news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Meta releases the most powerful open source model Llama 3.1

On the evening of July 23rd, Beijing time, Meta officially released the latest open source large model Llama 3.1 series, further narrowing the gap between open source models and closed source models. Llama 3.1 includes three parameter scales of 8B, 70B and 450B. Among them, the model with 450B parameters surpassed OpenAI's GPT-4o in multiple benchmarks and is comparable to leading closed source models such as Claude 3.5 Sonnet. In terms of specific details, the model context window of the three versions of Llama 3.1 has increased from 8k to 128K, a 16-fold increase, and supports 8 languages ​​at the same time. Among them, the Llama 3.1 -405B model used more than 15 trillion tokens for training, and in order to achieve this training scale, the team used 16,000 H100 GPUs. Officials said that the 405B model is the first Llama model trained at this scale.

Comment: Open source large language models mostly lag behind closed source models in terms of functionality and performance, "but now, we are ushering in a new era led by open source." Meta founder and CEO Zuckerberg also published a blog on the official website to promote the release. He said that Llama 3.1 will be a turning point in the industry, and most developers will begin to use open source mainly. Open source AI is the future direction of development. Nvidia senior research scientist Jim Fan posted a message on X to congratulate the Meta team. He mentioned, "The power of GPT-4 is in our hands. (This is) a truly historic moment.

OpenAI tests AI search function

OpenAI is testing a new set of search features with a small number of users, which will provide more timely information and links to cited materials when answering questions. This means that the company is launching a direct challenge to Google. This new feature, called "SearchGPT", will be released in a test version on a web browser to provide users with an independent search experience, and is expected to be added to ChatGPT in the future. OpenAI did not disclose the launch time of SearchGPT.

Comment: Previously, OpenAI has frequently announced that it will launch an AI search tool. On May 10th local time, OpenAI clarified the rumor that it will launch an AI search engine next week. But OpenAI's AI search tool is finally on the way. According to OpenAI, SearchGPT will be powered by the GPT-4 series of models and combined with real-time information from the Internet, and will also list the source and link of the answer. However, some manufacturers have launched AI search functions before, but the impact on the performance of search giant Google is not great. In the second quarter of fiscal year 2024, Google's advertising business revenue was US$64.616 billion, an increase of 11% from US$58.143 billion in the same period last year. How big the market for AI search is and whether it is a disruptive product remains to be seen.

EU's AI Act will come into effect on August 1

On August 1, the EU Artificial Intelligence Act will officially come into effect across the EU and will be implemented in stages over the next three years. This is the world's first comprehensive law regulating artificial intelligence. After 6 months, the general provisions and the ban on unacceptable risk AI systems will apply; after 12 months, some chapters on high-risk AI systems and general AI models will apply; after 24 months, except for some content such as the classification rules for high-risk AI systems, the rest will apply; after 36 months, some content such as the classification rules for high-risk AI systems will also apply.

Comments: The bill imposes some restrictions on many types of overseas companies doing business in the EU, and the impact extends to areas outside the region. The phased implementation still gives AI companies time to adapt and make adjustments, and the EU has yet to provide further detailed explanations and supporting facilities for some regulations. The bill divides AI systems into prohibited, high-risk, limited-risk, and minimum-risk, and each level has corresponding compliance requirements. There will be many regulations, which are complex and comprehensive. Companies that are willing to go overseas should still study them thoroughly to avoid being fined for non-compliance.

Google AI wins silver medal at IMO Mathematical Olympiad

On July 25th, local time, Google DeepMind announced on its official website that its artificial intelligence had completed an important challenge: using AI to solve the real questions of this year's International Mathematical Olympiad IMO, and was only one step away from winning the gold medal. The IMO competition that just ended last week had a total of six questions, involving algebra, combinatorics, geometry and number theory. The hybrid AI system proposed by Google got four questions right and scored 28 points, reaching the silver medal level. Behind this, AlphaProof and AlphaGeometry 2 developed by the DeepMind team jointly achieved a milestone breakthrough. Among them, AlphaProof is a reinforcement learning system for formal mathematical reasoning, and AlphaGeometry 2 is an improved version of DeepMind's geometry solving system AlphaGeometry.

