Financing Bulletin
Formation Bio宣布完成3.72 亿美元D轮融资：Formation Bio（原名为TrialSpark）由华人Benjamine Liu和Linhao Zhang共同创立，是一家专注于将人工智能应用于药物临床试验的创业公司。 本轮融资由Andreessen Horowitz 领投，老股东赛诺菲、红杉资本、Thrive、Emerson Collective、Lachy Groom，新投资SV Angel Growth和FPV Ventures等共同参投。
Hebbia raises nearly $100 million in Series B funding: Hebbia, a startup that uses generative AI to search large documents and return answers, values the company at between $700 million and $800 million and is led by Andreessen Horowitz.
Ng plans to continue to raise $120 million for his AI fund: the AI fund provides funding in the seed and Series A phases of the company's lifecycle, helps teams work in stealth, and connects Ng's professional network. The AI Venture Fund II will be smaller in size than the first round.
Non-stop Technology has received nearly 100 million yuan in A+ round of financing: Non-stop Technology is a 2B food robot RaaS service provider, focusing on AI digital kitchen solutions, and creating an online robot Botin Bota, which can analyze the operation data of catering stores in real time and provide guidance for its refined operation. This time, it was led by Huashan Capital, with the participation of the old shareholder Professor Gao Bingqiang Ecosystem Fund Future Technology, and the old shareholder Professor Li Zexiang's Clear Water Bay Fund and Professor Gan Jie's Zhixing No. 1 Fund continued to exceed the quota for three consecutive rounds. Centron Capital acted as the exclusive financial advisor for the follow-on financing. This round of funds will be used for product development and global delivery.
Enzyme Technology Receives Tens of Millions of Yuan in Angel Round Financing: Enzyme Technology uses AI technology to reduce the cost of enzymatic DNA synthesis. This round of financing was led by Linear Capital, followed by MiraclePlus and Danen Capital.
Shuocheng Technology received tens of millions of yuan in C1 round financing: Shuocheng Technology focuses on providing intelligent manufacturing services such as predictive maintenance of equipment and intelligent operation and maintenance through full-perception intelligent hardware and AI algorithms. The investor is Binfu Capital.
Ruichi Information Receives Strategic Investment: Ruichi Information is a high-tech enterprise focusing on the research and development of Android cloud, AI edge computing, cloud infrastructure and other products and solutions, based on ARM technology and unique software and hardware architecture design, to provide customers with cloud computing and big data as the core of products and solution services.
Constructor Secures Series B Funding: Constructor is a U.S.-based e-commerce service provider that provides AI search software, product discovery platforms, and recommendations and recommendations to the e-commerce industry.
Ora Closes $20 Million Funding Round: Ora Lab is an AI-powered blockchain project provider that aims to integrate AI into decentralized applications through its "on-chain AI oracle." Polychain, HF0 and Hashkey Capital participated. The new funding will allow the project to "continue to develop its technology and infrastructure to tokenize AI models and introduce decentralized AI to the Ethereum ecosystem."
MagicSchool AI 获 1500 万美元 A 轮融资：MagicSchool AI是一家AI教育技术平台。 本轮融资由Bain Capital Ventures领投，其他投资者包括Adobe Ventures和Common Sense Media、Replit创始人Amjad Masad、Clever联合创始人Tyler Bosmeny和Rafael Garcia，以及OutSchool联合创始人Amir Nathoo。
Yingteng Completes Angel Round Financing of Millions of Yuan: Yingteng is an AI technology research and development company, mainly engaged in AI basic software development, artificial intelligence application software development, computer system services and other businesses. The financing was led by Beijing Jilu, followed by a number of other investment institutions. The financing funds will be mainly used for AI technology research and development, scenario application deepening and market expansion.
