7 Reasons AI Needs Knowledge Management

7 Reasons AI Needs Knowledge Management

Cover image by DALLE-3 using a summary of this article as a prompt.

Much of the discussion about generative AI and the future of work focuses on the potential for AI to displace workers. Often vague references suggest that AI will create new jobs as did all the automation of the past. Like previous mechanical and digital automation, there are populations of experts who already understand the human labor required to make automation efficient and sustainable. Computers without programmers could do nothing. Grain combines without mechanics to clean and maintain their mechanisms would soon stall in the vast wheat fields they were intended to harvest.

Generative AI certainly has its inventors, those who know the secrets instilled in algorithms, as much as the algorithms permit. They know, more than the users of generative AI, at least, what mixes of data were used for training.

Once trained, many believe generative AI to be akin to a mythical beast that rises from an abyss. In this case, that abyss is the Internet, both light and dark. Generative AI follows the proclivities of all those who populated those vast repositories with facts, opinions, grievances, poetry, and many other outputs of humanity, written and visual.

To move beyond experiment—to prompt adoption and use by individuals and businesses—generative AI must be trusted. AI that randomly curses, one that reflects discriminatory positions, one that devolves into nonsense, raises doubt about the efficacy of AI as a legitimate partner in business or in life.

So, the developers of Large Language Models (LLM) institute various remediations to the aberrant AI behavior, reigning in its most outlandish assertions. Rather than reflect all of the Internet, LLM makers hone output to avoid culturally sensitive areas, adopt a shareable context, and attempt to curtail the most unconventional outbursts. They do this response management with manual overrides that sit between the LLM and the people or systems making queries.

The key word here is manual. While there may be some automation involved in writing what has become known as Guardrails, the choices about what needs to be ‘guardrailed’ are purely human. No generative AI is self-policing of its aberrant or offensive behavior.

The manual nature of guardrails requires knowledge management. Which guardrails are in place, their content and context, may need to be known to offer comfort to buyers, or to act as a baseline as expectations, politics or other contexts change.

Guardrails are not the only area where knowledge management needs to be applied to generative AI management. The following list outlines the most important areas where organizations need to apply knowledge management principals to generative AI development and deployment.

1. Guardrails. Guardrails need to be documented and searchable. They will likely become complex. Some commercial services may already have guardrails in place. Customers will need to understand those parameters so they do not write new guardrails that conflict with those already wrapped around a generative AI service. Organizations will need to include guardrails in their change management practices, as they may become irrelevant or require augmentation depending on business situations—such as acquiring a new firm or opening a branch office in a new country.
2. LLM Metadata. Those who have sifted through the dozens of models available to LMStudio for download will realize that the most important parameters to that search come in the form of a model’s ability to run on the target hardware. Neptune.ai is all about machine learning metadata. They understand that organizations need to know how models were built and evaluated, what dataset versions were used, and test metrics, just to name a few attributes of metadata that are crucial to understanding the nature, history, skills required to maintain them, and their best use cases.
   Model repository Hugging Face offers some search filters for sifting their repositories, but they don’t include the kinds of parameters Neptune.ai tracks. Hugging Face models often list important information in their associated narrative, but that means reading through the released notes of candidate LLMs to see how they were trained and perhaps if they can be trusted. But first, they need to be discovered. Hugging Face and others reprioritize and need to apply metadata so that they can keep their repositories relevant. A recent post on Medium from Ayush Thakur offers insight on Hugging Face models. The output of analysis like this should be part of the metadata associated with the models. (Read his work here: Top LLM Models on HuggingFace Outperforming GPT Variants on Medium.)
3. Context models, such as knowledge graphs, are often human-generated. Knowledge graphs are not new. They offer a semantic, often visual way to represent the relationship between categories of knowledge and their related data. Knowledge graphics are becoming a go-to approach to reaching more contextually correct results of generative AI by reducing its tendency to wander through and incorporate less relevant data that might skew output and create distrust. Knowledge graphs require human reflection, monitoring and maintenance.
4. Prompt management. People will want to share prompts. They want to remember them. Several organizations and individuals already publish prompt cheat sheets, a common form of knowledge sharing. People will also want to collaborate on prompts. The rise of prompt engineering as a discipline will be replete with knowledge management needs. Prompt will likely be managed as products, and entered into lifecycles that capture their creation, their evolution, and their retirement.
5. Model management and retirement. LLMs are new. They pop up out of the ether, seemingly like quantum particles making a random appearance. Unlike quantum particles, at some point the number of LLMs created daily will slow. Larger models will gain momentum—but technology isn’t going to stop. As LLMs evolve, newer models may replace older models, and those models may not include the same interfaces. New APIs, new guardrail syntax, and other factors may change—just like they do with any large enterprise system over time. Models will get retired and replaced. Detailed knowledge about how a model is embedded in the technical and business architectures of a business will matter. Some models may prove radical departures that prompt a reconsideration of all things built on earlier LLMs.
6. New knowledge or lack thereof. I was talking with a manufacturing engineer at San Diego Comic-Con. She had just shared her superpower: the ability to forecast the future, because of the predictive analytics available to her from the production line. She could determine which machine might fail and plan around the maintenance of that machine to minimize the impact on production. I asked her, “What happens when you get a new machine?” “I lose my foresight for a while,” she said because the new machine’s operations profile and how it impacts workflow isn’t yet known.
   Generative AI has the same issue. It only knows what it knows. It has limited ways of expanding its knowledge; currently, human-type learning isn’t in the mix. Understanding the edges of knowledge will be a knowledge management task for companies adopting generative AI. New discoveries in physics, for instance, might invalidate knowledge about how we understand and perceive the universe. With decades of consensus, emergent data and models will be challenged to find a footing.
7. Retrieval-Augmented Generation or RAG requires crafting content that can be used to augment the learning of an LLM, acting as in-the-moment training during a query. RAG helps retrieve more relevant answers. It requires the selection of content by people to inform the model. How RAG is used, the data that it employs, and lessons learned from the experience all represent classic knowledge management work.

There you have it: 7 reasons generative AI needs knowledge management. More will likely emerge as LLM developers share their practices and enterprises learn how to adopt and apply them. Humans created education because the accumulation of knowledge in context requires work.

Generative AI is riding on history. Keeping LLMs current, relevant, and useful will require them to learn in the piecemeal way humans learn about areas they have not yet encountered and new discoveries in their discipline areas. Reading and comprehending text quickly is a differentiating skill in human learning and machine learning, but it isn’t the only skill. Humans can overlay their learning with new knowledge when they discover new facts that replace old ones or when they have been deceived.

As fast as generative AI can learn, it remains to be seen if it can relearn as the world changes. Although knowledge management often purports to focus on knowledge retention, its real value derives from it as a platform for learning, creating, and innovating, not just remembering.

For more serious insights on AI, click here.