From self-driving cars shuttling through bustling urban streets to smart home systems making people’s lives more convenient and comfortable; from precise diagnosis and personalized treatment in the medical field to automation and intelligence in industrial production, AI’s footprints have spread all over the world every corner of society.
If an application is not integrated with AI, it seems to indicate that it has become "outdated". However, the combination of AI and applications does not happen overnight, and there is still a high "threshold".
Amazon Cloud Technology believes that "no one model can be applied to all business scenarios", and users need to select and select based on their needs during use. Based on this, Amazon Bedrock provides a series of leading basic models for customers to choose from.
Anthropic builds smarter Claude 3 model
Anthropic is an artificial intelligence security and research company dedicated to building reliable, explainable and controllable artificial intelligence systems. Anthropic recently launched a new generation of Claude 3 series models. Among them, the smartest Claude 3 Opus outperformed OpenAI's GPT-4 and Google's Gemini Ultra in benchmark tests, and has strong application potential.
In the field of generative artificial intelligence, Anthropic is building models through Amazon Trainium and Amazon Inferentia chips, and has begun to provide global Amazon Cloud Technology customers with long-term access to its future base models on Amazon Bedrock. At the same time, as Amazon completes its $4 billion investment in Anthropic, the cooperation between the two parties will further deepen.
Chen Xiaojian, General Manager of Amazon Cloud Technology Greater China Product Department
Recently, Amazon Cloud Technology held a generative AI media communication meeting in Beijing, focusing on the Amazon Bedrock + Anthropic Claude 3 model's ability to empower enterprises to build generative AI applications.
Currently, there are three models in the Claude 3 series, including the Claude 3 Haiku, with instant responsiveness; the Claude 3 Sonnet, an ideal balance between skill and speed; and the Claude 3 Opus, the most intelligent model designed to handle highly complex tasks . The first two models are already officially available on Amazon Bedrock, and Opus will be available soon. Users can choose the combination that best suits their application scenarios based on their business needs.
“Anthropic’s global popularity did not happen overnight. Claude 3 surpassed all existing models in standard assessments such as math questions, programming exercises and scientific reasoning. Users can use AI-driven responses to automate tasks and ensure high accuracy, such as in Optimizing experimental procedures in specific fields in the manufacturing industry, or auditing financial reports based on contextual data, etc. At the same time, a comparison can be seen that the prices of Sonnet and Haiku are much cheaper than equivalent intelligent models in the industry. This is also what attracts customers. One of the important reasons." Speaking of Claude 3's capabilities, Chen Xiaojian, general manager of Amazon Cloud Technology Greater China Product Department, fully affirmed it.
It is understood that Claude 3 currently has multi-modal capabilities. It can receive image-based input, its capabilities are roughly the same as other cutting-edge models, and its latency is lower than other multi-modal models.
In the face of complex enterprise-level applications, Claude 3 has undergone targeted training. It is very good at understanding pictures, charts, graphs, technical illustrations and optical character recognition (OCR), so it needs to process a large number of images, charts, reports and other visual assets. Excellent performance in enterprise use cases. In terms of speed, according to relevant evaluation results, the Claude 3 model is equivalent to the leading model in terms of image input capabilities, and the speed of Claude 3 Haiku is better than all leading models with equivalent capabilities.
Claude 3: Efficient and fast large models
So, is Claude 3 really that powerful?
To this end, Amazon Cloud Technology also conducted a series of field tests at the event and responded through the real performance of the Claude3 model in Amazon Bedrock.
The first is a test called "Needle In A Haystack (NIAH)", which is to insert a sentence from "The Three-Body Problem" into a complete text of "The Wandering Earth" to see whether the large model can be recognized in a short time. . It mainly examines the accurate recall ability of Claude 3 in 200K level ultra-long context.
Although this work seems simple, in order to actually do it, it will involve many complex logical functions. In an instant, the test results came out, and Claude 3 successfully recognized the inserted sentence "Don't answer, don't answer, don't answer!". It can even identify limitations of the test itself, such as discovering that a certain "target" sentence was obviously artificially added to the original text later.
Subsequently, Amazon Cloud Technology also demonstrated tests such as imitating the style of the novelist Gu Long, continuing to write novels, and code generation, analysis, and optimization. Claude 3 performed significantly better than similar models, and its practical value is very high.
Based on the above capabilities, Claude 3 can accurately identify e-commerce products through pictures, accurately capture product details based on the product model display, and generate more accurate product descriptions. It has been widely used in Amazon’s e-commerce business.
Build more powerful apps with Amazon Bedrock
With the powerful Claude 3, how can we use it conveniently?