(Welcome to add WeChat AIyanxishe2 to learn more about AIGC, financing, and chat with like-minded friends about the latest AI products)
Domestic Intelligence:
Baidu Wenxin Model 4.0 Turbo was released, with 300 million Wenxin Yiyan users:
At Baidu's WAVE SUMMIT deep learning developer conference, Haifeng Wang, chief technology officer of Baidu, announced the release of Wenxin Model 4.0 Turbo. This new version significantly improves response speed and retrieval capabilities. At the same time, the user scale of Wenxin Yiyan has reached 300 million, and the highest number of calls in a single day has reached 500 million, which is due to the strong support of the paddle platform.
Baidu launched the intelligent code assistant "Wenxin Quick Code", based on the Wenxin model, to achieve the scenario application of "help you think, help you write, and help you change", accelerate the development speed, and improve the speed of business iteration. Eighty percent of Baidu's tens of thousands of engineers are already using Wenxin Quickcode, and the code adoption rate has reached 44 percent. Wenxin Express supports more than 100 major programming languages and multiple IDEs, and is available in four versions, and Baidu promises not to store or analyze user code, ensuring data security, and ensuring that user code snippets are not used for other users' suggested code in accordance with privacy agreements.
Huawei and the Guangdong Meteorological Bureau promote AI weather applications and release the Galaxy AI network solution to lead the Net5.5G intelligent cloud network.
Huawei and the Guangdong Meteorological Bureau signed a framework agreement to deepen cooperation, aiming to jointly promote the application of modern technologies such as high-performance computing and artificial intelligence in the meteorological field.
In addition, Huawei launched the Galaxy AI network solution for Net5.5G intelligent cloud network. Addressing the complexity of network O&M, service experience assurance, and network security protection in the AI era, the solution introduces AI technology to empower the network to achieve L4 autonomous driving networks and highly intelligent ubiquitous security protection.
Zhihu released a new AI product "Zhihu Direct Answer", which supports functions such as asking questions and searching:
"Zhihu Direct Answer" is the productization of Zhihu's AI search function, which has been launched on the PC side. Based on the Q&A data of Zhihu creators, the product can provide "brief" and "in-depth" answer generation results, and support "finding content" and "finding people".
Duix, a silicon-based intelligent open-source AI digital human interaction platform, can quickly create and deploy realistic digital humans:
The platform is designed to help developers simplify the process of creating and deploying intelligent digital humans, providing a wealth of tools and support to deploy digital humans on a variety of end devices without the intervention of technical teams. Users can download a variety of digital human models, suitable for different industry needs, and the project has been open-sourced, which is convenient for developers to carry out secondary development and personalized customization. Functions include voice input, voice output, real-time interaction, and multi-terminal support. In addition, the platform also provides virtual assistant services, which can be applied to passenger services, automated customer service, intelligent consulting services, virtual hosts and other scenarios to improve user experience and service efficiency.
Zhiyuan Zhang Hongjiang said that AI systems should never deceive humans:
Zhang Hongjiang, founder and first chairman of the Beijing Academy of Sciences, said that "AI systems should never replicate and improve themselves," said Zhang Hongjiang, founder and first chairman of the Beijing Academy of Sciences, about the importance of international cooperation in AI assurance, as well as the opportunities and challenges facing AI technology in China. This red line is very important. When a system has the ability to replicate itself and improve itself, it gets out of hand. The second is deception. AI systems should not have the ability to deceive humans. ”
Kai-Fu Lee said that if AI takes my job, I can endorse the advertisement of hair growth supplement:
Kai-Fu Lee, CEO of Zero One Everything and chairman of Sinovation Works, said that his hair has not become less for so many years, and if AI replaces his job, he can endorse the advertisement of hair growth supplement. "Steve Jobs said don't make career plans in life, the world changes too fast, just follow your heart at every important decision. And this era of artificial intelligence has been the era that I have longed for for more than 40 years, and when I got my PhD at the age of 26, I didn't expect to do such a bold thing at the age of 62, and that is because the time is coming. I think AI is the ultimate human understanding of themselves. Kai-Fu Lee said that AI intelligence is expected to catch up with the doctor next year, and the singularity will explode in a few years.
