Interface News reporter | Chen Zhenfang
Interface News Editor | Song Jiannan
OpenAI held the second day of the spring launch.CasinoholdemGoogle is facing off with the new Igamo developer conference.
The event smelled of gunpowder since 1: 00 a. M. Beijing time on May 15. Google chose to "announce everything" at the meeting: it released and updated more than a dozen products in a row.CasinoholdemIncluding AI assistant Astra, text model Imagen3, Vincent video model Veo of standard Sora, and flagship large model Gemini.
When OpenAI lost its search and launched the latest flagship model GPT-4o, Google, which has long dominated search, not only redesigned AI search, but also launched AI map recognition assistant at the same time.
Gemini's new voice conversation function Live directly targets the GPT-4o of OpenAI. It can also inquire about the surrounding situation in real time through the mobile phone, and can follow up in time even if the conversation is interrupted.
In addition, Google browser Chrome will add GeminiNano. The latter is a lightweight version of the Gemini series and is mainly designed for mobile devices.
Google also said that another small model, Gemma2Casinoholdem.0 will be launched this summer, including the open source model PaliGemma, which can be used to tag photos and add titles to images. The Gemma model uses the same technology stack as the Gemini model, but has a smaller scale and is suitable for deployment in a resource-constrained environment.
To a large extent, the artificial intelligence competition is also a competition for smartphones. SameerSamat, Google's vice president of product management, made it clear that Google will further optimize its Android operating system through Gemini. This optimization will first be reflected in Google's own mobile phone Pixel.
Gemini is obviously the protagonist of this conference, especially in multimodal and long-context technology.
In the past few months, Gemini has launched Google 1, which can perform long-context previewsCasinoholdem.5Pro, a series of improvements have been made in translation, coding and reasoning. At present, the context length of Gemini 1.5Pro has been refreshed from 1 million token (the basic unit of text processing) to 2 million token, doubling in three months, indicating that the company is eager to "show off its muscles" to the outside world.
It has been a year since the advent of Gemini, this large multimodal model has been able to reason across text, images, video, code and so on. According to Google, 2 billion users and more than 1.5 million developers are using the Gemini model, which can be used to debug code, gain new insights and build the next generation of artificial intelligence applications.
In order to further demonstrate the various features of the model, Google has made a more detailed introduction to different scenarios such as search, photos, Android and so on.
For example, in terms of search, Gemini has brought it a comprehensive AI transformation. Users can query with updated, longer and more complex questions, or even search with photos. Google plans to launch a "AI Overview" search in the United States this week, which will be launched in other countries.
Google demonstrated the function of "asking for photos" on the spot. When users pay in the parking lot but forget the license plate number, they may search for keywords in mobile phone photos and browse a large number of passing photos to find the license plate. But now, by asking for photos, you can accurately tell frequent cars, triangulate them, and tell you the license plate number.
For example, you can ask the photo when your child learned to swim, or even let the photo tell you how your child's swimming is going.
Gemini is not only a chat robot, but also a personal assistant that can help users deal with complex tasks and take action. Gemini 1.5 Pro has also been introduced into Google's cloud computing service GoogleWorkspace. Google claims that Gemini can do all the steps needed to do the job. Take the return as an example, AI can search the receipt in the email, find the corresponding order number, automatically fill in the return form, and arrange pickup.
A big model is a math contest, and it takes a lot of math to train the most advanced models. In the past six years, the industry's demand for machine learning computing has increased 1 million-fold, and has increased tenfold every year. As an important player in the AI era, Google has also made a lot of efforts in infrastructure.
That night, Google released the sixth generation of TPU, an application-specific integrated circuit designed by Google to accelerate machine learning workloads, saying that Trillium is its highest-performing and most efficient TPU to date, with a 4.7-fold increase in computing performance compared to the previous generation of TPUv5e, and is scheduled to be available to customers by the end of this year.
Gemini is trained and served entirely on Google's self-developed fourth-and fifth-generation TPU, and other leading artificial intelligence companies, including Anthropic, have also trained their models on TPU.
But while Google "infuses" its various products with AI features, it means that users need to make more concessions to their personal data. In response, Google promised not to use user files on its platform to train Gemini or other artificial intelligence models.
Google CEO Pichai said that the day's press conference mentioned 121 times "AI", enough to show the importance of AI to Google. But in addition to emphasizing the importance, the expected counterattack against OpenAI did not bring a bigger surprise.