In order to simplify the tedious work, Amazon Cloud Technology provides the Knowledge Bases knowledge base function in Amazon Bedrock, which will become an enterprise-specific knowledge base. Although large models are powerful, they need to be combined with internal business knowledge to be integrated into the business. Knowledge Bases is just such a tool, providing comprehensive hosting support, simplifying the process for users to combine private data with large models, and easily building internal knowledge bases. With the Amazon Bedrock knowledge base, retrieval-augmented generation (RAG) is achieved by combining contextual information and provides more accurate and personalized responses. All retrieved information is accompanied by citations to ensure transparency and reduce misunderstandings.
Just like people use multi-step execution to break down complex tasks, the Agents function launched by Amazon Bedrock is also very good at handling multi-step complex tasks. This feature enables customers to use natural language to perform multi-step business tasks. The agent function uses the reasoning capabilities of the underlying model to decompose the problem and solve the user's problem in a step-by-step manner. Agent will use the reasoning function to decompose this requirement and execute it step by step. Agents can access the organization's enterprise systems, processes, knowledge bases and some basic building blocks, and then formulate logical steps to solve problems, determine which APIs to call and when to call them, and ensure that transactions are completed in the correct order.
At the same time, when users use large models, they must also ensure the security of the application. To this end, Amazon Cloud Technology proposed Responsible AI and provided the Guardrails function on Amazon Bedrock, which can provide protection for a variety of basic models and agents. Content filtering policies can be added to protect sensitive information and user privacy.
In terms of compliance, Amazon Bedrock provides comprehensive monitoring and logging capabilities to support governance and auditing needs.
"Getting through the last three kilometers of generative AI" is Chen Xiaojian's summary of how to implement generative AI applications. He said that Amazon Cloud Technology has comprehensive technical support resources, including architects, product experts, AI laboratories, data laboratories, rapid prototyping teams and professional service teams to assist users in completing the final challenges of generative AI engineering.
The iPhone era of generative AI has arrived
Generative AI is developing rapidly. What is the next most critical step?
Chen Xiaojian said that in the face of increasingly complex user needs, we actually still have a lot of work to complete, including the basic layer, model and integration with business.
At a fundamental level, the status quo of chips still faces challenges. Despite the rapid technological advancement of semiconductor chips, the rapid expansion of model parameter sizes has far exceeded the processing capabilities of existing chips. From millions of parameters to tens or tens of billions of parameters, the complexity of models is growing at an unprecedented rate. As a basic service provider, Amazon Cloud Technology needs to constantly explore how to match the underlying hardware capabilities with the complexity of the business and the complexity of large models to ensure that the development of hardware can keep up with the expansion of software scale. The B200 launched by Nvidia is far from the pinnacle of hardware, and there is still a lot of work to be done in the future.
At the model level, the capabilities demonstrated by Claude 3 are indeed impressive. However, it will take a lot of effort to truly integrate this capability with the business. Current models may have reached the level of PhD students, but there is still a long way to go before breakthroughs can be achieved at the level of university professors, academicians, or even Einsteins. Therefore, generative AI suppliers, including Amazon Cloud Technology, must continue to invest in research and development to improve the capabilities of their models.
As for the top-level business integration, we can see attempts to combine Amazon Q with BI, Amazon Connect intelligent customer service and other solutions. The integration of generative AI with various industries and scenarios will be a huge project. It is also necessary to think about how to make large models better serve all walks of life, providing more powerful model capabilities, more convenient use, and lower costs.
The iPhone era of generative AI has arrived. Today’s demo gave us a glimpse into the amazing things that generative AI can accomplish. But to realize this great vision, not only Amazon Cloud Technology needs to work hard, but the entire industry needs to make huge efforts. Generative AI has the potential to bring huge value to human society, but this requires us to explore and practice together.
A programmer born in the 1990s developed a video porting software and made over 7 million in less than a year. The ending was very punishing! High school students create their own open source programming language as a coming-of-age ceremony - sharp comments from netizens: Relying on RustDesk due to rampant fraud, domestic service Taobao (taobao.com) suspended domestic services and restarted web version optimization work Java 17 is the most commonly used Java LTS version Windows 10 market share Reaching 70%, Windows 11 continues to decline Open Source Daily | Google supports Hongmeng to take over; open source Rabbit R1; Android phones supported by Docker; Microsoft's anxiety and ambition; Haier Electric shuts down the open platform Apple releases M4 chip Google deletes Android universal kernel (ACK ) Support for RISC-V architecture Yunfeng resigned from Alibaba and plans to produce independent games for Windows platforms in the future