Sohu, the fastest AI chip in history, received $120 million in financing; Groq raised funds with a valuation of $2.5 billion; Mac version of ChatGPT is now available for download丨AI Intelligence Bureau
Latest update time：2024-06-27
Reads：
Financing Express
FINANCING NEWS
Groq is raising a new round of financing, with a valuation that could reach $2.5 billion:
According to multiple sources, Groq, an AI chip company founded by Jonathan Ross, a former Google executive and one of the inventors of TPU, may be valued at up to $2.5 billion in a new round of financing led by BlackRock, one of the world's largest asset management companies.
Bright Machines receives $126 million in Series C funding:
Bright Machines is an American developer of robot-driven software. Investors include Nvidia and Microsoft. Other investors in the Series C funding round include venture capital firm Eclipse Ventures, robot manufacturer Jabil, and BlackRock.
Etched receives $120 million in Series A funding:
AI chip startup Etched announced that it has raised $120 million in Series A funding, led by Primary Venture Partners and Positive Sum Ventures, and supported by institutional investors such as Hummingbird, Fundomo, Fontinalis, Lightscape, Earthshot, Two Sigma Ventures and Skybox Data Centers. The funds raised will be used to design and develop Sohu's new AI chip, which focuses on processing the Transformer architecture in AI reasoning.
Function Health receives $53 million in Series A funding:
Function Health is a health data integration company that provides health advice to users based on artificial intelligence and clinician opinions. The financing was led by a16z, followed by Wisdom Ventures and others.
TechWolf receives $42.75 million in Series B funding:
TechWolf provides an AI-based HR technology solution. This round of financing was led by Felix Capital, with participation from 20VC, Acadian Ventures, Fortino Capital Partners, Notion Capital, PMV, SAP, SemperVirens, ServiceNow Ventures, Stride VC, Workday Ventures, and well-known AI leaders such as Deepmind and Meta.
Norm Ai receives $27 million in Series A funding:
Norm Ai is an AI-driven regulatory compliance platform. This round of financing was led by Coatue, with participation from Bain Capital Ventures, Blackstone Innovations Investments, New York Life Ventures, Citi Ventures, TIAA Ventures, and Jefferson River Capital.
VidAU receives angel round financing:
VidAU is an AI video creation platform. This round was invested by River Jin Technology Limited, an AI overseas industry investment company.
(Welcome to add WeChat
AIyanxishe2
to learn more about AIGC and financing, and chat about new AI products with like-minded friends)
Industry News
INDUSTRY NEWS
Domestic Intelligence
DingTalk will be open to all AI large model manufacturers, and the first 7 companies will be connected:
DingTalk President Ye Jun announced that DingTalk will be open to all large model manufacturers to build "the most open AI ecosystem in China". Among them, MiniMax, Dark Side of the Moon, Zhipu AI, Orion Starry Sky, Zero One Everything and Baichuan Intelligence, six large model manufacturers with a scale of hundreds of billions of yuan, have announced their access to DingTalk, becoming the first large models to be connected to DingTalk after Tongyi Qianwen. In the future, users can directly use the products of seven large models including Tongyi on DingTalk.
Huawei releases AI network access "Open City Plan", 10,000 stations in five cities:
The first phase of the plan will empower 1,000 site engineers within six months to manage more than 10,000 sites in five cities, including Hangzhou, Guangzhou, Bangkok, Jinan, and Shenzhen. Huawei has proposed the "Intelligent Network" strategy to build a wireless intelligent body, reshape operation and maintenance, experience, and business, inject intelligence into 5G-A networks, and improve quality and efficiency for operators.
Honor releases the first end-to-end AI anti-fraud detection technology in the mobile phone industry:
Honor CEO Zhao Ming introduced that the technology can autonomously identify the image elements in the user's video call. If AI face-swapping is detected in the video, the user will be warned of the risk.
Chen Danqi's team's latest research creates AI "copyright shield":
Chen Danqi's team built an evaluation suite called CopyCat, which includes a dataset covering 50 popular copyright characters and tools for evaluating the similarity of generated content to copyrighted characters and the consistency of user intent. It aims to avoid copyright infringement issues in AI image/video generation models.
