Enterprise Generative AI Unfolded

Subscribe • Previous Issues

Ten Keys to Accelerating Enterprise Adoption of LLMs

By reviewing job postings in the US and analyzing recent reports on enterprise Large Language Models (LLMs) and Generative AI (GAI), I sought to understand the most critical enterprise requirements for these technologies. From this analysis emerged the following pillars, each representing a set of key capabilities and features crucial to accelerating the adoption of Generative AI and Large Language Models in enterprise settings.

1. Mastering Model Development & Optimization: Developing robust Generative AI and Large Language Models involves a series of complex tasks like preprocessing and cleaning training data, model training, performance evaluation, and refinement.  The ability to scale to large datasets and models, and applying model compression strategies are vital for building efficient and powerful models.  This is because GAI and LLMs require large models and a massive amount of data to train. Model compression strategies can also be used to reduce the size of models, which can make them more efficient and easier to deploy.

2. Emphasizing Customizability & Fine-tuning:  Tools that allow businesses to adapt pre-existing LLMs to their specific needs, which is crucial for successful adoption. Techniques like fine-tuning and in-context learning help tailor LLMs to better serve unique business use cases. A clear example of this would be adjusting a speech synthesis model to generate speech mimicking a specific person’s voice, providing a unique customer service experience. A new wave of startups like Lamini  enable developers to to train powerful custom LLMs on their own data and infrastructure.

3. Investing in Operational Tooling & Infrastructure: In a previous post, I discussed recent initiatives and tools to help companies serve and deploy LLMs efficiently and cost-effectively. Successful model usage requires a robust infrastructure that includes model hosting, orchestration, caching, and AI agent frameworks. These tools ensure efficient operation and maintenance of the AI system while enabling real-time learning post-deployment. Operational tooling includes logging, tracking, and evaluating LLM outputs, ensuring quality control and insights for ongoing improvement.

4. Prioritizing Security and Compliance: Protection of LLMs from adversarial attacks, compliance with legal requirements, and ethical usage are paramount for enterprise adoption. Features facilitating content moderation and legal compliance not only safeguard the business from legal repercussions but also protect its reputation.

Key features to help accelerate enterprise adoption of LLMs and generative AI.

5. Ensuring Seamless Integration & Real-world Scalability: APIs and plugins provide interfaces that allow LLMs to interact with other software components (Vector DBs, knowledge graphs, etc.), ensuring seamless integration into existing workflows. A typical example would be using an API for a custom LLM to integrate it into a customer service chatbot, enabling the enterprise to scale its customer service operations. Model deployment and scaling solutions are essential for handling real-world use cases and loads.

6. Promoting Collaborative Development: LLMs have expanded the number of personas involved in AI initiatives. Tools that facilitate cross-functional collaboration can help align technical decisions with business strategies, driving AI initiatives. This ability to work alongside product managers and cross-functional teams becomes crucial in a dynamic business environment.

7. Encouraging Experimentation & Innovation: Features like playgrounds for developers to experiment with LLMs and open-source models allow for widespread experimentation, fostering innovation. By leveraging freely available models, developers can experiment with different use-cases, feeding learnings back into their applications.

8. Maintaining Ongoing Evaluation and Enhancement: The effectiveness of LLMs can be maintained by regular monitoring and using tools that streamline the process of updating them.  By keeping the AI model up-to-date with fresh data, and advancements in AI technology, enterprises can ensure their models remain relevant and perform optimally. Regular updates could involve adding new features or capabilities, improving efficiency, or patching security vulnerabilities.

9. Adopting Data Tools for Generative AI:  Last year, there was a surge in articles and interest around “data-centric AI,” a concept that puts the focus on improving data quality over refining algorithms. While LLMs have taken center stage, data-centric AI tools remain essential for ensuring the quality and accuracy of these models. The quality, diversity, and relevancy of data used in training and fine-tuning foundation models greatly determine their performance and utility. On the other hand, the majority of existing data engineering tools have been designed with structured or semi-structured data as their primary focus. To fully leverage the potential of GAI and LLMs, enterprises will need tools specifically targeting this kind of data. For example, Unstructured is building data processing and data ingestion tools specifically for LLMs.

