Why do we need to tell an AI that it is a genius at X or a guru at Y – and yet sometimes, often, get excellent output without doing any of that?
Think of it as a map
Picture a language model as a vast map of everything it was trained on – mountains, deserts, seas, and everything in between.
When you ask a question, upload a file, provide a prompt, or tell the model that it “is” something, all of those actions have exactly one effect: choosing where on that enormous, hyperdimensional map it goes to generate the output.
Role-playing instructions simply narrow that landing zone, steering it towards training data that hopefully did come from an expert on the topic.
Saying you want to know about X leaves a broad scope – real data, forum posts, rumour, actual fairy tales. Saying you want to know about X from an expert on that topic narrows it to industry journals, formal studies and informational presentations. Much better.
The catch: context is a cup that overfills
Given the current technical limits of context, that focusing effect combined with a continuing conversation can lose or degrade the original target parameters – like a cup that got too full.
That is how you start getting bad data unexpectedly in the middle of an otherwise excellent session. It is also why even the strictest rules against socially problematic material break down the longer a conversation runs.
Conversation length and available context are inextricably intertwined, and the challenge for both AI services and AI users is finding the sweet spot – which varies by service, by topic and by need.
Think of it like a workbench. Some are smaller and some are larger, but all of them are limited, and the work may need breaking into smaller pieces to stop things falling off the edge.
Three techniques that follow from this
- Especially for complex work: ask an AI to help you improve your prompt, then use that prompt in a new session.
- Ask multiple AIs the same thing, or parts of the same thing, then put the results into a model with a very large context to combine and summarise.
- Use separate sessions to critique outputs.
In other words: given the inherent limitations of the current technology, it is usually best to work within those limitations strategically, and build something bigger out of the pieces.
Itwerx Corp is a service-disabled veteran-owned small business providing IT services across Seattle, Bellevue, Everett and Snohomish County. This is the kind of thing our business process analysis work deals with – talk to us about yours.
