Pika Labs, an AI video generation tool founded by girls after 1995, was screened last week. A four-person team received $55 million in financing and a valuation of $200 million after its establishment for half a year. At the same time, the drama of "father depends on women" was staged in A-shares, and the listed company where his father was located gained three daily limit after this tool exploded. The first set of Hanfu ready-to-wear garments designed entirely by AIGC in China appeared at the 17th Hangzhou Cultural Fair, and AIGC has a place in the film and television industry. With the help of unbounded AI, directors can intuitively generate what they want and give it to departments that need communication, which greatly saves communication costs.
finance
Generative AI startup Together AI received more than $100 million in Series A financing.
According to the AIGC open community report, On November 30th, Together AI, an open-source generative AI platform, announced in official website that it had obtained $102.5 million (about 730 million yuan) in Series A financing. Kleiner Perkins led the investment this time, followed by NVIDIA, Emergence Capital, NEA, Prosperity 7 and Greycroft.
Pika Labs, a generative AI startup, raised $55 million and launched Pika 1.0, a video generator.
According to reports, on November 29th, Pika Labs, a generative AI startup, raised $55 million in the previous seed round and seed round financing led by Nat Friedman and Daniel Gross, and the A round financing led by Lightspeed Venture Partners. Other investors include Adam D ‘Angelo (founder and CEO of Quora), Andrej Karpathy, Clem Delangue (co-founder and CEO of Hugging Face and partner of Factorial Capital) and Craig Kallman (CEO of Atlantic Records).
In addition, Pika Labs announced the launch of video generator Pika 1.0. It is reported that Pika 1.0 adopts a brand-new artificial intelligence model, which can generate and edit videos of different styles such as three-dimensional animation, animation, cartoons and movies.
HeyGen, the AI generation tool, completed $5.6 million in financing.
AI Xinzhijie News: On November 29th, HeyGen, an AI generation tool, announced on social platform X that it had obtained a new venture capital of $5.6 million, which was led by Conviction Partners of SarahGuo. The company’s current round of financing is valued at 75 million US dollars. In addition, HeyGen said that its ARR (annual recurring income) has increased from $1 million to $18 million within one year, and launched Instant Avatar(Avatar 2.0).
Biotechnology and AI startup Cradle received $24 million in Series A financing.
According to reports, on November 28th, following the completion of $5.5 million seed round financing last year, Cradle, a startup of biotechnology and artificial intelligence, received $24 million in Series A financing, which was led by investors Index Ventures, Kindred Capital (also a seed round investor) and individual investors such as Chris Gibson and Tom Glocer. The new funds will be used to develop teams and sales.
"Haina AI" completed tens of millions of A round financing, and Lenovo Venture Capital invested exclusively.
According to 36Kr’s report on December 1st, "Haina AI" recently completed tens of millions of A-round financing, which was exclusively invested by Lenovo Venture Capital. This round of financing will be used for talent recruitment, AI model research and development, and marketing system construction. It is understood that "Haina AI" is a vertical AI product of talent recruitment under Beijing Stars Shining Technology Co., Ltd., which was first launched in 2019, specializing in AI interview services to help enterprises complete recruitment interviews with AI technology.
large-scale model
Harbin Institute of Technology (Shenzhen) released the multi-modal large model "Nine Days"
According to the report of the webmaster’s home on December 4th, Harbin Institute of Technology (Shenzhen) recently released a multi-modal large language model named JiuTian-LION. By integrating fine-grained spatial perception and high-level semantic visual knowledge, it achieved the performance of the state-of-the-art in 13 visual language tasks, especially in the Visual Spatial Reasoning task, which achieved a 5% performance improvement.
Zhiyuan research institute officially opened the 70 billion parameter model Aquila2-70B-Expr.
