AI Singularity.com丨AI News丨AI Tool Recommendation丨AI Tool Tutorial
Provide 1000+ AI efficiency tools丨 AIEVA.CN 丨Welcome to visit
AI Singularity reported on July 19, 2024丨AI Information Morning Post
OpenAI发布GPT-4o mini，免费替代GPT-3.5
On July 18, local time, OpenAI released the latest large model GPT-4o mini on its official website, which is a small-sized version of GPT-4o and has multimodal inference capabilities for text, images, audio, and video. Consistent with GPT-4o, GPT-4o mini also supports 128K contextual window tokens.
According to the official evaluation report, the performance of GPT-4o mini is better than GPT-4, and it has about 80% of GPT-4o's capabilities. However, the price of API has dropped by 60%, with 15 cents per 1 million tokens of input and 60 cents per 1 million tokens of output.
According to the test data released by OpenAI, GPT-4o mini's mathematical, coding, visual and other reasoning capabilities are very good, and it is one of the strongest multimodal small models at present.
It is worth mentioning that starting today, GPT-4o mini has replaced GPT-3.5 Turbo to provide services to registered ChatGPT users for free, while GPT-3.5 has completed its historical mission to go offline from ChatGPT. At the same time, this is also the first model based on the instruction hierarchy released by OpenAI.
Microsoft's AI drawing design tool Designer has been launched on the mobile app
At present, Microsoft has released a preview version of the AI design app Designer on the iOS / Android platform, which allows users to create custom images, stickers, greeting cards, invitations, etc. through templates, as well as advanced features for AI editing and redesigning images, users can use the advanced features for free for 15 times, and Copilot Pro subscribers can get 100 times.
Last year, based on DALL· Originally available on the web and Microsoft's own Edge browser, Designer developed by E 3 is now available for free on Windows 11, iOS, and Android platforms for users with a personal Microsoft account.
除了独立应用,微软也在 Word、PowerPoint 中的 Copilot 助手添加了 Designer 支持,Copilot Pro 订阅者可以在 Office 内直接调用。
科大讯飞上线星火大模型Spark Pro-128K API调用
On July 18, iFLYTEK announced that the iFLYTEK Xinghuo API officially opened the long text context model - Spark Pro-128K, which supports 128K Tokens and the lowest price is 0.21 yuan/10,000 Tokens.
Different from traditional text processing models, long text models have more accurate text understanding and generation capabilities and stronger cross-domain migration capabilities, which can understand and generate more information at one time, and are suitable for tasks such as complex conversations, long-form content creation, and detailed data analysis, which can improve the boundaries of model problem solving.
On June 27 this year, iFLYTEK released the Xinghuo large model V4.0, and announced that its long text capability has been improved, and the industry's first "content traceability" function is the first in the industry to solve the illusion problem of long document knowledge Q&A. Specifically, in the reply of the Xinghuo model, the user will be provided with the content of which paragraph the reply of the marked model is referenced, so as to accurately trace the source and easily grasp the accuracy of the source of information collection.
According to the AI Singularity.com, the Spark model Spark Pro -128K has been open to all developers on the development platform for API calls, with a price of 0.21~0.30 yuan / 10,000 tokens, and individual users can directly receive 2 million tokens for free.
The well-known large model faces the 2024 college entrance examination with all subject results, and only two science courses can be taken
On July 17, the Shanghai Artificial Intelligence Laboratory released the test results of all subjects in the college entrance examination for 7 mainstream AI large models, according to the relevant person in charge of the open source evaluation system "Sinan" of the large model, "The current large model still has great limitations. The purpose of organizing AI large models to participate in the college entrance examination is to evaluate the real level of the current large models, identify problems, and continue to promote technological progress. ”
Following the release of the results of the large models participating in the 2024 college entrance examination "Language, Mathematics and English" last month, the Shanghai Artificial Intelligence Laboratory and Sinan Evaluation recently tested 7 AI models in all subjects, and the results have been announced: the scholar · Puyu 2.0 series Wenquxing large model (Puyu · Wenquxing), Ali Tongyi Qianwen large model Qwen2-72B, and GPT-4o are among the top three in liberal arts and science.
According to the score line of Henan Province, the scores of the top three "AI candidates" exceeded the first line of liberal arts and the second line of science respectively.
Among them, Ali Tongyi Qianwen model Qwen2-72B won the "top prize in liberal arts" in the AI college entrance examination with a score of 546 points, and Puyu Quxing became the first place in science with a score of 468.5. On the whole, the science scores of AI candidates are weaker than those of liberal arts, which reflects the general shortcomings of large models in mathematical reasoning ability.
Apple responds to unauthorized use of YouTube videos to train AI: for research purposes only
A few days ago, Bloomberg broke the news that Apple, Nvidia and other technology giants used a large number of high-quality content videos of well-known creators on YouTube without authorization, and converted the content audio of the video into subtitle document material for training their own AI models. Among them, we learned that Apple purchased relevant data from a third party to train the open-source model OpenELM.
Now that further follow-up is coming, Apple's PR staff responded that OpenELM, as an open-source model, does not provide technical support for any AI devices or machine learning functions, including its newly released Apple AI intelligence service "Apple Intellegent".
Apple emphasizes that the purpose of OpenELM models is to contribute to the research community and advance the development of large language models, which are for research purposes only.
According to Apple, since OpenELM is not part of "Apple Intellegent", Apple has never used unauthorized YouTube video subtitles as training data. Apple has said that the data used by "Apple AI" is public data collected with permission and web crawlers.
[This article is from丨AI Singularity Network丨The same name of the whole network account丨Welcome to search and follow]
AI Singularity丨Provide 1000+ AI efficiency tools丨 AIEVA.CN 丨Welcome to visit