I. eventsBsgblockchainsplitgame: may 15thBsgblockchainsplitgameGoogle held its annual Iamp O developer conference. Google updated and upgraded Gemini 1 at the conference.Bsgblockchainsplitgame.5 Pro version, and launched the Gemini 1.5 Flash lightweight small model. At the same time, Google released the video generation model Veo, the Sora of the standard OpenAI, and the advanced visual and dialogue response agent project called Astra. In addition, Google officially launched the "AI Overviews" search function, the search engine will directly summarize the search results. II. CommentsBsgblockchainsplitgameThe Google IWeiO developer conference focused on artificial intelligence, where Google announced the latest developments in Gemini AI and shared around the integration of Google's core business and generative artificial intelligence. Model product matrix and technical capability upgrade: (1) Gemini 1.5 Pro advanced model: the maximum support context window will be upgraded from 1 million Tokens to 2 million, supporting 35 languages at the same time. You can now analyze longer documents, code bases, video and audio recordings than ever before. (2) Gemini 1.5Flash lightweight model: the model meets the needs of users for low latency and low cost. Based on "distillation" technology, it is designed for large-scale services with faster speed and low cost of US $0.35 / million Tokens. It is suitable for summaries, chat applications, image and video subtitles, data extraction of long documents and tables, and other tasks that need to be processed quickly. (3) Gemma upgrade of open source large model: the number of Gemma 2 parameters that will be launched next month will reach 27 billion. At the same time, the visual language model PaliGemma is also added to the Gemma model series. (4) Imagen 3: compared with the previous generation, the text information can be identified more accurately, and the created pictures will be more consistent with the text description. (5) Video generation model Veo: you can generate video through text and image, with a resolution of 1080p and a video duration of more than 1 minute, formally calibrating the Sora model of OpenAI. Model enabling products and scenarios: (1) the first AI Agent product Project Astra: an AI agent based on a new audio overview function and a large Gemini model, which can achieve multimodal understanding and real-time dialogue. At the conference, Google demonstrated that AI assistants can use camera video to identify commands such as "what makes a sound" and "where are we now?" as well as voice interactions similar to smart glasses. (2) AI Overview (AI Overview) function: Google deeply integrates the Gemini model with its search engine, and the search engine will directly summarize the search results. At the same time, Google search will also have the ability of multi-step reasoning, which can deal with long-term problems with multiple constraints at one time, and support the new search form of "shooting video" search solution. (3) Android 15, which integrates Google's Gemini model, will provide more AI features, such as the ring-and-search feature already available on Samsung AI phones, and summarizing the contents of PDF files. Hardware: (1) the sixth generation TPU chip Trillium: the computing performance of the previous generation chip is 4.7 times higher than that of the previous generation chip. Cloud users can use the new chip from the second half of this year. At the same time, Google Cloud will use Nvidia's latest Blackwell architecture GPU in early 2025. Google has made remarkable progress in comprehensively upgrading its large model, AI applications and development tools, in which Google has gradually transformed its search engine into an AI agent portal through AI empowerment, which can deeply understand users' needs and provide efficient, multimodal and personalized services. Recently, the GPT-4o model also shows a great improvement in human-computer interaction ability. With the accelerated iteration of the technical capabilities of large models and the continuous expansion of AIGC application scenarios, the deployment of downstream applications is accelerated, and the industrial chain of large models is expected to accelerate the formation of commercial closed loops. It is recommended to pay attention to AI office, AI education, AI e-commerce and other areas that are easy to land quickly. At the same time, in terms of landing terminals, it is recommended to pay attention to the terminal reconstruction brought about by the improvement of the capability of large models, such as AIPC, AI mobile phones, smart home, intelligent cockpit and so on. Key companies: Jinshan Office, iFLYTEK, China Science Chuangda, Fluorite Network, China Science Dawn. Risk tips: technology development is not as expected, downstream demand is not as expected, industry competition is intensified, and so on. [disclaimer] this article only represents the views of a third party and does not represent the position of Hexun. Investors operate accordingly, at their own risk.
[disclaimer] this article only represents the views of a third party and does not represent the position of Hexun. Investors operate accordingly, at their own risk.