Anthropic, a competitor of OpenAI, announced the launch of Claude 3.5 Sonnet, the first product in the Claude 3.5 series. This model outperforms its competitors and its predecessor, Claude 3 Opus, in multiple evaluations.

According to foreign media reports from The Wall Street Journal, Apple is seeking cooperation with local Chinese companies, including Baidu, Alibaba Group, and Baichuan Intelligence, aiming to provide its "Apple Intelligence" service in the Chinese market. These companies have not yet made a public response.

What other hot topics in the AI industry, both domestically and internationally, are worth paying attention to in the past day? Let's take a look with Raven Jun.

/ 01 / Large Models

1) OpenAI Competitor Anthropic Releases the Most Powerful AI Model, Claude 3.5

Anthropic, a competitor of OpenAI, has released the AI model Claude 3.5 Sonnet, the first product in the Claude 3.5 series. This model outperforms its competitors and its predecessor, Claude 3 Opus, in multiple evaluations, while maintaining a speed and cost comparable to mid-range models. Claude 3.5 Sonnet has set new industry benchmarks in graduate-level reasoning, undergraduate-level knowledge, and coding abilities, achieving significant performance improvements.

2) Apple AI is Seeking Collaboration with Local Chinese Companies, Having Contacted Baidu, Alibaba, and Baichuan

According to reports from The Wall Street Journal, Apple is seeking cooperation with local Chinese companies, including Baidu, Alibaba Group, and Baichuan Intelligence, aiming to provide its "Apple Intelligence" service in the Chinese market. This move may be to address competitive pressures in the Chinese market. According to Counterpoint Research, iPhone's market share in China has dropped to third place. The aforementioned companies have not yet made a public response.

3) Investigation into Price Wars in Large Models: Some Customers' Invocation Volume Increased by 5000 Times, Changing the Logic of Large Model Deployment

In mid-May, over 15 well-known large model vendors reduced prices or offered free services, triggering a price war in the large model industry. Market feedback shows a significant increase in the number of new users and an expansion of business volumes for existing users. Whether in the internet or smart terminal industries, the invocation volume of large models has increased significantly. However, price reductions have also led some vendors to change their original independent research and development routes, focusing more on AI applications.

4) Groq Launches whisper-large-v3 Model, Supporting Speech Transcription and Translation, Openly Available for Free

Groq's newly launched Whisper Large-V3 model provides users with speech transcription and translation capabilities, allowing them to use the API in Playground or local projects. Users can experience high-speed transcription, supporting translation of multiple languages into English. The Whisper API is compatible with OpenAI standards, providing speech-to-text and translation functions, making it easy to integrate into applications. With superior performance, it utilizes the advanced "whisper-large-v3" model.

/ 02/ AI Applications

1) Kuaishou Launches New Features for Video Creation and Continuation Based on Images

According to insiders, Kuaishou's video model has launched new features called "Image-to-Video" and "Video Continuation." The Image-to-Video feature can generate a 5-second video based on an image, supporting the addition of prompts to control image movement. The Video Continuation feature allows users to continue a generated video for 4-5 seconds with a single click, supporting multiple continuations up to 3 minutes long. It also allows for video creation through fine-tuning of prompts. Additionally, the text-to-video feature has added new video size options of 9:16 and 1:1.

2) Apple Intelligence Has Too Many Device Limitations? Response from Apple Executives

Apple Intelligence is limited to iPhone 15 Pro/Pro Max as well as iPad and Mac devices equipped with M1 or later chips. Apple explained that this is because the inference and computation requirements of large language models are extremely high. Analyst Guo Minggui believes that whether Apple's intelligence is compatible depends on the device's DRAM size, rather than AI computing power.

3) Tencent Yuanbao has released a new version with access to WeChat search

Tencent Yuanbao has recently released a new version, mainly improving its ability to handle ultra long texts and AI search and parsing functions, as well as adding WeChat search access. This update has improved the efficiency of handling ultra long documents, as well as enriched file format support, chart generation, and image parsing functions. The new version has enhanced the search function and integrated with search engines such as WeChat search.

4) CNKI announces the launch of CNKI AI Academic Research Assistant 4.0

China National Knowledge Infrastructure (CNKI) has recently launched the AI Academic Research Assistant 4.0 version, which combines AI big model technology and high-quality data to improve the efficiency of literature retrieval, research, and academic creation. New features include controllable generation, literature expansion, scholar search, full-text translation, and academic expansion services. Highlighting upgrades is a question answering enhanced retrieval and scholar retrieval service.

5) WeChat Input Method launches the "Ask AI" function, with answers provided by WeChat Reading AI

WeChat Input Method has brought a new AI Q&A function. The AI Q&A answers are provided by WeChat Reading AI Q&A. Clicking on the link will lead to the page referenced in the text in WeChat Reading, allowing users to better understand the problem through context. At present, WeChat Input Method has not launched AI Q&A function in iOS and Android versions.

