New impetus for AI upgrading; anxiety about the application of large models, how does Baidu solve it; Wang Xiaochuan releases open source large model丨AI New Retail Morning News

AI Marketing Observation

《Preserving power, the new impetus for AI upgrade》

In the era of large models, storage not only requires large quantities and high quality, but also requires strong performance, good stability, and energy saving. Provide storage support measures for AI: By supporting all types of storage protocols, we can realize the intelligent adaptation and integration of multi-protocol data in multiple scenarios; by realizing direct data access by smart chips, we can improve the loading and processing of data sets during the training process. speed. (Source: "Yuanchuan Technology Review" WeChat public account)

《Apple’s AI Ambition: Internal Conflict, Counterattack and Challenge》

After Apple launched iOS 10 in 2016, it began a new AI counterattack, mainly reflected in two aspects: vertical integration and the combination of software and hardware. Apple has invested in multiple AI fields, such as speech recognition, facial recognition, etc., and integrated AI functions into its products and services. Measures: It is necessary to find a balance between protecting user privacy and improving computing power, while responding to the attacks of technology giants such as Google and Microsoft on large AI models. (Source: "Chuangye Bang" WeChat public account)

《RLHF no longer requires human annotation. Google proposes AI-based feedback generation: the effect is comparable to human annotation》

RLAIF is an AI-based feedback generation method that can replace human feedback in RLHF, so that large model training is no longer subject to human limitations. RLAIF uses LLM to generate feedback data, and then uses a modified version of the A2C algorithm to perform reinforcement learning and train the target model. RLAIF's performance on text summarization tasks is comparable to RLHF, and even better in some details. (Source: "Xi Xiaoyao Technology Talk" WeChat public account)

Large Model Application Observation

《How does Baidu solve the application anxiety of large models? 》

The large model industry is faced with the lack of application layer and loss of user interest, and needs to shift from "flirting" to innovation that can truly change life and work. Baidu Initiatives: The "Wenxin Cup" Entrepreneurship Competition was held to provide participating teams with the Wenxin Big Model API interface, Baidu Intelligent Cloud computing resources and investment support; the Wenxin Big Model was used to reconstruct Baidu Search and other series with over 100 million users. The product. (Source: "Alphabet List" WeChat public account)

《 Dialogue with Shen Dou: There are many large models on the market today, but most of them will disappear quickly》The large model industry needs to shift to something that can change lives and Innovation in work, and Baidu meets the needs of both C-side and B-side through Wenxinyiyan and Qianfan platforms. Baidu's advantages in the research and development of large model technology: Wenxin Large Model 3.5 has a 50% improvement in model effect, a 2x increase in training speed, and a 17x increase in inference speed, which significantly reduces the cost of inference and allows Baidu to host a larger number of users. (Source: "Geek Park" WeChat public account)

《baichuan Intelligent released Baichuan2, Wang Xiaochuan: The era of Chinese enterprises using LLaMA2 has passed｜Jiazi Discovery》

Features: Supports dozens of languages ​​including Chinese, English, Spanish, French, etc.; selects and screens high-quality vertical industry data based on trillions of Internet data; creates a super-large-scale content clustering system to complete hundreds of billions of hourly data cleaning and Filter weight; multi-granularity content quality scoring system. Prospects: A 100-billion-level parameter model benchmarking GPT-3.5 will be released in the fourth quarter of this year, and a super application is expected to be released in the first quarter of next year. (Source: "Jiazi Discovery" WeChat public account) [End]

Pictures come from the Internet and have been deleted due to infringement

The above is the detailed content of New impetus for AI upgrading; anxiety about the application of large models, how does Baidu solve it; Wang Xiaochuan releases open source large model丨AI New Retail Morning News. For more information, please follow other related articles on the PHP Chinese website!