Former Xiaomi backbone startup "Spiritual Era" raises tens of millions of yuan in funding for embodied intelligence; Zhang Qianchuan, head of MiniMax products, resigns; OpenAI releases the most powerful reasoning model o1丨AI Intelligence Bureau
Latest update time：2024-09-14
Reads：
Domestic Information
DOMESTIC NEWS
Former Xiaomi backbone startup "Lingzu Times" completed tens of millions of yuan in angel round financing
Lingzu Era, a new company producing core components for embodied intelligence, has completed an angel round of financing worth tens of millions of yuan. This round of financing was jointly invested by Yarui Zhiyou Scientist Fund, Inno Angel Fund, and Yiwei Venture Capital. The financing funds will be used for the continuous R&D and iteration of the "full series of integrated joint modules" and team building.
Lingzu Times, founded in November 2023, focuses on the industrial application of robot integrated joints and building a global robot hardware open source platform. The company focuses on high-performance, high-stability, small-volume, and lightweight "new integrated joint modules", and is committed to providing "nuclear power" for the automation industry including robots, accelerating the universal implementation of high-standard products in the industry. In the eight months since its establishment, Lingzu Times has successfully reached cooperation with industry giants such as Xiaomi, ByteDance, JD.com, Meituan, and Joyy Group, and has accumulated orders of more than 10 million RMB.
The core team of Lingzu Times used to work for Xiaomi and has nearly 100,000 robot integrated joints mass production experience. Lingzu Times founder and CEO Wang Bo was the general manager of Xiaomi robot joints and the chief designer of Xiaomi CyberGear; founder and COO Shao Yuanxin was the project leader of Xiaomi CyberDog and CyberGear. (Yiou.com)
Molecular Heart received hundreds of millions of yuan in Series A financing, focusing on AI protein technology platform
Molecular Heart is an AI protein design platform that uses top AI algorithms to bring revolutionary changes to protein discovery, optimization and design, in order to accelerate the entire process of drug development and promote innovation in all fields such as industrial and agricultural production, material design and environmental improvement. Molecular Heart recently completed a round A financing of hundreds of millions of yuan, led by Xienuo Investment and Shenzhen Capital Group, and followed by SenseTime Guoxiang Capital and Jiuyi Investment. This round of financing will be used to further expand the team of compound talents, further improve the AI protein basic model, AI protein optimization design platform MoleculeOS and other technical platforms, and accelerate the industrial implementation and commercialization of AI protein technology. (IT Orange)
MiniMax product manager Zhang Qianchuan recently resigned
Zhang Qianchuan, the product manager of China's large-scale model unicorn MiniMax, has stopped participating in the company's daily operations for personal reasons and has become a product consultant, which is actually the same as leaving the company. Prior to this, MiniMax co-founder Yang Bin also reportedly left the company.
Zhang Qianchuan worked at Baidu and Zhihu, and joined ByteDance in 2016, responsible for small game products, and later served as the head of the user product team of Toutiao. In early 2023, Zhang Qianchuan left ByteDance and announced his entrepreneurship, and eventually joined MiniMax.
At MiniMax, Zhang Qianchuan is mainly responsible for the C-end AI role-playing product "Xingye" and its overseas version "Talkie". According to media reports, Zhang Qianchuan has an important influence on the commercialization strategy of "Xingye". He personally led the team to plan many key gameplays. In mid-2024, he gradually faded out of the decision-making of "Xingye" and turned to participate in the productivity tool "Conch AI". (Intelligence Emergence)
Quark releases new CueMe intelligent conversation assistant, supporting thousands of genres and 20,000 words of long text
Quark has released a new intelligent conversation assistant CueMe, which users can now use through cueme.cn or the Quark App.
CueMe is independently developed based on the Quark large model. It has outstanding AI writing capabilities and supports content generation of different genres and lengths, with the longest content being 20,000 words. Quark said that CueMe has been fully pre-trained with corpus for different sub-genres and can support writing needs of thousands of different genres, including research reports, daily writing, course papers, new media copywriting, etc., and is still expanding.
In addition, CueMe also supports writing style customization, lowering the usage threshold through "common commands" and diversified questioning. (IT Home)
Oracle AI cluster equipped with 131,072 NVIDIA B200 GPUs: 24 teraflops of computing power
Oracle announced the launch of multiple GPU computing clusters that can provide AI training services through Oracle Cloud Infrastructure. The top one is equipped with more than 100,000 NVIDIA Blackwell GPUs.