The study found that only relevant keywords or descriptions can trigger the model to generate content that is highly similar to copyright characters. To this end, the research team proposed several strategies that can significantly reduce the risk of infringement, but completely preventing the generation of copyright characters remains a challenge.
SenseTime will release "Daily Update 5.5", which will significantly improve hybrid modal capabilities:
The hybrid modal capability of "Daily Update 5.5" will not only cover images and texts, but will also include multiple functions such as long documents.
Huawei's Wang Tao predicts that AI phones will account for 90% by 2030:
Wang Tao, executive director of Huawei and director of the ICT Infrastructure Business Management Committee, said that the combination of 5G-A and AI has the opportunity to bring about three changes to accelerate traffic growth, including changes in content generation, changes in interaction methods, and changes in mobile terminals. In terms of mobile terminals, he believes that this year's AI mobile phone shipments account for 11%, and are expected to reach 90% by 2030.
ByteDance releases MarsCode intelligent development tool, free for domestic developers
On June 26, ByteDance released Doubao MarsCode, an intelligent development tool built based on the Doubao big model, in Beijing. It has two product forms, namely programming assistant and Cloud IDE. MarsCode has functions such as question and answer, code completion, unit test generation, Bug Fix, etc. It is currently open to domestic developers for free.
Li Dongjiang, head of ByteDance's developer service team and Doubao MarsCode, said that AI is not a "competitor" that replaces developers, but a "good helper" for developers. The team hopes to create a software that improves developers' work efficiency and allows developers to have more energy and time to think and create.
International Intelligence
OpenAI postpones the release of voice assistant function, and the Mac version of the application is open to all users:
OpenAI has postponed the release of the GPT-4o voice assistant feature. This decision was made due to concerns about the safety and effectiveness of the product to ensure that it can handle requests from millions of users. However, OpenAI still plans to launch this voice feature to all paying users in the fall, and is actively developing video and screen sharing features.
At the same time, OpenAI's ChatGPT chatbot application for Mac is officially available for download to all users. The application not only natively supports Mac systems, but also provides a shortcut key launch function. In addition,
OpenAI has also reached a cooperation with Eli Lilly, an American pharmaceutical giant.
Generative AI will be used to develop new antibacterial drugs to address the problem of drug-resistant pathogens.
Challenging NVIDIA! The fastest AI chip in history, Sohu, was born, with inference performance 20 times higher than H100:
The fastest Transformer chip in history was born. Using Sohu to run Llama 70B can generate up to 500,000 tokens per second. The Sohu chip is designed for Transformer model reasoning acceleration. Its reasoning performance is ten times higher than B200 and twenty times higher than H100. 1 Sohu ≈ 20 H100 ≈ 10 B200. Etached, the company founded by these 00s guys after dropping out of Harvard, announced another $120 million in financing.
It is reported that Apple's A18 processor NPU performance is stronger than the M4 chip:
It is reported that Apple's iPhone 16 series will be equipped with A18 series processors, and the NPU performance is expected to surpass the M4 processor. The A17 Pro processor has an NPU performance of 35TOPS. The M4 processor is based on TSMC's second-generation 3nm process, has 28 billion transistors, and the NPU computing power is increased to 38TOPS.
Apple's new visual model 4M-21 handles 21 modes:
The 4M-21 vision model jointly developed by Apple and EPFL can process 21 modalities, including images, text, and structured data, improving cross-modal retrieval and generation capabilities; the model achieves unified processing by performing specific discrete tokenization on different modalities, and jointly trains on multiple data sets to enhance performance and adaptability.
Reddit issues ultimatum to AI companies:
Reddit plans to update its robot exclusion protocol to block unauthorized automated scraping of the platform. It won’t affect “good faith actors” like web archives and researchers, but appears to be a response to AI companies like Perplexity circumventing the robots.txt protocol, and Reddit wants all companies using automated proxies to access the platform to comply with its terms and policies.
