Text | Chen Sida
Editor｜Anita Tang
Weekly overview
Big events are happening this week. The biggest hot spot is undoubtedly the NVIDIA GTC conference. NVIDIA officially released the new generation AI chip "Blackwell B200" , which founder Huang Renxun called the most successful product in history.
Other players were not idle either. Musk’s AI startup xAI has open sourced the large model Grok-1 , which has 341 billion parameters. On the other side, OpenAI finally has new news, and it is rumored that it will launch GPT-5 in the middle of the year. More important news comes from hardware players - Apple is actively negotiating with Google and OpenAI, hoping to introduce large models to the iPhone.
The AI unicorns in Silicon Valley are quite turbulent. The founder of Inflection AI and co-founder of Google DeepMind is about to join Microsoft’s AI camp. Stability AI is still in turmoil . Following the previous collective resignation of the core technical team, this week the CEO also announced that he has left.
This week’s domestic stage belongs to “Long Context”. The Dark Side of the Moon officially announced that its Kimi smart assistant has supported 2 million words of ultra-long lossless context , which attracted widespread attention in the market. Subsequently, Alibaba Tongyi Qianwen updated it, opening up the 10 million word long document processing function for free, and Baidu and 360 Intelligence also successively Announced that long text processing capabilities of 2 million to 5 million words will be launched soon. Over the weekend, the general large model startup Step Star officially made its debut and released a preview version of the Step-2 trillion parameter MoE language large model.
Key Points
- OpenAI is expected to launch GPT-5 mid-year
- Apple discusses cooperation with Google and OpenAI, large models may enter iPhone
- Musk Grok-1 is open source with 341 billion parameters
- Reappearing Sora, Colossal-AI releases open source project "Open-Sora"
- Stability.ai releases Stable Video 3D
- Kimi intelligent assistant supports 2 million words of context and became popular in the circle
- Large model manufacturer "Step Star" makes its debut and has trained a large model with trillions of parameters
- Nvidia releases the most powerful AI chip
- Microsoft acquires Inflection AI "employment style", founder joins Microsoft
- Stability AI core team collapses, CEO resigns
- Apple hit by U.S. Department of Justice antitrust lawsuit
- United Nations passes first global resolution on AI
- Nvidia considers acquiring Israeli AI startup Run:ai
- Saudi Arabia plans to set up $40 billion fund to invest in AI
- "Little Nvidia" Astera Labs goes public, raising $600 million
- Suno officially releases V3 music generation model
- New Adobe research: Generate image from sketch in 0.11 seconds
Large model first line
OpenAI is expected to launch GPT-5 mid-year
According to The Business Insider, citing anonymous sources, OpenAI is planning to launch GPT-5 in the middle of this year, which is expected to be in the summer. In addition, people familiar with the matter revealed that some enterprise customers have already experienced the demonstration of the latest model and the accompanying ChatGPT enhancements in advance. A CEO who has experienced the GPT-5 version spoke highly of its performance: "It performed very well and brought significant improvements."
Apple discusses cooperation with Google and OpenAI, large models may enter iPhone
According to the latest report from Bloomberg, Apple is actively engaged in in-depth negotiations with Google and OpenAI, aiming to integrate the two companies’ generative large-scale language models into the new artificial intelligence functions of the iPhone, laying a strong foundation for the future iOS 18 system. frame. The official announcement is not expected until after this summer. At the same time, on March 23, according to the Wall Street Journal, Apple and Baidu have held preliminary negotiations on using Baidu’s generative AI technology in its Chinese devices. It’s unclear whether Apple is working with other Chinese generative AI companies.
Musk Grok-1 is open source with 341 billion parameters
On March 18, xAI, an AI startup owned by Musk, announced that the large model Grok-1 it developed was officially open to the public. Users can download basic model weights and network architecture information directly through magnet links. Grok-1 is a 314 billion parameter hybrid expert (MOE) model trained from scratch by xAI in October 2023 using a custom training stack based on JAX and Rust, far exceeding OpenAI's GPT model. However, the open source version this time is the original basic model of the Grok-1 pre-training stage and has not been fine-tuned for any specific application (such as dialogue).
