Can AI Actually Think? Meta, OpenAI, and the Rise of Reasoning Models

AI can write poems and pass the bar, but can it think?

While generative AI has dazzled us with fluent text and creative outputs, a new frontier is emerging: reasoning. The latest models from OpenAI and Meta are not just generating responses; they're engaging in internal deliberation, marking a significant shift from pattern recognition to genuine problem-solving.

~ Durnisa Baghirova, Technology Analyst at CIH

OpenAI’s o1: From Language to Logic

In December 2024, OpenAI introduced the o1 model, designed to enhance reasoning capabilities beyond its predecessor, GPT-4o. Unlike earlier models that primarily relied on pattern recognition, o1 employs a "chain-of-thought" approach, allowing it to process information more deeply before responding (OpenAI, 2024).

This advancement is evident in benchmark performances. On the International Mathematics Olympiad (IMO) qualifier, o1 achieved a score of 83%, a substantial improvement over GPT-4o's 13% (OpenAI, 2024). Additionally, o1 ranked in the 89th percentile in Codeforces programming competitions, showcasing its prowess in complex problem-solving tasks (OpenAI, 2024).

Meta’s Thought Preference Optimization: Teaching AI to Reflect

Meta, in collaboration with researchers from UC Berkeley and NYU, has developed Thought Preference Optimization (TPO), a novel training method aimed at enhancing the reasoning abilities of large language models. TPO encourages models to generate internal "thought steps" before producing a final answer, promoting more deliberate and coherent responses (Meta AI, 2024).

Experiments have demonstrated that TPO can significantly improve reasoning performance. For instance, models fine-tuned with TPO showed an 8.6% increase in math reasoning accuracy and a 25.9% increase in output length, indicating more thorough deliberation (arXiv, 2025).

The Shift from Bigger to Smarter Models

The evolution of AI is moving from scaling model sizes to enhancing cognitive capabilities. OpenAI's o1 and Meta's TPO represent a paradigm shift towards models that can reason, reflect, and make informed decisions. This transition is not merely about increasing computational power but about fostering genuine understanding and problem-solving skills in AI systems.

Conclusion: The Dawn of Reflective AI

We're entering an era where AI models don't just generate responses—they think before they speak. This progression towards reflective AI signifies a move closer to human-like intelligence, where machines can analyze, deliberate, and reason. As AI continues to evolve, the question isn't just about what it can do, but how thoughtfully it can do it.