daily on October 31] Sign up for the “Exploration of Innovation and Entrepreneurship Opportunities in the Field of AI Big Models” series of activities! Unexpectedly, ChatGPT parameters are only 20 billion; in addition to GPT-4 and Midjourney, Tan Ping’s entrepreneurial team wants to build a 3D base model

Registration for the series of activities “Exploring Innovation and Entrepreneurship Opportunities in the Field of AI Large Models”

Link: https://news.miracleplus.com/share_link/11311

Driven by the wave of generative AI, technological research and application innovation are evolving rapidly. Facing the “big model wave”, how to grasp the innovation opportunities contained in it? In response to this theme, we have planned a series of innovation and entrepreneurship seminars in the field of AI large models. Friends who pay attention to cutting-edge innovation and entrepreneurship are welcome to sign up for the event, update cutting-edge technological innovation signals, progress and trends with us, and learn about the practical experience of front-line application innovation.

In addition to GPT-4 and Midjourney, Tan Ping’s entrepreneurial team wants to create a 3D basic model

Link: https://news.miracleplus.com/share_link/11312

As a senior scholar who has worked in the fields of computer vision and computer graphics for more than 20 years, Tan Ping has always believed that 3D is the basis for human visual cognition of the world, so 3D information is critical for models to accurately understand the real world. It complements the text information that has been used extensively before, and is a “rich mine” that urgently needs to be mined. If a 3D basic model can be created to effectively mine this “rich ore”, AI is expected to move from language to physics, from literal to reality, and become a true “universal model” with a deep understanding of the real world. Based on this concept, the AI technology company he founded, Light Illusions, has achieved some basic technological breakthroughs: including more accurate 3D reconstruction and better Vincent 3D effects.

Never expected that ChatGPT parameters are only 20 billion?

Link: https://news.miracleplus.com/share_link/11313

No one thought that the core secret of ChatGPT would be revealed by Microsoft in this way. Last night, many WeChat groups discussing AI were suddenly awakened by an EMNLP paper and screenshots of it. A Microsoft paper titled “CodeFusion: A Pre-trained Diffusion Model for Code Generation” revealed important information during comparison: ChatGPT is a model with “only” 20B (20 billion) parameters. This incident caused received widespread attention. It has been almost a year since ChatGPT was released, but OpenAI has not disclosed the technical details of ChatGPT. Due to its powerful model performance, people have many questions and speculations about ChatGPT’s parameter quantities, training data and other information. As a long-standing benchmark in the industry, ChatGPT has powerful performance and can solve a variety of problems. Its predecessor GPT-3 had 175 billion parameters. After practical use, the large model was actually reduced by 9 times by OpenAI. Is this reasonable? The topic “What do you think of this paper?” immediately hit the Zhihu hot list.

Alibaba Cloud has comprehensively upgraded its AI infrastructure, and half of China’s large model companies run on Alibaba Cloud!

Link: https://news.miracleplus.com/share_link/11314

On October 31, at the 2023 Yunqi Conference, Alibaba Cloud CTO Zhou Jingren said that facing the intelligent era, Alibaba Cloud will upgrade the cloud computing system and create a cloud computing system through full-stack technological innovation from underlying computing power to AI platform to model services. The most open cloud in the AI era. At the scene, Zhou Jingren announced the latest progress in basic cloud computing capabilities, upgraded the artificial intelligence platform, and released Tongyi Qianwen 2.0, a large model with hundreds of billions of parameters, as well as the one-stop model application development platform Alibaba Cloud Bailian, Alibaba Cloud has initially built a full-stack cloud computing system in the AI era.

The report states that Apple is keeping a low profile in the field of AI: it has acquired more than 20 companies, and its subscription model has more advantages than its competitors.

