Working with AI: Embracing Agency and Experience

Janpha Thadphoothon

In the realm of collaborating with AI agents, it's imperative to steer clear of relying solely on zero-shot prompts. Rather, our approach should prioritize engaging with AI agents through purposeful and refined steps, fostering a sense of agency and drawing from experiential knowledge.

Image: DALLE-2

I'm talking, of course, about working with AI agents. The context of my essay here is ELT, or English language learning and teaching, which explicitly states that it is not just about teachers but learners of English as well. I must also note here that to collaborate with AI agents (LLMs), you may need to have some level of proficiency in the language and understand how the system operates. At the very least, your reading and comprehension skills should be strong.

Avoid Zero-Shot Prompts

Recently, I watched a YouTube video featuring a speaker engaging with an AI agent to gain insights into the future. In the video, the speaker initiated the interaction with a simple prompt: "Tell me about the future." However, this zero-shot prompt lacked specificity and context, leaving the AI agent without clear guidance on what aspects of the future to focus on or what type of information was desired.

One lesson I have learned and would like to share with you all is this - Avoid Zero-Shot Prompts. What do I mean by that? Zero-shot prompts involve providing minimal or no context to AI models and expecting them to generate relevant outputs. While this approach might seem efficient, it often leads to subpar results or misunderstandings due to the lack of context. By avoiding zero-shot prompts, we ensure that AI agents have the necessary information and context to deliver meaningful responses.

Here are a few examples and explanations of avoiding zero-shot prompts:

Example: Asking an AI language model, "Write a poem."

This is a zero-shot prompt because it provides no specific context or guidance for the AI model. Without any additional information about the theme, style, or mood desired for the poem, the AI model may produce a poem that doesn't meet the user's expectations or lacks coherence.

Example: "Give me advice."

Similar to the previous example, this prompt lacks specificity and context. Without additional details about the topic or issue for which advice is sought, the AI chatbot may provide generic or irrelevant advice that doesn't address the user's actual concerns.

Even in translation, an AI agent needs you to provide adequate context. Asking a language translation AI to translate a phrase without specifying the source or target language - should be avoided. AI can't read your mind. If it could, we would have to think harder about dealing with machines.

One may ask what's wrong with that. Without indicating the source language of the phrase to be translated and the target language into which it should be translated, the AI may struggle to accurately interpret and produce the desired translation. This lack of context can result in mistranslations or ambiguous outputs.

In each of these examples, avoiding zero-shot prompts involves providing additional context, guidance, or specificity to the AI model to ensure that it can generate relevant and meaningful outputs. This approach enhances the effectiveness and accuracy of interactions with AI agents, leading to better outcomes for users.

Embracing Agency

By taking an agentic approach, we actively participate in guiding the interaction with AI agents. This involves providing clear instructions, refining inputs, and actively steering the conversation towards desired outcomes. Through an agency, we maintain control over the interaction and ensure that AI agents align with our goals and intentions.

Embracing agency refers to actively taking control and responsibility for one's actions, decisions, and interactions. In the context of working with AI agents or any technology, it involves asserting one's authority to guide and shape the course of the interaction. Instead of passively accepting outcomes or relying solely on the capabilities of the AI system, embracing agency empowers individuals to actively participate in the process, providing input, direction, and feedback to ensure that the interaction aligns with their goals and intentions.

Embracing agency encompasses several key aspects:

Active Participation: Rather than being passive recipients of information or outputs from AI agents, individuals actively engage in the interaction by providing input, asking questions, and making decisions.

Taking Control: Agency involves asserting control over the interaction, steering it in a direction that serves one's objectives. This may include setting goals, defining parameters, and guiding the flow of communication.

Responsibility and Accountability: Embracing agency entails taking responsibility for the outcomes of the interaction. Individuals are accountable for their decisions and actions, as well as the impact they have on the overall outcome.

Alignment with Goals: Agency ensures that the interaction with AI agents is purposeful and directed towards achieving specific goals or objectives. Individuals prioritize actions and decisions that support their overarching intentions.

Feedback and Adaptation: Embracing agency involves providing feedback to the AI system and making adjustments based on the responses received. This iterative process allows individuals to refine their approach and optimize the interaction for better outcomes.

In essence, embracing agency in the context of working with AI agents empowers individuals to actively shape their experiences and outcomes, ensuring that technology serves as a tool to support their goals and aspirations rather than dictating the course of action.

Learning from Experiences

Learning from experience involves leveraging past experiences and knowledge to enrich interactions with AI agents. Rather than relying solely on theoretical concepts or abstract prompts, prioritizing experience allows individuals to draw upon real-world insights and nuances, thereby enhancing the quality and relevance of the dialogue with AI agents. This approach is particularly valuable in research contexts, where a deep understanding of the subject matter is essential for meaningful engagement with AI technologies.

We must not forget that, like humans, AI agents too are capable of learning from experiences. This highlights the profound capability of AI agents to learn and enhance their performance through experiences and training, akin to human learning processes. AI agents accumulate knowledge and insights over time by engaging with diverse interactions, tasks, and datasets, allowing them to recognize patterns, make predictions, and generate more accurate outputs based on their past encounters. Through iterative training cycles, AI agents adjust their internal parameters and algorithms, progressively refining their performance and capabilities. Incorporating feedback mechanisms, they learn from both successes and failures, adapting their behavior to optimize outcomes. This continuous learning process enables AI agents to adapt to changing conditions, refine their decision-making processes, and achieve tasks with greater efficiency and precision. Moreover, they can generalize from specific instances to novel situations, applying past knowledge to new contexts. Some AI agents are even capable of autonomous learning, autonomously seeking out new information and experiences to expand their knowledge base without direct human intervention. Overall, the capacity of AI agents to learn autonomously from experiences and training is fundamental to their versatility and effectiveness across various domains and applications.

