The Corporate Ladder & Generative AI Training: How Alex's Professional Journey Mirrors AI Training Techniques
In the rapidly evolving world of artificial intelligence, understanding advanced techniques can sometimes feel overwhelming, but what if we could unravel these complexities through a relatable narrative? In this blog post, we’ll journey alongside Alex, a fresh business graduate, and discover how his professional experiences draw uncanny parallels with the intricacies of generative AI models. From prompt engineering to fine-tuning and domain adaptation, let’s dive into the world of AI, all through the lens of Alex’s corporate adventures.
Alex is a recent graduate with a Bachelor’s degree in Business Administration from a top-tier university. During his college years, Alex acquired a strong foundation in business principles, analytics, and strategic thinking. Eager to apply this academic knowledge to the real world, Alex joins a multinational corporation called Cyberdyne Systems. Cyberdyne boasts a wealth of proprietary data, encompassing detailed customer profiles, years of sales data, and unique industry insights not readily available to the competition.
Prompt Engineering with Retrieval-Augmented Generation
In the first week at Cyberdyne, Alex isn’t just handed an overview of the company’s operations, he has to fend for himself. He’s equipped with an advanced internal search tool, allowing him to swiftly pull up pertinent company data as needed. Picture Alex in a crucial strategy meeting: A question about European market trends emerges. Rather than relying purely on his own learned knowledge and recall, Alex uses the tool to promptly retrieve a concise summary of recent sales trends, customer feedback, and competitor actions in Europe. Armed with this real-time data, Alex confidently weighs in on the discussion, anchoring his points in current company data while applying the skills he learned from his time spent at the university.
Generative AI Equivalent
This is where retrieval-augmented generation (RAG) comes into play in the AI realm. When faced with a query, the AI model doesn’t just lean on its pre-trained knowledge. It actively retrieves relevant details from a dataset (which could span a vast collection of documents or knowledge bases) and then formulates its response based on this retrieved insight. For instance, when prompted about strategies for the European market, a RAG-enabled model would first source pertinent data or insights from an internal, private company dataset and then construct a detailed answer grounded in that information.
As the months roll by, Alex gains deeper access to vast amounts of Cyberdyne’s proprietary data. Participating in specialized workshops, he dives into this data, deciphering customer feedback, scrutinizing sales patterns, and grasping product performance metrics. This granular knowledge empowers Alex to make decisions tailored precisely to the company’s nuanced needs.
Generative AI Equivalent
In the AI world, this mirrors fine-tuning. The model is trained on a subset of the company’s private data, allowing it to specialize and develop a deeper understanding of that domain. The model moves beyond mere general trends and short, concise data points, but instead grasps the subtleties and specifics due to this in-depth tuning across a large corpus of data points.
A year into his role, a new challenge beckons for Alex: leading Cyberdyne’s foray into South-East Asia. To gear up for this, Alex accesses a specialized set of proprietary data that chronicles the company’s past ventures in analogous markets, deciphers cultural purchasing trends, and analyzes regional competitor data. This exclusive information is pivotal, helping Alex not only fine-tune his strategies to resonate authentically with the local market but also look at things specific to this entirely new market.
Generative AI Equivalent
This adaptation in AI is embodied in domain adaptation. A model, proficient in one domain, is adapted using insights and data peculiar to a new domain. The model leverages its foundational understanding while acclimating to the intricacies of the new domain, informed by the private data.
Alex’s journey from a newly minted graduate to a seasoned professional at Cyberdyne Systems mirrors the progression of using generative AI models. Both tales underscore the invaluable role of proprietary data as a resource, paving the way for enhanced performance and specificity. Whether you’re scaling the corporate ladder or diving into the world of AI, adaptation, and continuous learning remain key.
My hope is that you can take this completely ficticious example and see the parallels to training generative AI models.
Many individuals I encounter commonly get the concepts of prompt engineering, fine-tuning, and domain adaption confused, or feel that a model has to be “trained” in order to extract value from its use. Each of these techniques has their use case. By gaining an understanding of the differences between each of the approaches to integrating private data into a generative AI model you gain the power to truly wield this technology free of misconceptions.
An overwhelming MAJORITY of AI use cases can be solved via prompt engineering (along with retrieval augemented generation) as a starting point. Once you start to realize the limitations of this approach then you should consider instruction based fine-tuning and finally domain adaptation.
Ironically(?) this order of operations is also representative of cost & complexity, with prompt engineering being the lowest and cheapest approach, followed by fine-tuning then domain adaptation.