๐ฐ๐ค๐ขAI News
OpenAI denies the imminent release of GPT-4.5
OpenAI CEO Sam Altman refuted rumors on Reddit that the company has no plans to release GPT-4.5, denying the leakage of the latest language model, and the screenshots show false information.
[AiBase feed]:
๐ฅ OpenAI CEO Sam Altman denied that GPT-4.5 was leaked and confirmed that the screenshots were fake information.
๐ It is unclear whether OpenAI will release GPT-4.5 or jump directly to GPT-5.
๐ OpenAI released GPT-3 in 2020 and GPT-4 in March 2023. At the same time, it is actively developing GPT-5.
OpenAI new research: GPT-2 can supervise GPT-4
Research has found that fine-tuning GPT-4 using the GPT-2 level model as a weak supervisor can significantly improve the generalization performance in natural language processing tasks and provide a solution for super AI alignment problems Provide new ideas.
[AiBase Feed]
๐ Innovation direction: Solve the challenge of super AI alignment problem by controlling large models through weak supervision of small models.
๐ Research results: GPT-2 level model fine-tuned GPT-4 achieved significant improvements in language processing tasks, demonstrating the feasibility of weak to strong generalization.
๐ฅ Research opportunities: Provide open source code and a $10 million funding plan to encourage researchers to conduct in-depth research in the field of super AI alignment.
Stability AI launches new membership model
Stability AI launches new membership model in dynamic 2023 to standardize commercial use, including free and PRO membership, while maintaining openness to source code and weight.
[AiBase feed:]
๐ผ New membership model: Stability AI launches a new membership model designed to promote commercial applications, expand the scope of enterprise deployment, and make the company model the cornerstone of building business.
๐ฐ Business model: including free personal, PRO membership and enterprise customized pricing, while maintaining openness to source code and weight, focusing on diversified open methods.
๐ Future Outlook: Free individual users can also enjoy the value of membership, including early access to new model releases, participation in public forums, and opportunities to showcase on the Stability AI channel. The founders are optimistic about the new modelโs appeal to startups and large enterprises and believe it will become a solid revenue base.
Intel releases AI accelerator Gaudi3
Intel releases Gaudi3 series AI accelerator, which uses advanced 5nm process and has superior performance. It is planned to be launched next year to compete with Nvidia's H200 accelerator card.
[AiBase Feed]
๐ Excellent performance: Intel Gaudi3 adopts advanced 5nm process, which increases bandwidth by 1.5 times, BF16 power by 4 times, and network computing power by 2 times.
๐ก Market competition: Gaudi3 plans to occupy a larger market share in 2024 and directly compete with NVIDIA's H200 accelerator card.
๐ฐ Cost advantage: With excellent performance and competitive overall cost, Gaudi3 is expected to achieve greater success in the market.
University of Technology Sydney has successfully developed a non-invasive system that converts brainwave signals into text
A research team at the University of Technology Sydney has successfully developed a portable, non-invasive system that uses an AI model to convert brainwave signals into text to treat stroke or paralysis. Patients offer new ways to communicate.
ใAiBase Feed:ใ
๐ง Without the need for surgery or other invasive methods, a system developed by the University of Technology Sydney can interpret brain waves and convert them into text.
๐ค has wide application prospects, especially for stroke or paralysis patients, providing them with a silent thinking communication method.
๐ uses an AI model called DeWave to record brain electrical activity by wearing a hat, making it non-invasive and convenient for daily use.
Microsoft expands Azure AI Studio and introduces Llama2 and GPT-4Turbo with Vision
Microsoft introduces Meta competitor Llama2 into Azure AI Studio to provide AI model as a service (MaaS), and also joins OpenAI's GPT-4Turbo with Vision to expand Azure cloud platform AI choose.
[AiBase feed:]
๐ Expanding AI services: Microsoft integrated Meta's Llama2 and introduced it into Azure AI Studio as a model-as-a-service, providing multiple open source Llama models to enrich the AI โโchoices of Azure cloud storage and service customers.
๐ Diversified AI options: Microsoft has added OpenAIโs GPT-4Turbo with Vision to Azure AI Studio to provide customers with more advanced AI tools, including image analysis and description capabilities.
๐ค Strategic diversification: Microsoft adopts a diversified strategy to expand its AI model library, not only providing models in cooperation with OpenAI, but also introducing open source models from competitors to meet different customer needs.
Ollama supports multi-modal models. Using
, the latest version of Ollama provides multi-modal model support for macOS and Linux users. By entering the command "ollama run llava" and downloading the llava-7B model, users can easily run Llama2, Code Llama and other models locally. , supports nearly twenty language model series.
official website address: https://top.aibase.com/tool/ollama
[AiBase feed:]
๐ Multimodal model support: The latest version of Ollama allows users to run multimodal models locally on macOS and Linux, providing more flexibility application scenarios.
