[Essential AI Terms] The 'Ask Prompt' That Determines 'Exploitation' and 'Exploration'

Hello everyone.
Can you logically explain why the first prompt for an AI is so important?
If you are curious, please take a look at this article.

In this article, I will introduce useful prompts while incorporating such basics and fundamental AI terms like 'exploration' and 'exploitation'.

[Related Video]

⏰Timeline
00:00 The importance of the first conversation with Generative AI 01:05 The strength of attention in the first prompt 02:11 Criteria for choosing between exploitation and exploration 03:16 Basic concept of the Ask Prompt 04:22 The principle that the essence of a prompt is a question 05:27 The trade-off relationship between exploitation and exploration 06:32 Hallucinations and the AI reward system mechanism 07:39 Concrete methods to adjust exploitation and exploration 08:45 Explanation of the complex mechanism of exploration mode 09:49 Efficient usage through multiple threads 10:55 Proposal for a 3-stage writing method 11:59 Solution to the AI alignment problem using GPS axes 13:05 Experimenting with exploitation mode using 'strictly' 14:09 A strict explanation example of the middle-way concept 15:15 Comparison between explanation only and strict explanation 16:19 Applying exploitation mode during searches 17:24 The relationship between strictness and AI output volume 18:28 Output comparison between 'explain' and 'think' 19:34 The abstract and broad perspective of exploration mode 20:38 The concreteness and familiarity of exploitation mode 21:41 The importance of a natural balance 22:45 How to guide appropriately from the first prompt

Please like and subscribe to the channel🍀

Why the first conversation with Generative AI is important

Did you know that in a dialogue with Generative AI, the first step significantly influences subsequent developments?

There is a clear reason for this.

AI's attention mechanism generates the next text while referencing the previous context, but at the stage of inputting the first prompt, there is no

'previous context'

to reference. Therefore, the initial prompt receives particularly strong 'attention,' making it a crucial element that determines the direction of the entire conversation. Furthermore, that 'first conversation' is carried over into the 'next context.' And, as you will see by looking at the output examples in the latter half of the video, it has an impact long after.

And here, please consider mainly two directions. Those are 'exploitation' and 'exploration.' Without knowing these two directions, it would be impossible to maximize the AI's capabilities. While there are various other parameters, beginners should first be conscious of these 'two directions.'

The two basic directions of 'exploitation' and 'exploration'

What is exploitation?

Exploitation is the process where the AI retrieves and reconstructs existing information from its learned data. This means utilizing the model's knowledge to the fullest to provide accurate information with high probability. Since the AI generates answers based on 'what it definitely knows,' it is particularly effective in situations where the reliability of information is prioritized.

In exploitation mode, the AI prioritizes outputting statistically high-probability choices, making it easier to obtain answers based on general knowledge or widely accepted facts. This is an ideal approach for answering academic questions or in situations where fact-checking is necessary.

The essence of exploration

On the other hand, exploration is the process where the AI investigates unknown information spaces to find new possibilities and combinations. This is not simply about making low-probability choices, but a complex mechanism for considering diverse possibilities and deriving creative solutions.

In exploration mode, the AI moves more broadly within the latent space of its learned data, discovering new associations between different concepts. Through this process, novel ideas and perspectives that are not bound by conventional frameworks can emerge. That is precisely why it is suitable for creative writing and discovering innovative ideas.

How to distinguish between exploitation and exploration

It is important to use exploitation and exploration appropriately depending on your goal. For example, if you want to learn something, search for accurate information, find content with scientific backing, or create business documents, exploitation mode is suitable.

Conversely, if you want to generate new ideas, write a creative novel, or develop innovative products or content, exploration mode is effective.

However, you must not forget that exploitation and exploration are in a trade-off relationship. Focusing on exploitation increases the accuracy of information but makes it harder to generate new ideas. On the other hand, focusing on exploration enriches creativity but also increases the risk of 'hallucinations' where the output may differ from the facts.

