news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Today's Financing Express

iFLYTEK to invest HK$400 million in Hong Kong and establish international headquarters

iFlytek has announced a five-year HK$400 million investment plan and has established its international headquarters in Hong Kong. The company said the investment plan will support it in forming a 150-person R&D team focused on the development of large language models and AI applications in areas such as intelligent speech, education and healthcare. iFlytek Vice President Duan Dawei said: "Our initial budget is HK$400 million. If everything goes well in Hong Kong, this number will increase." (South China Morning Post)

Bangke Intelligent received RMB 120 million strategic investment, and Tianyang Technology invested

Bengke Intelligence is an artificial intelligence company committed to innovation. It has multiple basic large language models, including general large models, code large models, multimodal image and text large models, multimodal speech large models, etc., which are suitable for rapid deployment by domestic enterprises.

Tianyang Technology signed the "Investment Agreement on Beijing Bangke Intelligent Technology Co., Ltd." with Benghe Juchuang, Bengke Consulting, Bengke Management, and Benghe Chuangzhi, investing 120 million yuan in Bengke Intelligent, subscribing to 870,000 yuan of new registered capital of Bengke Intelligent, and the remaining 119 million yuan was included in the capital reserve fund to obtain 8% of the equity of Bengke Intelligent after the completion of this capital increase.

Computer vision technology and product developer Gantu Technology received hundreds of millions of yuan in C2 round financing

Gantu Technology is a computer vision technology and product developer. Its main business is to apply computer vision technology to precision appearance inspection scenarios. Currently, Gantu Technology focuses on intelligent quality control and yield management in the field of high-end manufacturing. Through its independently developed underlying AI framework and core technology, it provides one-stop intelligent solutions for high-end manufacturing. Gantu Technology has completed a C2 round of financing worth hundreds of millions of yuan. This round of financing was supported by local government industrial funds and obtained credit support from several banks. (Investment Circle)

AI-driven B2B payments platform Slope secures $65 million in strategic equity and debt financing

Slope, an AI-driven B2B payments platform, has raised $65 million in strategic equity and debt financing from JPMorgan Chase. Y Combinator, Notable Capital, and Jack Altman and Max Altman’s new fund Saga also participated in the round.

Planned, a provider of end-to-end tours and activities solutions, raises $35 million in Series B funding

Founded in 2017, Planned is a "source-to-pay" service company that accelerates travel and activities through technology. The company combines human services with artificial intelligence to provide customized procurement and booking services, and its clients include PwC, Block, AWS and Instacart.

This round of financing was led by Drive Capital, with participation from Outsiders Fund and two other companies. In addition to announcing the financing, Planned also added Frederic Lalonde, CEO and co-founder of Hopper and former vice president of Expedia, to its board of directors.

AI-driven revenue cycle automation company Thoughtful AI raises $20 million in funding

Thoughtful Automation Inc., an AI-driven revenue cycle management startup focused on the healthcare industry, has launched three AI agents: CAM, EVA, and PHIL, which are used to handle claims processing, patient eligibility verification, and payment posting tasks, respectively. The round was led by Nick Solaro of Drive Capital, with participation from TriplePoint Capital.

GPU cloud provider SF Compute receives $12 million in funding led by Alt Capital

SF Compute, founded by Evan Conrad, who once worked at OpenAI and AI Grant startup accelerator, and his roommate Alex Gajewski, announced that it had received $12 million in financing led by Alt Capital, founded by the Sam Altman brothers, with a valuation of $70 million.

Due to the lack of sufficient computing power in the market, it is difficult for startups to obtain the large amount of semiconductors needed for AI. SF Compute hopes to help startups obtain these resources and create a computing power trading platform; currently, SF Compute has received 8,000 H100 orders to start the project. (Newin)

AI guide glasses .lumen completes 5 million euros in financing

.lumen is committed to creating life-changing technology. Founded by Cornel Amariei, its flagship product .lumen Glasses and underlying AI technology provide independence and safety for the visually impaired. .lumen's blind glasses use pedestrian autonomous driving (PAD AI) technology to mimic the functions of a guide dog. .lumen raised €5 million in a round of financing from equity management and investment company SeedBlink. (Zpotentials)

AI customer service Xinlian Times received 20 million yuan in angel round financing

Xinlian Times provides enterprises with intelligent products and services based on artificial intelligence technology, covering multiple fields such as intelligent customer service, intelligent marketing, etc. This round of financing was led by Heyi Capital Co., Ltd.

AI engine provider Artificial.Agency completes multi-million dollar seed round of financing

Artificial.Agency was founded in 2023 and focuses on providing game creators and studios with an AI-driven behavior engine. The engine is able to integrate runtime decisions into game dynamics to bring a dynamic experience to players. Investors include BDC Venture Capital, Kaya, Radical Ventures, TIRTA Ventures, and Toyota Ventures.