Be Experimental

Being experimental when working with AI agents involves adopting an exploratory mindset and actively testing different approaches, prompts, and sequences to optimize interactions and outcomes. Instead of adhering to a fixed set of procedures or relying on conventional methods, being experimental allows individuals to innovate, iterate, and discover new strategies for engaging with AI technology. Here's an explanation of this concept:

1. Exploratory Mindset: Being experimental requires approaching interactions with AI agents as opportunities for exploration and discovery. Rather than adhering to preconceived notions or rigid frameworks, individuals remain open to experimentation, curiosity, and the possibility of unexpected insights or outcomes.

2. Testing Various Prompts and Sequences: Experimentation involves systematically testing different prompts and sequences to gauge their effectiveness in eliciting desired responses from AI agents. This may involve varying the wording, context, tone, or structure of prompts to observe how AI models interpret and respond to different stimuli.

3. Exploring Different AI Models: Being experimental also entails exploring a diverse range of AI models and platforms to assess their capabilities and suitability for specific tasks or objectives. Individuals may experiment with different AI architectures, algorithms, or providers to identify the most effective solutions for their needs.

4. Integration of Multiple Models: Experimentation can extend to integrating multiple AI models or systems to collaborate on complex tasks or workflows. This may involve orchestrating interactions between different AI agents to leverage their respective strengths and capabilities, leading to synergistic outcomes that surpass the capabilities of any single model.

5. Iterative Learning Process: Being experimental is inherently iterative, involving a continuous cycle of testing, observation, and refinement. Individuals learn from their experiments, adapt their strategies based on insights gained, and iteratively improve the effectiveness of their interactions with AI agents over time.

6. Risk-Taking and Innovation: Experimentation requires a willingness to take risks and challenge conventional approaches in pursuit of innovation and improvement. Individuals embrace uncertainty and ambiguity, recognizing that experimentation is essential for pushing the boundaries of what is possible with AI technology.

It must be noted here that being experimental when working with AI agents involves adopting an exploratory mindset, testing various approaches, exploring different AI models, and embracing a continuous cycle of learning and innovation. By experimenting with prompts, sequences, and models, individuals can optimize their interactions with AI technology, unlock new capabilities, and drive advancements in their respective fields.

Sensible and Superior Steps

When we interact with AI agents should take sensible and superior steps, meaning that each action or input should be purposeful and contribute to the overall goal. This approach ensures efficiency and effectiveness in communication, guiding the interaction towards successful outcomes while avoiding unnecessary detours or errors.

In essence, effective collaboration with AI agents necessitates a departure from zero-shot prompts towards an approach that values agency and draws from experiential knowledge. By embracing these principles and taking sensible steps, we can maximize the utility and effectiveness of AI technologies in various domains.

 Unleashing the Power of Agents: Insights from Andrew Ng

Andrew Ng, the renowned computer scientist, co-founder of Google Brain, and founder of Coursera, recently delivered a compelling talk on the transformative potential of agents in AI. Ng expressed his strong optimism about the capabilities of agents, citing examples like GPT 3.5 and emphasizing their ability to reason and iterate, potentially even surpassing the capabilities of future models like GPT-4.

Source: https://images.app.goo.gl/cd9S9Nk2ywyuAaWq6

In his talk at Sequoia, a prominent venture capital firm, Ng highlighted the significance of an "agentic workflow" in contrast to the traditional non-agentic approach. He illustrated how agents, equipped with diverse roles, backgrounds, and tools, can collaboratively tackle tasks through iterative processes, leading to significantly improved outcomes compared to non-agentic methods.

Ng provided empirical evidence supporting the superiority of agentic workflows, particularly in coding tasks. He demonstrated how wrapping GPT 3.5 in an agentic framework outperformed GPT-4 and even achieved near-perfect accuracy when combined with reflection prompts. This highlights the immense potential of agents in enhancing productivity and problem-solving across various domains.

The talk also delved into specific design patterns for implementing agentic reasoning, including reflection, tool use, planning, and multi-agent collaboration. Ng emphasized the practical implications of these patterns, envisioning a future where AI systems autonomously navigate complex tasks with minimal human intervention, potentially revolutionizing industries.

Finally, Ng addressed the paradigm shift required in our expectations of AI response times. As agents become more sophisticated, users may need to adjust to longer wait times for AI-generated responses. However, advancements in technologies like hyper-inference speed, exemplified by projects like Grok, could mitigate this issue, enabling lightning-fast interactions between agents.

Ng's insights underscore the transformative potential of agents in AI applications. By embracing agentic workflows and leveraging emerging design patterns, organizations can harness the power of AI to achieve unprecedented levels of productivity and innovation.

As we dive into the world of AI agents, both individuals and businesses must get ready for the best possible results. We need to develop our skills, abilities, and mindset when working with AI. Avoid zero-shot prompts. Andrew Ng's insights shed light on how these agents can totally change the game, emphasizing the importance of working together with them to solve problems. To make the most out of AI agents, it's key to understand how they work and what they can do. This means being ready to adjust our expectations, like being patient when waiting for responses, and keeping up with new tech developments. With some preparation and flexibility, we can use AI agents to boost productivity and creativity across different fields.

Janpha Thadphoothon is an assistant professor of ELT at the International College, Dhurakij Pundit University in Bangkok, Thailand. Janpha Thadphoothon also holds a certificate of Generative AI with Large Language Models issued by DeepLearning.AI.