AI Weekly | Character AI sold to Google for over $2.5 billion; Jensen Huang talks to Zuckerberg: Open source is still the best for the Internet
2024-08-08
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Character AI sold to Google for over $2.5 billion
In the early morning of August 3rd, Beijing time, Character.AI announced that Google has agreed to pay model licensing fees to AI digital human unicorn company Character.AI, and will allow Character.AI co-founder and CEO, president and 30 Character employees engaged in model training and voice artificial intelligence to join Google to participate in the Gemini AI project. It is reported that investors' shares will be acquired at a valuation of approximately US$88 per share, and Google's overall acquisition value of Character.AI will exceed US$2.5 billion.
Comment: Although Character.AI is well-known, it has been short of financing for a long time. The last round of financing was in March 2023, when the well-known venture capital firm Andreessen Horowitz (A16z) led a $150 million Series A financing. Previously, the media revealed that Character.AI had difficulty in financing and was looking for mergers and acquisitions with Meta and other large companies. Musk's xAI was revealed to be one of the important acquirers, but was later denied by Musk.
Huang Renxun talks to Zuckerberg: Open source is still the best option for the Internet
On July 29th, local time, at the 51st SIGGRAPH graphics conference held in Denver, the United States, Nvidia founder and CEO Huang Renxun and Meta founder and CEO Zuckerberg held a fireside chat to discuss the future of artificial intelligence and simulation. During the conversation, Huang Renxun praised Meta's open source big model, and he and Zuckerberg both agreed that the open source route of big models should be adhered to. As the "head" of a representative company of open source big models, Zuckerberg generously admitted in the conversation that part of the reason for open source is that "we are later than some other technology companies, and when we built these facilities, we no longer have a competitive advantage." However, in the longer term, he believes that one day, every company will have its own AI-just like everyone has their own social media profile.
Comment: This is a clash of opinions between NVIDIA, the most important upstream computing power supplier in the large model industry, and Meta, the publisher of Llama 2, the most well-known large model in the open source field and regarded as an open source benchmark in the industry. Observing the AI industry, from the perspective of interests, NVIDIA will definitely become a firm supporter of large model open source, because only open source can attract more companies to invest in AI "gold mining", and NVIDIA, the largest provider of AI computing power, can also continuously "sell shovels".
Meta plans to pay Hollywood stars millions of dollars for their voices
On August 3, several sources familiar with the negotiations said that Meta is offering millions of dollars to Hollywood stars for their recordings and the right to use their voices in AI projects. According to people familiar with the matter, Meta has discussed developing a chatbot that can serve as a user's digital assistant or "friend", similar to Apple's Siri. For example, users can talk to a chatbot with the voices of Hollywood stars.
Meta is rushing to complete these transactions to buy more time for the development of AI tools, which it plans to showcase at the Connect 2024 conference in September. Although it is not yet clear how Meta will use these voices, this move shows the company's emphasis on and investment in the development of artificial intelligence technology.
Comment: Previously, OpenAI was involved in an infringement storm because one of the voices of the voice chat robot GPT-4o was similar to that of actress Scarlett Johansson, so Meta's negotiation to purchase the voice at a high price was necessary. However, it was reported that negotiations between Meta and some actor representatives stalled several times and then restarted because the parties could not reach an agreement on the terms of using the actor's voice. Meta hopes to obtain the right to use the voice multiple times within a fixed period of time for a single project, while the actor's representative is seeking stricter restrictions.
Nvidia may face antitrust investigation
It is reported that the US Department of Justice may launch an antitrust investigation into Nvidia. According to people familiar with the matter, the US Department of Justice has recently received complaints from Nvidia's competitors, such as AMD and AI chip startups, involving Nvidia's threats to punish customers who also purchase its competitors' products, and concerns about Nvidia's acquisition of startups to strengthen control over software used by AI developers. The US Department of Justice is collecting relevant information and conducting an investigation.
