Alibaba Cloud releases the Tongyi Qianwen 2.5 model, while Apple releases the AI PC chip M4.
OpenAI announced product update next Monday: it will not involve GPT-5 and search engines
On May 10th local time, OpenAI announced on social media that updates to ChatGPT and GPT-4 will be announced live next Monday. “It’s not GPT-5, it’s not search engines. We’re working hard to develop things that people think people will like,” said Sam Altman, CEO of OpenAI. Previously, there were frequent rumors that OpenAI would launch an AI search engine. After the news of not launching an AI search engine spread, Alphabet, the parent company of search engine manufacturer Google, saw a significant increase in its stock price during intraday trading on May 10th, with a narrower decline of 0.77%.
Comment: The form of AI search engines is not yet mature. Representative products include Bing search engine integrated with ChatGPT and AI search engine products from startup Perplexity, which have similar features such as generating responses and citation sources. Fu Sheng, Chairman and CEO of Cheetah Mobile, recently commented that AI search now focuses on analyzing comprehensive results, such as survey reports, and users may search for simple content or addressing, which AI search cannot meet. In fact, the search engine landscape has not undergone disruptive changes due to the arrival of big models, and Google’s global market share is still around 90%. The immaturity of product form may be the reason why OpenAI is not eager to release AI search engines, and how big models can change user search behavior still needs to be explored.
AI image software Remini has become popular with clay filters
During the May Day holiday, the AI imaging software Remini climbed to the top of the Apple App Store app list with its clay filter. According to data from Qimai, the download volume of Remini in iOS China has exceeded 1.7 million times in the week before and after May Day, with revenue exceeding $25000 (180000 RMB). At present, Apple users can directly download Remini from the App Store and try it out for free for a week. If they want to use its AI filter function, the subsequent package fee is 68 yuan/week and 548 yuan/year.
Remini was launched by Beijing Dagong Technology in 2019, starting with intelligent restoration of old photos. In 2022, it was acquired by Italian company Bending Spoons, which is also a legendary company. Its strategy is not to develop a product from scratch, but to acquire products with mature brands, user groups, and markets for “refurbishment”. In addition to Remini, Bending Spoons also acquired the domestic note taking software Everonote last year, as well as video editors Splice, Film, and event startup Meetup.
Comment: Although there has not been a super application in the field of AI yet, in the field of imaging, AI generated images have not become popular for the first time. Zhang Yi, CEO of iMedia Consulting, told First Financial that the popularity of Remini is closely related to the combination of popular culture and predictive gameplay with the product itself. Its presentation format is also close to users, hitting their curiosity. Secondly, the sharing of top anchors is also a driving force for its dissemination. Another important point is that AIGC has outstanding performance in the field of image and video. However, the wind of the image app blew one after another. Initially, there were AI avatars and AI face changes, but later there were AI comics and AI photos, all of which calmed down after a brief burst of popularity. It remains a question mark how long Remini will become popular.
Alibaba Cloud Releases Tongyi Qianwen 2.5 Big Model
On May 9th, Alibaba Cloud officially released the Tongyi Qianwen 2.5. Alibaba Cloud stated that the performance of this model comprehensively surpasses that of GPT-4Turbo. Alibaba Cloud stated that compared to version 2.1 of Tongyi Qianwen, the comprehension ability, logical reasoning, instruction adherence, and code ability of Tongyi Qianwen 2.5 have been improved by 9%, 16%, 19%, and 10%, respectively; Compared with GPT-4, in the Chinese context, Tongyi Qianwen 2.5 has multiple abilities that surpass GPT-4, including text comprehension, text generation, knowledge Q&A and life advice, chatting and dialogue, and security risks. With the release of Tongyi Qianwen 2.5, Alibaba Cloud stated that on the authoritative benchmark OpenCompass, Tongyi Qianwen 2.5 scored on par with GPT-4Turbo, marking the first time a domestic large model has achieved this score on this benchmark. Alibaba Cloud also released the latest open-source model Qwen1.5-110B with 110 billion parameters. Alibaba Cloud stated that this model surpassed Meta’s Llama-3-70B model in benchmark evaluations such as MMLU, TheoremQA, and GPQA.
Comment: “Benchmarking” GPT-4 is becoming the trend of domestic large models. In April of this year, Shangtang Technology announced the release of the newly upgraded Nissin SenseNova 5.0 large model, which comprehensively benchmarks GPT-4 in terms of comprehensive capabilities, significantly improving knowledge, mathematics, reasoning, and coding abilities. While accelerating technological catch-up, domestically produced large models are also continuously promoting industry implementation. According to data released by Alibaba Cloud, currently, the Tongyi Big Model has served over 90000 enterprises through Alibaba Cloud and over 2.2 million enterprises through DingTalk services.
