Tracing the Chain of Thought: The Future of Agentic Tasks with OpenAI's O1 and Insights from the National AI Prompt Design Challenge

By Kevin Shepherdson, Founder & CEO, Straits Interactive

The National AI Prompt Design Challenge 2024 that we spearheaded earlier this year was a resounding success, with 103 teams from various sectors in Singapore participating. The central objective of this competitive event was to equip participants with essential skills in prompt design and responsible AI development, while recognising impactful AI applications and tools that address various business challenges.

During the National AI Prompt Design Bootcamp, we introduced the Chain of Thought (CoT) prompting method to the participating teams. This powerful technique enables the creation of AI tools and chatbots by guiding the model through a structured, step-by-step reasoning process. During the Challenge, we introduced our Generative AI platform, Capabara, which leverages CoT in its system prompts to govern the behaviour of our AI apps, ensuring logical and coherent outputs.

I also emphasised to the participants that CoT prompting would form the basis of agents or agentic tasks, enhancing decision-making, transparency, and efficiency in AI systems.

Little did I expect that CoT would be featured so prominently in OpenAI's new model series, validating our focus and efforts in teaching this technique.

The Significance of OpenAI's O1 Model

What is OpenAI's O1?

Launched in September 2024, OpenAI's O1 model series is a groundbreaking advancement in AI reasoning and problem-solving capabilities. These models are initially intended to be preview models, designed to provide users - as well as OpenAI - with a different type of LLM experience than the GPT-4o model. The O1 models are trained via reinforcement learning to use Chain of Thought (CoT) reasoning without requiring extra prompting, significantly improving their performance in complex tasks.

Key Features of O1

1. Advanced Reasoning Skills: O1 models excel in solving complex problems, especially in STEM fields like mathematics, coding, and science.

2. Chain of Thought Integration: These models are designed to think out loud, producing a long internal chain of thought before responding to a user query.

3. Performance in STEM Fields: O1 models score in the 89th percentile in competitive programming tasks and exceed Ph.D.-level accuracy on physics, biology, and chemistry problems.

4. Improved Safety Features: In addition, they are harder to jailbreak and come with enhanced safety mechanisms to ensure adherence to safety and alignment guidelines.

How the Chain of Thought is Used in O1

In O1 models, CoT prompting guides the model through a structured process of reasoning. The prompts are designed to explicitly instruct the model to approach problems using step-by-step reasoning, often starting with phrases like "Let's think step by step." This approach ensures that the model breaks down complex problems into manageable steps, allowing it to produce more logical and coherent outputs.

Agents and Their Role

Agents in AI are systems that can autonomously perform tasks on behalf of users. These tasks often require decision-making, problem-solving, and the ability to act based on the environment and user inputs. Agents can be used in various applications, from customer service chatbots to complex data analysis tools. In 2025, we will see agents playing an even bigger role in AI applications, leveraging advanced techniques like CoT to enhance their effectiveness.

Implications in Agentic Tasks

The integration of CoT prompting in O1 models has several implications for agentic tasks:

1. Enhanced Decision-Making: By thinking through each step logically, O1 models can make more informed decisions, especially in complex or uncertain situations.

2. Improved Transparency: The ability to see the chain of thought behind a model's decision can enhance transparency and trust in AI systems.

3. Reduced Errors: The step-by-step reasoning approach reduces errors by allowing the model to correct its mistakes at each stage of the problem-solving process.

4. Increased Efficiency: By breaking down tasks into manageable steps, O1 models can handle complex problems more efficiently, reducing the need for repeated iterations.

Reflecting on the National AI Prompt Design Challenge

Reflecting on the success of the National AI Prompt Design Challenge, it is clear that our focus on CoT prompting has positioned participants at the forefront of AI innovation. The skills and knowledge gained through this challenge will undoubtedly empower them to create more advanced and responsible AI applications, contributing to the ongoing evolution of AI technology.

As we move forward, the integration of CoT prompting in AI development will continue to play a pivotal role in advancing the field. The National AI Prompt Design Challenge has helped drive this transformation forth, and I am excited to see how our participants and those who attend our Prompt Engineering courses will leverage these techniques to drive future innovations in AI.

With OpenAI's O1 models validating the effectiveness of the CoT technique, participants of the Challenge are well-equipped to lead the way in creating impactful AI solutions. The Capabara platform has shown that anyone with no programming skills can apply their own domain expertise and create their own chatbots and AI tools. With the importance of CoT as part of our Capabara tools, I am excited about the future roadmap to incorporate agentic tasks enabled by CoT. As we continue to explore the potential of CoT prompting, we can look forward to a new era of advanced, transparent, and efficient AI systems.