Generative AI company EdgeRunner AI completes $5.5 million seed round of financing

EdgeRunner AI aims to build secure, reliable, and transparent generative AI for the edge. The company has developed small, task-specific, ultra-efficient language models that can run without access to the internet, thereby improving data privacy, security, and compliance. This round of financing was led by Four Rivers Group, with participation from Madrona Ventures and strategic angel investors.

AI editing platform Edit Cloud receives £2 million in funding

Edit Cloud is a cloud-based AI editing platform provider headquartered in London, UK, providing a cloud production platform designed to create high-quality content through AI technology. It also brings together cloud-based tools that enable teams to work together efficiently. This round of financing was led by Edge, and angel investors included Simon Ward and Justin Cooke.

AI large model development platform Arcee.ai completes Series A financing

Founded in February 2023, Arcee.ai is committed to developing a domain-adapted language model system that aims to provide LLM tailored for specific fields and seamlessly integrate DALM with business operations to achieve data-informed decisions and valuable insights. The company's system not only helps customers obtain LLM, but also creates a trustworthy and authentic system. Investors in this round include Centre Street Partners, Emergence, Flybridge Capital and Journey Ventures. (Yiou)

Artificial intelligence company NobiKan completes D+ round of financing

NobiKan focuses on applying advanced artificial intelligence technology and digital twin technology to highly complex open scenarios, providing highly universal artificial intelligence products and industry solutions for rail transit operation and maintenance, smart energy, smart cities, smart environment and other fields. The investor in this round is Peikun Investment.

Enigma Robotics receives strategic investment from National Supercomputing Center in Wuxi

EnigmaRobotics is a startup founded in 2023 that focuses on the development and promotion of intelligent companion robots, multimodal integrated models and their applications. The National Supercomputing Center in Wuxi announced a strategic investment in Enigma Robotics.

AI-driven credit financing platform New Frontier Funding receives investment

New Frontier Funding is committed to using generative AI to help small and medium-sized enterprises find credit and debt financing. The company uses its proprietary data and fine-tunes OpenAI's language model for semantic search and agent workflows to reduce background work and accurately match borrowers with lenders. It has completed a round of growth capital financing from Homsher Family Office. The transaction amount was not disclosed.

Edge Innovation Receives Angel Round Investment

The primary application scenario of Edge Innovation is to quantify the data of mental patients during the diagnosis and treatment process by using multimodal sensors + AI technology, improve the efficiency of psychologists' consultation, and focus on indicators such as facial muscle changes, heart rate and breathing. Recently, it received angel round investment from Qiji Chuangtan. CEO Zhao Zihe graduated from the Hong Kong University of Science and Technology and is a serial entrepreneur. He has rich experience in robot algorithm development and product development, and has worked for DJI.

Galaxy General, a multi-modal large-scale robot developer, receives investment from Hong Kong Investment Corporation

Galaxy General is a multi-modal large-scale robot developer, focusing on manufacturing robots with embedded AGI and providing general-purpose robots to the world. Hong Kong Investment Management Co., Ltd., the "Hong Kong version of Temasek", announced its investment in Galaxy General, and the investment amount was not disclosed. In June, Galaxy General received 700 million yuan in angel financing, and investors included Meituan Dianping Strategic Investment, BAIC Capital, SenseTime Guoxiang Fund, iFlytek Fund and other top strategic and industrial investors. Guangyuan Capital served as the exclusive financial advisor and participated in early investment.

(Welcome to add WeChat AIyanxishe2 to learn more about AIGC and financing, and chat about new AI products with like-minded friends)

Today's big factory news

NVIDIA Mistral AI jointly released 12B parameter small model Mistral Nemo, crushing Llama 3 single 4090 can run

NVIDIA has collaborated with Mistral AI to release a new small AI model, Mistral NeMo, which has 12 billion parameters, supports 128K contexts, and beats similar models Gemma 2 9B and Llama 3 8B in multiple benchmarks. Mistral NeMo is designed to serve enterprise users and can easily customize and deploy enterprise applications that support chatbots, multilingual tasks, encoding, and summarization. The Mistral NeMo model has excellent performance, strong compatibility, and is easy to use, and can directly replace any system using Mistral 7B. The model uses the FP8 data format for reasoning, which reduces memory size and speeds up deployment while maintaining accuracy. Mistral NeMo also supports multilingual applications and has an efficient word segmenter Tekken, which improves the processing efficiency of multiple languages. In addition, Mistral NeMo is ready to run anywhere, such as the cloud, data center, or RTX workstation, and developers can try Mistral NeMo using mistral-inference.

