As large language models (LLMs) become more accessible, an important question remains:
How do we make them genuinely useful, reliable, and practical?
At Zorn Novum, we've been examining how instruction structure, context, communication styles, and verification methods influence the responses produced by LLMs.
One observation is that the way a task is communicated can significantly affect the usefulness of its output.
A vague instruction may produce a reasonable-sounding answer without actually addressing the intended problem.
A more structured request, with clear objectives, relevant context, and defined expectations, can produce a more useful result.
However, there's another important consideration.
A convincing response is not necessarily a correct response.
Understanding that distinction is essential when evaluating how LLMs might support research, business operations, and software development.
Our work has also involved examining common response failures, methods for improving clarity, and approaches to reducing unsupported conclusions.
We've compiled a more detailed internal field guide covering these topics, and we're interested in engaging with others who are exploring similar questions.
We'd particularly welcome perspectives from researchers, developers, business professionals, and others who regularly work with LLMs.
What challenges have you encountered? Which techniques have proven useful? Where do you believe current approaches fall short?
If you're interested in discussing our work or potentially providing feedback on our more detailed findings, we welcome a conversation.
At Zorn Novum, we believe meaningful technological progress comes from curiosity, critical thinking, and a willingness to challenge assumptions.