International Intelligence:
OpenAI CriticGPT code review model released, model annualized revenue of more than $1 billion, and reached a strategic cooperation with Time magazine:
OpenAI has launched a new CriticGPT model, which is built on GPT-4 and is specifically designed to review and identify errors in code generated by large language models, such as ChatGPT. CriticGPT uses reinforcement learning technology from human feedback to significantly improve the accuracy and efficiency of code review. Not only can it identify potential issues in the AI output, but it can also provide explanations to help developers improve the quality and security of their code.
On the business side, OpenAI has achieved annualized revenue of about $1 billion as of March by selling access to its AI models, according to the latest internal data from OpenAI and Microsoft. This achievement marks OpenAI's leapfrog in AI model sales over tech giant Microsoft, whose Azure OpenAI Service until recently reached the same annualized revenue level.
In addition, OpenAI has a multi-year content licensing agreement with the world-renowned Time magazine. OpenAI will be able to access Time magazine's archives and articles from over the past 100 years, which will be used to train OpenAI's AI models, such as ChatGPT.
Perplexity was instructed to use misinformation to scrape website data in violation of regulations, leading to an Amazon AWS investigation:
Startup GPTZero has found that more and more of the sources Perplexity links to are AI-generated, and even use outdated and incorrect information from those sources. Perplexity claims that its answers come only from "reliable sources". In addition, Amazon AWS is investigating Perplexity AI for allegedly scraping data using a bot hosted on AWS servers without permission and allegedly violating the robots.txt exclusion protocol.
Amazon hired the founder of Adept to increase the layout of AGI research and development:
The two parties have reached a technology licensing agreement that will see the co-founders and some of the team members of Adept join Amazon. Adept is focused on developing AI "agents" capable of performing a variety of software tasks, and its technology will support Amazon's expansion in the generative AI space. Adept is not shutting down business. Zach Brock, head of engineering, will take over as CEO, and the company will refocus on "agent intelligence-enabled solutions." Co-founder and CEO David Luan will work under the leadership of Rohit Prasad, a former Alexa lead who now leads the new AGI team. Adept is reported to have raised more than $415 million in funding and is valued at about $1 billion.
Rain AI Recruits Apple Chip Experts to Accelerate AI Chip R&D:
American chip startup Rain AI has poached Jean Didier Allegrucci, a chip executive who worked for Apple for 17 years. He will serve as the company's head of hardware engineering, leading the development of the next generation of breakthrough energy-efficient chips. Three weeks ago, Rain AI hired Amin Firoozshahian, Chief Architect of Meta's ASIC architecture team.
Fields Medal Winner Personally Tested GPT-4o, Classic River Crossing Problem Solving Failed:
Fields Medalist Timothy Gowers reveals how large language models make mistakes in dealing with the classic wolf-goat-cabbage crossing conundrum. Gowers came up with the "ratio" as a new benchmark, pointing out GPT-4o's mistakes on the simplest of topics, and Claude 3.5 was not immune. This phenomenon raises questions about whether large language models are really capable of reasoning and planning. Gowers also pointed out that current methods for evaluating large language models are flawed, proposing CheckMate, an interactive evaluation platform, and MathConverse, a scoring dataset. While large language models excel on a variety of benchmarks, they fail on real-world math and reasoning problems.
谷歌DeepMind开源AI模型 Gemma 2，单 A100 / H100 GPU 可运行：
Gemma 2 is available in 9 billion (9B) and 27 billion (27B) parameter sizes. The 27B model was trained with 13T tokens, and the performance was comparable to that of the mainstream model with twice the scale. 9B is 8T tokens, surpassing Llama 3 8B and others, all with 8192 contextual windows that are available in Google AI Studio. Gemma 2's architecture is designed to run fast on a wide range of hardware, including Google Cloud TPU hosts, NVIDIA A100 or H100 GPUs.