Comment: This breakthrough shows that general artificial intelligence (AGI) with advanced mathematical reasoning has the potential to open up new areas of science and technology. Nvidia senior scientist Jim Fan wrote that large models are mysterious, they can win silver medals in the Mathematical Olympiad, but they often make mistakes on questions like "Which number is bigger, 9.11 or 9.9?" However, AlphaProof and AlphaGeometry 2 are trained on formal proofs and domain-specific symbolic engines, so even if they are built on general LLMs, they are better at solving professional Olympiad problems.

Baichuan Intelligent completed RMB 5 billion Series A financing

On July 25, market sources said that the large model startup Baichuan Intelligence has recently completed its A round of financing. China Business News asked Baichuan Intelligence to verify the above information, and relevant people replied that Baichuan Intelligence has indeed recently completed its A round of financing, with a total financing amount of 5 billion yuan, and will start its B round of financing with a valuation of 20 billion. It is reported that the investors in Baichuan Intelligence's A round include Alibaba, Xiaomi, Tencent, Asia Investment Capital, CICC and other leading companies and market-oriented investment institutions, as well as state-owned industrial investment funds such as Beijing Artificial Intelligence Industry Investment Fund, Shanghai Artificial Intelligence Industry Investment Fund, and Shenzhen Capital Group.

Comment: In the past year and a half, domestic large-model unicorn companies including Baichuan, Zhipu, and Dark Side of the Moon have rapidly raised funds and grown, with valuations soaring to around 20 billion. Currently, the industry has initially formed a pattern of 6 unicorn large-model companies against more than a dozen base large-model companies such as BAT. Some investors believe that there are few new investment opportunities at the basic model layer, and the opportunities in the next two to three years will be in AI applications.

Google's parent company's Q2 net profit increased by 29% year-on-year, and many indicators exceeded expectations

On July 24, Beijing time, Google's parent company Alphabet (NASDAQ: GOOG, GOOGL) released its second quarter financial report for fiscal year 2024 ending June 30. The financial report shows that Alphabet's second quarter revenue was US$84.742 billion, a 14% increase from US$74.604 billion in the same period last year, and a year-on-year increase of 15% at a constant exchange rate; net profit was US$23.619 billion, a 29% increase from US$18.368 billion in the same period last year. Both revenue and net profit exceeded market expectations.

In terms of specific business, Google's advertising business is the main support, with total revenue of US$64.616 billion, an increase of 11% from US$58.143 billion in the same period last year. The highly anticipated cloud business had quarterly revenue exceeding US$10 billion for the first time, with revenue of US$10.347 billion in the second quarter, an increase of 29% from US$8.031 billion in the same period last year, exceeding market expectations. Management said in the earnings call that cloud revenue was boosted by AI demand.

Comment: Although many indicators in the financial report are impressive, Google's stock price experienced several fluctuations after the release of the financial report. It first rose by 2% and then fell. At the close of July 25, the company's stock price fell by 2.99%. Behind the fluctuations are two different market expectations. Investors are delighted with the booming cloud business and search engine advertising business, but they are also concerned about when Alphabet's huge investment in AI will pay off.

Zhipu launches AI-generated video function Qingying

On July 26, Zhang Peng, CEO of Zhipu AI, announced at Zhipu Open Day that the AI ​​video generation function Ying was officially launched on Zhipu Qingyan, and it only takes 30 seconds to generate a 6-second video. From now on, all C-end users can experience the AI ​​video generation capabilities of text and image through Ying. The Ying API has been simultaneously launched on the big model open platform bigmodel.cn. Enterprises and developers can experience and use the model capabilities of text and image videos by calling the API.