โ๏ธ Model selection and running: Users can easily run Llama2, Code Llama and other models by inputting "ollama run llava" and downloading the llava-7B model, and drag and drop image input problems.
๐ Quantization level and performance trade-off: Ollama supports multiple language model series and different "tags". Users can choose the quantization level according to their needs to weigh model accuracy and running speed.
๐ค๐๐ป๐ก Large model dynamic
Alibaba image generation video model I2VGen-XL code release
Alibaba released the image generation video model I2VGen-XL in November, and open sourced its code and model as scheduled. The model passed 35 million orders Data training on lens-text-video pairs and 6 billion text-image pairs improves the semantic accuracy and detail continuity of the generated videos.
code address: https://github.com/damo-vilab/i2vgen-xl
[AiBase summary:]
๐๏ธโ๐จ๏ธ Basic stage and optimization stage: The I2VGen-XL model is divided into a basic stage and an optimization stage, through layered coding The processor maintains semantic coherence and integrates short text to enhance video details.
๐ Model optimization data: The research team optimized the I2VGen-XL model by collecting data on approximately 35 million single-shot text-video pairs and 6 billion text-image pairs, improving the semantic accuracy, detail continuity, and clarity of the generated videos. Spend.
๐ Code open source address: The code and model of Alibaba image generation video model I2VGen-XL have been open sourced on GitHub, providing researchers and developers with resources that can be explored and used.
Yuanxiang open sourced the XVERSE-65B-Chat large model
Yuanxiang announced the open source XVERSE-65B-Chat large model, providing powerful and unconditionally free commercial tools. Developers can log in to the official website or mini program to experience it.
Github: https://github.com/xverse-ai/XVERSE-65B
[AiBase feed:]
๐ Open source power: Yuanxiang released XVERSE-65B-Chat, which is the earliest free commercial model with the largest parameters in China and was evaluated in SuperCLUE Ranked first in the country in total open source scores.
๐ง Excellent performance: Compared with other models, XVERSE-65B has stronger understanding, generation, logic and memory capabilities, and can handle more diverse and more difficult tasks.
๐ Resource link: Developers can obtain the XVERSE-65B-Chat model through Github, Hugging Face, ModelScope and other platforms.
Shanghai Jiao Tong University and Baidu jointly released the Magnolia Science Model Version 2.0
Shanghai Jiao Tong University and Baidu Intelligent Cloud cooperated to release the "Magnolia Science Model Version 2.0" including "Legal Open Source" and "Chemical Synthesis 2.0". The "Legal Open Source" model has performed well in the legal field, surpassing similar Chinese general models and Chinese legal models.
[AiBase feed:]
๐ Shanghai Jiao Tong University and Baidu jointly released version 2.0 of the Magnolia Science Model, including the fields of law and chemistry.
๐ง "Magnolia Science Model - Legal Open Source" is based on domain pre-training and integrates legal knowledge to surpass similar models.
๐ This release marks new progress made by both parties in the field of AI for Science and sets a new example for in-depth cooperation between schools and enterprises.
Google launches the generative AI medical model MedLM
Google releases the MedLM generative AI medical model. Based on Med-PaLM2, the accuracy of the US Medical Licensing Examination reaches 85%. It plans to integrate the Gemini model to serve the global medical industry.
[AiBase Feed]
๐ Google MedLM model, specially designed for the healthcare industry, achieved 85% accuracy through the US Medical Licensing Examination.
๐ฅ MedLM is based on Med-PaLM2, which is 18% higher than the first generation. Google plans to integrate the Gemini model to expand its artificial intelligence capabilities.
๐ MedLM serves all aspects of the medical industry, including hospitals, drug development, chatbots, etc., and has been tested in multiple organizations and gradually put into production.
๐ค๐ฑ๐ผAI application
Spotify tests AI playlist function
Spotify is testing the function of creating playlists based on AI technology and user prompts, responding to user input through ChatGPT, demonstrating the AI-driven playlist generation process.
ใAiBase feed:ใ
๐ต Spotify confirmed to test the prompt-based AI playlist feature, allowing users to create playlists using AI technology and prompts.
๐ค The video shows the process of users creating a playlist using ChatGPT in the Spotify application through the "Your Library" option. AI responds to the user's prompts and generates the playlist.
๐ Spotify confirmed the test, but did not disclose the technical details, working principle, or commit to the official launch time.
