Unpacking OpenAI’s New “o1” Model (aka Strawberry) and Chain-of-Thought (CoT)

Recently, OpenAI introduced a new model dubbed “o1,” and it is making waves. Aimed at addressing complex logical reasoning and multi-step problems, the “Strawberry” model seems to utilize a technique called chain-of-thought (CoT). This technique breaks down the reasoning process into smaller, more manageable steps, enhancing the model’s ability to generate accurate responses.

While exciting for general purpose AI applications, this update also highlights the need for custom AI tools tailored to specific business needs. Broad-audience models often lack the precision and effectiveness required for specialized tasks. Plus, the “black box” nature of the o1 model, in which the internal workings remain undisclosed, prevents users from understanding how the model reaches conclusions. Domain-specific solutions, on the other hand, offer the transparency, flexibility, and precision needed to meet users’ goals in the most effective way possible.

What exactly does CoT mean?

It’s like when you’re helping your kid solve a math problem. Instead of having her jump straight to the final answer after reading the problem for the first time, you tell her to take it step-by-step. This step-by-step approach will probably result in a better answer.

A detailed example showing the CoT technique in action:

At Magid, we’ve had the CoT methodology integrated into our AI tools for quite some time. For example, it powers our Content Analysis feature, which is meticulously engineered to evaluate journalistic content for balance, fairness, and writing quality. 

Instead of a “black box,” our CoT implementation is highly transparent to our clients. It’s also designed for specific use cases, rather than in a generic, “one-size-fits-all” manner.

Below we demystify some of the misconceptions there are around CoT and outline reasons for why we think transparency and built-for-purpose solutions are critical in CoT applications.

CoT leads to more effective AI results, but should be engineered for specific purposes.

Before we go further, let’s address a common misconception around AI tools being able to “think.” They can’t. Here’s a helpful quote from Clement Delangue, Co-founder & CEO of HuggingFace, the open and collaborative platform for AI builders:

“An AI system is not ‘thinking,’ it’s ‘processing,’ ‘running predictions’… just like Google or computers do.

Giving the false impression that technology systems are human is just cheap snake oil and marketing to fool you into thinking it’s more clever than it is.”

That noted, CoT improves the frequency of correct answers by methodically breaking down inquiries into a series of steps before arriving at a final answer. But this doesn’t come for free. There are trade-offs as a result of the improved outcomes, namely in:

• Latency (time to get an answer)
• Computational costs (costs of running the model)

At Magid, we apply CoT methodology to target specific tasks. This targeted approach mitigates the trade-offs concerning latency and computational costs. 

A CoT example: Magid’s Content Analysis tool

Our Content Analysis tool uses CoT to dissect journalistic content, evaluating it across various dimensions to determine whether it upholds the highest journalistic standards. 

Our Journalism Advisory Board, comprised of industry experts and research professionals, continuously outline the key factors that constitute quality journalism. Then, through CoT and other techniques, Collaborator AI evaluates each piece of client content based on those standards, providing immediate feedback and recommendations for improvement.

CoT is most effective with transparency and control.

Transparency and control in CoT lead to:

• Improved user understanding and output effectiveness. Transparency allows users and developers to understand the AI’s “decision-making” process. Users can easily collaborate with developers to tweak the system to perform better for their specific tasks. This is particularly important in fields, such as journalism, where precision and reliability are paramount.
  
• Trust. We’ve also found that transparency is critical for building user trust. If people can understand how decisions are made, they are more likely to feel reassured by the process and confident about the results.

CoT is powerful, but must be used for the right reasons.

The CoT technique, undeniably, is a powerful tool in enhancing AI’s problem-solving capabilities. OpenAI’s “Strawberry” model is a noteworthy advancement, especially for general public applications. Yet, for specific, mission-critical tasks, having transparency and control is essential. 

At Magid, our AI solutions reflect this philosophy, providing tailored, well-understood, and efficient applications. Embracing AI’s future means balancing innovation with responsible, use-case-driven implementations—and that’s our priority with the Collaborator AI tools.

—

By unpacking these nuances, we aim to keep our clients informed and empowered about AI, knowledgeable about landscape updates and best practices, while also enjoying the transformative benefits.