Here is your guide to exploring the world of artificial intelligence every day, and every day we present you hot content in the field of AI, focusing on developers, helping you gain insight into technology trends and understand innovative AI product applications.
Fresh AI products Click to learn:https://top.aibase.com/
What's new in AI
Stanford pushed the Octopusv2, a large model that can run on mobile phones, to surpass GPT-4 and become an overnight hit
【AiBase Summary:】
⭐️ Stanford University launched Octopusv2, a 2 billion parameter model that can run on devices such as mobile phones, surpassing GPT-4 in accuracy and latency, and reducing context length by 95%.
⭐️ With the advent of the era of device-side AI agents, Octopusv2 innovates the function token strategy in the development process, and the training method improves the inference speed and performs well.
⭐️ The Octopus-V2-2B performed well in the performance evaluation, with a 168% increase in speed, injecting new vitality into the development of on-device AI.
Address: https://arxiv.org/abs/2404.01744
Model homepage: https://huggingface.co/NexaAIDev/Octopus-v2
The post-90s guy used AI to "resurrect" the deceased, and 1,000 orders were sold a year
【AiBase Summary:】
⭐️Zhang Zewei, a young man born in the 90s, uses AI technology to create a digital image for the deceased, and has received more than 1,000 orders.
⭐️ His team allows customers to interact with their digital avatars by recreating the faces and voices of the deceased.
⭐️ What makes this business unique is that AI technology generates a response to a digital image of the deceased, without the need for a human to act as one.
Details: https://www.chinaz.com/2024/0407/1608319.shtml
OpenAI adds new features for developers that allow them to build custom models
【AiBase Summary:】
⭐️ Developers can use OpenAI's new capabilities to build custom models specific to the needs of the organization, business domain, and tasks.
⭐️ Custom models include a specialized knowledge base, specific data understanding, performing a task, or a specific input response.
⭐️ OpenAI provides features such as fine-tuning APIs, custom training model programs, and assisted fine-tuning services to help developers build custom models.
Details: https://openai.com/blog/introducing-improvements-to-the-fine-tuning-api-and-expanding-our-custom-models-program
OpenAI transcribed over a million hours of YouTube videos to train GPT-4
【AiBase Summary:】
OpenAI trained GPT-4 using YouTube video transcription
AI companies are challenged with high-quality training data
⚖️ The company's handling of data issues involves areas of ambiguity in copyright law
AI video understanding ceiling, the new MiniGPT4-Video swept SOTA!
【AiBase Summary:】
⭐ The MiniGPT4-Video framework can understand complex videos and create poetry captions.
⭐ It supports the processing of temporal visual data and text data, and is good at understanding the complexity of video.
⭐ In multiple benchmarks, MiniGPT4-Video has improved significantly, providing powerful interpretation capabilities for video captions, ads, and more.
Thesis: https://arxiv.org/pdf/2404.03413.pdf
DeepMind Launches Gecko: Focus on Document Retrieval Performance Comparable to 7x Larger Models
【AiBase Summary:】
Gecko is a general-purpose text embedding model that focuses on document retrieval, semantic similarity, and classification tasks.
Gecko achieves powerful search performance by extracting knowledge from LLMs and incorporating them into search engines.
On the large-scale text embedding benchmark, the 256-dimensional Gecko outperforms the 768-dimensional existing models.
Address: https://arxiv.org/pdf/2403.20327.pdf
Microsoft is investing $100 billion to boost generative AI, and the stock is on track to soar
【AiBase Summary】
Microsoft's deployment of custom chips based on Arm designs is expected to bring growth to Arm Holdings.
Arm Holdings has benefited from the growth of AI chips, and the Microsoft project could further boost its performance.
Microsoft may reduce its dependence on other companies by customizing chips, improve performance and reduce costs, which is expected to drive Arm revenue growth.
Musk's friends will help xAI raise $3 billion
【AiBase Summary:】
The investor has a close relationship with Musk and plans to help xAI raise $3 billion.
xAI competes with competitors such as OpenAI and Anthropic to step up its development in the highly competitive field of artificial intelligence.
The competition for AI talent is fierce, with xAI and other competitors competing to attract and retain them.
The next big leap forward in AI is understanding emotions, and the first conversational AI with emotional intelligence is coming
【AiBase Summary:】
⭐️ HumeAI released a conversational AI with emotion recognition capabilities that can detect 53 emotions.
