"AI or die"
For at least a brief period of time, the cloud computing industry really started to worry that Amazon Web Services would fall behind.
At that time, Microsoft bet on OpenAI, the scenery was unparalleled, behind the exciting story was a big gamble, even Bill Gates warned Satya Nadella not to invest in OpenAI, and then the phenomenal popularity of ChatGPT made all doubts disappear, and the combination of OpenAI GPT and Microsoft Azure seemed to win.
Image source: Unsplash
Before the "little-known" OpenAI exploded, Google has always been a benchmark in the field of artificial intelligence, and OpenAI has only achieved something by standing on Google's shoulders, such as the Transformer architecture was first launched by Google, and the emergence phenomenon was discovered by Google researchers. If AI is the biggest variable in the future of cloud services, then Google Cloud must be the strongest contender, ahead of Microsoft Azure.
I have to take my hat off Microsoft Azure and Google Cloud, which have accelerated the pace of AI to change everything, and have also shifted the focus of cloud services to AI. This is an "AI or die" race, where cloud giants must not only have AI capabilities, but also have more prominent differentiation advantages than their direct competitors.
When the outside world is generally worried about Amazon Web Services, Adam Selipsky obviously does not think so, Amazon Web Services' huge infrastructure and thriving partner ecosystem around the world are its long-standing leading barriers, and the outside world is looking forward to the new hole card of Amazon Web Services.
Those who believe that Microsoft Azure and Google Cloud can surpass Amazon Web Services with large models, or close the gap with them, also have the expectation that once Amazon Web Services incorporates AI into the robust service ecosystem it provides to customers, any lost ground will be quickly regained.
This was the background when Amazon Bedrock was born in April 2023. Compared with OpenAI and Google PaLM, which are mainly promoted by Microsoft and Google, Amazon Web Services has taken a different strategy.
At first, Amazon Bedrock focused on multi-model access, and Amazon Web Services' basic models (Amazon Titan) and various third-party models (such as AI21 Labs, Anthropic, Stability AI, etc.) can be accessed through APIs. Also released are code generators and cloud computing power for training and inference.
Since then, Amazon Bedrock has been updated with new base models, base model vendors, and agents capabilities that provide a variety of rich features such as fine-tuning, knowledge base, agents, model evaluation, and seamless integration with other Amazon Web Services workloads.
Image credit: pixabay
To sum up, there are three major aspects: simplifying the selection: diversified selection of industry-based models, simplified customization: continuous pre-training, fine-tuning, RAG, etc., and simplified integration: Agent on Amazon Bedrock.
最新一次的重磅更新，亚马逊云科技数据和人工智能副总裁Swami Sivasubramanian形容其为“Significant new capabilities make it easier to use Amazon Bedrock to build and scale generative AI applications – and achieve impressive results”。
Major updates, easier to use, and remarkable results have become the keywords of this update of Amazon Bedrock.
Looking down, the large model is still a cloud platform
It has been a year since Amazon Bedrock was released to the public, and in the past year, the large model has been constantly iterating, and even iterating too fast, so that enterprise customers do not know how to make the application, and the answer given by Amazon Bedrock is that the cloud vendor is not only responsible for the large model, but is responsible for the final business effect.
According to Swami Sivasubramanian, Amazon Bedrock focuses on the key areas that customers need to build production-ready enterprise-grade generative AI applications at the right cost and speed.
Through a series of updates, Amazon Bedrock has created an end-to-end platform that is responsible for the customer's business performance, and the industry has once again focused on Amazon Web Services.
First of all, the Llama 3 8B and Llama 3 70B versions of Llama 3 have been added, and the strongest open source model is officially GA. Llama 3 8B excels at text summarization, classification, sentiment analysis, and translation for resource-constrained and edge device scenarios. The Llama 3 70B excels in content creation, conversational AI, language understanding, R&D, enterprise, accurate summarization, granular classification/sentiment analysis, language modeling, dialogue systems, code generation, and instruction tracing.
Amazon Bedrock also teased Amazon Titan Text Embeddings V2, which offers 256, 512, and 1024 vector space sizes, prioritizing cost reduction while retaining 97% accuracy for RAG use cases and outperforming other leading models.
and soon support for Cohere's Command R and Command R+ Enterprise FM. Command R+, Cohere's most powerful model, is optimized for long-context tasks, while Command R is optimized for large-scale production workloads.
It is worth noting that Amazon Bedrock's model evaluation tool is fully available, which can evaluate models based on specific use case metrics (such as relevance, style, and brand voice), and evaluate, compare, and select the best model for their application, reducing the time to evaluate the model from weeks to hours, and providing an evaluation report for more model users.
