Science and Technology Weekly ｜ 360 released Zhinao 4.0 Weilai reduced the price by 30,000.
Industrial trends
01 360 released Zhinao 4.0, which has the ability of cross-modal generation.
On June 13th, 360 Company held the application conference of 360 Zhinao Big Model, the cognitive general big model "360 Zhinao 4.0" was unveiled, and the 360AI digital man was officially released.
In addition, Zhou Hongyi said that the key to overtaking China’s self-developed large model curve in the GPT era is to develop multi-modes. 360 Zhinao has the ability of cross-modal generation, and has the ability of text, image, voice and video processing, which can be applied to scenes such as Wensheng map, Wensheng map and Wensheng video. At the same time, 360AI Digital People Square was released, providing more than 200 roles, including digital celebrities, digital employees and other categories.
At the meeting, the multi-modal function of 360 Zhinao "Wensheng Video" was first launched in China, and any text script can generate video, which is not limited by professional skills and materials, showing the ability of "making something out of nothing".
Zhou Hongyi said that the core innovation of 360 in the big model lies in AI digital people, and digital people are the future.The most important application entrance of large model. 360 will redefine "digital people" based on the ability of 360 wisdom brain model. Different from traditional digital people who can only output content according to established scripts, 360AI digital people are characterized by having memories, people’s designs and personalities, and being able to reproduce their way of thinking and life experiences.
02 can generate video with one click, and Gen2 is open for free trial.
Recently, Runway, which is known as the video version of Midjourney and provides technical support for Stable Diffusion and Instant Universe, has developed an AI video editing tool Gen, which has been upgraded from the first generation in February to the second generation, and has been officially released from internal testing.
In February of this year, Runway released a model named Gen-1, which can transform an existing video into a new video by applying any style specified by text prompts or reference images.
Just over a month later, Runway officially released the more powerful Gen-2 on March 20th, completing the transformation from "editing existing videos" to "Wensheng videos", and providing multiple functions such as text+reference images to generate videos, still pictures to videos, and video style migration. Now, after the internal test in March, Gen2 finally opened a free trial-although the video it generated is only 4 seconds at present, and it has a total usage quota of 105 seconds.
In terms of pricing, the standard edition member of Gen2 is 15 dollars/month, the Pro edition member is 35 dollars/month, and the annual membership can get 20% discount.
According to reports, Twitter user Steve Mills has experienced its potential ability to generate MV in the Gen-2 beta. This user generated a total of about 500 seconds of video, and finally edited it into a 140-second MV.
The author estimates that the whole production will cost at least $30 at the price of the public version. Coupled with the previous study and exploration stage, this price will be higher.
Cristóbal Valenzuela, CEO and co-founder of Runway, once said when Gen-1 was released: "We have witnessed the explosion of image generation model, and I firmly believe that 2023 will be a year of video generation model."
Chatgpt update, which can generate visual animation with one click.
On June 13th, local time, OpenAI announced a major update to its large-scale language model API (including GPT-4 and GPT-3.5-turbo), including adding function calling function and reducing the use cost. After updating, the cost of embedded model is reduced by 75%, and the tokens input length of GPT-3.5-turbo is extended to 16,000.
Since the end of March, ChatGPT update announced the launch of plug-in function, the number of third-party plug-ins has increased from 10 to more than 70. With the support of these plug-ins, ChatGPT can not only generate web pages and help users find jobs, but also visualize data directly according to their needs. ……
At present, the new version of the model has gradually begun to iterate. The latest versions of gpt-4-0613, gpt-3.5-turbo-0613 and gpt-4-32k-0613 with extended context length all support function calls. Gpt-3.5-turbo-16k does not support function calls, and provides four times the context length, which means that one request can support about 20 pages of text.
In addition, the old model began to be gradually abandoned. Applications using the initial version of gpt-3.5-turbo and gpt-4 will be automatically upgraded to the new version on June 27th. Developers who need more time to transition can also manually specify to continue using the old version, but all requests for the old version will be completely abandoned after September 13th.
04 Nature refuses to post the generated pictures and videos.
On June 13th, local time, Nature magazine announced in an editorial that in order to ensure the integrity, transparency and morality of the research, no pictures, videos or illustrations created by using generative AI tools will be published, unless they are articles related to artificial intelligence.