Comments: In the past few months, there have been large video models such as Kuaishou Keling, Luma AI, and Runway Gen-3. Among the major models, Qingying's generation time is not outstanding, but it can be used for free. The reporter found that it took 1 minute to queue up on the first day of Qingying's launch, and the generation process was relatively smooth. From the technical base, Qingying's base video generation model is CogVideoX, which can integrate the three dimensions of text, time, and space. The model refers to Sora's algorithm design and is also a DiT architecture.

OpenAI could lose $5 billion this year

It is reported that OpenAI is one of the fastest-growing companies in history, but it may face soaring operating costs. Based on previously undisclosed internal financial data estimates and analysis by relevant people, ChatGPT maker OpenAI may lose up to $5 billion this year. If this estimate is correct, OpenAI, which was recently valued at $80 billion, will need to raise more funds in the next 12 months or so.

Comment: It is no secret that large model manufacturers have high operating costs, and the majority of the costs are concentrated in computing power costs. If there are no good commercial products to bring in revenue, some manufacturers may find it difficult to sustain the model of continuous financing and continuous burning of money. It is worth looking forward to what kind of commercial model OpenAI's upcoming AI search tool will have.

Tongyi Big Model will become the Olympic AI Big Model application technology provider

Alibaba Tongyi Big Model will become the technical provider of Olympic AI big model applications. A series of AI applications will be unveiled at the upcoming Paris Olympics. These include 360-degree live broadcast special effects, black and white image AI color restoration, carbon emission reduction assistance and other applications supported by Alibaba AI technology. The Paris Olympics will also launch the first big model application, which will be technically supported by Alibaba Tongyi Qianwen and open to official commentators of the International Olympic Committee to assist in the commentary of various events.

Comment: The application of artificial intelligence technology in the Olympic Games is entering a new stage. Previously, the International Olympic Committee (IOC) released the Olympic AI Agenda in April 2023. The core goal of the Olympic AI Agenda is to optimize the event process, improve safety, enhance audience interaction, promote sports culture, and promote the popularization and development of the Olympic movement around the world through artificial intelligence technology. With the application of artificial intelligence technology in the Olympic Games, more people will recognize the capabilities of artificial intelligence.

Cybersecurity firm Wiz rejects $23 billion acquisition offer from Google

Earlier this month, sources revealed that Google's parent company Alphabet was planning to bid $23 billion to acquire the cybersecurity startup Wiz and had been in deep negotiations with the latter. On the 23rd, it was reported that Wiz CEO Assaf Rappaport had announced in an internal email that the company would reject Google's acquisition proposal.

It is reported that Assaf Rappaport said in an internal email that the company will seek to go public as planned and achieve the goal of $1 billion in recurring revenue per year. "It is difficult to reject such an offer, but with our outstanding team, I am confident in making this choice."

Comment: Although Google's cloud business has continued to grow in recent years, it faces competitive pressure from leaders such as Microsoft and Amazon, and Wiz's joining could have provided a boost to Google. Wiz's decision is a setback for Google, which has been seeking to enhance its capabilities in cybersecurity as a key part of its strategy to expand its market share in the cloud computing sector.

YOOZOO AI Playmate Product Begins Testing

YOOZOO Interactive (002174.SZ) recently launched the testing of its first AI playmate product, "Codename Xiaoyoujiang". At the Global Game Industry Conference on July 26, Dai Qi, Chief Product Officer of YOOZOO Interactive, said that "Codename Xiaoyoujiang", as a game-exclusive assistant, provides players with customer service answers and game strategies, while taking into account personalized companionship to create a diverse interactive experience. YOOZOO Interactive founded the AI ​​Innovation Institute last year, focusing on the two major areas of "AIGC" and "AI empowering game globalization". The company has used generative tools to assist in the creation of game animation scripts, storyboards, refinement, dynamics, dubbing, etc.

Comment: Artificial intelligence is showing more and more potential to transform the game industry. Domestic leading game companies such as Tencent and NetEase are also studying how to use this technology to improve game development efficiency and reduce development costs. In some specific links such as painting, large models are surpassing mature painters; in terms of cost, AI also has more advantages. More importantly, AI technology is still moving forward, and its transformation of the game industry has just begun.