ElevenLabs launches AI text-to-speech app for iOS:
AI voice startup ElevenLabs has launched a text-to-speech app, ElevenLabs Reader, which uses AI to convert various text contents, such as articles, PDF files, ePub, etc., into natural, smooth, high-quality speech. The app is currently free to download and use for three months for iOS users (Android version to be launched), initially supporting English, and plans to expand to more than 29 languages in the future (the Chinese iOS version is expected to be launched on July 11).
Output value exceeds $100 billion for the first time, AI will drive server GPU shipments to 4.82 million units in 2024:
The DIGITIMES Research Center report pointed out that the global server GPU output value will exceed $100 billion in 2024, of which high-end server GPU output value will account for more than 80%, and shipments will reach 4.82 million units, with Nvidia accounting for 92.5% and AMD accounting for 7.3%. Generative AI is still in its early stages of development, and cloud server suppliers and global large companies are actively deploying, further increasing Nvidia's shipments.
Signapse uses AI sign language to reshape the world of communication for the deaf:
Signapse's core advantage is to provide highly realistic sign language translation services that can convey emotions and make the deaf feel understood and respected. After the user enters the text, the AI system converts it into a sign language video. After GAN technology optimization, the user can receive high-quality sign language translation services. This process not only ensures the accuracy of the translation, but also ensures the natural and smooth movement of the sign language.
Japan has developed a human face with living skin, and its smile is hard to describe:
Japanese researchers have developed a new type of artificial skin that can repair itself and simulate the healing process of human skin. This skin is fixed to the robot using V-shaped hooks by drilling tiny holes in the robot's skeleton to keep the skin smooth and flexible. In addition, the researchers have reconstructed the way human skin changes when smiling, and achieved the "smile" effect by connecting a sliding silicone layer between the artificial skin and the robot's face.
Akamai reports that bots account for 42% of all Internet traffic:
According to a report by Akamai Technologies, robot traffic accounts for 42% of total Internet traffic, of which 65% is malicious traffic. The e-commerce sector is the most affected. Although some robot traffic is beneficial to enterprises, it has a negative impact on user experience overall. It is mainly used for web crawlers, information collection, and the creation of counterfeit websites.
Udio responded to the lawsuit filed by the record company, saying that the model does not copy copyrighted works:
Udio
????
In response, they said that their music model is learned from a large amount of recorded music, with the goal of developing an understanding of musical ideas and generating music that reflects new musical ideas. They are not interested in copying the content in the training set, and have implemented state-of-the-art filters to ensure that the model does not copy copyrighted works or artists' voices. They believe that generative AI will become the mainstream of modern society.
Sora's first commercial film debuts in Cannes! 3 million netizens watched:
Sora's first commercial film, "The Origin of Toys R Us," was produced by director Nik Kleverov using OpenAI's Sora technology and attracted the attention of nearly 3 million netizens. The film incorporates the slow-motion effect that Sora loves. Some netizens questioned its authenticity and consistency, and questioned AI-generated commercial videos.
More international information
Civitai joins the Open Model Initiative:
As a community-driven project for the open source development of AI models for image, video, and audio generation, Civitai has joined forces with organizations such as Invoke, ComfyOrg, LAION, and others to collaborate on the development of open source AI model technology that meets standards.
Google launches Gemini AI sidebar for Gmail, other apps:
Google is rolling out the Gemini side panel for Docs, Sheets, Slides, Drive, and Gmail, and has also launched Gemini for the Gmail app on Android and iOS.
Notion released a new feature, sites:
This feature allows users to publish Notion pages directly as websites. It is free and supports theme customization, website icon settings, Google Analytics and other features. It lowers the threshold for building a website, but the degree of customization may be insufficient compared to large model technology.
ChatGPT writing style has penetrated more than 10% of scientific abstracts:
AI text generation has led to the increase of certain style vocabulary, which has an impact on scientific writing. About 15% of abstracts in PubMed subgroups in countries such as China and South Korea are generated using ChatGPT.
AI Products
AI PRODUCTS
Product Hunt hot list, customer feedback analysis Platform Insights Hub
Survicate's Insights Hub is a powerful customer feedback analysis platform that helps companies quickly understand customer satisfaction, effort, and loyalty by integrating multi-channel feedback, automatically classifying, and extracting key insights. In addition, the tool supports a variety of survey templates and integration options, and the AI assistant function can mine valuable information from existing feedback. Survicate attaches great importance to data security and ensures that customer information is strictly protected.
