Skip to content Skip to sidebar Skip to footer

Managing AI Risk with VUCA and Knowledge Engineering Part 5 – COMPLEXITY

Courtesy of Joe Glick, Chief Data Scientist, RYLTI

Part 1 addressed why VUCA is appropriate for managing AI risk. On the issue of complexity DARPA’s 2005 “Get Real” project calculated the potential interactions of one million agents as 10^300,000 to stress that exhaustive modeling of any significant element of the real world is not technically feasible.  Their recommendation was to research real-world reasoning methods in imitation of the human mind, which filters out the vast majority of ingested data and uses what is needed from relevant domains. Computing relevance has two key challenges:

  1. Identifying relevant information domains for the analysis purpose
  2. Identifying relevant data required to perform the analysis

Steven Eppinger, Professor of Innovation at MIT, used to teach a course in complexity management at the Graduate School of Engineering.  The example used for practical application was designing a new jet engine, but in the final test, he asked us to apply the principles to a pizza recipe. It helped focus on what was relevant and necessary and we used that at Pfizer to explain complexity modeling during a business transformation. Part 4 discussed how differently a chef or a healthcare researcher would answer the question: what makes food taste good? I asked ChatGPT and got a list of attributes for the interaction between food and the senses – flavor, aroma, texture, etc.  As Part 1 explained, complexity is invisible to an AI algorithm because:

  1. It can only see what it has been engineered to find.
  2. It works within the narrow limits of the training data.

The LLM response is a classic example of both points above. If we implement the principles discussed in Part 2, independent real-world reasoning agents that apply small/wide data methods to real-world data, we can build expert knowledge graphs to comprehensively define the relevant domains and data. Those models can be used to test AI outputs for correctness and completeness and identify gaps that need to be addressed.  Based on the output, engineers can make needed adjustments to training or explore other options to correct the issues.

Sometimes, the precise definition of relevant domains and data is insufficient to validate the AI outputs because we do not fully understand activities that we can observe and measure. A frustrating example is photosynthesis. It doesn’t appear complex and is the most common process on the planet, but we can’t reproduce it in the lab.  Achieving that could eliminate food and energy shortages worldwide. So if we have to validate a model that is too complex to understand, what can we do? Model and validate the underlying premises. The biomimetic digital twin ecosystem described in Part 3 can enable the exploration of complex systems objectively and test the conclusions stochastically.

With LLMs, however, the fourth VUCA risk driver, Ambiguity, is a serious issue.

Leave a comment