10. Enhancing User Experience & Accessibility: This pillar is built on a user-friendly interface and API, detailed documentation, and the transparency and explainability of AI systems. These elements foster user trust and facilitate user adoption. Transparency is also important for compliance reasons as companies need to know how the AI is making decisions, especially in sensitive sectors like healthcare, finance, or legal.

As enterprises consider the adoption of Generative AI and Large Language Models, these pillars provide a framework for driving successful implementation and utilization. Each pillar represents a set of features and capabilities that collectively enable enterprises to tap into AI’s enormous potential.  A good place to start is with the open source project Ray. Its vast array of accompanying libraries, which encompass data processing, model training and tuning, and model serving, have established themselves as the cornerstone of platforms being built by leading AI and LLM teams.

Data Exchange Podcast

1. The Future of Graph Databases.  With LLMs on the rise and knowledge graphs on the radar of many teams, I caught up with Emil Efrem, co-founder and CEO of Neo4j, the most popular graph database in the market.

2. Delivering Safe and Effective LLM and NLP Applications. David Talby (CTO of John Snow Labs), walks me through LangTest, a new open-source Python library designed to help developers deliver safe and effective language models.

Speaking of enterprise adoption, here’s an overview of our database of the Top AI startups. Many trailblazers are honing their AI prowess with a distinct focus on specialized application areas.

Exploring the Boundaries of LLMs: Learning from Versatile Developer Tools

To facilitate enterprise adoption of LLMs, we need to make it easy for teams to experiment and build simple apps that use them. A new crop of developer tools imbued with a suite of features meticulously designed to simplify the nuances of LLMs, are fast becoming indispensable for developers. They offer the flexibility to switch between different LLMs with minimal code changes, providing much-needed agility in development. They ensure compatibility with a wide array of programming languages and allow seamless integration with other tools and data sources, broadening the scope of what can be achieved with LLMs. 

In addition to providing functional packages (or “skills”) to streamline common tasks, these tools incorporate capabilities such as call chaining for context maintenance, integration with external services and data, orchestration for decision-making, and prompt abstraction for intuitive interactions. They harness the power of semantic memory management to maintain coherence in generated text and can process multimodal data, including text, images, and audio, for richer applications. 

A prototypical example is LangChain – an open source project that facilitates rapid exploration and prototyping, functioning as an ideal platform for ideation and quick testing of concepts. Its versatility enables developers to switch between different LLMs, hosted either on-premise or through external cloud services, with minimal code changes. Furthermore, it integrates seamlessly with numerous existing tools and data sources, amplifying its usefulness and application breadth. LangChain’s undisputed value in prototyping and exploring the possibilities with LLMs underscores its popularity.

As best I can tell LangChain is used mainly for exploration and its use in production environments is limited. More generally, we’re still in the early phases of building out AI developer tools, you should be exploring and learning from a variety of related open source projects. Experimenting with a diverse set of frameworks like LangChain, LlamaIndex, Griptape, LLM, Aim, and others, enables you to learn the strengths and weaknesses of different approaches and architectures. This understanding serves as a valuable guide in developing robust tools that can help in safely and efficiently productionizing your LLM-enabled applications.

Which brings me to Semantic Kernel (SK), an open-source software development kit designed for developers looking to integrate AI & LLM services into their applications using programming languages like Python and C#. What I’ve found particularly enjoyable about SK in my LLM experiments is its simple programming model, which facilitates the seamless combination of AI models and plugins to craft new user experiences. Furthermore, its connectors allow easy addition of memories and models, and its ability to integrate AI plugins extends application capabilities.

Semantic Kernel enables function composition by letting users combine native and semantic functions into a single pipeline.

While the documentation is not yet as extensive compared to some other projects, Microsoft is actively working to enhance it. A standout feature of SK is its proven application in real-world production environments, with deployments not only within Microsoft but also among large-scale Microsoft Azure clients. In fact, SK has already been successfully used for tasks such as enabling chats over proprietary data and building AI agents to accomplish specific tasks. Another project I’ve taken a liking to is Griptape, which targets enterprise grade applications. I’m enjoying using Semantic Kernel and Griptape, and I highly recommend giving them a try.

Join us at the AI Conference in San Francisco (Sep 26-27) to network with the pioneers of LLM development and learn about the latest advances in Generative AI and ML.

If you enjoyed this post please support our work by encouraging your friends and colleagues to subscribe to our newsletter:

%d bloggers like this: