miracle coder
Generative AI (GenAI) models have become the focus of attention, captivating our imagination with their remarkable capabilities. Whether it’s crafting email responses or generating computer code, these large language models (LLMs) hold immense potential. However, they also face a significant challenge: hallucinations. Hallucinations refer to instances where GenAI produces responses that include fabricated data, often resulting in inaccuracies. In this article, I explore the reasons behind hallucinations, their impact on user trust, and strategies to mitigate them.
Understanding Hallucinations
One noteworthy example of GenAI hallucination is Google’s Bard chatbot incorrectly claiming that the James Webb Space Telescope had captured the world’s first images of a planet outside our solar system. These inaccuracies result from various factors such as limited data or overfitting. While some hallucinations can be negligible, others can have significant consequences.
As users of GenAI, how can we recognize when to trust the output from GenAI?
In the realm of human decision-making, our intuition plays a vital role. Let’s draw parallels between GenAI hallucinations and trusting our gut. There are two scenarios where I would not blindly trust the output from GenAI:
- Limited Data: GenAI struggles with insufficient data, just as our intuition can fail where we have limited experiences.
- Overfitting: GenAI can overfit to its training data when the model is trained for too long on a sample data or when the model is too complex. It can rely on these past experiences, even when they don’t align with the current context. Just as our intuition clings to familiar patterns, GenAI can sticks to its training data.
Building Trust
To improve my trust in GenAI results, I use these steps when formulating my questions.
- Write clear and specific instructions
- Use delimiters to clearly indicate distinct parts of input.
- Request structured output. This helps the model to produce more consistent outputs.
- Prompt the model to simply tell me when it doesn’t know the answer to a question instead of trying to come up with an answer.
- Use few shot prompting
- Give the model some initial examples in the prompt to guide it towards generating a more accurate response.
- Give the model time to think
- Ask the model to specify steps required to complete a task.
- Instruct the model to work out its own solution before reaching a conclusion.
Conclusion
GenAI holds a bright promise, but trust is contingent upon recognizing the limits within which GenAI operates. By understanding the root causes of GenAI hallucinations and drawing insights from intuition, we can navigate this exciting frontier while maintaining confidence in it’s application.
Whether it’s GenAI or our intuition, trust is a delicate balance—one that requires vigilance and thoughtful consideration.