Apple releases AI PC chip M4
On May 7th local time in the United States, Apple released a new iPad Pro equipped with a new generation of AI PC chip, Apple Silicon M4, specifically designed for AI and based on ARM architecture. Apple called it an “extremely powerful artificial intelligence chip” that can quickly isolate the subject and background in videos.
It is reported that the new M4 adopts TSMC’s second-generation 3nm technology, with 28 billion transistors and supports a new series connected OLED display engine. Its CPU performance is 50% faster than M2, and GPU performance is 4 times higher than M2. It is equipped with a new NPU (new neural engine) and supports 38 trillion AI computing processing power per second, which can be up to 60 times faster than Apple’s A11 chip’s neural network engine (6000%). Based on the M4 chip and new product design, the new iPad Pro runs 4 times faster (400%) than the M2 iPad Pro and 10 times faster (1000%) than the original iPad Pro.
Comment: Apple CEO Tim Cook said, “This is the most important day since the launch of the iPad.”. Although Apple has not announced the price of the new iPad Pro, it will undoubtedly be the most expensive tablet Apple has released so far. During Apple’s latest quarterly earnings conference call, Cook told investors that the company will continue to invest in artificial intelligence, “believing that the transformative power and strong advantages of artificial intelligence will make us stand out in this new era.”
Apple phone or equipped with ChatGPT
There are reports that Apple is about to reach an agreement with OpenAI to introduce OpenAI technology into the new generation iOS operating system. Both parties are negotiating terms to use the ChatGPT feature in Apple’s next-generation iPhone operating system. Previously, there were rumors of negotiations between Apple and Google to include a Gemini chatbot in their products. Recently, Apple also announced that it will deploy self-developed chips in cloud servers to launch AI functionality.
Comment: Based on multiple reports, Apple’s products will adopt self-developed and third-party collaborations for the large parameter model. The large parameter model will be developed in collaboration with other manufacturers, while the small parameter model will be self-developed. In addition, Apple will also promote the implementation of AI at the chip level. In addition to the MM1 model with 3 billion parameters, Apple focuses more on the layout of end side models, with the end side scenario based small model ReALM having a minimum parameter of 80 million. These end side small models may not need to be connected to the internet and can perform calculations on end side devices.
“AI Godmother” Li Feifei Entrepreneurship, Targeting “Space Intelligence”
Famous Chinese American artificial intelligence scientist and Stanford University professor Li Feifei recently founded a new AI company. According to foreign media reports, the new company founded by Li Feifei is mainly engaged in the research and development of “spatial intelligence”, using humanoid visual data processing technology to enable AI to perform advanced reasoning. It is reported that the company has completed a seed round of financing, with investors including a16z and RadicalVentures. Li Feifei did not speak to the public, but her personal LinkedIn page has been updated accordingly. The resume column at the top shows that she has been in a new position since January 2024, with related projects anonymously displayed as “something new” and work status as “full-time”.
The “space intelligence” that the new company aims to do is consistent with the AI application direction frequently mentioned by Li Feifei on many occasions recently. She gave a keynote speech at the TED conference last month, introducing that space intelligence is to understand the relationships between objects and make new discoveries or predictions from them. This is a computer vision intelligence that is more advanced than traditional visual recognition. It is a research field that integrates research results from multiple fields such as natural language models, robots, and computer vision. Machines can perform more complex visual reasoning like humans, and then take more realistic actions.
Comment: Li Feifei is a pioneering figure in the field of AI, known as the “AI Godmother”, and therefore his entrepreneurial trend has also received much attention. Li Feifei has gained great fame in the field of AI by developing a large-scale image dataset called ImageNet, which helped to create a new generation of computer vision technology that can reliably recognize objects and is also one of the underlying technologies of ChatGPT. Years ago, when asked about the key figures in the direction of technology for the next 25 years, Li Kaifu mentioned the name “Li Feifei” without hesitation.
Former head of Miao Ya founded Mu Yan Zhiyu, with a valuation of over 100 million US dollars
According to Tianyancha, Mu Yanzhiyu, a new company founded by the former head of AI product Miaoya, has completed a Pre-A+round of financing of 120 million yuan, led by well-known investment firm Hillhouse Capital. It is reported that less than half a year after its establishment, Mu Yan Zhiyu has completed three rounds of financing with a valuation of over 100 million US dollars, and other investors include Gao Rong Capital, Zhipu AI, etc.
In 2023, Alibaba’s AI product, the Miao Duck Camera, became popular. The person in charge of this product is Zhang Yueguang, a senior product expert from Alibaba’s entertainment team. At the end of 2023, Zhang Yueguang resigned to start a business. According to Tianyancha, his new company “Muyan Zhiyu” was registered in December 2023.
Commentary: In addition to the general model of burning money, the industry believes that there is still a great opportunity for the implementation of the AIGC application layer. Entrepreneurship in this field is also in full swing, and the successful development of the popular AIGC product Miaoyang’s resume may be a great attraction for Zhang Yueguang to investors. However, currently there is no more external product information available from Mu Yan Zhiyu. Public information shows that Mu Yan Zhiyu is an artificial intelligence application software developer whose main business direction is in the entertainment industry, and new products are being developed.