However, Nature does not completely exclude artificial intelligence. The magazine still allows the use of text generated by generative artificial intelligence in articles, as long as it is clearly stated in the method or acknowledgment section. In addition, all sources of data, even data generated with the assistance of artificial intelligence, must be provided by the author. However, Nature firmly refuses to accept any generative artificial intelligence tool as the author of the paper.
···
Frontiers of science and technology
01 Meta Releases the First Humanoid Model
Meta announced on June 13th local time that it will open some components of a new "humanoid" artificial intelligence model to researchers, which can analyze and complete unfinished images more accurately than existing models.
This model is called I-JEPA, which uses the background knowledge of the world to fill in the missing parts of the image, instead of inferring only from nearby pixels like other generative artificial intelligence models. This method adopts the humanoid reasoning method advocated by Yann LeCun, the chief artificial intelligence scientist of Meta Company, which helps to avoid the common mistakes in the image generated by artificial intelligence, such as the problem of having an extra finger.
Zuckerberg, CEO of Meta, said that sharing the model developed by Meta with other researchers can help the company promote innovation, find security vulnerabilities and reduce costs. He told investors in April this year: "For us, if the industry standardizes the basic tools we are using, then we can benefit from the improvement of others."
02 DeepMind releases new AI system to discover faster algorithms.
Researchers at Google DeepMind, an AI research laboratory owned by Alphabet (GOOG), released a new AI system on Wednesday, which can improve the efficiency and sustainability of computing.
DeepMind said that this latest breakthrough, published in the scientific journal Nature, focuses on finding faster computer algorithms. The computer algorithm isThe foundation is used by enterprises for trillions of times every day.
This AI lab in London is famous for pioneering AI models, such as AlphaFold and mastering complex Go.AlphaGo, which named its new AI system AlphaDev. Based on AlphaZero (iteration of AlphaGo), the system uses reinforcement learning to find faster algorithms for computer science functions such as sorting and hashing. Reinforcement learning is a form of machine learning, in which computers learn and make their own strategies.
Colin Murdoch, chief business officer of DeepMind, said, "This means that we can use fewer resources to do the same amount of calculations."
···
Traffic driving
01 Huawei won the trademark "Wen Jie" and intended to build a car alone?
Trademark electronicsThe system shows that a total of 21 applications for trademark transfer of "Wenjie" have been approved. The transferor is Beijing Yongan Shida Trade Co., Ltd. and Beijing Kechuang Huida Trade Co., Ltd., and the transferee is Huawei Technologies Co., Ltd. The receipt time of the transfer application is March 7 and 8 this year.
The specific categories of trademarks involve categories 12, 9, 7, 39, 37, 42, 2, 35, 3, 4, 16, 38, 11, 1, 5, 21, 27 and 41, of which categories 12 and 9 respectively represent automobiles and vehicles, electronic products and scientific instruments.
According to the first financial report, Huawei responded that "Huawei does not build cars alone, but builds good cars with car companies" has not changed.
In addition, according to the science and technology innovation board Daily, the successful transfer of the trademark "Wen Jie" by Huawei means that Huawei will have the dominant position in the trademark, and it will become the unified brand name of Huawei’s intelligent car selection model in the future.
02 The price of the whole department was reduced by 30,000 yuan, and the price of ES8 fell below 500,000 yuan.
On June 12th, Auto (hereinafter referred to as "") announced that it would adjust the price of all new cars, and the starting price of all models of the company was reduced by 30,000 yuan.
After adjustment, the starting price of ET5 fell below 300,000 yuan to 298,000 yuan; The brand-new ES6 just listed in May this year also dropped to 338,000 yuan, and the price of the brand-new ES8 launched at the end of last year also fell below 500,000 yuan.
This adjustment is not limited to the price, and the car rights of Weilai users have also been lowered.
Specifically, the rights and interests of the first owner of a new car purchase user are adjusted to a 6-year or 150,000-kilometer warranty for the whole vehicle, of which the three-power system can enjoy a 10-year unlimited mileage warranty; Free power exchange will no longer be used as the basic car rights, users can choose to pay a single fee for power exchange, and then a flexible power supply package will be launched.
For the users who have already picked up the car, Weilai promises that the rights and interests of the first car owner will remain unchanged. At the same time, when buying a new car from Weilai, the old users can choose to transfer the car rights of the original car to the new car, or choose to give up their car rights to deduct the car purchase price of 30,000-50,000 yuan. Li Bin, CEO of the company, explained that the old users can choose to give up their rights when buying a new car, which means that they can get 60,000-80,000 yuan cheaper.