⭐️ HumeAI is committed to understanding and responding to user emotions, enabling interaction through voice features.
⭐️ APIs are provided to allow users to train their own AI models, which are widely used in health, customer service, and other fields.
Official website address: https://dev.hume.ai/docs/expression-measurement-api/overview
Kingsoft Office WPS365 will release one-stop AI office products
【AiBase Summary:】
⭐ WPS365 will emphasize improving user efficiency and experience
⭐ The suite includes content creation tools and collaboration software
⭐ The core idea is to unify tools, collaboration, and management
AI applications
Infinity AI: Enter the script to generate a movie with one click It also provides a digital human cloning feature
【AiBase Summary:】
⭐ Goal: You only need to enter the content of the script, and you can generate a movie with one click, and the official demo has been released
⭐ The technical team successfully cloned the image of the CEO to show the product functions, and proposed that in the future, the small team is expected to win the Oscar with the help of AI.
⭐ Provide the clone yourself service, users can train custom AI models by recording videos, and generate video content containing their own voices and facial expressions.
Official website: https://top.aibase.com/tool/infinity-ai
Online experience address: https://studio.infinity.ai/
Detailed tutorials and videos: https://qqi2gjmnk4.feishu.cn/wiki/HTmRwpZ1hiRONpkZ3SIce89ynuc?fromScene=spaceOverview
Google pushes Scenic: recognizes video content and generates detailed descriptions
【AiBase Summary:】
SOTA models and baseline models are available to support rapid prototyping of large-scale visual models.
Develop with JAX and Flax with support for image, video, audio, and multimodal combination models.
Recognize video content and generate detailed descriptions, providing feature-rich baseline models and datasets.
Product Entry: https://top.aibase.com/tool/scenic
CameraCtrl: Enables Text to Generate Video for Lens Control AnimateDiff is supported
【AiBase Summary:】
⭐ Lens control is crucial in video generation
⭐ Plug-in lens modules are implemented by training the lens encoder
⭐ Videos with different datasets enhance the controllability and generalization of lens control
Product Entry:https://hehao13.github.io/projects-CameraCtrl/
Lixel CyberColor: Automatically generate infinitely cinematic 3D scenes
【AiBase Summary:】
⭐️ LCC uses Multi-SLAM and Gaussian sputtering technology to generate cinematic 3D scenes.
⭐️ XGRIDS' Multi-SLAM algorithm and 3DGS technology create photorealistic large-scale 3D models.
⭐️ XGRIDS PROVIDES LCC PLUGINS AND SDKS THAT SUPPORT UNLIMITED REPLICATION OF 3D CONTENT ACROSS MULTIPLE PLATFORMS.
Official website address: https://xgrids.com/lcc
The AI speech recognition tool Universal-1:38 seconds can process 60 minutes of audio faster than fast Whisper
【AiBase Summary:】
⭐️ Universal-1 provides accurate and robust multilingual speech-to-text capabilities
⭐️ Universal-1 improves accuracy and speaker recognition through timestamp estimation
⭐️ AssemblyAI leveraged the Conformer RNN-T architecture and Google Cloud TPUs to build an efficient Universal-1 model
Product Entry: https://top.aibase.com/tool/universal-1
InstantStyle: Text-generated image style reference for SD to maintain a consistent style
【AiBase Summary:】
⭐️ Content and style are effectively separated, achieved through simple but powerful techniques.
⭐️ The application of CLIP's global features clarifies the decoupling of style and content.
⭐️ Different layers of the network capture different semantic information to achieve better style preservation.
Product Entry: https://top.aibase.com/tool/instantstyle
————
每日midjourney prompt:性感电商模特
Image source note: The picture is generated by AI, and the image is authorized by Midjourney
A beautiful Chinese Girl，Detailed facial details，long hair ，dark hair， wearing a short pink satin dress. She is sitting on concrete stairs in front of a tiled building facade. The lighting is soft and natural， with sunlight casting shadows --ar3:4--v6.0--style raw
A beautiful Chinese girl with detailed facial details, long hair, dark hair, wearing a short pink satin dress. She sat on a concrete staircase in front of the front of the tiled building. The lighting is soft and natural, and the sunlight casts shadows