【 January 15-16 Big Model Daily Collection 】 Microsoft TaskWeaver Open Source Framework: Collaborate with data analysis and industry customization to create top-level Agent solutions; Microsoft Copilot membership is here: $20 per month, everyone can use it; Zhipu releases the GLM-4 family bucket: performance improvement of 60%, multimodal, 128K long text, all-around All Tools and GPT stores; New record for open source models: models that surpass Mixtral 8x7B Instrument are here

Microsoft TaskWeaver Open Source Framework: Collaborating with Data Analysis and Industry Customization to Build Top tier Agent Solutions

Link: https://news.miracleplus.com/share_link/15912

Data analysis has always been an important tool in modern society, helping us to understand the essence, discover patterns, and guide decision-making. However, the process of data analysis is often complex and time-consuming, so we expect to have an intelligent assistant to assist users in directly “talking to data”. Thanks to the development of Large Language Models (LLM), intelligent agents such as virtual assistants and Copilot have emerged, and their performance in natural language understanding and generation is impressive. Unfortunately, when dealing with complex data structures (such as DataFrame, ndarray, etc.) In terms of introducing domain knowledge, existing Agent frameworks still face difficulties, which is precisely the core requirement in data analysis and professional fields. To overcome this bottleneck, Microsoft has launched TaskWeaver – a code first Agent framework. TaskWeaver can cleverly transform users’ natural language requests into executable code, and supports rich data structures, dynamic plugin selection, and domain specific planning processes.

Superior to human doctors in 24 scenarios, the Google team developed a diagnostic dialogue model based on self game theory

Link: https://news.miracleplus.com/share_link/15915

The core of medicine lies in doctor-patient dialogue, and skilled medical history collection paves the way for accurate diagnosis, effective management, and lasting trust. Artificial intelligence (AI) systems capable of conducting diagnostic conversations can improve the accessibility, consistency, and quality of care. However, learning the professional knowledge of clinical doctors is a huge challenge. The AI teams of Google Research and Google DeepMind have developed AMIE (Artificial Medical Intelligence Explorer), an artificial intelligence system based on Large Language Model (LLM) optimized for diagnostic conversations. AMIE utilizes a novel self play based simulation environment with an automatic feedback mechanism that can expand learning across different disease conditions, specialties, and backgrounds. “To our knowledge, this is the first conversational artificial intelligence system designed to be the best for diagnostic dialogue and obtaining clinical medical history,” said Alan Karthikesalingam, a clinical research scientist at Google Health and co-author of the paper.

The Shanghai Artificial Intelligence Laboratory team has developed an AI driven projection tomography scan with multi-core fiber optic unit rotation function

Link: https://news.miracleplus.com/share_link/15916

Optical tomography has become a non-invasive imaging method that provides three-dimensional insights into subcellular structures, enabling a deeper understanding of cell function, interactions, and processes. The traditional optical tomography method is limited by the limited illumination scanning range, resulting in resolution anisotropy and incomplete imaging of cell structures. To overcome this problem, researchers from the Shanghai Artificial Intelligence Laboratory, Technische University ä t Dresden in Germany, and the Max Planck Society used a compact multi-core fiber optic cell rotator system, which helps to accurately manipulate cells inside microfluidic chips, enabling full angle projection tomography with isotropic resolution. In addition, researchers have demonstrated an artificial intelligence driven workflow for tomography reconstruction, which can be a paradigm shift from traditional computational methods that typically require manual processing to fully autonomous processes. The performance of the proposed cell rotation tomography method was validated through cell models and three-dimensional reconstruction of HL60 human cancer cells.

Microsoft Copilot membership is here: $20 per month, everyone can use it

Link: https://news.miracleplus.com/share_link/15917

Last September, Microsoft announced a new plan to add the Copilot capability of the large model to Office and Windows 11. Today, this capability has finally reached all users. On January 15th, Microsoft officially released Copilot Pro, with a monthly subscription fee of $20. It can unlock access to large model capabilities in commonly used office software such as Word, Excel, and PowerPoint, while prioritizing access to the latest OpenAI models. On top of it, you can also build your own Copilot GPT based on requirements.

Zhipu Releases GLM-4 Family Bucket: Performance Improved by 60%, Multimodal, 128K Long Text, Versatile All Tools and GPT Store

Link: https://news.miracleplus.com/share_link/15918

Domestic large-scale model player Zhipu AI has submitted the latest transcript – releasing the fully self-developed fourth generation base large-scale model GLM-4, and all updated and iterated capabilities have been fully launched. As the only large model company in China that fully benchmarks its product line with OpenAI, Zhipu accumulated a large number of achievements this year: according to official statements, the performance of GLM-4 has improved by 60% compared to GLM-3, approaching GPT-4. The GLM-4 All Tools, launched simultaneously, can achieve various model capabilities such as user intent understanding, complex instruction planning, and free invocation of the GLM ecosystem. It is worth mentioning that Zhipu has released GLMs, providing AI intelligent agent customization capabilities for all developers. Personalized GLM intelligent agents can be created with a simple prompt command.

MiniMax’s first MoE large language model in China has been fully launched

Link: https://news.miracleplus.com/share_link/15919

Today, after half a month of internal testing and feedback from some customers, MiniMax has fully released the large language model abab6, which is the first MoE large language model in China. Under the MoE structure, abab6 has the ability to handle complex tasks with large parameters, and the model can train enough data per unit time, resulting in a significant improvement in computational efficiency. Improved abab5.5’s ability to handle more complex and fine-grained model output scenarios.

New record for open source models: models that surpass Mixtral 8x7B Instrument are here

Link: https://news.miracleplus.com/share_link/15920

From Llama, Llama 2 to Mixtral 8x7B, the performance records of open source models have been constantly refreshed. Due to the fact that Mistral 8x7B outperforms Llama 2 70B and GPT-3.5 in most benchmark tests, it is also considered an open source option that is very close to GPT-4. In a recent paper, Mistral AI, the company behind the model, announced some technical details of the Mistral 8x7B and launched the Mistral 8x7B – Instrument chat model. The performance of this model significantly exceeds that of GPT-3.5 Turbo, Claude-2.1, Gemini Pro, and Llama 270B chat models on human evaluation benchmarks. In benchmarks such as BBQ and BOLD, Mixture – Instrument also shows less bias. But recently, the Mixture Instrument has also been surpassed. A company called Nous Research announced that their new model, Nous Hermes-2 Mixture 8x7B, trained on Mixture 8x7B, has surpassed Mixture Instrument on many benchmarks and achieved state-of-the-art performance.