For those who have access, what has been your experience?
2/2

********
Want to get better at writing AI prompts?
Reach out!

calendly.com/lapis-a...

www.lapisconsults.com/
9/9

And if you can't figure out how to phrase something in the positive, just ask the LLM to rephrase it for you!
8/9

The most effective approach is to tell the model what you DO want rather than what you DON'T want. Providing positive instructions and examples tends to work much better than prohibitions.
7/9

4. Training data patterns: Models are trained on data where negations are often followed by examples of the negated concept, creating a learned association.
6/9

3. Negation complexity: Understanding negation requires a higher level of logical processing. It's easier for the model to recognize "do X" than to fully process "don't do X" as a prohibition.
5/9

2. Attention mechanisms: Modern LLMs use attention to focus on relevant parts of input. By mentioning a forbidden topic, you've directed attention toward it, increasing its activation weight.
4/9

1. Semantic priming: When you tell an #LLM "don't do X," you're actually making X highly salient in the context. The model has now processed and activated the concept of X, making it more likely to appear in the output, not less.
3/9

The phenomenon, a.k.a. the "Don't think of an elephant" effect, happens for several reasons:
2/9

Glad you enjoyed it!
The voice is truly mind-blowing and so simple to implement 😁