It uses a total of 131,072,000 B200 GPU accelerator cards, with FP8 floating point and INT8 integer peak performance of up to 2.4 ZFlops. The basic node is the NVIDIA GB200 NVL72 liquid-cooled cabinet, each with 72 built-in GPU accelerator cards. Different cabinets are interconnected through the NVLink bus with a bandwidth of 129.6TB/s. Oracle said that NVIDIA will not mass-produce Blackwell GPUs until the first half of next year. Currently, companies such as WideLabs and Zoom have begun to adopt Oracle's new cluster service. (Fast Technology)
Yuanxiang releases the largest open source MoE model in China
Yuanxiang XVERSE released China's largest MoE open source model: XVERSE-MoE-A36B. The model has a total parameter of 255B and an activation parameter of 36B, achieving a "cross-level" leap in model performance of 100B. At the same time, the training time is reduced by 30%, and the inference performance is improved by 100%, which greatly reduces the cost per token.
MoE (Mixture of Experts) is the most cutting-edge hybrid expert model architecture in the industry. It combines expert models from multiple sub-fields into a super model, breaking the limitations of the traditional scaling law. When the model scale is expanded, the computing cost of training and reasoning will not be significantly increased, and the model performance can be maximized. At the same time, in multiple authoritative evaluations, Yuanxiang MoE has significantly surpassed many similar models, including the domestic 100 billion MoE model Skywork-MoE, the traditional MoE overlord Mixtral-8x22B, and the 314 billion parameter MoE open source model Grok-1-A86B. (Synced Pro)
AI chefs are about to take up their posts, Beijing issues the first robot food business license
It is reported that on September 12, Beijing issued the first "Embodied Intelligent Robot Food Business License", which allows intelligent robots with perception and learning capabilities to successfully enter the catering market and serve consumers.
It is understood that considering the regulatory gap after intelligent robots enter the catering field, the Haidian District Market Supervision Bureau has also taken the lead in the country to formulate management standards and 47 regulatory projects for "catering service intelligent robots" to ensure the safety of robots. By then, the chefs busy frying French fries, fried chicken nuggets, and fried rice in the back kitchen will no longer be real chefs, but robot chefs with AI functions. (Beijing Daily)
It is understood that considering the regulatory gap after intelligent robots enter the catering field, the Haidian District Market Supervision Bureau has also taken the lead in the country to formulate management standards and 47 regulatory projects for "catering service intelligent robots" to ensure the safety of robots. By then, the chefs busy frying French fries, fried chicken nuggets, and fried rice in the back kitchen will no longer be real chefs, but robot chefs with AI functions. (Beijing Daily)
International News
FOREIGN NEWS
OpenAI releases the most powerful reasoning model o1, opening a new chapter in AI general complex reasoning
In the early morning of September 13, the US company OpenAI released a new model technology product o1, including two versions o1-preview and o1-mini. The former has advanced reasoning capabilities and has significantly improved its ability to reason about mathematics, programming, science and other problems, with performance close to that of a Ph.D. in physics, chemistry and biology; the latter is a more compact model optimized for code generation. This is the "Strawberry" project with a powerful model with advanced reasoning capabilities that was previously rumored.
It should be noted that o1's current chat experience is still relatively basic.
Unlike its predecessor GPT-4o, o1 currently cannot browse the web or handle file analysis functions.
Although it has image analysis capabilities, this function is temporarily closed pending further testing.
In addition, o1 has a message limit - currently o1-preview is limited to 30 messages per week, and o1-mini is limited to 50 messages per week.
Starting today, both o1-preview and o1-mini versions are available on ChatGPT Plus/Team and API interface channels, and enterprise and education users will receive priority access early next week.
(Titanium Media AGI)
AI godmother Fei-Fei Li's AI startup World Labs completes $230 million in financing
On September 14, according to foreign media reports, the newly established company World Labs said on Friday (September 13) that it has raised $230 million to build a large world model (LWM) that "perceives, generates and interacts with the 3D world."
The company said in a post on LinkedIn on Friday:
Our goal is to elevate artificial intelligence models from a two-dimensional pixel plane to a complete three-dimensional world - virtual and real world - giving them the same rich spatial intelligence as us.
World Labs was founded by artificial intelligence pioneer Fei-Fei Li and computer vision and graphics technology experts Justin Johnson, Christoph Lassner and Ben Mildenhall, according to the post. In a blog post announcing the company’s founding and funding on Friday, World Labs said that Spatial Intelligence LWM will allow artists, designers, developers and engineers to imagine and create virtual spaces with physics, semantics and control.
