The Weight of Complexity in Decision Making

Imagined by Bing Image Creator from Designer with the Prompt: “Weighted scales stacked upon one another varying in weight on either side”

Humans are constantly “weighted” by decisions that inform our every action and dictate our past, present, and future.

I quote “weighted” intentionally as decisions always involve tradeoffs, doing the desired action often involves giving up the opportunity to do the other in that exact moment. We as humans naturally calculate the value of each decision (think “Expected Value” from statistics”) based on a number of internal and external criterion ranging from the intangible such as happiness to tangible such as money. Computing these expected values “weighs” the possible tradeoffs to a point where a decision can be made and the action taken. In the moment, weight weighs down on us as we compute the value and try to evaluate the tradeoff, often with too many variables to assign values to (think n-dimensional space) and can lead to emotional burden as well (Imagine making a life-changing decision, for example). In retrospect, after taking said action, a postmortem or reflective view can inform whether the correct decision was made, leading to reevaluation in any direction, that is how we learn.

Machine Learning systems, or Artificial Intelligence, is programmed to function similar to humans.

1. Training Data: Past experiences in the form of real (or synthetic) data create the learning and creates that mental model. In the case of Conversational Intelligence engines — Text-to-speech, speech-to-text, natural language understanding —training data can consist of millions of examples or hours of audio recordings. For Generative Intelligence engines, such as those powering Large Language Models, input is billions of text-based documents!
2. Models: After ingestion and training (notwithstanding model tuning which will be discussed shortly), the mental models are created to allow for interpolation or evaluation. Keeping in line with the bifurcation earlier, Conversational Intelligence engines generate a smaller subset of variables to consider within the model versus a Generative Intelligence model creating variables within 70B+. Imagine trying to weight 70 Billion different influences on the outcome in a split second, that is the beauty of AI!
3. Learning: At the heart of the success of AI models comes the ability to improve over time, similar to humans, however, its a more manual process. Tooling capabilities allow for AI designers and data scientists to review the inputs and outputs of a model, make edits or corrections as required, and input that new information into the model to be retained. As an example, for interactive voice response applications, a conversational designer may look at the utterances or phrases said by the caller and the intent that was identified. Should there be a mismatch, it will be corrected and added into the dataset to be retrained. With Generative Intelligence engines, that comes in the form of a term called “Reinforcement Learning through Human Feedback” or “RLHF”. RLHF effectively assigns scores to the response on a sliding scale to reward the system for its response, similar to treat training a dog where the machine, or dog, optimizes to earn the highest reward and thus will generate a response appropriate to get the prize. (This is highly simplified and there are multiple forms of AI training but mostly on the quantitative side of Machine Learning.)

Again, these Artificial Intelligence models are simply probability and statistics, using the training data to “weight” outcomes based on computed expected values, similar to how humans make decisions, just in a fraction of time.

Simply Put, Its Probability and Statistics

Behind the models are data scientists and speech scientists whose job (simplified) is to “weight” the models. The pun is accurate — these experts select a multitude of weights that comprise of the training, modeling, and learning of these systems:

1. Training — given a complete dataset or corpus of knowledge/documents, data scientists, both on the conversational and quantitative side of Machine Learning, partition the data into various groups — a training group for training the data, a validation set for validating the model, and a test set for assessing the accuracy of the model. Playing with these weights often changes the final accuracy and optimizing for the perfect weight is critical to the success of the role (and using optimizers to reduce the error but that is beyond the scope here).
2. Model — often within applications, multiple models could be used to make a decision. For language-based models, should the outcome of a generalized model be compared with a topic specific model? What weights should be assigned to each model such that the output is valuable to the user? This is a common practice within conversational design both on the underlying model and overall output.
3. Learning — going back to the dog treat example, speech scientists assign value to correct answers or penalties for incorrect answers as feedback into the engine (Generative Intelligence). This helps to optimize the model over time to be a subject matter expert with additional real data, other than the initial ground truth it was trained on which could be out of date or not holistically representative of the overall breadth of asks the model should be trained on.

Comparing to humans, as we experience life and gain more knowledge, experience, and feedback, we too change the weights of our own internal mental models.

Taking one level higher comes parameters of models, specifically as we have seen within large language models. For example, one could play with temperature to make the model more creative or precise depending on the use case — brainstorming versus research respectively, p-values that penalize the model for repeating words too often, and even how many works or tokens to generate. These can be tuned by application designers to obtain a desired output.

Comparing to humans, we may have guardrails that are imposed on us that effect the outputs of our actions — ranging from commitments, rules of differing environments, and much more. Imaging sitting at a fine dining restaurant versus a casual one, there are different rules tuning your behavior parameters.

In summary, yes Artificial Intelligence is simply probability and statistics, however, its “weighted” heavily by a multitude of parameters or conditions that effect its behavior. Humans, when inventing Artificial Intelligence, draw upon our common knowledge, ourselves, in its modeling.

There is immense complexity to AI that are still unknown such as how the models create internal weights within a neural network and what those weights mean (as opposed to the confidence score output of a Conversational Intelligence engine), but at least we can appreciate the intricate balance involved in creating AI engines, AI models, AI powered applications using those models.