[January 20-21 Big Model Daily] Study on the joint review of Tsinghua, Xiaomi, Huawei, vivo, Ideal and other institutions, the first to introduce personal LLM intelligent agents, and divide the industry’s large models into five levels of intelligence. Self-reward: Meta lets Llama2 give it to itself Fine-tuned by yourself, the performance surpasses GPT-4

A joint review by Tsinghua University, Xiaomi, Huawei, Vivo, Ideal and other institutions, firstly mentioning personal LLM intelligence and dividing it into five levels of intelligence

https://news.miracleplus.com/share_link/16400

Presumably at least one of these wake words has been uttered by your mouth and successfully called out an intelligent personal assistant (IPA) that can help you navigate, tell jokes, add schedules, set alarms, and make calls. It can be said that IPA has become an indispensable standard for modern smartphones. A recent review paper believes that “personal LLM agents will become the main software paradigm for personal computing in the AI era.” This personal LLM agent review paper comes from many domestic universities and corporate research institutes, including Tsinghua University, Xiaomi, Huawei, Huantai, vivo, Yunmi, Li Auto, Beijing University of Posts and Telecommunications, and Suzhou University. The article not only sorts out the capabilities, efficiency and security issues required by personal LLM agents, but also collects and organizes the insights of domain experts. In addition, it also pioneeringly proposes a 5-level intelligence level classification method for personal LLM agents. The team has also created a literature library on GitHub and published relevant literature, which can also be jointly maintained by the IPA community and updated with the latest research and development progress.

Pika, Gen-2, ModelScope, SEINE…which one is better in AI video generation? This framework will tell you once you test it

https://news.miracleplus.com/share_link/16401

AI video generation is one of the hottest fields recently. Various university laboratories, Internet giant AI Labs, and start-up companies have joined the AI video generation track. The release of video generation models such as Pika, Gen-2, Show-1, VideoCrafter, ModelScope, SEINE, LaVie, and VideoLDM is even more eye-catching. To this end, NTU, together with Shanghai AI Lab, CUHK and Nanjing University, launched VBench, a comprehensive “evaluation framework for video generation models” to tell you: which video model is strong and where each model is strong.

NVIDIA’s new dialogue QA model is more accurate than GPT-4, but it was criticized: unweighted code has little meaning

https://news.miracleplus.com/share_link/16402

For more than a year, ChatGPT and its successors have sparked a paradigm shift in building question-answering (QA) models in the production and research communities. Especially in practical applications, QA models become the first choice in the following situations. Recently, in a paper by NVIDIA, researchers proposed a white-box conversation QA model ChatQA 70B with GPT-4 level accuracy. They employ a two-stage instruction tuning approach along with a RAG-enhanced retriever for conversational QA, a rigorous data management process.

Roche and GRCEH teams develop interpretable machine learning methods for analyzing immune synapses and functional characterization of therapeutic antibodies

https://news.miracleplus.com/share_link/16403

Therapeutic antibodies are widely used to treat serious diseases. Most of them alter immune cells and act within the immune synapse. Important cell-cell interactions that guide humoral immune responses. Although many antibody designs have been generated and evaluated, high-throughput tools for systemic antibody characterization and functional prediction are lacking. A research team from the German Research Center for Environmental Health and Roche has developed a comprehensive open source framework scifAI (Single Cell Imaging Cytometry AI) for imaging flow cytometry. (IFC) data for preprocessing, feature engineering, and interpretable predictive machine learning.

A photo, customized portrait pictures for deep learning giants

https://news.miracleplus.com/share_link/16404

Topic-driven text-to-image generation usually requires training on multiple data sets containing the topic (such as characters, styles). Representative work in this type of method includes DreamBooth, Textual Inversion, LoRAs, etc., but this type of scheme is due to The need to update the entire network or require long-term customized training is often not very effectively compatible with existing models in the community, and cannot be applied quickly and cost-effectively in real scenarios. However, the current embedding methods based on single image features (FaceStudio, PhotoMaker, IP-Adapter) either require full-parameter training or PEFT fine-tuning of the Vincentian graph model, which affects the generalization performance of the original model and lacks integration with community pre-trained models. compatibility, or the inability to maintain high fidelity. In order to solve these problems, researchers from the InstantX team proposed InstantID. This model does not train the UNet part of the Vincent graph model, but only trains pluggable modules. There is no need for test-time tuning during the inference process, and it hardly affects text control capabilities. In this case, high-fidelity ID retention is achieved.

AI automatically finds “high-energy moments” when watching videos | Byte & Institute of Automation, Chinese Academy of Sciences @AAAI 2024

https://news.miracleplus.com/share_link/16405

ByteDance and the Institute of Automation of the Chinese Academy of Sciences have proposed a new method that uses AI to quickly detect highlight segments in videos. It is extremely flexible to the length of the input video and the length of highlights expected to be extracted. The relevant papers have been included in AAAI 2024.

Large models reward themselves: Meta lets Llama2 fine-tune itself, and its performance surpasses GPT-4

https://news.miracleplus.com/share_link/16406

In the field of large models, fine-tuning is an important step to improve model performance. As the number of open source large models gradually increases, people have summarized many methods of fine-tuning, some of which have achieved good results. Recently, researchers from Meta and New York University used the “self-reward method” to allow large models to generate their own fine-tuning data, which brought a new shock to people. In the new method, the author fine-tuned Llama 2 70B in three iterations, and the resulting model outperformed a number of existing important large models on the AlpacaEval 2.0 rankings, including Claude 2, Gemini Pro and GPT-4.

Ultraman raises billions of dollars to build global fab network and manufacture its own AI chips

https://news.miracleplus.com/share_link/16407

According to Bloomberg, OpenAI CEO Sam Altman has recently raised billions of dollars in funding for an artificial intelligence chip company, hoping to establish a global network of wafer factories. )” and plans to cooperate with unnamed top chip manufacturers.

Stability AI is back: The new demo effect of video generation is amazing, netizens: The consistency is outstanding

https://news.miracleplus.com/share_link/16408

Stable Diffusion is coming back? Stability AI CEO Emad Mostaque’s latest tweets and four videos have aroused countless imaginations. Many netizens suspect that this is a demo of a new version of Stable Video Diffusion. Because from the effect point of view, the picture clarity, consistency and smoothness are all amazing.

Two DeepMind employees resigned to create an AI model that may raise US$200 million in the first round, and Flexport received another US$260 million.

https://news.miracleplus.com/share_link/16409

Flexport CEO Ryan Petersen announced on X that it has raised a new round of financing of $260 million from Shopify. What’s more interesting is that this financing is an uncapped convertible note, which is a convertible bond with no valuation upper limit. It is generally used for investment in early-stage entrepreneurial projects and is a very beneficial financing method for entrepreneurial companies. The adoption of this approach in a late-stage company such as Flexport is probably a support for the various problems Flexport has faced in the past year. In the past year, its business volume has dropped significantly due to the epidemic, causing the company to face a series of problems. Problem, Flexport eventually went through two 20% layoffs.