The 2.6 billion parameter (2.6B) model will be released soon, small enough to run locally on mobile phones.
Developer ecstasy! Meta's newly released LLM Compiler achieves 77% auto-tuning efficiency:
This is a compiler optimization tool based on large language models. The tool enhances its understanding of compiler intermediate representations, assembly language, and optimization techniques by training on a corpus of 546 billion tokens of LLVM-IR and assembly code. In testing, LLM Compiler achieved 77% of the optimization potential of auto-tuning searches, significantly reducing compilation time and improving code efficiency. In terms of disassembly, the LLM Compiler is able to convert x86_64 and ARM assembly code back to LLVM-IR with a 45% success rate for round-trip disassembly.
Resemble AI 发布下一代深度伪造检测模型 Detect-2B，准确率高达94%：
Using a series of pre-trained submodels and fine-tuning techniques, the model is able to perform an in-depth examination of an audio clip to determine if it is AI-generated. Detect-2B is able to correctly detect deepfake audio in six different languages with at least 93% accuracy, and is able to predict the AI-generated probability of the audio, eliminating the need to retrain the model each time you listen to a new clip. Detect-2B's submodel consists of a frozen audio representation model and adaptive modules inserted into key layers that distinguish between real audio and AI-generated audio by identifying unexpected sounds left behind in recordings.
OpenAI CEO 奥特曼预测AGI或在十年内实现全球 GDP 翻倍：
Sam Altman emphasized that the development of AI is not a one-time thing, but a gradual process. Despite the widespread attention that ChatGPT has received from its launch, most applications have yet to change radically. In the coming years, the changes will be even more significant as more applications are built on top of AI models. Altman predicts that AGI could double global GDP, which will be a huge productivity driver. He believes that as people adopt these tools, AI will bring unprecedented economic and social benefits.
Microsoft AI executives say content on the open web is free to copy, but there are gray areas:
When asked if "AI companies are actually stealing global intellectual property", Microsoft AI CEO Mustafa Suleiman replied: "I think the social contract for content that is already on the open web has been fair use since the 90s. Anyone can copy and recreate...... That's what 'free software' is, that's how I understand it. "There's also a separate case when a website, publisher, or news organization makes it clear that it can't scrape or crawl its content for any reason other than indexing it for others to find it." It's a gray area that I think will gradually be addressed in the courts. ”
Support open source! Zuckerberg slammed closed-source competitors for trying to "create God":
Zuckerberg is convinced that there won't be "one" for AI in the end, and he emphasized the value of open source – putting AI tools in the hands of more people. "I don't think AI technology is something that should be kept private — then only one company can use it to build a centralized, monolithic product that they want," he said. Zuckerberg said that in the development of AI, it is necessary to create many different AIs to reflect people's different interests. When it comes to companies that are building closed-source AI platforms, he doesn't see it as the way to create the best experience for people.
Bill Gates urges environmentalists not to "worry too much" about AI electricity consumption:
In response to the current problem of accelerated energy consumption caused by AI systems, Bill Gates launched a "defense", arguing that AI technology will eventually "offset" its power consumption, and not to "worry too much" about the huge amount of power needed to run the next generation of AI systems, as big tech companies such as Microsoft compete to invest tens of billions of dollars in large new data centers.
More international information:
The nation's top 5 machine learning PhDs post that labs don't have H100 GPUs: GPU resources in academia are unequal, and many researchers need to compete for computing resources. Universities like Princeton and Harvard have a large number of H100 GPUs, while other institutions may not have a single light. PhDs in the same lab even often need to compete for GPUs.
The first Sora-like open-source reproduction solution to generate AI-generated video on NVIDIA RTX 3090 graphics cards: Open-Sora can generate text-based video on NVIDIA RTX 3090 GPUs, up to 240p resolution and up to 4 seconds long. It takes about 2 seconds to generate a 30-second video and a 4-second video about 60 seconds. For a 424x240 output, a 4-second video output is close to 10 million pixels.
Microsoft's $13 billion investment deal in OpenAI faces EU scrutiny: Regulators will ask more questions about Microsoft's competitors and customers about Microsoft's exclusivity clauses with OpenAI and whether they could have a negative impact on competition.
AI helps college students get higher scores and be hard-to-spot: Professor Scarfe's team used GPT-4 to generate test answers and submit them on behalf of 33 fake students. Scored unknowingly, AI-generated answers to undergraduate psychology coursework go undetected 94% of the time, and the average score is higher than the student's true score.
Smart glasses with GPT-4o and camera are coming: Solos will launch a smart glasses called AirGo Vision, which is equipped with OpenAI's GPT-4o AI model and camera that is able to recognize objects and answer users' questions about the items they are looking at, is compatible with Google Gemini and Anthropic's Claude AI models, and has LED notification lights, the specific price and release date have not yet been announced.
Audi and ChatGPT are working to inject intelligent voice assistants into 2 million cars: the service is expected to launch in July. Car owners will be able to interact with their cars through natural language, enabling voice control of infotainment, navigation, and air conditioning systems. ChatGPT is even capable of answering general intellectual questions.
Product Hunt 热榜，AI 智能化 API 客户端ApyHub Fusion
ApyHub Fusion is an innovative API client that integrates AI technology and aims to revolutionize the API development process. The platform borrows from Notion's intuitive interface design, integrating the process of building, testing, and documenting APIs into a single intelligent workspace.
Fusion's core strength lies in its AI-driven intelligence that anticipates user needs and optimizes workflows. It enables real-time team collaboration and simplifies the complexities of API development. The platform's modular test system and seamless documentation integration capabilities greatly improve development efficiency. Fusion is compatible with data import from major API clients and supports multiple platforms, including MacOS, Windows, and the web.
?https://apyhub.com/product/fusion?ref=producthunt
GitHub Trending 热榜，AI短视频生成MoneyPrinterTurbo
MoneyPrinterTurbo is an open-source project based on a large AI model that aims to automate the short video generation process. The tool supports one-click generation of high-definition short videos by entering topics or keywords, including video copy, materials, subtitles, and background music.
The project provides a web interface and API interface, and supports a variety of speech synthesis services and subtitle generation modes. It is based on the MoneyPrinter project refactoring, adding new features such as video transition effects and length options. It can be deployed via Docker or manually, or you can use a one-click start package for quick experience.
?https://github.com/harry0703/MoneyPrinterTurbo
Developer Recommendation, Gif Author Creates Meme Meme Generator Takes Twitter by Storm:
Fabian, the founder of Glif, built a meme meme generator in a few minutes, which can quickly generate humorous and sharp memes, becoming a new generation of "mouth substitutes". The AI-transformed memes show a new form of creative expression, and users can quickly create personalized memes through Glif.
?https://glif.app/@Hanwei/glifs/clxv3atsq00009wq1iwsmw1ks/source
Cutting-edge technology
1.Visual Sketchpad ：AI 视觉推理能力
The University of Washington, the Allen AI Institute, and the University of Pennsylvania have teamed up to launch Visual Sketchpad, an innovative framework for empowering multimodal language models with visual reasoning. At the heart of the project is the combination of multimodal language models with visual drawing capabilities, enabling them to generate assisted sketches for more effective thinking and problem-solving when solving visual reasoning tasks such as geometric problems, computer vision tasks, etc. Unlike previous paradigms of text-chained reasoning and tooling, Visual Sketchpad allows models to draw lines, boxes, markers, and more, closer to the human way of sketching, aiding in the inference process.