Comment: AI giants are under close regulatory scrutiny. In June, there was also news that US federal regulators had reached an agreement to allow an antitrust investigation into Microsoft, OpenAI, and Nvidia for their dominance in the artificial intelligence industry. In fact, in addition to selling AI chips, Nvidia is also strengthening its voice in the field of AI through investment. This year, Nvidia has participated in no less than 16 financings or acquisitions, including the acquisition of Shoreline.io and Run:ai. The two acquired companies are respectively targeting enterprise developers and developing software to optimize the performance of graphics card clusters.
Intel to cut 15% of its workforce, suspend dividends
On August 1, local time, Intel released its second quarter financial report for fiscal year 2024. In the second quarter, Intel's revenue was US$12.8 billion, a year-on-year decrease of 1%; its net loss was US$1.6 billion, and its net profit in the same period last year was US$1.5 billion, a year-on-year loss. On the same day, Intel's stock price fell 5.5% to US$29.05 per share. On August 2, Intel's stock price fell by more than 20%. Announced along with the results was Intel's plan to cut spending and the news that dividends would be suspended starting in the fourth quarter of 2024. Intel said it would implement comprehensive spending cuts, including a 15% layoff.
Comment: Intel's poor second-quarter performance is related to its investment in AI PCs. In a conference call after the release of the earnings report, Intel management pointed out that profitability in the second quarter was lower than expected, partly due to the company's decision to speed up the production of Ultra AI PC CPUs. Intel has previously stated that investments made to define and promote the AI PC category will put pressure on profits in the short term. It is worth noting that although AI PCs have good growth expectations, the implementation of large models on PCs is still in its early stages.
AI image generation platform LiblibAI completes hundreds of millions of yuan in financing
On July 29, AI image generation platform LiblibAI announced that it had completed three rounds of financing in the past year, with a total amount of hundreds of millions of RMB. This financing scale is the largest financing amount in the domestic AI image track, and is jointly participated by several top venture capital institutions: the angel round was invested by Source Code Capital, Gaorong Ventures and Jinshajiang Ventures; the second round was led by strategic investors; the latest round was led by Mingshi Capital; and old shareholders continued to support it for multiple rounds. LiblibAI said that the financing amount will be mainly used in three aspects: building a large-scale computing power middle platform to support user reasoning and training computing power needs; research and development of plug-ins, fine-tuning models and control capabilities based on image models; support and operation of developer ecology and original model authors.
Comments: Founded in May 2023, LiblibAI is one of the earliest AI image generation platforms in China. According to the announcement, LiblibAI has accumulated nearly 10 million professional AI image creators, more than 100,000 original models, and produced and shared more than 230 million AI images. The core members of LiblibAI graduated from famous domestic and foreign universities such as Tsinghua University, Peking University, and Carnegie Mellon University, and have work experience in Internet companies such as Tencent and Alibaba.
AI robotics company Stardust Intelligence receives tens of millions of dollars in financing
On July 31, a reporter from China Business News learned from AI robot company Astribot that the company has completed tens of millions of dollars in Pre-A round of financing, led by Matrix Partners, followed by industrial capital such as Daotong Investment and Qinghui Investment, and old shareholder Yunqi Capital, with Huaxing Capital serving as the exclusive financial advisor. This round of financing will be used for top talent recruitment, R&D investment, commercial deployment and other work.
Comment: Financing for humanoid robot companies continues. In July, Zhuji Power, a general humanoid robot startup, also completed its Series A financing. Other humanoid robot companies that received financing this year include Galaxy Universal Robots, Kepler Exploration Robotics, Yushu Technology, and Xingdong Jiyuan. Judging from the background of Stardust Intelligence, the company's founder Lai Jie once worked in Tencent Robotics Laboratory and Baidu's "Xiaodu Robot". The company expects to complete the release of the first generation of products in the near future and complete commercialization within 2024.
Apple uses Google chips to train models
On Monday local time, Apple stated in a technical paper that the two artificial intelligence models supporting Apple Intelligence were pre-trained on Google's cloud chips. Google's self-developed chip is TPU (tensor processing unit), which is mostly used within Google and not sold to the outside. This may mean that Google's self-developed chips are more widely used, and large technology companies may actively look for alternatives to Nvidia's GPU (graphics processing unit) outside of Nvidia. On July 30, Nvidia's stock price fell more than 7% to close at $103.73 per share.
Comment: Nvidia's competitors have never stopped challenging it. Google has its own chips (tensor processors), and Apple is also a chip manufacturer, with A series chips and M series chips. In addition to Nvidia chips, Microsoft also uses AMD's MI300 chip. In addition to using Google chips to train models, Apple previously introduced that it will use A17 Pro and M series chips to drive Apple Intelligence in terminal applications.
Vidu, an AI video generation tool under Shengshu Technology, is launched globally
On July 30, Shengshu Technology announced that its AI video generation tool Vidu was launched globally and can be accessed on the web at www.vidu.studio, with the fastest inference speed of 30 seconds. In April this year, Zhu Jun, deputy dean of the Institute of Artificial Intelligence at Tsinghua University and chief scientist of Shengshu Technology, released Vidu at the 2024 Zhongguancun Forum. Shengshu Technology said that the model can generate videos with a maximum length of 16 seconds, which is fully comparable to Sora. This time, Vidu opens two core functions: text-generated video and image-generated video, providing two length options of 4s and 8s, with a maximum resolution of 1080P.
Comments: From the evaluation, Vidu's performance in terms of basic requirements for AI videos, including consistency, content content, and fluency, is almost perfect. The generated images are basically stable and smooth with details in place. At the same time, the videos generated by Vidu are beautiful, textured, and generated quickly, but there are also some shortcomings, including occasional penetration and adhesion of the images. For text comprehension, Vidu can basically understand and meet most of the requirements, but there are still occasional inadequate responses to prompts.
Google Gemma 2 series adds 2 billion parameter model
On July 31, local time, Google announced that the Gemma 2 series added a small model, Gemma 2B (2 billion parameters). In the large model arena LMSYS Chatbot Arena, Gemma 2B scored higher than GPT-3.5-Turbo. Released with Gemma 2B are the security content classifier model ShieldGemma and the model interpretability tool Gemma Scope. Gemma 2 is Google's open source model. In June this year, Google released the Gemma 2 series, including Gemma 2 27B and 9B.
Comment: Google is also working on models with very small parameters. In July, OpenAI released a new lightweight large model, GPT-4o mini. Google also released a lightweight model, Gemini 1.5 Flash, in May this year. The Claude 3 series released by Anthropic in March includes the lightweight Claude 3 Haiku. Small models consume less energy to run, are suitable for end-side applications, and there is room for improvement in model capabilities.
AMD raises its data center GPU sales forecast for this year
On July 30th local time, AMD released its second quarter financial report for fiscal year 2024. In the second quarter, AMD's revenue was US$5.835 billion, a year-on-year increase of 9%, and its net profit (Non-GAAP) was US$1.126 billion, a year-on-year increase of 19%. In terms of business departments, the data center division, including CPUs, GPUs, etc., had revenue of US$2.834 billion, a year-on-year increase of 115%. The client division, which includes personal computer sales, had revenue of US$1.492 billion, a year-on-year increase of 49%. AMD CEO Lisa Su said in a conference call that AMD's data center GPU sales are expected to be US$4.5 billion in 2024, higher than the US$4 billion expected in April this year.
Comment: AI has catalyzed AMD's second-quarter performance. In the second quarter, the revenue growth of its data center division was mainly driven by a significant increase in Instinct GPU shipments and strong sales growth of the fourth-generation AMD EPYC CPU. Su Zifeng also revealed that after the delivery of AMD's MI300 chip, an accelerator for AI and high-performance computing, the chip's revenue exceeded $1 billion in the second quarter. Microsoft has increased its use of MI300 chips as computing power support for GPT-4 Turbo. In addition, AMD has also accelerated its layout of AI through investment this year.
(This article comes from China Business Network)
