ai weekly report｜openai releases new model o1 series; apple intelligence chinese version will be launched next year
2024-09-16
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
openai releases new model o1 series of "strawberry project"
the new openai project, previously known as the "strawberry project", was launched in the early morning of september 13th, beijing time. openai introduced on its official website that a new series of reasoning models will be available, which will be used to solve difficult problems. these models will spend more time thinking before responding, and can reason through complex processes to solve more difficult scientific, coding and data problems than before. chatgpt plus users and team users will be the first to experience the new model, but the new model also has a limit on the number of messages sent.
comments: with the chain of thoughts, this big model can think longer like humans to solve difficult problems. the performance of the new series of models after the update is similar to that of a phd student completing challenging benchmark tasks in physics, chemistry, and biology. openai mentioned that the new model training used reinforcement learning, which means that the big model is looking for a combination of more technologies to solve problems that were originally difficult to solve.
apple intelligence chinese version will be launched next year
in the early morning of september 10, the iphone 16 series was officially released. when introducing the iphone 16 series, apple emphasized the new apple intelligence.
apple introduced that apple intelligence will be tested in the us english version first, and the localized english version will be launched in december this year. at the same time, the chinese, french, japanese and spanish versions will be officially launched next year, and then the chinese version of the iphone will experience apple's ai function.
comments: apple intelligence was first unveiled at wwdc2024. with the help of a generative ai big model, apple intelligence supports intelligent writing and can proofread the user's text until the tone and wording are just right. the tool can be used almost anywhere in writing, including third-party applications. more importantly, siri has been greatly upgraded. when processing complex instructions, apple combines the capabilities of chatgpt. siri can not only accurately understand user intentions, but also provide more intelligent responses, which is a significant improvement over previous versions.
dark side of the moon founder yang zhilin: openai's release of o1 marks a paradigm shift in ai
on september 14, yang zhilin, founder of dark side of the moon, mentioned in a sharing session at tianjin university xuanhuai college that the next paradigm of (ai) is to be done through reinforcement learning. recently, openai will release o1, marking the migration from the left paradigm to the right paradigm, because the left paradigm does not have enough data. in his opinion, the "left paradigm" is to predict the next token and scaling (scaling law), and the "right paradigm" is reinforcement learning and scaling.
comments: after openai released the o1 series, some industry insiders said that this was a victory for the combination of large model technology and reinforcement learning. originally, large models relied on using a lot of data in training, but now there is less and less natural data available. reinforcement learning allows ai to "play with itself" to generate data, and ai learns and improves its own capabilities in the process. with the paradigm change, more and more computing has shifted from training to reasoning.
the agency predicts that apple intelligence will not boost sales soon
apple held its fall conference in the early morning of september 10th, beijing time, launching the new iphone 16 mobile phone. apple's ios 18 system will also provide the apple intelligence intelligent system. analysts at market research firm counterpoint research predict that the impact of apple intelligence on apple's shipments and revenue growth will be more obvious in 2025 and beyond, because the introduction of ai is phased in terms of functions and regions, and iphone revenue is expected to remain stable this year. idc analyst ryan reith believes that consumers also have other options such as gpt to use ai functions, so it is difficult for apple intelligence to directly drive new consumer purchases.
comments: there are different opinions among various institutions on whether iphone 16 will start apple's "super sales cycle". investment bank wedbush previously released a report saying that the iphone 16 series will start the largest upgrade cycle in apple's history, and the initial shipments of the iphone 16 series are expected to exceed 90 million units. at present, after the e-commerce platform opened the pre-sale of iphone 16, iphone 16 pro max became the "out-of-stock king" of the series, and it was sold out in seconds on multiple platforms. the subsequent sales volume needs to be observed.
honor ceo zhao ming talks about apple ai: making the industry sober
on september 6, local time, honor ceo zhao ming announced the global release of the industry's first cross-application open ecological intelligent body at the 2024 consumer electronics show in berlin, germany, and announced that honor ai agent will be commercially available on the next-generation honor all-around flagship honor magic7 series. in an interview after the meeting, zhao ming told reporters that apple's direction for ai will help the entire industry to be more sober, and mobile phone manufacturers cannot simply rely on integrating third-party ai services to label their products as ai phones. a true ai phone should have the ability to recognize human-computer interaction like ai agent, which is the core value of end-side ai.
comment: as the era of massive ai applications approaches, ai applications will show a fragmented trend, just like the mobile app ecosystem. how to effectively call these ai capabilities on the terminal side will become a major challenge facing mobile phone manufacturers. mobile phone manufacturers should think deeply about how to use ai to empower hardware and reconstruct the operating system to achieve deeper innovation.