Reappearing Sora, Colossal-AI releases open source project Open-Sora
Following the launch of the Sora training inference reproduction process, which has a cost reduction of 46%, the Colossal-AI team has fully open sourced the world's first Sora-like architecture video generation model "Open-Sora 1.0", covering the entire training process, including data processing, all training details and model weights, and join hands with global AI enthusiasts to promote a new era of video creation.
A snapshot of urban bustling generated by Open-Sora 1.0
Stability.ai releases Stable Video 3D
Stability.ai releases Stable Video 3D, leveraging its multi-view consistency to optimize 3D Neural Radiation Fields (NeRF) and mesh representations to improve the quality of 3D meshes generated directly from novel views, providing coherence from any given angle view and have proficient generalization skills. Stable Video 3D quality and multi-view are significantly improved, outperforming other open source alternatives such as the previously released Zero 123XL.
Stable Video 3D generation effects
"Kimi Smart Assistant" supports 2 million words of context and became popular in the circle
On March 18, the large model manufacturer "Dark Side of the Moon" officially announced that its Kimi smart assistant has supported 2 million words of ultra-long lossless context and will start internal product testing from now on. On the afternoon of the 21st, the APP and mini-programs of Kimi, a large model application owned by Dark Side of the Moon, could not be used normally. The Dark Side of the Moon stated that it observed that Kimi's system traffic continued to increase abnormally, far exceeding the expected resource planning. Catalyzed by Kimi's popularity, the stock prices of Huace Film and Television, Zhangyue Technology, Zhongguang Tianze, and Foxit Software-related concept stocks continued to rise.
Large model manufacturer "Step Star" makes its debut and has trained a large model with trillions of parameters
On March 23, at the 2024 Global Developer Pioneer Conference, the general large-model startup company Step Star was officially unveiled. The Step-1V 100-billion-parameter multi-modal large model developed by Step Star ranked first in the multi-modal model evaluation list of "OpenCompass", China's authoritative large-scale model evaluation platform. Dr. Jiang Daxin, founder and CEO of Stepstar, officially released a preview version of the Step-2 trillion-parameter large language model at the conference. The model adopts MoE architecture, focuses on the exploration of deep intelligence, and provides API interfaces for some partners to try.
Big event
NVIDIA releases new generation AI chip
From March 18th to 21st, NVIDIA held a GTC conference in San Jose, USA, to release the next-generation chip architecture Blackwell. This GPU platform is also the most successful product in NVIDIA's history according to Huang Renxun. According to reports, the Blackwell GPU is named after mathematician David Harold Blackwell and also follows the Hopper architecture previously launched by Nvidia. Blackwell GPUs contain 208 billion transistors and can support AI models with up to 10 trillion parameters.
Microsoft acquires Inflection AI "employment style", founder joins Microsoft
On March 19, Microsoft officially announced that Mustafa Suleyman and Karén Simonyan of artificial intelligence startup Inflection AI and most other employees will join Microsoft AI to focus on consumer-oriented AI products and research. On March 22, according to people familiar with the matter, Microsoft has agreed to pay approximately US$650 million to Inflection AI, mainly in the form of a license agreement, so that Inflection AI’s models can be sold on Azure cloud services. This means that in the name of hiring core team members plus "model licensing fees", Microsoft has only essentially completed the valuation of Inflection, which was once ranked third in the AI rankings (after OpenAI and Anthropic). Acquisitions of companies.
Stability AI core team collapses, CEO resigns
On March 23, local time, unicorn AI company Stability AI said it announced the resignation of company CEO Emad Mostaque. Emad Mostaque announced on the social media platform X that he will focus on decentralized artificial intelligence (Decentralized AI) after his resignation. Earlier this week, the core R&D team resigned en masse. Stability AI is famous for developing Stable Diffusion, a large model of Vincentian graphs. It was founded at the end of 2020 and was valued at US$1 billion in 2022.
