We are very excited about the release of LLM models, including its chat adaptations like ChatGPT. LLMs have achieved tremendous improvement in the accuracy of language-based predictions. Accuracy has always been a focus in NLP, which has been improving over the years aided by many breakthroughs, including the use of deep learning/neural networks for predictions, vector representations (word2vec, glove, etc.), high dimensional embeddings, transformers, transfer learning, and so on.

‍

For almost all the use-cases we solve using NLP over last decade, we leveraged both probabilistic and deterministic algorithms to ensure its reliability in a real-life enterprise production functions. NLP, as it's based on probabilistic models, (yet) cannot be used by itself as a reliable solution for business-critical use cases. We can safely use it for use cases like classifying customer interactions, reviews, extracting entities, etc. which are tolerant to a few miss classifications or wrong predictions. However, for use cases where we have a monetary value or impact on people/businesses, we need deterministic algorithms or processes in place to ensure accurate processing.

‍

For e.g., Assume a simple business process in which we send checks/cash to users based on a handwritten or image of a form they submitted. The system automatically extracts the amount from the submitted form and issues a check. To ensure there is no leakage, we add conditions like human in-loop verification for amounts greater than $1000 and/or check users' past transactions, designation etc before the amount is processed.

‍

The above use case is a simple one and enterprises have thousands of simple and complex use cases, where the impact will be critical to the business if it's incorrectly processed. The tolerance for error depends on individual scenario and trade-offs.

‍

LLMs, irrespective of all the hype it's chat version generated, hasn't changed the above except that the chances of error has significantly been reduced. Since we cannot rely on "chances", we still need deterministic algorithms that can be clearly explained and some times humans in loop to complete business processes. We believe this would remain true as long as we use probabilistic methods in NLP, but then predictions are probabilistic unless we know the states of each of the features that has an impact on the value being predicted, right?

‍

‍

We believe a combination of LLMs, specialized language models, deterministic algorithms or checks and humans will be key to process automation. A lot of manual work that involves straightforward reading or writing can be automated. But anything that needs creativity or decisions within a context will require a more reliable process devoid of hallucinations.

‍

Fun question: Will you cross a river which is average 4ft deep, if you are 6ft tall?

‍

‍