Rajul Rana serves as the Chief Technology Officer at Orion Innovation.
getty
Every decade or so, a technology comes along and disrupts everything around us. It changes the way we work, the way we play and the way we engage customers. In the ’80s, the personal computer revolutionized our world, then the internet in the ’90s, smartphones in the 2000s and cloud computing in the 2010s. Now, we are on the brink of another significant technological transformation driven by generative AI (GenAI).
AI has been around for decades. So, why is there a sudden buzz around GenAI?
This surge can be attributed to a confluence of factors coming together in the past two years. Advances in raw computing power, the exponential growth in available data, plummeting AI training costs and breakthroughs in machine learning algorithms, such as the transformer model, now deliver unparalleled sophistication. OpenAI was quick to harness these trends; they wrapped the technology in an approachable interface and created a killer app that started the revolution and has gained millions of users since its launch.
Foundational Models
GenAI is based on foundation models (FM) that allow it to understand context and relevance within the content it processes, making it incredibly powerful and versatile. The transformer architecture represents a paradigm shift in how AI processes text, offering unprecedented language understanding.
Building FMs involves two key stages: unsupervised training on vast amounts of content (text, images, etc.) and fine-tuning for specific tasks, such as Q&As. The process is complex and resource-intensive. Selecting what FM to use is the first step in building a GenAI solution.
Once you select an FM, you need to decide how you will build your GenAI application.
We have developed numerous GenAI use cases and have identified four solution patterns that enterprises can choose based on cost, complexity and accuracy.
Pattern 1: Out-Of-The-Box LLM API-Based Solutions
The most straightforward approach to generative AI involves leveraging readymade APIs from large language model (LLM) providers like OpenAI, Anthropic and others. This method is straightforward and doesn’t require extensive development or data science skills. It simply requires building a front-end application that calls these LLM APIs.
User prompts are the primary mechanism to provide context in this approach. The better the prompt, the better the output. This has led to prompt engineering as a discipline in which prompt engineers can try different inputs to arrive at the “optimal” prompt that elicits the “best” answer.
While this pattern is quick to implement and requires minimal coding, accuracy is low as the LLMs have a limited "context window." A context window refers to the length of text an LLM can process and respond to for any given instance in a prompt.
This pattern is best for generic use cases that don’t require enterprise context.
Pattern 2: In-Context Learning
This popular pattern, also known as a RAG (retrieval-augmented generation) pattern, involves using cloud-based APIs from LLM providers while feeding them with enterprise context.
This pattern also leverages an API from the LLM provider, but there is a backend application that can take the enterprise content, break it down, vectorize it, store vectors in a database and, when a user query occurs, provide appropriate enterprise context through vector search as input to the model.
Unlike Pattern 1, the prompt engineering is done by the backend application, not by users. Pattern 2 balances simplicity and effectiveness, making it a go-to choice for many without finetuning existing models or building new ones from scratch.
This approach is a lot more accurate than Pattern 1 and can be built by regular application developers without machine learning expertise. However, this pattern has limitations due to the limited context window, particularly for domain-intensive use cases.
Pattern 3: Fine-Tuning Existing Models
An enterprise can fine-tune existing models with domain-specific content if it wants to go further with more accuracy and speed.
Fine-tuning is the process of adjusting model parameters. It involves training an existing open-source model on your labeled dataset, allowing it to change the model’s weights based on your data. Clean, curated datasets can speed up the training process and boost accuracy. This approach requires data science, machine learning expertise and meticulous data labeling of your content.
Based on our experience, we’ve found that it typically takes a few months to train an existing model on your domain-specific content. It’s an iterative process and requires continuous evaluation and testing to ensure that accuracy is improving. If not, you need more relevant content or better labeling of content.
While this solution pattern is complex and expensive, it gives an organization a competitive advantage with higher accuracy and performance by tailoring models to specific tasks.
Pattern 4: Building Your Own Foundation Model
The most complex pattern involves building a custom model from scratch. It’s a significant investment and commitment but well worth it, especially if you have terabytes of unique data, and the GenAI use case can allow you to differentiate from and leapfrog over your competitors.
It requires vast data resources and advanced data science and infrastructure management expertise. While this option may seem out of reach for most companies today, as open-source tools emerge and GPU chips become more affordable, overall costs will come down, boosting the ROI for building models from scratch.
Selecting The Right Pattern
When choosing the appropriate pattern for your use case, consider the balance between cost, complexity and accuracy.
Most of our customers start with building proof of concepts using Pattern 2. Given its balance of effectiveness and feasibility, it’s no wonder that this pattern is the most popular. However, the field of GenAI is maturing rapidly, and I expect Pattern 3 (fine-tuning models) will gain prominence over time.
Each technological leap in the past 30 years has redefined our world. Fail to jump, and you’re left behind. Choose one of the four solution patterns that align best with your company’s readiness, capabilities and goals. The key is to start experimenting, learning and continuously improving.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
{Categories} _Category: Inspiration,*ALL*{/Categories}
{URL}https://www.forbes.com/sites/forbestechcouncil/2024/01/12/how-to-embrace-the-power-of-ai-four-patterns-to-fuel-your-ai-strategy/{/URL}
{Author}Rajul Rana, Forbes Councils Member{/Author}
{Image}https://imageio.forbes.com/specials-images/imageserve/65a03cc043b4372e7aa58497/0x0.jpg?format=jpg&height=600&width=1200&fit=bounds{/Image}
{Keywords}Innovation,/innovation,Innovation,/innovation,technology,standard{/Keywords}
{Source}Inspiration{/Source}
{Thumb}{/Thumb}