According to 36Kr’s report on November 30th, Lin Yonghua, vice president and chief engineer of Beijing Zhiyuan Artificial Intelligence Research Institute, announced at the 2023 Artificial Intelligence Computing Conference that Aquila2-70B-Expr (Heterogeneous Pioneer Edition), a 70-billion-parameter large model, was officially opened, which was the first large model to complete training based on NVIDIA’s mixed resources and Tian Zhi Xin’s mixed resources.
According to reports, the Aquila2-70B-Expr large model training was completed in NVIDIA mixed resources (A100 cluster+A800 cluster) and Tianzhi Zhixin mixed resources (BI-V100 cluster+BI-V150 cluster) based on FlagScale v0.2.
Inspur information released the 100 billion-level open source model "Source 2.0"
According to the report of titanium media on November 27th, Inspur Information officially released the 100 billion-level open source model "Source 2.0". Source 2.0 adopts LFA (localized filtering-based attention) self-attention mechanism, which can effectively capture local information and short sequence information, so that the model can more accurately grasp the strong semantic association between contexts and learn human language habit paradigm and programming ability.
Fudan Insurance Team released a special model of "Insurance Smart Cool" in the insurance field.
According to the science and technology innovation board Daily reported on December 2nd, Fudan Insurance Red Man Festival and "Insurance Smart Cool" big model conference was held, and the insurance vertical big model "Insurance Smart Cool" developed by Fudan Insurance Team was released at the conference. According to reports, "Insurance Smart Cool" is a specialized large language model in the insurance field that provides professional, intelligent and comprehensive digital services for all kinds of users in the insurance scene.
Alibaba Cloud Open Source Tongyi Thousand Questions 72 Billion Parameter Model
According to the news of Jin Shi on December 1st, Alibaba Cloud Open Source Tongyi asked about 72 billion parameter model Qwen-72B, 1.8 billion parameter model Qwen-1.8B and audio model Qwen-Audio. It is reported that in this open source model of Alibaba Cloud, in addition to the pre-training model, the corresponding dialogue model is also derived synchronously, and a 4-bit/8-bit quantized model is provided for 72B and 1.8B dialogue models, which is convenient for developers to reason and train.
Tencent, Nanyang Technological and other open source charts, alpaca model ChartLlama.
According to the heart of the machine, recently, Tencent, together with Nanyang Technological University and Southeast University, proposed ChartLlama. The research team created a high-quality chart data set and trained a multi-modal large-scale language model focusing on chart understanding and generation tasks. ChartLlama combines multiple functions such as language processing and chart generation, and provides a powerful research tool for researchers and related professionals.
DeepSeek, a subsidiary of Magic Square Quantification, released 67B model.
AI Xinzhijie News: On November 29th, Magic Square Quantization, a well-known private equity giant, officially announced that its new organization "DeepSeek", which explores AGI (General Artificial Intelligence), officially released the universal large language model "DeepSeek LLM 67B" after releasing the Coder code model in early November. The model has been completely open source, and the service has been fully open for internal testing.
According to DeepSeek, compared with the open source equivalent model LLaMA2 70B, DeepSeek LLM 67B performs better in nearly 20 public evaluation lists in Chinese and English, especially in reasoning, mathematics and programming.
Stability AI introduced the Stable Diffusion XL Turbo model.
According to the report of IT House on November 30th, Stability AI recently launched the stable diffusion XL Turbo (SDXL Turbo), which is an improved version of the previous SDXL model, claiming to adopt "antagonistic diffusion distillation technology" to reduce the iterative steps of image generation from the original 50 steps to one step, and it is said that "only one iterative step can generate high-quality images".
It is reported that the biggest feature of the Stable Diffusion XL Turbo model is the above-mentioned "one iteration to generate images", which claims to be "instant text to image output" and can guarantee the quality of images. The experimental results show that the Stable Diffusion XL Turbo can greatly reduce the computational requirements, while still maintaining good image generation quality. The word iteration of this model is better than the LCM-XL with four iterations, while the Stable Diffusion XL Turbo with four iterations can beat the previous stable diffusion XL with 50 iterations. On the A100 GPU, it takes only 207 milliseconds to calculate the image with 512 x512 resolution.
