Re:Inventing With Generative AI — AWS Has Rung In The Public Cloud Race 2.0
Generative AI is reshaping the public cloud landscape, marking the onset of a new race for dominance among Amazon, Microsoft, and Google. As Swami Sivasubramanian, VP of Data and AI at AWS, puts it, “Generative AI is at a tipping point, with the potential to redefine customer experiences across industries.” This statement underscores the need to understand the profound impact generative AI could have in the not so far future.
This GPT-4 Turbo-generated topic map illustrates generative AI’s diverse applications across industries, highlighting its transformative potential. The above topic map, with nodes sized by ‘betweenness’ (overall importance of the node) and colored by modularity (topic grouping), offers a glimpse into how generative AI could affect our personal and professional lives, driving transformative outcomes in various sectors. As an example, let’s take a look at medicine.
Personalized Medicine Enhancement through Generative AI
Drug Discovery and Development: Generative AI can analyze vast datasets of molecular structures and biological interactions, accelerating the discovery of new drugs and therapies. It can predict the efficacy and safety of compounds, reducing the time and cost of clinical trials.
Genomic Analysis for Personalized Treatment: AI can process and interpret complex genomic data, identifying genetic markers associated with diseases. This enables healthcare professionals to tailor treatments based on a patient’s genetic makeup, improving treatment outcomes.
Predictive Diagnostics: By analyzing medical records and diagnostic images, generative AI can identify patterns that might be missed by human eyes. This leads to earlier detection of conditions like cancer, allowing for timely intervention.
Customized Treatment Plans: AI algorithms can consider a patient’s medical history, lifestyle, and preferences to recommend personalized treatment and wellness plans. This approach ensures that each patient receives care that is best suited to their unique needs.
Virtual Health Assistants: Generative AI can power sophisticated virtual health assistants, providing patients with instant, personalized medical advice and improving ongoing patient engagement and monitoring.
In essence, generative AI in personalized medicine not only promises more effective treatments but also paves the way for a future where healthcare is more predictive, preventive, and personalized. This example underscores the transformative power of generative AI across various sectors, highlighting its potential to profoundly impact our lives and well-being.
Generative AI As The New Differentiator
If we expect generative AI-driven capabilities to become the key differentiator for product and service offerings in the future, we can assume that organizations will select the cloud platform that makes it easiest and most cost effective to add these capabilities to their current portfolio. Achieving this goal will not be easy for Amazon, as both of its competitors should be considered as heavyweights in the generative AI arena. Microsoft has invested approximately $13 Billion in OpenAI, the market leader in generative AI. The fact that the OpenAI platform runs entirely on Azure further demonstrates the close ties between both companies, even before we start speculating about Microsoft’s role in reinstalling OpenAI CEO Sam Altman after he was fired by the board. On the other side, Google got caught out cold by the launch of OpenAI’s ChatGPT platform, as the performance of Google’s own generative AI called Bard, can only be described as disappointing. While I cannot explain how this type of ‘sucker punch’ could have happened to Google, we also know that Google is behind some of the most successful analytics, machine learning and AI products and services such as TensorFlow, Google Cloud AI and Machine Learning Services, Waymo, and of course, Google Search.
Now let’s take a look at the generative AI related highlights Amazon presented at re:Invent 2023 in Las Vegas.
The Bedrock of Amazon’s Generative AI Strategy
Amazon Bedrock offers six managed foundation models (Claude, Llama, Jurassic, Stable Diffusion, Command, and Amazon’s own Titan) via a unified API for customers to base their generative AI apps on. However, GPT-4, today’s most prominent large language model is not available via Bedrock. Developers who want GPT-4 can consume it on Microsoft’s cloud in the form of Azure OpenAI Service. While Amazon mentioned that Bedrock already has 10.000 enterprise users so far, there is no information available regarding how deep these deployments go. Therefore, it is fair to assume that Bedrock adoption pales in comparison to the currently 100 million GPT users.
Fine Tuning Bedrock For Custom Capabilities
There are two types of fine tuning offered by Amazon for LLMs (Anthropic is not yet supported):
1. Continuous Model Pre-Training Through Unlabeled Data
Amazon Bedrock’s approach to customization allows users to tailor large language models to their specific needs and objectives. This flexibility is particularly beneficial for industries with unique jargon, specialized knowledge, or specific communication styles. The stream of pre-training data can include a variety of sources, such as internal company documents, industry-specific literature, customer interaction logs, or any text corpus that reflects the language and context relevant to the user’s business. By continuously feeding these tailored datasets into the model, Amazon Bedrock ensures that the language models evolve and adapt, becoming more aligned with the user’s operational context and providing more accurate and relevant responses. This method of customization not only enhances the model’s effectiveness in specific scenarios but also contributes to a more personalized and efficient AI-driven experience.