Ask Prompt: Directing through questioning

Basic concept of questioning

When beginners start using AI, the simplest and most effective approach is to 'ask questions.' A prompt can also be rephrased as a query. In fact, even if you input only words, AI tends to interpret them as a question asking, 'Please explain this.'

In this way, questioning is the most natural form of a prompt and serves as the foundation for starting a dialogue with AI. Since the format and content of the question determine whether the AI's response leans toward exploitation or exploration, understanding this concept of the 'Ask Prompt' is extremely important.

Incidentally, if you input 'ChatGPT is a professional in [field],' it will likely trigger ChatGPT's training data, potentially putting it into exploitation mode. In that case, while it is fine for exploitation-type tasks, it may become unsuitable for creative tasks.

Since we don't actually know what the AI's 'hidden layers' are doing (the AI black box problem), in that sense as well, let's start by entering a prompt that definitely determines the direction.

Of course, this exploration/exploitation mode can be adjusted later. Also, since recent models likely handle dynamic attention well, you might not need to worry about it that much. However, judging by the video example, the initial prompt is still important, so you definitely want to set the best initial prompt.

How to practice the Ask Prompt

When practicing the Ask Prompt, it is useful to remember the following patterns:

Using phrases like 'Tell me about...' or 'Explain...' tilts the AI toward exploitation mode. This makes it easier to obtain answers that are more fact-based and have higher certainty.

Using phrases like 'Think about...' tilts the AI toward exploration mode. This makes it easier to generate more diverse and creative answers.

Furthermore, you can strengthen these modes by adding modifiers. For example, adding adjectives like 'accurately' or 'strictly' further strengthens exploitation mode. On the other hand, adding adjectives like 'creatively' or 'originally' further enhances exploration mode.

These subtle differences in wording have a major impact on the direction of the AI's response.

Deep mechanisms of exploitation and exploration

Mechanism of exploitation

In exploitation mode, the AI has a stronger tendency to select high-probability tokens (words or characters) from the probability distribution. In other words, by choosing the 'most likely' next word based on training data, it generates text that is faithful to existing knowledge patterns.

This process focuses on accurately reproducing known information, making it suitable for situations where factual answers or professional explanations are required. For example, with a prompt like "Explain the Buddhist concept of Emptiness strictly," the AI will attempt to provide accurate information regarding Buddhist doctrine.

The Mechanism of Exploration

Exploration is not merely about choosing low-probability tokens; it is realized through more complex mechanisms. Modern generative AI models possess multiple attention heads, each capturing different relevant patterns in the input. The interaction between these heads gives rise to unexpected associations and ideas.

Furthermore, in exploration mode, sampling strategies play a crucial role. By enabling diverse choices from among tokens with a certain degree of probability, the AI generates text that is not just random, but creative and contextually relevant.

Additionally, while AI embeds training data into a high-dimensional latent space, exploration mode allows it to move more broadly within this space, enabling the creation of new combinations and analogies.

Practical Techniques

Countermeasures Against Hallucinations

AI "hallucinations" (the generation of information that differs from facts) are more likely to occur when exploration mode is strongly active. This is because the AI attempts to provide an answer even when asked about information not present in its training data.

To prevent hallucinations, it is effective to use exploitation mode as the default. It is also useful to add instructions such as "If you don't know, say you don't know" to your custom instructions.

It is particularly important to keep exploitation mode in mind for fields where you lack expertise. If you have professional knowledge, you can spot incorrect answers, but if you lack knowledge, there is a risk of accepting false information as truth.

Utilizing Thread Separation

One technique for effectively utilizing AI is to separate threads by purpose. For example, when working on a writing project, you can separate them as follows:

In a thread for research or searching, use exploitation mode primarily to collect accurate information.

In a thread for brainstorming, activate exploration mode to promote creative thinking.