????https://survicate.com/insights-hub/?ref=producthunt
HuggingFace Hot List, Python Lexical Search BM25S:
BM25S is a Python lexical search library that is based on the BM25 algorithm and combines scipy's sparse matrix technology to achieve high-performance search. BM25S strikes a balance between the complexity of Elasticsearch and the ease of use of Rank-BM25, providing users with a fast and easy-to-use tool. One of its highlights is its tight integration with Hugging Face Hub. In addition, it supports multiple BM25 variants, enhancing its applicability. It is worth noting that BM25S is not intended to replace existing search tools, but as a supplementary option, bringing new possibilities to the field of lexical search in the Python ecosystem.
????https://huggingface.co/blog/xhluca/bm25s
Developer Recommendation
1. revid.ai makes your short videos go viral
revid.ai is a platform that helps users quickly create engaging short video content for mainstream social media such as YouTube, Instagram, and TikTok. Its core is to provide users with creative inspiration and script suggestions by analyzing thousands of popular short videos, helping them to create highly engaging works. Users only need to provide text or links, and AI can automatically generate complete short videos, adding sound effects, animations and other elements.
????https://www.revid.ai/?ref=producthunt
2. AutoStudio: AI-driven multi-round interactive image creation
The AI framework does not require training and integrates a large language model and stable diffusion technology. Through the collaborative work of four major intelligent agents (theme manager, layout generator, supervisor and painter), it creates a coherent multi-theme image sequence. The innovative P-UNet structure and theme initialization technology enhance the theme perception ability, retain details, and improve the consistency of the image.
????https://howe183.github.io/AutoStudio.io/
3.Claude Chinese prompt library
The Chinese version of the Anthropic prompt library provides a variety of AI-generated prompts optimized for different tasks and needs, including applications in personal and business fields, as well as custom prompts submitted by users.
????https://docs.anthropic.com/zh-CN/prompt-library/library
4. Video-Infinity: can speed up the generation of long videos by 100 times
Video-Infinity proposed a distributed reasoning pipeline that uses multi-GPU parallel processing to solve the bottleneck problem of long video generation. The pipeline uses two core technologies, clip parallelism and dual-range attention, to efficiently generate long videos in a distributed manner. With the setting of 8 Nvidia 6000 Ada GPUs (48G), it only takes 5 minutes to generate a 2300-frame long video, which is 100 times faster than before.
????https://video-infinity.tanzhenxiong.com/
5. Four new features in ComfyUI:
They are GTS, iPNDM, ComfyUI-ODE and CFG++. GITS is a new scheduler based on AYS for cases where the number of steps is less than or equal to 20. iPNDM is a new sampler based on the Adams-Bashforth method, with two versions: ipndm and ipndm_v, the latter supports variable step size. Although studies have shown that ipndm performs better in experiments, the author's personal tests tend to favor ipndm_v. ComfyUI-ODE is a custom node that adds an additional ODE parser to ComfyUI, suitable for SD3 and SDXL generation. CFG++ is integrated into ComfyUI as a SamplerEulerCFG++ node to improve CFG.
????https://www.reddit.com/r/StableDiffusion/comments/1dohy20/quick_overview_of_some_newish_stuff_in_comfyui/
6.MegActor: AI portrait animation
MegActor, the open-source AI portrait video generation framework of Megvii, solves the identity leakage problem with an innovative synthetic data generation framework, uses foreground-background segmentation and CLIP coding technology to maintain the stability of the background, eliminates the impact of facial details on the final effect, and enables the model to be trained only on public datasets. After 200 hours of V100 training, MegActor achieved results comparable to commercial models.
????https://megactor.github.io/
Stay tuned for the latest updates tomorrow!
AI Intelligence Bureau is looking for intelligence partners to collect exclusive valuable clues! If you can provide information about the latest AI achievements, industry insider information, and unique products, please add the operation WeChat account:
AIyanxishe2
and note the industry position.
//
Recent Hot Articles
京公网安备 11010802033920号