<img src="https://imageproxy.pbkrs.com/https://img6.donews.com/img/2024/08/08/img_pic_281723109450.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://img6.donews.com/img/2024/08/08/img_pic_281723109450.png"/>
Editor | Yang Bocheng
Caption | IC Photo
At the 2024 I/O Connect China Google Developer Conference held by Google recently, AI large-scale models diversification became the focus of market attention.
Around APP software development, Google launched three different specifications of Gemini models. Google stated that Gemini Nano is the most efficient model, suitable for handling tasks on devices.
It is reported that Gemini 1.5 Flash is Google's fastest and most economical model to date, suitable for handling high-capacity tasks. Gemini 1.5 Pro, open to all developers, supports a context window of 2 million tokens. To reduce computing power consumption, both Gemini 1.5 Pro and 1.5 Flash have launched context caching functionality.
Considering that developers may need greater flexibility and control, Google introduced the Gemini sister large model Gemma. The newly released Gemma2 adds two specifications of 9 billion and 27 billion parameters compared to Gemma. Among them, the 27 billion parameter version has been optimized to support a single NVIDIA GPU on Google Cloud and a single TPU on Vertex AI.
Currently, the related large models of Gemini have been integrated into development tools such as Android Studio, Chrome DevTools, Project IDX, Colab, VS Code, IntelliJ, and Firebase, helping developers write, debug, test code, generate documents, and understand code libraries.
Taking Flutter as an example, the Xiaomi SU7 supporting application is built based on Flutter. In addition to the existing Flutter foundation, Google has released Flutter 3.24 and Dart 3.5. The biggest highlight of the new version is the early preview of the &#34;Flutter GPU&#34; new API. Through the built-in Flutter SDK, developers can use Dart code to access the GPU, thereby improving image rendering capabilities.
To facilitate developers' use, Google has launched multiple Packages. For example, Flutter_Scene can directly import 3D projects to enhance the gaming experience.
Furthermore, Google also released an early preview version of Android Studio ON IDX, which is different from the original Android Studio in that it supports running entirely in the browser. To ensure the reliability, compliance, and security of building applications with AI, Google has introduced development components such as Firebase AI Monitoring Information Center and Checks AI Safety
In the wave of open-source AI large models globally, Google has launched the open-source large model project Project Oscar. However, in the initial stage, Project Oscar only supports 93,000 code submissions and 2,000 developers for the Go programming language project.
Focused on web development, with Web GPU, WASM, and Gemini integrated into Chrome, Google has introduced the new Speculation Rules API, which enables instant navigation through searches, eliminating lengthy page loading times. The View Transitions API, tailored for single-page applications, enhances the page transition experience. When combined, they ensure seamless page transitions.
To ensure the efficiency of web developers, Google has timely introduced and optimized the Chrome DevTools, which issue warnings and error prompts when problems occur on developer websites. This application has been integrated into Gemini.
In the realm of next-generation native Android app development, Google has launched multiple new products. These include the device-side AI model Gemini Nano and system service AI Core; Kotlin Multi Platform for sharing business logic code across mobile, web, server, and desktop platforms; and Kotlin Multi platform support for multiple Jetpack libraries such as DataStore, Room, and ViewModel.
The Android Device Streaming testing platform, in collaboration with smartphone manufacturers like Xiaomi, OPPO, OnePlus, and Samsung, facilitates terminal testing for developers and is currently in the Beta stage. Gemini in Android Studio has been included in the stable version of Android Studio, adding code generation and conversion features, as well as AI privacy settings for controlling data sharing.
In the cloud business domain, Google's new cloud journey features five main characteristics: first, a new paradigm of cloud development, with the newly launched Vertex AI functionality enabling context caching and grounding. Second, a self-contained flexible expansion, with over 150 new models introduced, including the Gemini series, Gemma open-source models, Anthropic Claude models, Meta Llama models, and the Hugging Face model library.
Third, a cross-cloud journey breaking barriers, with the optimization of PostgreSQL databases and the BigQuery Omni feature supporting cross-cloud interconnection, federated queries, and multi-&#34;cloud&#34; collaboration. Fourth, easy access to powerful features, with the introduction of automation and intelligent default settings, allowing the setup of a complete cloud infrastructure in just 45 minutes, including networking, authentication, and logging.
Fifth, AI intelligence assistance, with the new Gemini Code AssistIDE plugin providing code generation, completion, interpretation, and testing capabilities
Gemini in Databases provides intelligent SQL generation and database operations.
Releasing multiple large models around developers undoubtedly demonstrates Google's determination to accelerate the commercialization of large models. However, Google's AI capabilities, especially in terms of output and retrieval, may still need improvement.
In June of this year, when overseas users asked Google &#34;how much glue to use when making pizza,&#34; the answer provided by Google AI search was: According to a related article from Business Insider in May 2024, Google AI search results suggest adding 1/8 cup, or 2 tablespoons, of white non-toxic glue to the pizza sauce to prevent the cheese from slipping.
The author of the article, KatieNotopoulos, stated that the glue did not significantly change the thickness of the sauce, and the pizza appeared with an appealing orange color. After verification by overseas media The Verge, it was found that the screenshot was not fabricated. As glue cannot be added to pizza, this answer also raised doubts among overseas users about Google AI's retrieval capabilities.
During our testing of Gemini, we found that on one hand, Gemini's ability to generate text and images has been disabled, and Gemini currently does not support generating videos. On the other hand, Gemini's logical reasoning and mathematical computation capabilities still need improvement.
When we provided Gemini with the 2024 math questions from the college entrance examination and specifically emphasized that they were multiple-choice questions, the correct answers to the three questions were (BD), (ACD), (ABC), but Gemini's answers were (AC), (AD), (AB) respectively. Although the answers for question 10 and question 11 by Gemini included the correct options, all options for question 9 were incorrect.
Source: Gemini Official Website
Source: 2024 College Entrance Examination Math Questions
Source: Gemini Official Website
In addition, as Google's large models continue to increase, it is facing a tricky problem of rapidly increasing carbon dioxide emissions from data center power consumption.
Google's related environmental report points out that by 2023, Google's data center power consumption alone will increase by 17%, with carbon dioxide pollution generated by power consumption increasing by 13% to 14.3 million tons compared to the same period in 2022, roughly equivalent to the annual carbon dioxide emissions that 38 gas-fired power plants may emit.
Not only Google, but Microsoft's greenhouse gas emissions for the 2023 fiscal year are about 30% higher than in 2020. Regarding how to reduce carbon emissions in the future, Google's related report indicates that as we further integrate artificial intelligence into our products, the increase in energy demand due to the increased intensity of artificial intelligence computation, and emissions related to the expected increase in our technical infrastructure investments, reducing emissions may be challenging.
At the current rate of development of AI large models, future electricity demand may double. As the United States, which has the most data centers globally, local residents are concerned about the sharp increase in electricity demand from artificial intelligence overwhelming the power grid, which may prolong the existence of coal and natural gas plants compared to other methods.
The existence of various issues such as commercialization, stable output quality, decarbonization, etc., poses challenges even for global giants like Google in the short term

