Are LLMs trustworthy and can they do reason as they claim?

🚩Are LLMs trustworthy and can they do reason as they claim? This was the main topic of the expert panel discussions at #DecisionCAMP 2024.

Certainly, there was a wide range of ideas on using LLMs as part of decision management practice and automation efforts and #DecisionIntelligence space. To give you a quick glimpse:
* Creating decision models for new projects
* Understanding a required change in existing models
* Generating reports based on the logs of executed decisions
* Reducing the barrier of the adaption of new techniques and modeling (e.g. DMN)

However, everyone on the panel agreed that the LLMs are good in certain areas and untrustworthy and unreliable in many others.

➡️What does that mean to you?

1️⃣ LLMs-driven applications, i.e., #chatGPT and other forms of LLMs use cases, cannot (and should not) make critical decisions, such as life-dependant ones.

2️⃣ LLMs driven based on user interface WILL reduce the need for IT consultants

3️⃣ Calling out to LLMs model (or external service providers) in the middle of a decision model is a valid scenario; however, it requires a form of consensus decision model between multiple outcomes e.g. running #businessrules, traditional #MachineLearning and #LLMs together and a decision model makes the consensus on what's acceptable and why.

💡Today's Expert Panel on #DecisionCAMP was interesting. One thing for sure is that #LLMs created so much buzz and hype around #AI that even decision modeling and decision automation experts cannot ignore it!

I don't know what the future holds as many big techs invested in LLMs and yet could not make the ROI from their investments. I believe LLMs cannot be reliable, not because of lack of data, parameters, or quality of data for that matter, but because of their underlying architecture.

🚀I strongly believe that LLMs hallucinations are not defects but are by-products of #LLMs' underlaying transformer architecture. So no matter how much we dislike them and how much #data, #RAG, fine-tuning and money we throw at them, hallucinations and unreliability WILL NOT GO AWAY!

Posted here.

Published September 20th, 2024 at 07:30 am