Designing for LLMs - Designing for Imperfection

Machine learning tools used to be like dogs trained solely to do a special tasks. Prone to distraction, unable to do much more than they are explicitly told. Large Language Models have broken that paradigm. The state-of-the-art model, GPT-4 generalizes across tasks, performing well on many aspects of human reasoning and across many different fields.

But LLMs still come with inherent limitations. They can produce outputs that deviate from reality, exhibit biases from their training data, and display high variability. Their stochastic nature make them near unpredictable to date, yielding variable response even to identical queries. In software applications, we must address these challenges and create tools that embrace these imperfections or take advantage of them.

Designing a tool for imperfection means creating a tool or system that acknowledges and accommodates the possibility of mistakes and variability in the AI’s output.

Guide the users through mistakes

Users need to understand that responses aren’t always accurate and can vary widely. Just as meteorologists analyze complex weather patterns to make predictions, LLMs process vast amounts of data to generate responses. However, just like weather forecasts can be subject to change due to the unpredictable nature of atmospheric conditions, LLMs too can produce responses that may deviate from the desired outcome or exhibit variability.

When users are made aware of the inherent limitations of LLM-based tools, they can approach the generated responses with a balanced perspective. To enhance this interaction, clear instructions and guidelines become our trusty navigational compass. Users receive specific guidance on how to sail smoothly and reach their desired outcomes. Educating them about input preferences, like preferred formats or specific phrasing, boosts the accuracy and relevance of the responses. Highlighting potential pitfalls and challenges equips users to avoid errors and misunderstandings along the way.

Be liberal in what you accept from others, even if you aren’t conservative in what you do

LLM-based tools can gracefully respond to user input by offering suggestions or requesting clarifications. It's a dynamic and collaborative dance, where the tool and user work together to disambiguate the intended meaning. This adaptability allows the tool to refine its responses.

To give the user more control and fost confidence, provide them with multiple outputs, allowing them to review and choose the most appropriate one or even explore an entirely new set of options. Users can evaluate and select the response that aligns best with their needs.

As we venture further into the realm of LLM-based tools, it is crucial to approach their imperfections with a mindset of innovation and opportunity. We can design systems that embrace these limitations, developing tools that accommodate mistakes and variability, while empowering users to actively participate in the refinement process.