The recent widespread adoption of large language models (LLMs) by individuals and leading corporations, with tools like ChatGPT, Bard, and numerous open-source projects, has demonstrated the significant value of these models and the shift we are currently experiencing, both in terms of the wow factor and commercial applications.

LLMs serve as valuable starting points for tasks such as text summarization, answering straightforward questions, coding assistance, text comprehension, task automation, and potentially much more.

So far, LLMs have mainly been implemented through chat-like interfaces or given the freedom to generate free-form paragraphs in word processing applications like Notion or Microsoft Word.

However, it is essential to address the potential downsides of these models, and taking a design perspective might be helpful, particularly when it comes to managing expectations.

One of the major limitations of LLMs is that they can generate text that appears linguistically accurate but lacks factual accuracy or context, necessitating subsequent editing or fact-checking. This inconsistency persists despite continuous advancements in these models, including issues related to model transparency, training data accuracy, and neutrality.

Furthermore, chat-like interfaces or even a simple button unintentionally tend to glorify the output, making it challenging for users to anticipate these problems, even if they are aware of them.

What if we approached interactions with LLMs differently?

Perhaps by treating them as preliminary idea generators and acknowledging the need to scrutinize and dismiss a significant portion of their output.

One proposed approach is to structure the responses of these models, especially when it comes to text generation, in a manner that simplifies the verification of their output and to utilize them primarily for tasks that prioritize quantity over quality, such as brainstorming.

NodePad is an experiment that explores the effectiveness of this approach. The application relies on a canvas and the concept of nodes or blocks, encouraging users to engage with the fundamental components of their writing, starting from a thought, a sentence, or even a very short paragraph.

Generating paragraphs is not possible in NodePad; instead, the application of LLMs is guided by predefined prompts that focus on specific tasks like brainstorming and questioning only.

The output consists of short nodes that expand upon the original topic.

It is important to emphasize that not all output may be useful, particularly in the context of brainstorming and generating questions.

The output is self-contained within nodes that can be easily deleted.

There is also an element of serendipity, encouraging users not to search for specific information but rather to embrace a mindset of exploration.

Finally, NodePad includes features such as node connections, highlighting interesting nodes, arranging them, and the ability to download a simple text file representing these nodes.