Google released the Translatotron 3 model, which can bypass the text conversion step.
According to IT House news, Google officially introduced and launched a new AI model named Translatotron 3, which can realize simultaneous speech-to-speech translation without any parallel speech data.
Google launched the Translator Otron S2ST system in 2019 and the second version in July 2021. In a paper published on May 27, 2023, it announced that it was deploying new methods to train Translatotron 3. According to the researchers, Translatotron 2 has provided excellent translation quality, speech robustness and naturalness, while Translatotron 3 has realized "the first completely unsupervised end-to-end model of direct speech-to-speech translation".
app; application
Amazon has launched a number of AI tools, including Titan series AI models.
According to venturebeat’s report on November 30th, following the launch of a new chat robot named Amazon Q, an upgraded AI system processor Trainium2, and a partnership with NVIDIA, Swami Sivasubramanian, vice president of data and artificial intelligence of Amazon AWS, announced a series of new AI tools at the re: Invent conference yesterday. It includes three "Titan" series of generative AI models: Titan Image Generator, Titan Text Express and Titan Text Lite. In addition, Amazon Bedrock has also been upgraded, which can provide corporate customers with access to most models on the market, including Jurassic of AI21, Claude 2.1 of Anthropic, Llama 2 of Meta and Stable Diffusion.
The first set of Hanfu ready-to-wear garments designed entirely by AIGC in China appeared at the 17th Hangzhou Cultural Fair.
AI Xinzhijie News, from November 23rd to 27th, the 17th Hangzhou Cultural and Creative Industry Expo was successfully held in Hangzhou. During this period, the first set of Hanfu ready-made clothes designed and created entirely by AIGC in China was unveiled at the exhibition. The design of this Hanfu is derived from the outstanding work "Decorating Rhyme Xizi" of the 2023 Mengxi Cup Song Yun Cultural Innovation Competition, which was created by contestant Li Chao using unbounded AI. The work was inspired by four traditional colors of China-Juyi, Yangfei, Qinglian and Cuiwei, and was created by Unbounded AI and Song Yun Hanfu models with meticulous figures, crossed watercolors and other LORA.
ByteDance launched the large model product "ChitChop" overseas.
According to Tech Planet’s report on November 29th, Byte launched a large model product called "ChitChop" overseas, which was developed and operated by POLIGON, and its overseas social product Helo was also operated by this company. At present, it has launched an independent App and a web version.
It is reported that ChitChop is an artificial intelligence assistant tool, which can provide users with up to 200+intelligent robot services, and serve users’ work and life by providing creative inspiration and improving work efficiency. It is worth noting that this product is similar to Tik Tok Group’s AI product "Goku", which is a collection of AI tools based on Lark Language Model.
Meizu released a large model Aicy AI, which supports many AI functions such as question and answer, painting and so on.
According to the report of Fast Technology on November 30th, Meizu today officially announced the release of Flyme 10.5 system and its own big model Aicy AI. According to reports, Aicy AI is an encyclopedia that can answer questions immediately. Aicy with vast knowledge can answer various questions such as natural science, life knowledge, health knowledge, emotional question and answer. In addition, Aicy also supports creative inspiration to generate paintings, which can generate various styles of paintings such as realistic style, two-dimensional style and ink style. In addition, the gallery also added AI photo function. After the user uploads the photo, AI can generate its photo.
Google DeepMind discovered 2.2 million new crystal materials by using the AI tool GNoME.
AI Xinzhijie News: On November 30th, Google DeepMind showed the AI tool GNoME in the journal Nature, and introduced the related applications of AI in materials science. It is reported that DeepMind has discovered 2.2 million kinds of new crystals using GNoME, of which 380,000 kinds are stable materials, which can be manufactured in the laboratory and are expected to be used in batteries or superconductors. DeepMind claims that it will take 800 years to calculate these materials by manpower alone.