6) The Fudan open-source project Hallo has been adapted to the ComfyUI plugin

The Hallo project is an open-source project that generates speech videos based on audio and images, with a high installation threshold, providing more possibilities and fun for rendering and other processes. It adopts an end-to-end diffusion paradigm and introduces a layered audio driven visual synthesis module to achieve alignment accuracy between audio input and visual output, generating natural speech videos.

7) Universal Music collaborates with AI music company SoundLabs to customize voice cloning models for singers

Universal Music Group collaborates with AI music technology company SoundLabs to launch the MicDrop feature, allowing artists to customize personalized voice models, have full control, overcome language barriers, and protect artist rights. This revolutionary technology brings music creation into a new creative space, promoting the application and development of AI in the music field.

/03/Investment and Financing Intelligence

1) An AI news reader developed by a former Twitter engineer, receiving $10.9 million in financing

Article received a $10.9 million Series A financing led by Lightspeed Venture Partners, which also includes global media company Axel Springer as investors. Article is a startup founded by former Twitter engineers Sara Beykpour and Marcel Molina, who use artificial intelligence technology to create personalized news platforms.

2) AI video startup HeyGen raised $60 million with a valuation of over $500 million

HeyGen successfully raised $60 million in Series A financing, with a company valuation of over $500 million. Its profitability is strong, with annual revenue increasing from $1 million to over $35 million, and its customer base covering small businesses to Fortune 500 companies. HeyGen plans to expand product supply and invest in enterprise security, AI ethics, trust, and security.

3) Former GitHub CTO raised $400 million in entrepreneurial financing to become an AI programmer, valued at $2 billion

Poolside.ai, a generative artificial intelligence company headquartered in Paris, is raising $400 million in funds with a valuation of $2 billion. Bain Capital Ventures and DST are currently negotiating on the current round. Founder&CEO Jason Warner was previously the CTO of GitHub and led the engineering departments of Heroku and Canonical.

4) Unveiling Ilya's New Company: Backed by 5 tons of GPU, Doing Nuclear Power Level Safety

The news of Ilya and others founding a new company, SSI, has attracted industry attention, and Ilya has expressed her focus on nuclear safety. It is understood that they invested $100 million in 2023 to establish an AI computing cluster called Andromeda, and used this computing power to exchange for equity in AI startups. This cluster has a large amount of computing infrastructure, including nearly 3000 Nvidia H100 GPUs, with GPUs alone weighing nearly 5 tons.

5) SoftBank is ready to fully bet on AI? Sun Zhengyi's Oath of Unsuccessful Success: Cheng Ren Reveals "New Investment Direction"

SoftBank CEO Sun Zhengyi is preparing to target new technology investments towards AI. Sun Zhengyi said that even if it is a failure, there is no other choice but to try. SoftBank did not disclose specific investment details, but will mainly expand its power generation business in the United States to power artificial intelligence projects; We are also seeking up to $100 billion in funding to invest in a chip company. At present, SoftBank under Sun Zhengyi has accumulated a cash reserve of 6.2 trillion yen.

/04/AI Infrastructure

1) Meta releases the latest RAG evaluation benchmark, with the recognized strongest GPT-4 scoring only 40 points

Meta released the RAG evaluation benchmark, and GPT-4 only scored 40 points (on a percentage scale) with RAG. There is still room for improvement in display technology; RAG technology attempts to solve the problem of hallucinations when LLM generates answers by enhancing the combination of LLM and external knowledge; The CRAG evaluation benchmark design includes multiple tasks and evaluation methods, aiming to comprehensively test the performance of RAG systems in diverse and dynamic question answering scenarios.

2) GaussianCube: High quality 3D generated modeling with a performance leap of 74%!

The field of 3D generative modeling has made breakthrough progress, with Gaussian Cube technology surpassing traditional NeRF and revolutionizing 3D modeling. This technology uses density constrained Gaussian fitting algorithm to simplify the modeling process and achieve high-precision fitting. The experimental results showed a performance improvement of up to 74%, demonstrating its enormous potential.

picture

3) The hottest AI role-playing traffic has reached 20% of Google searches, processing 20000 inference requests per second

Founded by Transformer author Noam Shazeer, Unicorn Character.ai processes 20000 AI inference requests per second, achieving 1/5 of Google search traffic in 2024. The founder revealed a unique secret to reasoning optimization, which quickly sparked industry discussion. Specifically, as follows:

Efficiently utilizing graphics memory, reducing the number of attention parameters by 20 times; Cleverly using state caching, 95% of requests do not require recalculation; Direct quantitative training, zero inference loss, and saves memory.