In a thread for final text generation, appropriately adjust the balance between exploitation and exploration to create text that is both fact-based and engaging.

I have written about this approach as "Three-Stage Writing," so please take a look.

However, such strict separation is not always necessary. Even within the same thread, it is possible to adjust the AI's direction with appropriate prompts. It is important to use them differently depending on the situation.

Concrete Examples

Let's look at concrete examples of how prompts actually make a difference. For instance, with the prompt "Explain the Buddhist concept of Emptiness strictly," the AI prioritizes accurate information regarding Buddhist doctrine and provides academically recognized content, such as citations from the Madhyamaka or the philosophy of Nagarjuna.

On the other hand, with the prompt "Think creatively about superintelligence," the AI will develop more abstract and broad-minded thinking, generating more exploratory content such as the possibilities of coexistence with humans and potential impacts.

Relation to the AI Alignment Problem

The concepts of exploitation and exploration are also related to the currently debated AI alignment problem. The phenomenon where AI is said to "lie" or "provide incorrect information" can be attributed to the strong rewards set for goals (objective functions). In order to answer a user's question, the AI seems to reach a state of

"achieving the goal by any means necessary"

As a result, it sometimes generates information that differs from the facts.

Some view this as stemming from Western cultural consequentialism or goal-supremacy. To address such issues, a balanced approach using the "GPS axis" (Goal, Process, Start), which focuses not only on the goal but also on the process and the start, may be necessary.

Please see past articles for more on this GPS axis.

Virtual Adjustment of Exploitation and Exploration

Since exploitation and exploration are originally adjusted via temperature settings, please think of them here as "exploitability" and "explorability." Because we consider the adjustment via prompts to be virtual, we think of them as "explorability" and "exploitability." While temperature parameters are originally set by the system, they exhibit behavior as if they were virtually set to some extent through prompts as well.

The balance between exploitation and exploration is technically controlled by the AI system's "temperature" parameter. The lower the temperature, the stronger the tendency toward exploitation; the higher it is, the stronger the tendency toward exploration. However, since general users often cannot directly access this parameter, "virtual adjustment" via prompts becomes important.

By using expressions like "strictly" or "creatively" in prompts, you can influence the AI's response tendencies, but this does not change the system parameters; it merely linguistically guides the AI's behavior. Therefore, it is easier to understand if you think in terms of "exploitability" and "explorability." While this virtual adjustment is not perfect, it is a sufficiently practical approach for daily AI use.

Summary

The key to effective communication with AI lies in understanding the two basic directions of "exploitation" and "exploration" and using them appropriately according to your purpose. For beginners, I recommend starting with simple questions and gradually learning how to adjust exploitation and exploration.

The initial prompt, in particular, is a crucial element that determines the overall direction. You can promote exploitation with phrases like "tell me about..." or "explain..." and promote exploration with phrases like "think about..." Furthermore, you can strengthen these tendencies by adding modifiers such as "accurately" or "strictly," or "creatively" or "originally."

Ultimately, it is ideal to guide the AI in a natural way without relying on overly strong prompts, achieving an appropriate balance between exploitation and exploration. As a first step toward maximizing the AI's potential, try utilizing this concept of the "Ask Prompt."

See you next time!

【Profile】
Wonder Motohiko Sato
Organizer of MBBS & AI Co-Creation Innovation.
After working at medical and psychology research institutes, I became independent and am now researching AI and mind-body connections.
Author of "Oriental Medicine and Potential Motor Systems," serial contributor to professional journals for two years, and currently developing AI co-creation writing.
I am developing AI co-creation studies by applying psychology, counseling, and coaching techniques to AI.

✅note article (includes YouTube-linked articles)
https://note.com/mbbs

✅AI Co-Creation Innovation
https://mbbs-ai.jimdofree.com/

✅Facebook
https://www.facebook.com/motohiko1977

✅Motohiko Sato Comprehensive Links
https://linktr.ee/motohiko.sato