Video redrawing tool DomoAI can convert videos into animations without SD videos with one click
DomoAI is a free artificial intelligence art generator. Through simple operation and diverse preset models, users can convert text into high-quality art in 20 seconds. products to create quickly and maintain a consistent painting style.
Official website address: https://top.aibase.com/tool/domoai
[AiBase feed:]
๐จ Creative release: DomoAI helps users quickly personalize through short text prompts, such as describing an old wizard or a girl swimming underwater. creation.
๐ Community interaction: Provide a community platform where users can get support in Discord, making DomoAI an art creation community that interacts and develops with users.
๐ Efficient creation platform: DomoAI provides users with an efficient and interesting art creation platform with the speed of converting text into artwork within 20 seconds, simple operation and rich preset models.
Visual Electric releases multiple image combination and redraw functions
AI image generation tool Visual Electric launches two major functions, allowing image creators to easily combine multiple images and redraw them, improving the flexibility of the creative process. Designers can generate each body separately and then combine them through the redraw function to achieve more intuitive creative implementation.
official website address: https://top.aibase.com/tool/visual-electric
[AiBase feed:]
๐ฅ Multi-image combination: Visual Electric allows users to combine multiple generated images, providing designers with more flexibility nature, supporting staged creation.
๐จ Custom style: Using several pictures, users can quickly customize the image generation style, similar to the Lora training method, expanding creative possibilities.
๐ Intuitive creative implementation: The new features introduced make the image generation process more flexible and intuitive, allowing designers to realize their creative ideas more easily and making the creative process more fun.
Instagram launches generative AI background editing tool
Instagram launches generative AI background editing tool, allowing users to customize unique picture backgrounds through various prompts to promote interactive experiences.
ใAiBase feed:ใ
๐จ Users can customize the background through prompts such as "walk the red carpet". After
๐คณ is published, other users can participate and interact, easily sharing unique image stories.
๐ Generative AI technology has gradually become the key to creative expression and user interaction on social media.
๐จโ๐ป๐ก๐ฏ Focus on developers
Google develops real-time rendering of large-scale three-dimensional scene technology SMERF
The SMERF technology launched by the Google team can create realistic three-dimensional scenes in real time in a room of up to 300 square meters, supports smartphones and notebooks, and has 60fps real-time Rendering and full six degrees of freedom navigation.This technology uses hierarchical model division and distillation training strategies to solve the performance and quality problems of rendering large 3D scenes and provide a more realistic and smooth 3D experience.
Project address: https://smerf-3d.github.io/
[AiBase feed:]
๐ Real-time rendering of large scenes: SMERF technology can render realistic three-dimensional scenes in a 300-square-meter room in real time, supporting smooth navigation at 60fps.
๐ฎ Efficient memory usage: Using hierarchical model and distillation training, it improves processing efficiency and rendering speed, and can run smoothly even on devices with limited memory.
๐ฑ Popularity and Reality: Through ordinary smartphones and notebooks, users can get a free three-dimensional experience close to photorealism.
AI generated front-end code project "Coffee"
Through the artificial intelligence tool "Coffee", front-end developers can quickly generate, edit and maintain the React code base with zero dependence and zero settings, significantly improving development efficiency.
code address: https://github.com/Coframe/coffee
[AiBase feed:]
๐ The innovative tool Coffee: Coffee uses artificial intelligence technology to support the rapid generation and editing of React code libraries without additional dependencies, making front-end development more efficient. Efficient.
๐ ๏ธ Unified development experience: Whether creating new components or editing existing components, Coffee provides the same development experience, generating clear and maintainable code that meets production standards.
๐ Future expansion plans: Coffee plans to expand support for other popular front-end frameworks, including Vue, Svelte, etc., to broaden its scope of application.
Google releases NeRFiller, which uses 2D images to complete 3D scenes
Google and researchers from the University of California, Berkeley, collaborated to launch the NeRFiller framework, which repairs missing 3D scenes through 2D images. It uses grid priors and joint multi-view completion strategies, significantly Improve repair effect and reconstruction efficiency.
will be open source soon: https://github.com/ethanweber/nerfiller
paper: https://arxiv.org/abs/2312.04560
[AiBase summary:]
๐ Multi-view consistency completion: NeRFiller uses grid prior and joint multi-view completion strategies, which are provided to the completion model through a 2x2 grid shape to increase the consistency of the repair effect.
๐3D scene integration iterative optimization: NeRFiller integrates the 2D image completion results into a globally consistent 3D scene through an iterative method, improving the geometry and consistency of the 3D scene.
๐ Improved reconstruction efficiency: Test data shows that NeRFiller performs better in multiple evaluation indicators such as PSNR and SSIM than the original data, and the reconstruction efficiency is increased by about 10 times.