Titanium Media App observed that at this stage, customers do not want to be bound by a model, the basic model is evolving, and the application scenarios are also being debugged, and customers want to have more choices, lower costs, and easier to debug different models to test their business effects, and the model evaluation tool is necessary and appropriate.
The very useful Amazon Titan Image Generator is now generally available, allowing customers to generate images efficiently and cost-effectively using natural language, with invisible watermarks on each image, helping to reduce intellectual property risk.
Second, how to make it more secure for enterprises to customize models and integrate them into specific business use cases, Amazon Bedrock implements custom model import, where customers can now import and access popular open model architectures, including Flan-T5, Llama, and Mistral, to build custom models as fully managed application programming interfaces (APIs) in Amazon Bedrock.
This allows customers to take models that have been customized on Amazon SageMaker or other tools, easily add them to Amazon Bedrock, and have seamless access to their custom models after automatic validation, just like the models provided by Amazon Bedrock, which is a complete reflection of the openness of Amazon Bedrock.
Finally, security, Amazon Bedrock's Guardrails is now generally available to help customers prevent harmful content and manage sensitive information within their applications. Customers are able to define content policies, set application behavior boundaries, and implement protections against potential risks. It is understood that it can help customers block up to 85% of harmful content compared to the protection provided by FM itself on Amazon Bedrock.
GenAI continues the first principles of Amazon Web Services
I have to say that it is the turn of the industry to continue to pay tribute to Amazon Web Services, and Big Brother is still Big Brother.
Amazon Bedrock turned the tide in just one year, and the industry is still focused on the technological progress of the big model itself, but more concerned about whether the big model can be used by me, which involves huge infrastructure, engineering capabilities, ecosystems, etc.
Rather than choosing to just make big models, Amazon Web Services is incorporating AI into the existing ecosystem in a different way, and that's what they do best, and tens of thousands of customers are using Amazon Bedrock to build and scale large model applications.
Amazon Web Services has built a complete generative AI full-stack layout, and enterprises build a three-tier architecture for generative AI applications, including: infrastructure for basic model training and inference, tools for building with basic models, and out-of-the-box applications built with basic models.
At the underlying computing layer, on the one hand, Amazon Web Services continues to provide computing instances from NVIDIA, including Amazon EC2 instances of the latest NVIDIA Grace Blackwell GPUs, and on the other hand, Amazon Web Services' self-developed chips, including inference chips Inferentia and Trainium series.
The middle tier is Amazon Bedrock along with various additional features. The upper layer is applications including Amazon Q, Amazon Q can be connected to the enterprise's own data, information and systems, and can be customized according to the customer's business, Amazon Q, marketers, project managers and sales representatives in the enterprise can use Q to customize conversations, solve problems, generate content, take action, etc., it is reported that Amazon Web Services is about to bring further updates.
If the large model has changed the rules of the game in the cloud computing industry, then people in the industry will find that the rules of the game formulated by Amazon Web Services are still playing a role at a broader level.
OpenAI GPT and Google PaLM are compelling enough, but they also create new competitive issues for Microsoft Azure and Google Cloud, and there are many model companies on the market, and Amazon Bedrock has established a common starting point for these model companies, so that customers can experience different models equally.
In the view of Titanium Media App, whether Amazon Web Services is doing cloud or generative AI, its basic concept and logic are the same, so Amazon Web Services does not have to sway to think about how to do and what to do, but directly follow the first principles to make the most suitable Amazon Bedrock for Amazon Web Services.
In the early years, Andy Jassy's concept of cloud computing was to "subdivide IT infrastructure into the smallest units, so that programmers can choose and combine them with maximum freedom", and Amazon Web Services did exactly that, and it is still the same when it comes to large models.
Flipping through Swami Sivasubramanian's comments on the launch of Amazon Bedrock, "We live in a very exciting time for machine learning. I might say that every year, but this year it's even more special because these large language models and foundational models can really support so many use cases where people don't have to form separate teams to build task-specific models. The speed of machine learning model development will really increase. ”
"But in the next few years, unless we make these models more accessible to everyone, you're not going to get to the end state you want. That's what we did in the early days of machine learning in SageMaker, and that's what we need to do in Bedrock and all of its applications. He said.
Artificial intelligence is a more complex technology stack than IT, and Amazon Web Services continues to deliver on its promise to give any developer in the enterprise the freedom to build generative AI applications without having to focus on complex machine learning or underlying infrastructure, making AI a critical leap to the promised land of cloud computing.
(This article was first published in Titanium Media App, by |.) Zhang Shuai, Editor | Liu Xiangming)