Xiaomi Xiaoai large model will be fully upgraded soon: all free, and will be supported by mobile phones, tablets, and TVs by the end of this month

The Xiaomi Xiaoai large model will be fully upgraded, and all for free. The upgraded model will be smarter, support intelligent Q&A, creation and other functions, and improve the chatting experience. At the end of July, mobile phones, tablets, TVs and other devices will support the new model. The version for mobile phones and tablets is V6.126.5, and the version for TVs is V4.30.1, and the memory capacity needs to be more than 1G. There will be upgrade support for screenless speakers at the end of August, and upgrades for speakers with screens at the end of October. (Fast Technology)

Apple releases DCLM-7B, an open-source model with performance exceeding Mistral-7B

Apple released the DCLM-7B open source model on Hugging Face. The performance of this model has surpassed Mistral-7B and is approaching Llama 3 and Gemma. The open source resources of the DCLM-7B model include model weights, training code, and pre-trained datasets. The research team proposed a new DCLM benchmark for evaluating the performance of large language models, especially in the multimodal field. The DCLM benchmark uses a standardized experimental framework, including a fixed model architecture, training code, hyperparameters, and evaluation to find the data organization strategy that is most suitable for training high-performance models. The DCLM-7B model uses a pre-training scheme based on the OpenLM framework, and achieves a 5-shot accuracy of 64% on the MMLU benchmark, which is comparable to Mistral-7B-v0.3 and Llama 3 8B, but the amount of computation required is only 1/6 of Llama 3 8B.

Google and Ray-Ban develop smart glasses with Gemini AI model

Google has approached EssilorLuxottica (the company behind the Ray-Bans brand) to collaborate on the production of Gemini smart glasses. EssilorLuxottica has previously worked with Meta to launch two generations of Ray-Ban Meta smart glasses, and the latest news is that Meta is planning to spend billions of dollars to acquire about 5% of EssilorLuxottica's shares. (The Verge)

Today's Product News

Product Hunt Hot List, Flow Studio

Flow Studio is a tool developed by the Flow GPT team that can convert text into high-quality short videos. The platform was developed by Lifan Wang, Sam Xu, Qianhua Ge, Jay Dang, and Luke Pioneero, and was released on Product Hunt on July 18, 2024. The highlight of Flow Studio is that it can automatically generate a complete video with a story, dubbing, background music, and sound effects through a single text prompt, greatly simplifying the video production process. Flow GPT has received high praise from users, with an average score of 4.9/5 stars.

Founder Jay Dang studied computer science at the University of California, Berkeley. He is the founder of FlowGPT, Markit AI, and LUUM, and has also worked as a data scientist and independent researcher at C. Light Technologies, Inc. and Glaucomark.

?https://flowgpt.com/flow-studio?ref=producthunt

GitHub Trending Hot List, Langflow, a framework for building multi-agent and RAG

Langflow is a visual framework designed to help developers build multi-agent and RAG applications. The project is based on Python, open source, fully customizable, and supports different language models and vector storage. Users can install Langflow via pip, and need to ensure that the Python version installed in the system is at least 3.10. The project provides detailed documentation and deployment guides.

?https://github.com/langflow-ai/langflow

Special attention

Yi Tay, former senior researcher at Google Brain and co-founder of Reka AI, explains: Why haven’t we seen more encoder model extensions after BERT?

Yi Tay is the co-founder and chief scientist of Reka AI, which has raised $100 million in funding. He worked on large language models and AI research at Google Brain, and was deeply involved in most of Google's large language model and multimodal work from 2020 to early 2023. He has won best paper awards at ICLR and WSDM conferences, and is a guest lecturer in the CS25 course at Stanford University.

Yi Tay published the first in a series of blogs on the X platform, aiming to explore model architectures in the era of large language models. Various architectures including Transformer encoders, encoder-decoders, PrefixLM, and denoising objectives are discussed. Yi Tay cites a common question of why we have not seen more encoder model expansion after BERT, and the fate of encoder-decoder or encoder-only models. He also questioned the effectiveness of the denoising objective and shared his thoughts in the blog.

The blog first mentioned the confusion of people in the field of natural language processing in recent years about the disappearance of encoder models, and the development of models such as BERT and T5. It emphasized the differences and connections between encoder-decoder models, encoder-only models, and decoder-only models, and pointed out the characteristics of the PrefixLM architecture. It further explained the concept of denoising targets, including BERT-style "in-place" denoising and T5-style sequence-to-sequence denoising, and discussed the advantages and disadvantages of denoising targets.

Yi Tay also analyzed the computational cost of the encoder-decoder model and why BERT-style models are gradually being phased out. He pointed out that denoising objectives are often complementary to causal language models and play a role in pre-training large language models. In addition, the role of bidirectional attention mechanisms in models of different sizes is also discussed. Finally, Yi Tay summarized the advantages and disadvantages of the encoder-decoder architecture and emphasized the importance of understanding inductive biases and pre-training strategies, as well as why the BERT model was replaced by the more flexible T5 model.

?https://www.yitay.net/blog/model-architecture-blogpost-encoders-prefixlm-denoising

Stay tuned for the latest updates tomorrow!

Leifeng Network