Apple was hit with an antitrust lawsuit by the U.S. Department of Justice, evaporating $110 billion in market value
According to Reuters, on March 21, local time, U.S. Attorney General Merrick Garland said at a press conference that day that the U.S. Department of Justice and the attorneys general of more than a dozen states had filed an antitrust lawsuit against Apple, accusing Apple of The company used its control of Apple product hardware and software to monopolize the mobile phone market, harming the interests of consumers, developers and rival companies. Affected by this news, Apple's stock price fell 4.09% that day, its market value evaporated by more than 110 billion US dollars (approximately 800 billion yuan), and its total market value fell back to 2.65 trillion US dollars.
United Nations passes first global resolution on AI
On March 21, local time, the United Nations General Assembly voted to adopt the first draft resolution on artificial intelligence (AI) to ensure that this new technology can benefit all countries, respect human rights and be "safe, reliable and trustworthy" technology . It is reported that the United States is the sponsor of this draft resolution, and China participated in the co-sponsorship.
Financing dynamics
Nvidia considers acquiring Israeli AI startup Run:ai
According to the Israeli Economist, on March 17, it was revealed that Nvidia was negotiating to acquire Run:ai, an Israeli artificial intelligence infrastructure platform, with a transaction value that may reach US$1 billion. Run:ai was founded in 2018 by CEO Omri Geller and CTO Dr. Ronen Dar. In March 2022, Run:ai raised US$75 million in Series C, led by Tiger Global Management and Insight Partners.
Saudi Arabia plans to set up $40 billion fund to invest in AI
The Saudi Arabian government plans to create a fund of about $40 billion to invest in artificial intelligence (AI) technology, according to three people familiar with the plan. If successfully established, this fund will make Saudi Arabia the largest AI investor in the world. People familiar with the matter said that in recent weeks, representatives of the Saudi Public Investment Fund (PIF) have discussed potential partnerships with financial institutions such as Andreessen Horowitz (a16z), the most successful venture capital firm in Silicon Valley, including how the fund will operate and what a16z can do. role.
"Little Nvidia" Astera Labs goes public, raising $600 million
On March 20, chip manufacturer Astera Labs listed on the Nasdaq in the United States, raising US$600 million. Taking advantage of AI, investment banks packaged Astera Labs into the concept of "Little Nvidia", which made Astera Labs highly sought after by the market. One of Astera Labs' core products is data and memory connection semiconductors, which can effectively improve the efficiency and speed of connections between software and hardware. Its customers include industry giants such as Amazon and Microsoft. Its closing price on the first day of listing was US$62.03, an increase of 72.31% from the issue price. Based on the closing price, the company's market value was approximately US$9.459 billion.
new gadgets
Suno officially releases V3 music generation model
AI music generation startup Suno officially released Vincent Music Model v3, which can create a complete two-minute song in seconds. The tool can be accessed through its free standalone website, or through Microsoft Copilot, a third-party Suno-enabled plug-in. Users only need a simple text description to create professional-quality music.
Suno interface
Experience address: https://app.suno.ai
Frontier Research
New Adobe research: Generate image from sketch in 0.11 seconds
On March 19, teams from CMU and Adobe published a paper on arXiv, proposing an image reasoning method. According to reports, this method solves two limitations of existing conditional diffusion models: slow inference due to the iterative denoising process and reliance on model fine-tuning on paired data. To address these issues, the authors introduce a general method to adapt single-step diffusion models to new tasks and domains through adversarial learning objectives, which can not only leverage the internal knowledge of pre-trained diffusion models, but also achieve efficient inference. For a 512*512 image, the generation time is 0.29 seconds on the A6000 and 0.11 seconds on the A100.
Screenshot of paper
Paper address:
https://arxiv.org/pdf/2403.12036.pdf
Trial address:
https://huggingface.co/spaces/gparmar/img2img-turbo-sketch
👇🏻 Scan the QR code to join the "Zhiyong AI Exchange Group"👇🏻
Welcome to pay attention
