
Автор: Ayal Steinberg, Carlo Appugliese, Paul Hake, Simon Bisson
Издательство: O’Reilly Media, Inc.
Год: 2025-04-29
Язык: английский
Формат: pdf, epub
Размер: 10.1 MB
Technologists and business leaders alike will explore the fundamental capabilities of Generative AI, high-impact use cases, key risks and challenges, technical integration considerations, and how to start building the necessary teams and culture for AI-enabled innovation. Whether you're just starting to explore this transformative technology or looking to scale its impact, you'll learn concrete strategies for harnessing the power of Generative AI to create measurable value for your business.
Most applications of Generative AI, like ChatGPT, consist of more than just one LLM. An LLM can generate text, but if you ask it about images or audio the LLM doesn’t have the appropriate capabilities, so it needs to use tools just like we do when we do things. Tools allow the LLM to handle multiple input and output types, making it multimodal. If you ask ChatGPT to create an image, it will use an LLM to parse and summarize your input, building a prompt for another model, DALL-E, which handles the actual image generation. This is all done via function calling so that ChatGPT can create a seamless end user experience. Most LLMs like GPT-4 have been fine-tuned to detect when a function needs to be called and output JSON containing the necessary arguments to call the function. The functions that are being called via function calling will act as a tool for your LLM.
The AI tools space is constantly evolving, making it impossible to provide a comprehensive and up-to-date taxonomy. The vast majority of popular tools are open source or based on open source. Even proprietary tools often leverage open source under the hood. Open source is by far the most important vehicle for both usage and innovation, and with the current pace of development it’s hard to justify not using the important tools listed next. Here is a way to group the tools and some examples of each category, along with a brief description:
AI base foundational frameworks:
Used to build models from the bottom up; for example, create a transformer model from scratch using neural network building blocks. These are usually implemented as Python libraries: PyTorch, TensorFlow, Keras, and JAX.
AI development and orchestration:
Used to speed development of AI systems to avoid having to code everything in a base language like Python, with modules and functions to streamline connecting prompts together to orchestrate AI applications like RAG or chatbots. Examples include LangChain, LlamaIndex, LangSmith, and LangGraph.
Cloud development and deployment platforms:
Platforms supporting development and deployment of AI models; they host and run the other tools listed in this section as a service. Examples include Azure AI Foundry, AWS Bedrock/SageMaker, Google Vertex AI, and LangSmith.
Development tools:
Provide orchestration, pipelines, and UIs. Examples include Streamlit, Gradio, LangSmith, LangChain Agents, and CrewAI. A new generation of tooling supports no- and low-code AI: AutoGen and Copilot Studio.
Vector databases:
Store and manage vectorized data for retrieval such as with RAG systems. Examples include Milvus, Pinecone, and Chroma. Vector indexing is now a common feature in most databases, allowing you to quickly extend existing stores.
Contents:
Скачать Generative AI for Business
