From Chatbots to Thinkbots: The Rise of Reasoning AI

AI is no longer just predicting words—it’s learning to think. The shift from pattern-based AI to reasoning AI is happening now. Unlike past large language models that primarily rely on statistical correlations (pattern recognition), reasoning AI analyzes, plans, and solves problems step by step. The transformation to reasoning AI models started with OpenAI’s o1 model (Sept 2024) and continued with its o3 model (Dec 2024). This article aims to help you understand reasoning AI models and the use cases it promises. I'll also use a real world example to show you how the current reasoning AI models performs.

Pattern-based AI Models

These AI models don’t “think” before answering. They work by recognizing patterns in data and making predictions. These models can summarize text, generate responses, and complete sentences, but they struggle with deep problem-solving.

Examples of Pattern-based AI:

1. OpenAI's GPT-3 (2020): Developed by OpenAI, GPT-3 is a language model that generates human-like text based on input prompts. It excels at tasks like text generation and summarization but lacks deep reasoning capabilities.
2. OpenAI's GPT-4 (2023): An advancement over GPT-3, GPT-4 offers improved language understanding and generation. However, it still primarily relies on pattern recognition and does not possess true reasoning abilities.
3. Google Gemini 1.0 (2023): Released by Google, Gemini 1.0 focuses on language tasks and exhibits enhanced performance in text processing. Despite these improvements, it does not engage in step-by-step logical reasoning.
4. Anthropic's Claude 1, 2 (2023): Developed by Anthropic, the early Claude models are designed for natural language understanding and generation. While effective in producing coherent text and analysis, they lack structured reasoning capabilities.

Reasoning AI Models

Reasoning AI models are more advanced. They think through a problem before responding. They use logic, break tasks into steps, and make smarter decisions. These models are used in areas like coding, science, and research.

Examples of Reasoning AI Models

1. OpenAI's o1 and o3 Models (2024):

These models plan responses before giving answers. They think through math problems, logic puzzles, and coding challenges. OpenAI's o1 introduced structured reasoning with chain-of-thought planning in September 2024. OpenAI’s o3 model, announced in December 2024, improved reasoning depth, accuracy, and efficiency on top of o1.

You can access o1 through ChatGPT Plus and o1-pro through ChatGPT Pro. The state-of-the-art reasoning model, o3, is not yet available for public use.

After you enter a prompt, o1 will think for a while before responding. See example below. o1 Pro will think even longer (could be a few mins).

2. DeepSeek-V3

DeepSeek is a Chinese AI start-up. Their latest model, DeepSeek-V3, is designed for reasoning, coding, and problem-solving. DeepSeek-V3 is open-source, meaning developers can use it freely. It was trained in less than 2 months at a relatively low cost of $5.57 million. You can try DeepSeek-V3 on their website (link). When you open DeepSeek chat window, to access the reasoning model, you need to click "DeepThink" icon below.

3. Google's Gemini 2.0 Flash Thinking (2024):

Gemini 2.0 Flash Thinking Mode is an experimental model that's trained to generate the "thinking process" the model goes through as part of its response. You can access it via Google AI Studio.

Theoretical Use Cases

Reasoning AI models like OpenAI's o1 Pro promise to solve complex problems, including assisting scientific discovery and tackle the hardest problems humanity faces.

?? Use Case 1: AI-Assisted Scientific Discovery

?? What if AI could think like a scientist—analyzing research, generating new hypotheses, and designing experiments faster than any human ever could?

A biotech company is developing a new cancer drug. Scientists must analyze thousands of research papers, clinical trials, and molecular interactions to identify promising compounds. Traditional AI can summarize papers, but it can’t reason through experiments or predict new findings.

How Reasoning AI Can Help:

• Synthesizes Research Papers – AI can scan thousands of studies, identifying key patterns and contradictions.
• Generates New Hypotheses – It can propose potential drug interactions, reasoning through biological pathways.
• Designs Better Experiments – AI can suggest optimized test conditions, reducing lab costs and speeding up trials.
• Finds Hidden Anomalies – It can detect overlooked genetic links that human researchers might miss.

?? Use Case 2: AI-Led Discovery of New Physics & Materials

?? What if AI could discover entirely new materials—ones that humans might never think to test?

A research lab is searching for next-generation superconductors—materials that could revolutionize energy, computing, and transportation. But there are billions of possible atomic structures, and testing each one experimentally is impossible. Today’s AI can assist with simulations, but it can’t reason through scientific unknowns like a human physicist.

How Reasoning AI can Help:

• Analyzes Scientific Data & Competing Theories – AI can compare studies, identifying gaps in existing knowledge.
• Proposes Novel Material Structures – AI can predict never-before-seen atomic arrangements that might work.
• Simulates Experiments – It can test new materials in virtual quantum simulations, predicting real-world properties.
• Refines Theories in Real Time – As new data comes in, AI can adjust its hypotheses dynamically, like a scientist would.

Real World Complex Math Test

Most of us don’t have cutting-edge scientific problems for AI to solve — but can it handle real-world math? Here’s an example challenge for you to try:

Scenario: You invest $100 every Tuesday in S&P 500 index fund, Vanguard S&P 500 ETF (VOO) starting from January 6, 2020. Using actual VOO price data from the past five years (available for download here), ask AI to calculate how much your investment is worth today.

Prompt: If I invested $100 every Tuesday in the S&P 500 starting five years ago on January 6, 2020, how much would my investment be worth today? Below is Vanguard S&P 500 ETF (VOO) daily historical data for the past 5 years. Use this data to calculate the exact amount—no estimates.

I tested this prompt across different AI models: 1) pattern-based models like GPT-4o, Claude 3.5; 2) reasoning models like OpenAI's o1, o1 Pro, DeepSeek-V3, and Gemini 2.0 Flash Thinking. Here are the results.

• GPT-4o: After the prompt, GPT-4o ran the Python code and give me the answer: the portfolio would now be worth $37,412.30.
• Claude 3.5: Claude couldn't do it. It ran into prompt-length issues.
• o1 and o1 Pro: o1 and o1 Pro didn't give me an exact answer. It says it's around 40K.
• DeepSeek-V3: DeepSeek-V3 can't do this question. It ran into prompt-length issues when the large volumes of data were fed in.
• Gemini 2.0 Flash Thinking: same as DeepSeek-V3.

Key Takeaway:

For everyday tasks such as text summarization or generating quick responses, pattern-based models continue to perform exceptionally well. Meanwhile, reasoning AI will become a powerful solution for tackling multi-step problems, although current limitations like prompt length constraints can pose challenges in real-world scenarios. As these models evolve, we can expect significant improvements, unlocking their potential for even more complex and sophisticated applications.

==

I write about cutting-edge tech like AI at FutureTech Yanyan and sometimes about self-development. Since starting in March 2022, the newsletter has grown to over 1,500 subscribers. Thank you all! ???I hope to keep delivering content you enjoy. If you know someone who would benefit, please share!