?https://visualsketchpad.github.io/
2. DigiRL Device Control AI Agent Training Method
Researchers at the University of California, Berkeley, and others have launched DigiRL, an innovative autonomous reinforcement learning method designed to train device-controlled AI agents in real-world environments. The technology significantly improves the performance of AI in complex graphical user interface control tasks by fine-tuning pre-trained visual language models in two stages. Key highlights include:
• Offline reinforcement learning is used to initialize the model, followed by offline-to-online reinforcement learning
• Build a scalable, parallelized Android learning environment with a VLM evaluator
• On the Android-in-the-Wild (AitW) dataset, the success rate of a 1.5B parameter VLM trained on DigiRL increased from 17.7% to 67.2%
• Significantly outperformed existing best practices, including AppAgent with GPT-4V (8.3%) and CogAgent with 17B parameters (14.4%)
?https://digirl-agent.github.io/
3.SciPhi-AI推出了开源RAG引擎R2R
R2R is a tool designed to bridge local LLM experiments with scalable, production-ready retrieval enhancement generation. It provides developers with a comprehensive and up-to-date RAG system, built around RESTful APIs for ease of use. R2R supports multimodal input, including text, files, images, audio, etc., while also providing hybrid search, graphical RAG, application management, client-server interaction, configurability, extensibility, and more.
?https://github.com/SciPhi-AI/R2R?utm_source=uwl.me
4. Director3D：文本到3D生成技术
Xiamen University, Shanghai Artificial Intelligence Lab has launched Director3D, a powerful open text-to-3D generation framework designed to generate real-world 3D scenes and corresponding camera tracks. By using the Trajectory Diffusion Transformer to model the camera trajectory distribution for the text description, and the Gaussian-driven Multi-view Latent Diffusion Model to model the image sequence distribution for a given camera trajectory and text, Director3D is able to produce high-quality 3D scenes that are consistent with the text description. In addition, Director3D further optimizes and refines the generation of 3D scenes by introducing SDS++ losses.
?https://imlixinyang.github.io/director3d-page/?utm_source=uwl.me
5.StreamingT2V: A new breakthrough in AI long video generation technology
Georgia Institute of Technology in Oregon and UIUC have launched StreamingT2V, the latest version of the open-source project that supports the generation of high-resolution long images to video, providing users with two frame rate options: 24fps and 12fps. This technology utilizes the Conditional Attention Module and the Appearance Preservation Module to achieve consistency between video clips and long-term scene feature preservation, capable of producing high-quality videos up to 2 minutes long.
The project uses a random mixing method that allows the video enhancer to be applied continuously in an autoregressive process, resulting in an infinite length of video. Experiments show that StreamingT2V has excellent performance in generating high-motion videos, which solves the problem that existing methods are easy to cause video stagnation. This technology is highly generalizable and is not limited by a specific text-to-video model.
?https://streamingt2v.github.io/
Big bull insights
Andrej Karpathy's Presentation: From Academia to Social Reinvention
Speaking at the UC Berkeley Hackathon, Andrej Karpathy noted that the AI field is undergoing unprecedented transformation, from small-scale academic discussions to impacting the entire socio-economic landscape. Karpathy emphasized that large language models are becoming the new computing core, similar to the role of traditional CPUs. He predicts that AI technology will expand from the digital realm to the physical world, profoundly impacting infrastructure. In the future, multiple AI entities may collaborate to complete tasks and automate a large number of tasks. Karpathy also draws on the sci-fi films "Her" and "I, Robot" to explore the potential direction of AI and the ethical and societal challenges it poses.
? https://www.youtube.com/watch?v=Tmrq914yLck
Stay tuned for updates tomorrow!
The AI Intelligence Bureau is recruiting intelligence partners to gather exclusive value clues! If you can provide information about the latest achievements of AI & industry insiders & unique products, please add the operation WeChat ID: AIyanxishe2 to note the industry position.
Leifeng.com
![On implantation, who is so unscrupulous as "Zhang Weiguo's Summer"? Advertising everywhere but no one is disgusted[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)