fei-fei li founded an ai company that received $230 million in financing, with nvidia and others participating
on september 14, world labs, a spatial computing startup led by "ai godmother" fei-fei li, announced that it had raised $230 million. the company's initial funding was jointly led by andreessen horowitz, new enterprise associates and radical ventures, and other investors included amd ventures, intel's intel capital and nvidia's nventures. this round of financing has valued world labs at more than $1 billion, and fei-fei li's goal is to use this money to build cutting-edge ai technology.
comment: fei-fei li's background makes her highly respected in this field, and her research in computer vision and machine learning has had a profound impact. it is reported that world labs' vision is to develop ai models that can understand and interact with the 3d world. the "huge world model" they are developing will provide support for professionals such as artists, designers, developers and engineers.
eu's top privacy regulator launches investigation into google's palm2 compliance
eu regulators said on september 12 that they are investigating an artificial intelligence model from google due to concerns about whether it complies with the eu's strict data privacy regulations. ireland's data protection commission said it has launched an investigation into google's palm2 large model, and its investigation is examining whether google has assessed whether palm2's data processing may pose a "high risk to individual rights and freedoms" in the eu. this is part of the eu's broader efforts in artificial intelligence regulation, with other regulators in the 27-nation bloc reviewing how artificial intelligence systems handle personal data.
comment: with the development of artificial intelligence technology, the protection of personal data has received increasing attention from regulators in various countries, especially in regions such as the european union, which have extremely high requirements for data privacy. earlier this month, irish regulators took action against musk's social media platform x. x company finally agreed to stop using european users' personal information to train its artificial intelligence chatbot grok, and promised to delete personal data from public posts of european users collected from its social media platform. in addition, the meta platform also suspended plans to use content posted by european users to train the latest version of its large language model after pressure from irish regulators.
kuaishou announced that it will co-create aigc movies with 9 directors including jia zhangke
on september 9, kuaishou announced the launch of the keling ai director co-creation plan, which will unite 9 directors to use aigc technology to produce movie-level content and jointly launch 9 aigc short films. it is understood that the directors participating in this plan include li shaohong, jia zhangke, ye jintian, xue xiaolu, yu baimei, dong runnian, zhang chiyu, wang zichuan and wang maomao.
comments: 9 directors will rely on the technical capabilities of keling ai to produce 9 aigc short films with no restrictions on themes. all 9 short films will be generated by keling ai. the film directors completely rely on the video to generate a large model, and the group deeply participates in the creation of film-level content. this is the first time in china.
humanoid robot company "accelerated evolution" completes 100 million yuan pre-a round of financing
the humanoid robot company "accelerated evolution" recently completed a pre-a series financing of 100 million yuan. this round of financing was jointly funded by beyond times, minsheng international, zhongguancun science park and icanx fund, and followed by old shareholder inno angel. the funds will be mainly used to build a universal body and improve software development capabilities. in terms of commercialization, accelerated evolution has received orders for humanoid robots and plans to mass produce and deliver them in small batches within the year. at the same time, the company has also launched a new round of financing to further promote the implementation of technology and products.
comments: humanoid robots have experienced an explosive growth in the past two years, and a number of robot startups have completed financing this year. cheng hao, ceo of accelerated evolution, believes that the current craze for humanoid robots is driven by ai, and general artificial intelligence requires new carriers, especially carriers that can have an impact on the physical world. therefore, the goal of accelerated evolution is to promote the transformation of humanoid robots from specialization to generalization.
the industry is paying attention to the evolution of large models: combining multiple models and technologies
many senior industry insiders recently discussed the limitations of llm (large language model) and how to move from the existing llm to agi (general artificial intelligence) at the gain summit world artificial intelligence summit hosted by the saudi data and artificial intelligence authority. simon see, global director of nvidia's artificial intelligence technology center, said that current artificial intelligence is "narrow" and it is not feasible to rely on a single technology. the industry is now working on combining llm with other knowledge and technologies such as new symbols and calculus for understanding and reasoning. antoine blondeau, co-founder and managing partner of alpha intelligence captal, also believes that ai will not be a single model, but may be a combination of multiple models, and machines will eventually learn to observe, prove or refute, and generalize like humans.
comment: including openai, the industry is solving the problem of poor numerical and reasoning ability from the operation mechanism of ai systems. in fact, it is using various technologies and integrating multiple models to make up for the shortcomings of llm's operation mode, and to some extent make llm's thinking mode more like humans. llm's original method of consuming a lot of computing power and data in training and quickly predicting the next token is not really intelligent, and llm needs to become more like humans.
(this article comes from china business network)