Ali International released three AI design ecological tools.
According to the science and technology innovation board Journal reported on December 1st, at the 6th China International Industrial Design Expo, Ali International released three design ecological tools: Duiyou, PicCopilot and Luban AI. According to reports, these three products have the functions of AI painting, AI model creation, AI image and video processing, and have served hundreds of thousands of businesses and covered 500,000 designers.
Freepik introduces Pikaso real-time AI drawing tool, which adopts LCM drawing technology, and simple lines can be drawn.
According to the news of the webmaster’s home on December 1st, recently, Freepik, a well-known gallery platform, released its innovative product-Pikaso real-time drawing tool, which combined LCM technology and million-level gallery, bringing users a brand-new creative experience. Pikaso adopts the advanced LCM drawing technology, combined with Freepik’s millions of authorized galleries, making online real-time drawing possible.
Korean media: Samsung Galaxy Book 4 series notebooks were released on December 15th, and will support local running of Gauss AI model.
According to Yonhap News Agency reported by IT House, an insider revealed that Samsung Electronics will launch the Galaxy Book 4 series notebook computer equipped with Intel’s next-generation processor Core Ultra on the 15th of this month, claiming to be the world’s first AI notebook.
According to the report, Galaxy Book 4 appeared one and a half months earlier than the previous generation. The reason why I chose to release it earlier this time is because Samsung needs to embody the symbolic significance of this new product as the "first AI notebook". Galaxy Book 4 is expected to be equipped with its self-developed AI model "Samsung Gauss", and the device does not need to transmit the collected information to the central server, which means that the built-in "Gauss" of the notebook will support local operation.
Asus will release the first Intel Core Ultra processor AI notebook
AI Xinzhijie News, ASUS ASUS announced at the Weibo today that the launch conference of 2024 ASUS Core Ultra AI PC thin and light notebook will be held at 15:00 on December 15th, when new ASUS Lingyao series notebooks will also appear. Asus said that this is the first latest Core Ultra laptop.
view
Bill Gates: Generative AI has reached its limit, and the next breakthrough is interpretable AI.
According to the science and technology innovation board Journal reported on November 27th, Bill Gates said that many people inside OpenAI, including Altman, believed that GPT-5 would be significantly superior to GPT-4. But he believes that there are many reasons to believe that the current generative artificial intelligence has reached its limit. The next breakthrough Gates thinks is explainable AI, but it is not expected to be realized until the next decade (2030-2039).
Xu Zongben, Academician of China Academy of Sciences: At present, large-scale model research is far from scientific.
According to the interface news report, Xu Zongben, an academician of China Academy of Sciences, said at CCF China Software Conference that as the trend of the new wave of artificial intelligence development, the revolutionary influence of big model on scientific research paradigm, production mode and industrial model should not be underestimated, and it is an inevitable choice to invest in big model research. However, he also said that the large-scale model research is still engineering, which is far from scientific.
He believes that software will become the first breakthrough field of artificial intelligence. "Software has language, language has grammar, and grammar has strict standards. As long as artificial intelligence can be standardized and has logical boundaries, it can do well in the software field."
Hugging Face Co-creation released a forecast in 2024: Open source LLM will reach the best level of closed source LLM.
AI Xinzhijie News, on November 27th, Clement Delangue, co-founder and CEO of Hugging Face, an AI open source community, made six predictions on the development of the industry in 2024, including: a popular AI company will close down or be acquired at a very low price; Open source LLM will reach the best level of closed source LLM; AI has made major breakthroughs in the fields of video, time series, biology and chemistry; Volkswagen will further care about the economic and environmental costs of AI; Most of the content of a popular media will be generated by AI; 10 million artificial intelligence developers on Hugging Face will not lead to an increase in unemployment.