GOOG

GOOGL

Google unveiled a variety of Gemini models at the 2024 I/O Connect China developer conference, including Gemini Nano, Gemini 1.5 Flash, and Gemini 1.5 Pro. Gemini 1.5 Pro and 1.5 Flash have introduced on-device caching to reduce computational power consumption. In addition, Google also introduced the sister model Gemma, with two new specifications of 9 billion and 27 billion parameters. The Gemini-related models have been integrated into multiple development tools to assist developers in writing, debugging, and testing code. Furthermore, Google also launched Flutter 3.24 and Dart 3.5, with the highlight being the early preview of the "Flutter GPU" new API, which enhances image rendering capabilities

- Google's 2024 I/O Connect China developer conference highlighted the market's focus on AI large models' diversification. Google introduced three Gemini models for app development, including the efficient Gemini Nano and the fast Gemini 1.5 Flash. Gemini 1.5 Pro, open to all developers, supports a 2 million token context window and features context caching to reduce computational power consumption.
- In the realm of web development, Google introduced the Speculation Rules API in Chrome for instant navigation, along with the View Transitions API for seamless page transitions. To enhance developer efficiency, Chrome DevTools now offer warnings and error prompts when issues arise, integrated into Gemini.
- Google's cloud initiatives emphasize AI integration, with features like Vertex AI and Gemini Code AssistIDE plugin for code generation. However, challenges like carbon emissions from data center power consumption pose hurdles for Google and other tech giants, raising concerns about future energy demands and environmental impact.

Open source, diversification, commercialization, Google is challenging large models