The post reads: Over time, we hope to train models that are increasingly powerful and capable, and that can be applied to a variety of fields and work alongside humans. According to the post, World Labs' round of financing was led by Andreessen Horowitz, NEA, and Radical Ventures. (Bianniushi)
French AI programming startup Poolside is reportedly valued at nearly $3 billion
French AI programming startup Poolside is currently negotiating a new round of financing, which will bring its valuation to $3 billion before the release of its first product. People familiar with the matter revealed that the company will raise nearly $500 million in the new round of financing. Poolside's existing investor Bain Capital is negotiating to lead this round of investment. (Titanium Media AGI)
AI robots bring goods: Amazon tests advertising in Rufus chatbot
Amazon has begun experimenting with advertising in its newly launched shopping chatbot Rufus. Amazon said that in the future, Rufus users will see some sponsored ads related to searches during conversations. These ads will not only appear on the conversation interface, but may also be accompanied by text generated by Rufus to help users better understand the content of the ads. Rufus' advertising test aims to increase brand and product exposure, while also hoping to provide users with more abundant shopping options.
This change may affect the future model of online shopping, allowing users to access more potential purchase options while enjoying chat services. (AIbase)
Google launches DataGemma: Improve AI accuracy and reduce hallucinations based on trusted data sources
Google has launched DataGemma, a new version of its open-weight Gemma model based on real-world statistics in Google Data Commons.
Google's Data Commons is a repository that brings together more than 240 billion data points collected by trusted organizations such as the United Nations and the Centers for Disease Control and Prevention. By leveraging this massive statistical data set, DataGemma, based on Gemini, is able to significantly improve model accuracy and ensure that its output is based on authentic and trustworthy real-world information. At the core of the DataGemma approach are two key technologies: Retrieval Interleaved Generation (RIG) and Retrieval Augmented Generation (RAG). These two methods reduce hallucinations by basing the model on real-world data during the generation process. (IT Home)
Apple shares first-ever Image Playground images, created by an executive for his dog
Apple has publicly shown off an example of one of its upcoming Apple Intelligence features, Image Playground, in action for the first time, which generates cartoon-style illustrations based on text prompts. The example image was created by Apple senior vice president of software engineering Craig Federighi to celebrate his dog Bailey's birthday.
Apple describes Image Playground as a "fun" feature that can generate original images in seconds. Apple reportedly plans to launch Image Playground image generation and Genmoji custom emoji generation in iOS 18.2, which is expected to be released in December.
At the same time, Apple plans to launch the first batch of Apple Intelligence features in iOS 18.1, which is expected to be released in October. New features include writing tools for generating text and summarizing text, as well as the ability to provide notification summaries. (IT Home)
Cloudera Launches New AMP to Accelerate Enterprise AI Deployment
Cloudera announced the launch of several new machine learning project accelerators (AMPs) designed to shorten the time to value for enterprise artificial intelligence use cases. The new features are mainly to provide enterprises with cutting-edge artificial intelligence technologies and examples in Cloudera, helping enterprises integrate artificial intelligence and drive more impactful results. AMP is an end-to-end project based on machine learning (ML) that can be deployed with one click directly from the Cloudera platform. Each AMP covers industry-leading practices and can solve complex ML challenges through workflows. The latest AMPs and updates include: Fine-Tuning Studio, RAG using knowledge graphs, PromptBrew, and chatting with documents. (Global TMT)
Google's Gemini AI app adds Gems and managers
Google has added Gems and Gem Manager features to its Gemini AI app on iOS and Android platforms. Users can now view Gems-related chats on the home screen and create and save personalized Gems in the Gem Manager to customize the answers of the Gemini app. This update is designed to improve the user experience and allow users to use AI assistants more effectively. (AI Digital Intelligence Source)
Adobe announces Firefly video AI model will be available this year
Adobe plans to launch the Firefly video model this year, allowing users to generate dynamic videos using text descriptions or image still frames. It is reported that the Firefly video model previewed by Adobe supports the generation of video clips through professional prompts. Users can verbally describe various camera mechanism prompts (such as "50mm focal length, F0.95 aperture") or use various scene description words (such as "basketball players wearing suspenders") to generate content that meets user needs.
Over the past year, the Adobe community has generated more than 12 billion AI images through the Firefly text model. The company plans to further increase the use of the model, while emphasizing that all of its models are only trained on "content that the company has the right to use" and will not be trained using user content. (IT Home)
Meta admits to using public content posted by adult users of Facebook and Instagram in Australia to train AI models
According to Australia's ABC News, Meta admitted that it collected public data of all Australian adults on its platform. However, unlike in EU countries, Australia did not provide users with an "opt-out" option because there are no relevant laws and regulations.
Australian Senator David Shoebridge said, "Meta has made it clear today that if Australia had the same laws, then Australians' data would also be protected. The government's failure to act on privacy protection means that companies like Meta will continue to monetize and exploit children's images and videos on Facebook. (IT Home)
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