[Collection of Big Model Daily on March 6] Just now, OpenAI officially issued an article to refute Musk, revealing screenshots of email exchanges over the past eight years; has Claude 3 been made self-aware? There is a sensation in the AI community, we bought a membership and took a real-life test; Big manufacturers compete in social networking again: QQ tests AI chat partner, Douyin launches AI social APP “conversation stove”

ICLR 2024 | Providing a new perspective for audio and video separation, Tsinghua University’s Hu Xiaolin team launched RTFS-Net

Link: https://news.miracleplus.com/share_link/20373

Audio-Visual Speech Separation (AVSS) technology aims to separate the target speaker’s voice from the mixed signal through facial information. This technology can be used in applications such as smart assistants, remote conferencing, and augmented reality to improve voice signal quality in noisy environments. Traditional audiovisual speech separation methods rely on complex models and large amounts of computing resources, and their performance is often limited especially in noisy backgrounds or multi-speaker scenarios. In order to break through these limitations, methods based on deep learning began to be researched and applied. However, existing deep learning methods face the challenges of high computational complexity and difficulty in generalizing to unknown environments. In order to overcome these challenges, researchers from the team of Associate Professor Hu Xiaolin of Tsinghua University proposed RTFS-Net: a new audio-visual speech separation model. RTFS-Net greatly reduces the computational complexity and number of parameters of the model while improving separation performance through compression-reconstruction. RTFS-Net is the first audiovisual speech separation method with less than 1 million parameters and the first time-frequency domain multi-modal separation model to outperform all time-domain models.

Using structure-based mutation preferences for protein design, a team from the University of California, MIT, and Harvard Medical School developed an unsupervised approach

Link: https://news.miracleplus.com/share_link/20374

Current state-of-the-art protein design methods often rely on large neural networks with hundreds of millions of parameters, and it is not clear which residue dependencies are critical for determining protein function. Researchers at the University of California, Massachusetts Institute of Technology, and Harvard Medical School have shown that the amino acid preferences of individual residues can Explains most and sometimes almost all of the combined mutation effects in the 8 data sets (R^2 ~ 78-98%); therefore, very few observations (~100 times the number of mutated residues) accurately predict “retained” “held-out variant effects” (Pearson r > 0.80). Assuming that the local structural context around residues is sufficient to predict mutation preferences, the team developed an unsupervised method called CoVES (Combinatorial Variant Effects from Structure) that can use structure-based mutation preferences for protein design. Test results demonstrate that CoVES outperforms not only model-free approaches, but also complex models used to create functional and diverse protein variants. CoVES provides an efficient alternative for identifying complex models of functional protein mutations.

Just now, OpenAI officially issued a document refuting Musk and revealing screenshots of email exchanges over the past eight years.

Link: https://news.miracleplus.com/share_link/20375

https://openai.com/blog/openai-elon-musk

The epic battle between the hottest technology company OpenAI and the world’s richest man Musk has reached a new level. Just now, OpenAI used a long article “OpenAI and Elon Musk” to officially refute all of Musk’s accusations. The title is concise, but the content is quite eye-catching. OpenAI directly posted screenshots of emails between the founding team members and Musk over the past eight years, and repeatedly reiterated OpenAI’s unremitting pursuit of its founding mission. The article begins: “OpenAI’s mission is to ensure that AGI benefits all humanity, which means both building safe and beneficial AGI and helping to create widespread benefits. We are sharing what we have learned in achieving our mission, and Some facts about our relationship with Musk. We intend to reject all of Musk’s claims.” At the same time, the exposure of some “mysterious information” in the content of the email has once again heightened people’s curiosity about the future of mankind mentioned by Musk. , what exactly does it depend on?

Claude 3 was played out of his self-awareness? There was a sensation in the AI community. We bought membership and took a real test.

Link: https://news.miracleplus.com/share_link/20376

On Monday, Anthropic released a new generation of large model series Claude 3. GPT-4, which has been far ahead for almost a year, finally has a strong opponent. The power of Claude 3 is not only reflected in various benchmark tests, but it also seems to have achieved some magical breakthroughs. Yesterday, Anthropic prompted engineer Alex Albert to post results on Twitter claiming that Claude 3 had “self-awareness,” causing quite a stir in the AI community.

Alibaba International Station: AI is driving the foreign trade industry to become fully intelligent

Link: https://news.miracleplus.com/share_link/20377

On March 6, at the “2024 New Trends in Cross-Border E-commerce Growth Conference” jointly organized by CCBEC and Alibaba International Station, Alibaba International Station revealed that in just half a year after its launch, 12,000 foreign trade companies have already ordered AI foreign trade products. , the number of products released by AI has reached one million, and the search volume of these products overseas has increased by 37% through AI optimization. This means that the “application era” of AI has arrived first in the foreign trade industry.

Big companies compete in social networking again: QQ tests AI chat partner, Douyin launches AI social APP “Chatbox”

Link: https://news.miracleplus.com/share_link/20378

Social networking is still an outdated track. Both QQ and Douyin have recently launched AI social functions and products. QQ’s AI dialogue function has been tested, called “AI Chat Partner”, an AI dialogue service jointly developed by QQ, Dream Island, and Hunyuan Assistant. Douyin, on the other hand, has launched an AI social APP called “Hua Lu”, built by Byte’s Flow department. QQ’s internally tested AI chat partner smart service is actually an AI character interactive chat product. It currently includes different types of virtual characters such as companions, stories, challenges, MBTI, celebrities, etc. Users can choose a virtual character that suits their type and chat in real time like they usually do with friends. It is worth noting that since each reply of a virtual character will incur a computing power cost, ordinary QQ users can chat with these virtual characters for 100 messages per day, and super member users can chat with 400 messages per day.