AI Controls

Creating precisely what you need with AI requires understanding how to control the AI's behavior. In this guide, we'll explore the various controls available in Chibi and how they affect your results. We'll start with the fundamentals and build up to practical applications.

The Foundation: How AI Generates Text

Before diving into specific controls, it's essential to understand how AI generates text. Modern AI models generate content one token at a time (tokens are roughly equivalent to word parts). For each token, the AI:

  1. Looks at all the tokens that came before

  2. Calculates probabilities for what might come next

  3. Chooses the next token based on these probabilities

  4. Repeats this process until completion

This token-by-token generation process is crucial to understand because each control we'll discuss influences how the AI makes these choices.


Core Controls

Temperature: Balancing Creativity and Consistency

Temperature is perhaps the most important control, affecting how the AI chooses between possible next tokens.

What It Does:

  • Low temperature (0.0 - 0.3): The AI consistently picks the most probable tokens

  • Medium temperature (0.4 - 0.7): The AI balances between probability and variety

  • High temperature (0.8 - 1.0): The AI is more likely to pick less probable tokens

Best Used For:

  • Low: Factual writing, technical documentation, structured formats

  • Medium: General content, blog posts, creative but focused writing

  • High: Brainstorming, creative writing, generating diverse ideas

Practical Example: When asked "What's a good breakfast?", the AI might respond:

  • Low temperature: "A balanced breakfast consists of protein, whole grains, and fruit."

  • High temperature: "Start your day with a vibrant symphony of golden honey-drizzled granola, paired with tangy berry compote and creamy Greek yogurt!"

TopP (Nucleus Sampling): Controlling Token Selection Range

While temperature affects how the AI chooses tokens, TopP determines which tokens it considers in the first place.

What It Does:

  • Low TopP (0.1 - 0.3): Considers only the most probable tokens

  • Medium TopP (0.4 - 0.7): Considers a moderate range of tokens

  • High TopP (0.8 - 1.0): Considers a wider range of possible tokens

Best Used For:

  • Low: When you need very focused, precise outputs

  • Medium: General purpose content creation

  • High: When you want the AI to consider less common but potentially interesting options

Important Note: TopP and temperature work together. A high temperature with low TopP still limits creativity because the AI has fewer tokens to choose from.


Advanced Controls

Understanding Penalties

Chibi offers two distinct penalties that help control repetition in different ways. Both can range from -2.0 to 2.0, with 0.0 as the default (no effect).

Frequency Penalty: Managing Token Frequency

The frequency penalty looks at how often tokens appear in the input and adjusts their probability in the output accordingly. The more a token appears, the stronger the effect.

How It Works:

  • Positive values (0.1 to 2.0): Discourage the use of tokens based on how frequently they appear in the input. The more a token appears, the less likely it becomes.

  • Zero (0.0): No effect on token frequency

  • Negative values (-2.0 to -0.1): Encourage the reuse of tokens based on their frequency in the input. The more a token appears, the more likely it becomes.

Example: Let's say your input contains the word "technology" several times. With a:

  • Positive frequency penalty: Each appearance of "technology" makes it increasingly less likely to be used again

  • Negative frequency penalty: Each appearance of "technology" makes it increasingly more likely to be used again

  • Zero frequency penalty: The frequency of "technology" in the input doesn't affect its probability in the output

Presence Penalty (Repeat Penalty): Managing Token Presence

The presence penalty looks only at whether a token has appeared in the input at all, regardless of how many times. It applies a flat adjustment based on presence alone.

How It Works:

  • Positive values (0.1 to 2.0): Discourage the reuse of any token that appears in the input, regardless of how often it appears

  • Zero (0.0): No effect on token reuse

  • Negative values (-2.0 to -0.1): Encourage the reuse of tokens that appear in the input

Example: If your input contains the words "artificial" and "intelligence" once each, with a:

  • Positive presence penalty: Both words become less likely to be used again, by the same amount

  • Negative presence penalty: Both words become more likely to be used again, by the same amount

  • Zero presence penalty: The presence of these words doesn't affect their probability in the output

When to Use Penalties

It's best to start without penalties (0.0) and only add them when you observe specific issues:

  1. Use frequency penalty when:

    • You notice certain words or phrases being overused proportional to their appearance in the input

    • You want to encourage (negative values) or discourage (positive values) the AI to build upon patterns in the input

  2. Use presence penalty when:

    • You want to broadly encourage (negative values) or discourage (positive values) the AI from reusing any tokens from the input

    • You need a simpler approach to managing repetition that doesn't scale with frequency

Best Practices for Penalties

  1. Start without penalties (0.0 for both)

  2. Observe the AI's natural output with your prompt and context

  3. If you notice issues with repetition:

    • For scaling repetition issues (words being used more as they appear more), try small adjustments to frequency penalty first (try 0.1 or 0.2)

    • For general repetition issues (any reuse of input tokens), try small adjustments to presence penalty first (try 0.1 or 0.2)

  4. Use negative values cautiously, as they can lead to excessive repetition

Remember: These penalties are advanced tools. Most use cases work well with default values (0.0). Only adjust them when you have a specific repetition issue to solve.


Practical Recommendations

For Different Content Types:

Technical Writing:

  • Temperature: 0.1 - 0.3

  • TopP: 0.4 - 0.7

  • Repeat Penalty: 0.0 - 0.1

  • Frequency Penalty: 0.0 - 0.1

Blog Posts:

  • Temperature: 0.4 - 0.6

  • TopP: 0.6 - 0.9

  • Repeat Penalty: 0.0 - 0.2

  • Frequency Penalty: 0.0 - 0.2

Creative Writing:

  • Temperature: 0.7 - 0.9

  • TopP: 0.7 - 1

  • Repeat Penalty: 0.0 - 0.3

  • Frequency Penalty: 0.0 - 0.3

Common Pitfalls to Avoid

  1. Over-constraining the AI:

    • Using low temperature AND low TopP

    • Setting both penalties too high

    • Result: Stilted, unnatural text

  2. Under-constraining the AI:

    • Using high temperature AND high TopP

    • Setting no penalties

    • Result: Potentially incoherent or rambling text

  3. Mismatched Settings:

    • High temperature with low TopP

    • Very high penalties with low temperature

    • Result: Conflicting goals leading to poor output


Best Practices for Getting Started

  1. Start with moderate settings:

    • Temperature: 0.5

    • TopP: 0.98

    • Repeat Penalty: 0.0

    • Frequency Penalty: 0.0

  2. Adjust one control at a time and observe the changes

  3. Keep notes about which combinations work well for different types of content

  4. Remember that different AI models may respond differently to these controls

Think of these settings like this:

  • Write naturally, using your full vocabulary (TopP: 0.98)

  • But be thoughtful and somewhat consistent in your word choices (Temperature: 0.5)

  • Without worrying about repetition rules (Penalties: 0.0)

For getting started, these settings are great because:

  1. It lets you see the AI's natural capabilities

  2. Makes it easier to understand how temperature affects the output

  3. Produces more consistent and predictable results

  4. Helps you establish a good baseline before experimenting with penalties


Conclusion

Understanding AI controls is crucial for getting the most out of Chibi. Start with the recommended settings and experiment gradually to find what works best for your specific needs. Remember that these controls work together as a system, and finding the right balance is key to achieving your desired results.

The beauty of Chibi is that once you find settings that work well for a particular type of content, you can save them in roles or actions for consistent results. Take time to experiment and find the combinations that work best for your specific use cases.

Was this helpful?