Open AI’s latest ‘Strawberry’ will actually think before it acts!

If we had one bone to pick with artificial intelligence (AI), it has to be the fact that the models sometimes sound like that friend of yours with verbal diarrhoea. They tend to trip over themselves to say something, nevermind processing things or fact checking. But what if we had access to AI that took more time to ‘think’ before responding to us?

In a much-anticipated release, OpenAI will launch its latest artificial intelligence model, o1, in 2 weeks, marking the debut of a new class of reasoning AI models. After months of speculation and whispers about "Strawberry," the official release of o1 represents OpenAI’s bold step toward crafting models capable of reasoning through complex problems faster than humans. Alongside o1, the company is also introducing o1-mini, a lighter, more cost-effective version of the new model, ideal for tasks like coding and problem-solving.

Why the excitement around o1?

OpenAI has positioned o1 as more than just an upgrade — it’s a leap forward. The model excels in solving intricate, multi step problems, such as math and coding challenges, by mimicking human-like reasoning and explaining its thought process along the way. This shift toward reasoning-based AI marks a significant evolution in how these models are trained and used.

o1 promises improved accuracy, with a dramatic reduction in hallucinations — when AI models generate false or misleading information — compared to previous models. Though not perfect, OpenAI’s Research Lead, Jerry Tworek, confirms that this new model hallucinates less frequently than its predecessors.

How is o1 different from previous models?

The training of o1 differs significantly from its predecessors, like GPT-4o. While earlier models were trained to mimic patterns, o1 uses reinforcement learning, a technique where the system learns through rewards and penalties, refining its problem-solving ability with each iteration. Additionally, o1 employs a "chain of thought" process, mimicking human cognition by breaking down problems into logical, sequential steps.

Bob McGrew, OpenAI’s Chief Research Officer, highlights that o1 outshines its predecessors in math-related tasks. In tests against the International Mathematics Olympiad, GPT-4o could only solve 13% of problems correctly, while o1 successfully tackled an impressive 83%. This makes it a powerful tool for scientific research, especially in fields like physics, chemistry, and engineering, where rigorous reasoning and complex problem-solving are essential.

What does this mean for users?

Starting today, ChatGPT Plus and Team users gain immediate access to o1-preview and o1-mini. Enterprise and Edu users will have access next week, and OpenAI has plans to make o1-mini available to free-tier ChatGPT users, though no release date has been set. However, for developers eager to integrate this advanced reasoning model into their applications, the cost will be steep—o1-preview costs $15 per million input tokens and $60 per million output tokens, making it more than triple the price of GPT-4o.

The future of AI reasoning: challenges and opportunities

While o1’s reasoning capabilities are a huge leap forward, there are still hurdles to overcome. The model is more expensive and slower to use than its predecessors, and it is not yet optimised for web browsing or processing files and images. Despite these limitations, OpenAI believes o1 represents a foundational step toward AI models that can perform complex tasks autonomously, inching closer to human-like intelligence.

This release also raises hopes for future advancements in fields beyond tech. With its ability to tackle advanced benchmark tests in physics and chemistry, o1 could pave the way for breakthroughs in medicine, engineering, and beyond. As McGrew aptly put it, “Reasoning is the critical breakthrough that could unlock unprecedented capabilities in AI.”

A step toward autonomous agents?

For OpenAI, the development of o1 is more than just improving existing AI models — it’s about moving toward a future where AI can act as autonomous agents. These agents would be capable of making decisions, taking actions on behalf of users, and solving real-world problems. This vision of AI goes beyond simple pattern recognition and into the realm of reasoning and decision-making, which could revolutionise industries from healthcare to engineering.

However, the journey toward autonomous decision-making agents is still in its early stages. o1 may not yet be fast or affordable enough for widespread use, but it represents a significant milestone in AI research. McGrew noted that solving complex reasoning tasks is essential to progress toward human-like intelligence in AI systems.

The release of o1 marks a transformative moment in AI development. As OpenAI pushes the boundaries of what reasoning models can achieve, the implications for industries relying on complex problem-solving are immense. From coding competitions to scientific research, the capabilities of o1 demonstrate a powerful shift toward AI that thinks — almost like a human. As OpenAI continues to refine and expand its models, we can expect o1 and its successors to play an ever-growing role in shaping the future of technology, science, and industry.

As we look toward a future where AI reasoning models become more commonplace, the release of o1 sets the stage for exciting advancements in both practical applications and AI research. The era of reasoning AI is upon us — and it’s only just beginning.