Lin Yonghua, Beijing Zhiyuan Institute of Artificial Intelligence: There is a three-year gap between the large-scale model training performance of domestic AI chips and that of foreign countries.
According to the science and technology innovation board Journal reported on November 29th, Lin Yonghua, vice president and chief engineer of Beijing Zhiyuan Artificial Intelligence Research Institute, said that at present, only a few of China AI chips are close to NVIDIA A100/A800, and most of them are less than 50%. In addition, there is a huge ecological gap. There are more than 40 AI chip companies in China, but the total market share of AI chips in China does not exceed 10%. The software of each AI chip is different and the ecology is very fragmented.
Ma Yun: The era of AI e-commerce has just begun, which is an opportunity and a challenge for everyone.
According to the science and technology innovation board Journal reported on November 29th, it was learned from many Ali insiders that Ma Yun rarely participated in the discussion and made a speech on Alibaba’s intranet in response to employees’ discussion on Pinduoduo’s financial report and e-commerce last night. Ma Yun said, please give more constructive comments and suggestions, especially innovative ideas. He believes that everyone in Ali today is watching and listening, and he firmly believes that Ali will change and Ali will change. All great companies are born in winter. The era of AI e-commerce has just begun, which is both an opportunity and a challenge for everyone. Ma Yun also said that he would like to congratulate Pinduoduo on his decision-making, implementation and efforts in the past few years. "Everyone has been a cow, but those who can reform for tomorrow and the day after tomorrow, and organizations that are willing to pay any price and sacrifice are respectable. Back to our mission and vision, Ali, come on! "
Yang Likun, chief scientist of Meta, refutes Huang Renxun: Super intelligence will not come soon.
According to the report of IT House on December 4th, NVIDIA CEO Huang Renxun recently announced that Super Artificial Intelligence (AI) will overtake human beings in five years. In this regard, Yann LeCun, the chief scientist of Facebook’s parent company Meta and a pioneer of deep learning, has the opposite view. He believes that super intelligence will not come soon.
Yang Likun said that it will take decades for the current artificial intelligence system to achieve a certain sense of human-like perception. By then, these artificial intelligence systems with common sense will be more powerful, and will no longer be limited to summarizing mountains of texts in a creative way. Regarding Huang Renxun’s point of view, Yang Likun commented: "I know Huang Renxun, the CEO of NVIDIA, who has benefited a lot from the artificial intelligence craze. This is an artificial intelligence war, and it is providing weapons. "
Research report
Research: GPT-4 beat the professionally tuned MedPaLM 2 model in medical problems.
According to the report of Webmaster’s House on December 4th, Microsoft researchers demonstrated the outstanding performance of GPT-4 in medical knowledge testing, especially when combined with advanced hint engineering technology, its performance surpassed that of MedPaLM2, which was professionally adjusted.
The results show that, compared with time-consuming and laborious optimization and model training, applying more effective cue engineering to mainstream general models may be a better way to achieve more accurate results. Medprompt method adopts a variety of hint engineering techniques, including the thinking chain reasoning generated by GPT-4 and generating multiple answers with separate scores, and then returns the answer with the highest score to the user. Although this method will increase the cost of reasoning because more tags are generated, the results show that it may be worth considering to combine the leading general model (such as GPT-4) with advanced hint engineering technology to evaluate the latest performance standards.
According to the research, the energy consumed to generate an AI image is equivalent to fully charging a mobile phone.
According to the news of Webmaster’s House on December 1st, according to a new study conducted by researchers from AI startup Hugging Face and Carnegie Mellon University, every time you use AI to generate images, write emails or ask questions to chat bots, it will cause a certain burden to the earth.
In fact, using a powerful AI model to generate an image consumes the same energy as fully charging a mobile phone. This study first calculated the carbon emissions generated by using the AI model for different tasks. However, they found that the energy consumption of using AI model to generate text is obviously low. Generating 1,000 texts only consumes 16% of the energy equivalent to the charge of the mobile phone.