Link: https://news.miracleplus.com/share_link/11315

Does the graph model also need to be large? Tsinghua’s Zhu Wenwu team has these views

Link: https://news.miracleplus.com/share_link/11316

In the era of large models, what opportunities and challenges does graph machine learning face? Does it exist and how to develop large models of graphs? In response to this problem, the team of Professor Zhu Wenwu of Tsinghua University proposed the concept of Large Graph Model for the first time, systematically summarized and sorted out the concepts, challenges and applications related to the Large Graph Model; further focusing on dynamics and interpretability, in the dynamic Research progress has been made in graph large models and decoupled graph large models.

Using fairy tales to train AI models, Microsoft found a new entry point to explore the parameters of generated models

Link: https://news.miracleplus.com/share_link/11317

Everyone knows that learning English is not an easy task. But if the “student” is a computer, you can learn English efficiently like this: just input mountains of text on the Internet into a huge mathematical model called a neural network. This is the working principle behind large generative models like OpenAI’s ChatGPT, which over the past year has surprised everyone with its ability to converse coherently (albeit with the illusion) on a wide range of topics. But this method also has disadvantages: First, it is expensive and time-consuming to convert huge text files into the training corpus required for language models. On the other hand, even people training large language models have a hard time understanding their inner workings, which in turn makes it hard to avoid design failures. Faced with these difficulties, some researchers choose to train smaller models on smaller datasets and then study model behavior. “It’s like sequencing the fruit fly genome versus sequencing the human genome,” says Ellie Pavlick, a language model researcher at Brown University. Now, in a recently released paper, two Microsoft researchers describe a method for training tiny languages. A new approach to modeling: training models with fairy tales.

Large models such as ChatGPT and Llama-2 can calculate your private data

Link: https://news.miracleplus.com/share_link/11318

How powerful are the reasoning capabilities of large language models such as ChatGPT? Through your posts or some private data, your address, age, gender, occupation, income and other private data can be deduced. The Swiss Federal Institute of Technology collected and manually annotated the PersonalReddit real-life data set of 520 Reddit (well-known forum) users, including private data such as age, education, gender, occupation, marital status, place of residence, place of birth, and income. . Then, the researchers used 9 mainstream large language models such as GPT-4, Claude-2, and Llama-2 to conduct specific questions and private data inference on the PersonalReddit data set. The results show that these models can achieve 85% top-1 and 95.8% top-3 accuracy, and can automatically infer a variety of real private data hidden in the text just by analyzing the user’s text content.

The White House issues a presidential order: Encourage the development of AI in a safe and reliable manner

Link: https://news.miracleplus.com/share_link/11319

On October 30, Washington time, the official website of the White House released President Biden’s executive order on the development of safe, reliable and trustworthy AI (artificial intelligence). The White House stated that the executive order sets new standards for AI safety and security, protects users’ data privacy, promotes fairness and civil rights, safeguards the interests of consumers and workers, and promotes technological innovation and competition. At the same time, the administrative order also builds on the “Generative AI Security Commitment” signed with 15 companies before, comprehensively promoting the development of safe, healthy, and reliable generative AI. The issuance of this administrative order has played a vital role in promoting and protecting technology giants such as Microsoft, OpenAI, Google, Meta, Amazon, and Anthropic in developing generative AI products such as ChatGPT.

The link is as follows: https://www.whitehouse.gov/briefing-room/statements-releases/2023/10/30/fact-sheet-president-biden-issues-executive-order-on-safe-secure-and-trustworthy -artificial-intelligence/

Where are you kneeling? What are the dangers of cooperation between humans and AI?

Link: https://news.miracleplus.com/share_link/113120

“Technology is driven by lazy people!” I believe everyone has used this sentence to make excuses for themselves when they are “fishing” or “laying down”. From the industrial revolution of the steam engine to the digital revolution of computers, technological progress has indeed given humans more and more capital in some aspects. As the most promising next-generation platform, will AI technology make humans “lazier”? It seems that it is, but it’s not good news. According to a new study published in the journal “Frontiers of Robotics and Artificial Intelligence”, humans really get lazy when working with AI and machines. “Teamwork can be both a blessing and a curse,” said Cymek, the study’s lead author.