2. Task-Specific Training Through Labeled Data
In addition to broad customization through pre-training data, Amazon Bedrock offers a specialized approach by leveraging task-specific labeled data. This targeted strategy is instrumental for applications requiring high precision, such as sentiment analysis, medical diagnosis, legal document analysis, or customer service interactions. Task-specific labeled data consists of carefully annotated datasets where each entry is tagged with specific information relevant to the task. For instance, in sentiment analysis, texts are labeled with their corresponding emotional tone, like positive, negative, or neutral. In medical applications, patient records could be annotated with diagnosis codes. This data, when fed into the language models, refines their understanding and response accuracy in these niche areas. The process of training with task-specific labeled data not only sharpens the model’s capabilities in those domains but also significantly reduces the likelihood of errors in output, leading to more trustworthy and reliable AI solutions. The use of such finely-tuned data sets within Amazon Bedrock underscores the platform’s commitment to providing highly specialized and effective AI tools tailored to specific industry needs and challenges. But how does AWS address privacy concerns resulting from users having to upload training materials that could be confidential?
Federated Learning for Customer Privacy
Federated learning is a machine learning technique that trains an algorithm across multiple decentralized devices or servers holding local data samples, without exchanging them. This approach is beneficial for privacy-preserving and efficient computation. In federated learning, the model is sent to the device where the data resides. The model is trained locally, and only the model updates (not the data) are sent back to a central server. The server then aggregates these updates from multiple devices to improve the model. Federated learning is supported by Bedrock, but is also offered by Azure (see article) and Google Cloud (see article).
Building Generative AI Apps With Bedrock, Lambda and Step Functions
The integration of AWS Step Functions, Lambda and Bedrock streamlines the process of building and scaling generative AI applications. An essential aspect of this integration is the ability to create complex generative AI applications using prompt chaining. This technique involves passing multiple smaller and simpler prompts to the foundation model rather than a single, detailed prompt. By creating a state machine that calls Amazon Bedrock multiple times for each smaller prompt, tasks can be run in parallel, and their responses unified into a single result using an AWS Lambda function. While this is critical functionality, you can achieve very similar results with Azure Logic Apps and Google Cloud Composer.
Fully Managed Retrieval Augmented Generation (RAG)
Amazon Bedrock Knowledge Base offers a fully managed Retrieval Augmented Generation (RAG) experience, enabling users to securely connect foundation models to their company data for enhanced, context-specific responses. This feature manages the entire RAG workflow, including vector store setup, data embedding and querying, and provision of source attribution and short-term memory for production applications. Knowledge Bases handles the creation, storage, and management of vector embeddings, which represent the textual data semantically. Two new APIs, `RetrieveAndGenerate` and `Retrieve`, facilitate the process by handling embedding and querying tasks, and providing options for further processing and developing custom text generation workflows. Users can define custom chunking strategies for effective data retrieval and select from available vector database options like Amazon OpenSearch Serverless, Pinecone, and Redis Enterprise Cloud for storing vector data. Additionally, Knowledge Bases can be integrated with other generative AI tools and applications, such as Agents for Amazon Bedrock or the LangChain retrieval plugin, to build sophisticated AI assistants and other generative AI applications.
Responsible AI: Guardrails for Bedrock
Guardrails for Amazon Bedrock is a tool designed to enhance the safety and responsibility of generative AI applications by allowing users to implement customized safeguards. It includes features like Denied Topics for specifying undesirable content, Content Filters for managing harmful content, and an upcoming PII (Personal Identifiable Information) Redaction feature for protecting personal information. These guardrails provide additional control over foundation models, ensuring AI interactions align with company policies and responsible AI practices. Integrated with Amazon CloudWatch for monitoring, these guardrails ensure that user interactions with AI are safe, relevant, and compliant with privacy and regulatory standards.
Agents for Bedrock
Agents for Amazon Bedrock, now generally available, accelerates generative AI application development by orchestrating multistep tasks using foundation models. It automates the prompt engineering process, handling tasks like order management or claim processing, and integrates company-specific information for natural language responses. Enhanced capabilities include improved control over task orchestration and better visibility into the reasoning process, allowing developers to view and refine the steps taken by agents. The platform also offers customizable prompt templates for preprocessing, orchestration, and postprocessing, enabling precise control over user interactions. This tool is optimized for specific tasks, with focused instructions and actions, ensuring efficient and accurate performance by the agents.
Final Thoughts
Did Amazon deliver enough new products and a clear enough vision to successfully defend its leadership position in the public cloud market? Having closely observed re:Invent 2023 in Las Vegas I can only conclude that AWS is all in on generative AI. Cynics may say that there is some “gen AI washing” going on, but I do not think that this is a fair statement, as infusing this new technology into the massive AWS portfolio obviously needs some realignment on the marketing side too.
I strongly believe that OpenAI has a significant advantage in terms of the quality of its generative AI services. The 100 Million users making daily queries are a giant advantage for OpenAI when it comes to rapidly improving model performance (accuracy and speed) through reinforcement learning. Of course we all thought that Google would benefit too much from its users’ daily web searches for anyone to overtake their search performance, but that was before OpenAI came and annihilated this seemingly insurmountable advantage. OpenAI’s strength boats well for Microsoft’s Azure platform.
Counting out Google would also not be a wise move, as the company is now in “recovery mode” from the ChatGPT-shock event and determined to get back to the top of search and AI.
All in all, it will be exciting to watch this race unfold over the next few quarters.
