Following Microsoft's example, Google has also merged its hardware and operating system divisions, OpenAI has set up a Japanese branch, and Wenxin Yiyan has more than 200 million users......

Written by | Neocortex Group

Edit | Jeff Wang

On Friday, U.S. time, technology stocks led by Nvidia were sold off by investors, of which Nvidia fell 10%, close to two CATL, and Nvidia's most important competitor in the field of AI chips, AMD, fell 5.44%. In addition, Meta fell 4.13%, and Amazon, Microsoft, Google, and Apple also fell more than 1%. As a result, the tech-heavy Nasdaq Composite fell 2.1%. Just over a month ago, driven by the huge demand for AI chips in the generative AI wave, Nvidia's market capitalization just exceeded $2 trillion, and its market value has risen more than fivefold in more than a year, becoming the third largest technology company in the United States.

So, is this the end of this AI wave? Investment institutions have given the direct cause of Nvidia's flash crash: Supermicrocomputer, which has a close relationship with Nvidia, did not give forward results as usual when it released its latest quarterly report on Friday, a move that caused investors to worry about its performance, causing the company's stock price to plummet by 23%, and also brought down Nvidia's stock price.

Chaowei Computer is a server manufacturer, and its core business is to package Nvidia's AI chips into AI servers and sell them to customers, and the two companies are less than 15 minutes away by car, and engineers often work together. It can be said that as long as you buy a Chaowei server, it is equivalent to buying an NVIDIA chip. In the past year, Chaowei's share price has even risen more than Nvidia, rising more than ten times in more than a year, with a market value of up to $66 billion.

It seems that the culprit is naturally Chaowei Computer, which did not announce its forward-looking results as usual, but Nvidia's "joint sitting" actually reflects investors' anxiety about the future of generative AI.

Meta finally released its open-source model Llama 3 this week, with significantly improved inference and code capabilities, and the performance of the strongest yet-to-be-released version, 400B, is said to be very close to GPT-4. Add to that the recently released Claude 3 and Command R+, and there are already four or five models that are close to GPT-4's capabilities. The optimistic view is that more competitors will accelerate the productization of AI and make generative AI more efficient for humans, while the pessimistic view is that the speed of AI improvement is slowing down, and GPT-4 may be close to the upper limit of the current model capabilities, which can improve efficiency in some aspects, but it is not enough to become a universal assistant (Copilot).

The following is a summary of the smart news worth paying attention to in the past week, produced by the "Neocortex" team.

a

Key Points

Overseas Edition

Following Microsoft's lead, Google has also merged its hardware and operating systems divisions;

OpenAI opens a branch in Tokyo and will release a GPT-4 model optimized for Japanese;

Meta发布第3代最强开源模型Llama 3;

Mistral AI plans to raise at $5 billion valuation;

After the decline in sales, Tesla laid off 10% of its global workforce;

In order to commercialize, Boston Dynamics finally changed the Atlas robot to electric drive;

Domestic chapter

The number of users of Wenxin Yiyan exceeded 200 million, and the cost was only 1% of the original;

In the first quarter of this year, Chinese mobile phone manufacturers performed well, with Xiaomi shipments up 34% and Transsion up 85%;

With the generative AI dividend, Mobvoi is about to go public;

Huawei Pura 70 series mobile phones went on sale;

Overseas Edition

Following Microsoft's lead, Google has also merged its hardware and operating systems divisions

On April 18, local time, Google announced the establishment of a new department of "Platforms & Devices" (Platforms & Devices), which mainly integrates AI into all Google Pixel series hardware products, and less than a month ago, Google's main rival in the field of AI Microsoft just announced the integration of the Windows operating system and Surface hardware departments, and appointed Pavan, the former head of Surface Davuluri serves as the division head of the two businesses and reports directly to Rajesh Jha, executive vice president and head of Microsoft Experience and Devices. At that time, the reasons given by Microsoft were similar to Google's, which were to speed up the implementation of AI and provide users with AI solutions that unify software and hardware.

The new Platform & Hardware team will be led by Rick Osterloh, who will be responsible not only for all of the hardware products in the Google Pixel family, but also for operating systems and software platforms such as Android, Chrome, ChromeOS, Photos, and more. Rick Osterloh previously served as Senior Vice President of Devices and Services, overseeing all of Google's hardware efforts such as phones, speakers, VR headsets, laptops, smartwatches, and more.

Osterloh said the key to the integration was to streamline the process – one team for AI research (Google DeepMind) and another for AI products (platform and hardware) was reorganized. "A team with only one leader and one goal does everything faster," says Osterloh. By merging teams, Google can more directly and efficiently integrate AI into all of its products.

OpenAI has opened a branch in Tokyo and will release a GPT-4 model optimized for the Japanese language

On April 15, OpenAI announced the opening of its first branch in Asia in Tokyo, Japan, which will focus on working with the Japanese government, local businesses, and research institutions to develop secure AI tools that meet Japan's unique needs and unlock new opportunities. OpenAI promises to provide local businesses with GPT-4 custom large models optimized specifically for Japanese, which can improve performance in Japanese text and run up to 3x faster than GPT-4 Turbo. At the same time, the model improves the performance and cost-effectiveness of translating and summarizing Japanese texts. The Tokyo branch is OpenAI's third overseas branch after London and Dublin. Tadao Nagasaki, former president of Amazon Web Services (AWS) Japan, will serve as president of OpenAI Japan, leading OpenAI Japan's commercial and marketing partnerships.

Meta发布第3代最强开源模型Llama 3

On April 18, local time, Meta released the open-source large model Llama 3, including two versions of pre-training and fine-tuning, with 8B and 70B parameters, respectively. Meta said that this is only the first round of Llama 3 releases, and that there will be multilingual, multimodal and longer context windows to follow, including the strongest version that is still being trained and has a model parameter size of up to 400B.

What are the highlights of Llama 3?One is the enhanced reasoning and coding skills. Zuckerberg said that even if the user didn't directly ask the code-related question, training the model in code would help it answer the question more accurately, and that the chat interaction required more comprehensive and in-depth thinking, "which actually turns into an inference problem." Meta optimizes Llama 3 through extensive code training.

The second is to embed more functionality into the model. For example, where Meta AI was previously not part of Llama's base model by allowing Meta AI to collaborate with Google, Bing, and other tools to obtain real-time information, Llama 2 is mainly done by hand coding, and now it does not have to manually code everything so that the large model can use Google search to do these tasks independently.

Mistral AI plans to raise money at a $5 billion valuation

French startup Mistral AI is recently raising its next round of funding from investors at a valuation of $5 billion. According to people familiar with the matter, it is unclear which investors are involved in the deal, but Mistral AI has been in discussions with funds in the United Kingdom, France and the United States, and the upcoming deal could boost Mistral AI's valuation to around $5 billion to $7 billion. In December 2023, Mistral AI closed a 450 million euros (about $487 million) financing at a valuation of $2 billion, with participation from a16z, Nvidia, and others.

Mistral AI is currently one of the most well-known AI startups in Europe, headquartered in Paris, France. The company's founders are three former employees from DeepMind and Meta, one of whom is Lample, who led Meta's large model team to complete the development of LLaMA. Last Tuesday, Mistral AI open-sourced its latest large language model, Mistral 8x22B, with a total of 176 billion parameters and a context window of 65,000 tokens.

Tesla lays off 10% of its global workforce after sales decline

On April 15, Tesla CEO Elon Musk issued an email to all employees, announcing that more than 10% of the global workforce will be cut, which will affect more than 14,000 employees. Musk said in an email that the company had made the difficult decision to lay off employees in order to reduce costs and increase productivity.

Tesla's recent business situation is not optimistic, and layoffs have become a means to reduce costs and increase efficiency. In early April, Tesla's latest delivery report showed that a total of 386,800 Model 3 and Model Y were delivered in the first quarter, lower than Wall Street's expectation of 450,000 vehicles and down 8.5% compared with the same period last year. It's also the first time in four years that Tesla's quarterly deliveries have declined, and the last time a drop in sales was announced was in the second quarter of 2020, when the pandemic forced Tesla to halt production.

In order to commercialize, Boston Dynamics finally changed the Atlas robot to an electric drive

On April 16, Boston Dynamics posted a video on YouTube announcing that it would no longer develop the hydraulically driven humanoid robot Atlas. A day later, Boston Dynamics spoke out again, saying that it would launch an all-electric humanoid robot, and still named it Atlas.

Boston Dynamics said it would say goodbye to the hydraulic Atlas, which is mainly for better commercialization. Although hydraulic drives are superior in terms of both movement and flexibility, a hydraulic Atlas is expected to cost up to $2 million to manufacture, making it more difficult to commercialize. On top of the cost, the hydraulic path is difficult to maintain and not very friendly to the business environment in which it is put in – when the Atlas parkour fails, the hydraulic fluid spills everywhere.

The use of electric drive scheme not only significantly reduces the cost, but also makes its response and maintenance simpler. For applications that require high precision and high sensitivity, the electric drive system allows for finer control of the movements. This may be the main reason why humanoid robots on the market, such as Elon Musk's Optimus Prime, OpenAI-invested Figure 01, and Zhiyuan's Expeditionary AI, generally use electric drives.

Domestic chapter

The number of users of Wenxin Yiyan exceeded 200 million

On Tuesday, April 16, Baidu held the Create 2024 Baidu AI Developer Conference in Shenzhen, at which Baidu CEO Robin Li introduced the application achievements of Baidu's generative AI in the past year, including the application status of Wenxin Yiyan and Wenxin large models, the three major development tools, and Baidu's new developments in intelligent driving. Its Chinese heart is a chatbot officially released by Baidu in March 2023, benchmarking OpenAI's ChatGPT. In the past year, the number of product users has exceeded 200 million, the average daily API call volume has also exceeded 200 million, the number of customers served has reached 85,000, and the number of AI native applications developed using the Qianfan platform has exceeded 190,000.

At present, Wenxin Yiyan is the most popular generative AI product in China, but the pressure from competitors is not small at all. Data from Alcpb.com shows that the Wenxin Yiyan app and website totaled 14.9 million visits in March, while Kimi, developed by AI Unicorn Moon's Dark Side, totaled 12.6 million visits in the same month. It can be seen that Wenxin Yiyan's lead is not large, and Kimi's growth rate is faster, its visits in March increased by 321.6% compared to February, while Wenxin Yiyan's visits increased by more than 48% in the same period.

In the first quarter of this year, Chinese mobile phone manufacturers performed well, with Xiaomi shipments up 34% and Transsion up 85%

According to data from market research agency IDC, global smartphone shipments in the first quarter of 2024 reached 290 million units, a year-on-year increase of 7.8%, which is the third consecutive quarter of growth, and the growth momentum is mainly from Chinese manufacturers.

In the first quarter of this year, Samsung shipped about 60.1 million smartphones, down 0.7% year-on-year, returning to the top with a global market share of nearly 21%, ending Apple's brief lead at the end of 2023.

The third to fifth largest companies in terms of shipments are all Chinese manufacturers, namely Xiaomi, Transsion and OPPO. Xiaomi is recovering strongly from the sharp decline it has experienced over the past two years, with shipments up 33.8% year-on-year, while Transsion has become a stable presence in the top five by aggressively expanding internationally, with shipments up 84.9% year-on-year. Transsion has three mobile phone brands, namely Tecno, Itel and Infini, which are popular in emerging markets such as Africa, South Asia and Latin America.

With the generative AI dividend, Mobvoi is about to go public

On April 16, Mobvoi, an AI company, announced that it would open its IPO today and end on the 19th, and plans to officially list on the main board of the Hong Kong Stock Exchange under the stock code "2438" on April 24. According to the prospectus, Mobvoi will achieve revenue of 398 million yuan, 500 million yuan, and 507 million yuan from 2021 to 2023 respectively. In terms of profit, the adjusted net profit was -73 million yuan, 109 million yuan and 18 million yuan respectively. By business, Mobvoi's revenue is mainly composed of AI software (including AIGC solutions and AI enterprise solutions) and AIoT hardware. Among them, AIGC solutions have grown rapidly in recent years.

Mobvoi was initially positioned as an "artificial intelligence company with voice interaction and intelligent hardware as the core", and the TicWatch smart watch is the company's flagship product. However, it turns out that smart hardware is not a very big market, and Mobvoi is also adjusting its business direction, gradually expanding from C-end smart hardware to B-end AI solutions. As one of the earliest companies in China to land a large model, Mobvoi developed the large model "UCLAI", that is, the general Chinese language AI, as early as 2020, and launched the AI dubbing assistant "Moyin Workshop" and its overseas version "DupDub" to provide AI dubbing services. In 2023, "UCLAI" will be upgraded to the "Sequence Monkey" large model, and its products will be updated. Today, the company's positioning has changed to "an artificial intelligence company with generative AI and voice interaction at its core".

Huawei Pura 70 series phones go on sale

On the morning of April 18th, Huawei Terminal announced that Huawei Pura 70 series mobile phones were officially launched for sale, which is the second time that Huawei terminals have sold flagship mobile phones without holding a press conference after last year's Mate 60 series. The price of Pura 70 starts at 5499 yuan, the price of Pura 70 Pro starts at 6499 yuan, the price of Pura 70 Pro+ starts at 7999 yuan, and the price of Pura 70 Ultra starts at 9999 yuan. The Pura 70 series is equipped with HarmonyOS 4.2, the Pura70 Ultra is equipped with a super-focused telescopic camera, and the Pura 70 Pro+ and Ultra versions support both Tiantong satellite calls and Beidou image messaging.

-END-