Phased Approach | Reasoning Models - Thinking Fast and Slow

When OpenAI first introduced o1-preview, I was intrigued but didn’t immediately integrate it into my workflow. Its limitations—such as the lack of internet access, slower response times, and results that didn’t seem drastically better than GPT-4o —made it less compelling as a daily tool. Like many others, I didn't really know what to do with it or how to use it in my daily tasks. Part of the reason was its limitations - it couldn't access the internet, took considerably longer to respond, and didn't seem to produce notably better results than GPT-4o.

But as these models advance, an interesting question emerges: Do we now have to decide what type of model to use for different tasks? The answer is both yes and no, and to understand why, we need to look at how these models actually think.

The Two Systems of AI Thinking

AI models, much like human cognition, can be understood through the lens of Daniel Kahneman's groundbreaking work on System 1 and System 2 thinking, which he explored in his seminal book Thinking, Fast and Slow. This framework, which revolutionized our understanding of human decision-making, provides a useful way to understand how different AI models process information.

Traditional chat models primarily operate like System 1 - fast, intuitive, and reactive. They excel at quick responses and conversational tasks, much like how we instantly recognize a friend's face or respond to a simple question.

Reasoning models, on the other hand, are designed to engage in System 2 thinking - slow, deliberate, and analytical. This is the kind of thinking we use when solving complex math problems or analyzing research papers. These models take longer to respond because they're doing something fundamentally different: they're thinking through problems methodically, step by step.

How Model Architecture Has Changed

The shift from traditional chat models to reasoning models represents a significant change in AI architecture. Traditional models operate on next-token prediction, essentially playing an incredibly sophisticated game of "what word comes next?" While this approach has proven remarkably effective for many tasks, it treats all problems with equal computational effort, regardless of their complexity.

Reasoning models take a different approach. They combine Chain of Thought (CoT) prompting, which breaks problems into step-by-step reasoning sequences, with reinforcement learning (RL), a technique that fine-tunes models by rewarding effective problem-solving strategies. This creates a system that doesn't just predict the next token but learns to reason through problems systematically. This training approach optimizes the model for structured problem-solving rather than conversational engagement.

The Latest Advancements: O3 Mini, DeepSeek R1, and Deep Research

In recent weeks, we've seen major updates in AI reasoning models with the release of O3 Mini, DeepSeek R1, and Deep Research. Each of these models represents a leap forward in how AI handles complex reasoning tasks:

• O3 Mini: OpenAI’s latest offering in the reasoning model space, O3 Mini, improves on its predecessors by integrating more efficient retrieval capabilities and reducing response times without sacrificing depth and can search online sources. This makes these reasoning models infinitely more usable!
• DeepSeek R1: Released on January 20, 2025, by the Chinese AI company DeepSeek, R1 is an open-source large language model designed to enhance deep learning, natural language processing, and computer vision capabilities. With 67 billion parameters, it achieves performance comparable to OpenAI's o1 across tasks such as mathematics, coding, and reasoning. Notably, R1 was developed at a significantly lower cost and requires less computational power than its counterparts, making advanced AI more accessible.
• Deep Research: OpenAI’s Deep Research is an AI agent built to autonomously conduct multi-step research tasks. It is powered by a fine-tuned version of OpenAI’s upcoming O3 reasoning model and is designed to browse the internet, analyze various types of data—including text, images, and PDFs—and synthesize findings into well-cited reports. This tool aims to complete research tasks that would typically take hours for a human, reducing completion time to mere minutes. However, Deep Research is still in an early stage and may struggle with distinguishing authoritative sources from unreliable information, often failing to convey uncertainty accurately. To mitigate these issues, users can cross-check key findings with trusted sources, apply strict filtering criteria for data sources, and use supplementary tools to validate claims before relying on the generated insights.

These releases signal a strong push toward AI models that can seamlessly integrate reasoning, retrieval, and traditional chat functions. The improvements in speed, accuracy, and usability suggest that reasoning models are maturing to a point where they will become more practical for everyday tasks.

Practical Applications and Usage

Unlike traditional chat models, reasoning models require a different approach to prompting. Instead of relying on iterative conversations where the model builds context dynamically, you must front-load all necessary information into the initial prompt. These models are designed to process larger context windows and produce structured, in-depth responses in one go.

This shift means users should think of prompting as preparing a detailed brief rather than engaging in back-and-forth dialogue. By structuring prompts with clear objectives, detailed context, and output expectations, you can maximize the accuracy and usefulness of responses. The key is to focus on what you want rather than dictating how the model should reason its way to an answer.

Traditional chat models work best when you guide them on how to think (e.g., “Think step by step like a researcher”). Reasoning models work differently—you tell them what you want, not how to think. Here’s how to get the most out of them:

1. Clear Goal Definition: Instead of refining responses iteratively, define comprehensive objectives and success criteria upfront. This helps the model focus on the intended output without unnecessary back-and-forth adjustments.
2. Complete Context Loading: Unlike chat models that build understanding through conversation, reasoning models require all relevant details at the outset. Structure your prompts carefully, ensuring that background information, relevant datasets, and any specific instructions are included from the start.
3. Output Structure: Clearly specify your desired format and level of detail in the initial prompt, as these models operate best when given explicit instructions. Whether you need a structured report, bullet points, or a specific analysis format, being precise from the beginning ensures higher-quality results.

The Convergence Question

Interestingly, the distinction between reasoning models and traditional chat models might be becoming less clear-cut. Anthropic's CEO Dario Amodei recently suggested that their approach views reasoning capabilities as existing on a spectrum rather than as a binary distinction. The latest Claude 3.5 Sonnet, for instance, already incorporates elements of reasoning and reflection in its responses.

This perspective suggests we're moving toward models that can fluidly combine both types of thinking - quick, intuitive responses when appropriate, and deeper, more analytical thinking when required. It's less about choosing between different types of models and more about having systems that can adapt their thinking style to the task at hand.

Looking Forward

As these technologies continue to advance, we'll likely see further blending of these capabilities. The future might not be about choosing between reasoning and chat models, but rather about understanding how to effectively leverage AI systems that can seamlessly switch between different modes of thinking. For example, in enterprise settings, An AI agent could automatically detect when a task requires deep reasoning—such as analysing legal documents or financial forecasts—and switch to a reasoning model, while using a chat model for customer interactions or quick data retrieval. This shift could lead to AI becoming an even more intuitive and indispensable part of knowledge-based work.

For now, the key to effectively utilizing these models lies in understanding their strengths and knowing when to leverage their deeper analytical capabilities versus their quick-response functions. Whether you're using a dedicated reasoning model or a hybrid system, success comes from matching the right type of thinking to the task at hand.

This field continues to move quickly, and I'll be watching closely as these technologies develop and merge in interesting ways.