In the first quarter, domestic AIGC and AI industry application financing exceeded that of new energy
On May 9, 2024, at the Yangtze River Unicorn Summit, Teng Binsheng, Professor of Strategy, Vice Dean, and Director of the Global Ecosystem Research Center for the New Generation Unicorn at Yangtze River Business School, stated that AI emerging companies have become a new force in China’s unicorns: in the first quarter of 2024, the total financing amount for AIGC and AI industry applications was nearly 20 billion yuan, exceeding that of new energy. Teng Binsheng pointed out that from the distribution of unicorns, China is often more related to automobiles, and the proportion of AI in the United States is the highest. This means that in terms of venture capital and capital flow, China still seizes the opportunities that can truly create returns within a few years. But he also stated that there are huge advantages to doing AI in China. Within a few weeks after ChatGPT was released, there were no less than 10 large models produced domestically, with a fast response speed. This rapid response ability focuses on the AI track, which has accumulated various execution abilities for decades. Li Haitao, Dean, Professor of Finance, and Distinguished Dean and Chair Professor of Changjiang Business School, stated that by 2032, generative AI is expected to form a market with a scale of 1.3 trillion US dollars, and the proportion of technology expenditure will expand from less than 1% to 10% to 12%.
Comment: From the perspective of investment mentality in fields such as AI, Teng Binsheng mentioned when discussing the distribution of domestic and foreign unicorns that this “eagerness for quick success and instant benefits” is reflected in a previous sharing by Zhu Xiaohu. The so-called sharing is what Zhu Xiaohu, the managing partner of Jinshajiang Venture Capital, mentioned: “Today, everyone relies on profits and dividends. This year, the second tier only looks at dividends, and the first tier also looks at dividends. My assumption is to get back the principal through dividends within 5 years.”. Based on the different environments of entrepreneurship and investment in large-scale models at home and abroad, there have been other discussions in the industry recently. Some entrepreneurs believe that it is still necessary to start from reality and combine products to create large-scale model products that can see commercial returns. Domestic entrepreneurs are more suitable for this path. At present, whether in China or abroad, the field of large models is still waiting for popular applications to emerge. It is not yet a consensus whether it is better for domestic startups to combine products to implement large models or invest in basic research and development of large models. But when a large amount of capital is invested in the field of large models, the market undoubtedly hopes to see commercial return opportunities faster.
The Trend of Chip Manufacturers and Collaboration with Large Model Application Developers
MediaTek recently held the first Dimensity Developer Conference, during which it launched the “Dimensity AI Pioneer Program” and provided relevant developer solutions to support large model manufacturers in the implementation of end-to-end AI technology and innovation of end-to-end generative AI applications. Zhang Li, Senior Director of Ecological Development at MediaTek’s Wireless Communication Business Unit, stated in a post meeting interview that some large model manufacturers have seen the transition from app applications to a large model application ecosystem, and have reached a stage where computing power is required from cloud chips and terminal chips.
Comment: Chip manufacturers are increasingly valuing cooperation with developers of large model applications. Behind this, large model applications require significant computing power, and the computing power of chips that can be carried by end devices is limited. In order to facilitate deployment and optimize computing power, large model manufacturers are gradually coming together with chip manufacturers. It is understood that currently, most of the large models that can be run on mobile devices have around 7 billion parameters, and the supported application scenarios for large models are limited. They are not yet suitable for conducting large model voice conversations in the event of a network outage. In addition to large model manufacturers demanding computing power from chips, the industry is also promoting the miniaturization of large models. By using more data for training, models can achieve better results, in order to promote the implementation of related applications on mobile devices.
UK autonomous driving company Waveve completed a $1.05 billion financing
On May 7th local time, British autonomous driving company Wave announced the completion of a $1.05 billion Series C financing, led by SoftBank Group and co invested by NVIDIA and Microsoft. As of now, Wave has completed three rounds of financing, with a cumulative financing amount exceeding 1.308 billion US dollars. Wave stated in a statement that this new financing will be used to develop the first autonomous driving solution for mass-produced vehicles and will launch the first generation of autonomous driving AI models. British Prime Minister Sunak stated in a statement that this investment is the largest investment received by a British artificial intelligence company to date.
Comment: Wave has applied neural network reinforcement learning to cars since 2017, and the company calls it the first company to develop and test end-to-end deep learning auto drive system on public roads. Wave’s AV2.0 auto drive system, like Tesla’s FSD system, adopts an end-to-end solution, replacing the perception, planning and control links of traditional autopilot with data-driven machine learning. The large amount of financing obtained by Wave also indicates that end-to-end autonomous driving solutions are more favored by capital.