OpenAI Unveils ‘Strawberry’: A Leap Forward in AI’s Problem-Solving Capabilities

To go back to this article, head to My Profile and then click on Saved stories to view them again

OpenAI Reveals Strawberry, a Novel AI Model Capable of Logical Reasoning Through Complex Challenges

In its most recent innovation, OpenAI transitioned from its previous strategy of expanding model sizes, notably with the introduction of GPT-4 last year, to a new methodology. The organization unveiled today a groundbreaking model designed to logically navigate through a variety of challenging issues, demonstrating a considerable leap in intelligence over current AI models without the need for extensive scaling.

The latest creation, known as OpenAI o1, is equipped to tackle challenges that leave current AI models, including OpenAI's top-tier model, GPT-4o, at a loss. Unlike the typical large language models that generate responses instantly, this model processes the issue by deliberating on it, mirroring human thought processes, until it reaches the correct conclusion.

"Mira Murati, OpenAI's chief technology officer, shared with WIRED that they view this as the latest breakthrough in these frameworks. She emphasized that it significantly excels in handling highly intricate reasoning challenges."

OpenAI has internally dubbed the latest model "Strawberry," and according to the company, it's designed to complement GPT-4o, not replace it.

Murati mentions that OpenAI is in the process of developing its upcoming flagship model, GPT-5, noting that it will be significantly more extensive than the previous version. Although the firm continues to support the idea that increasing size enhances AI capabilities, GPT-5 is expected to incorporate the logic technology unveiled today. Murati explains, "We're operating under two main concepts: the concept of scaling up and this novel concept. Our aim is to integrate these two approaches."

Large Language Models (LLMs) generate their responses through extensive neural networks that have been trained with enormous amounts of data. They demonstrate impressive skills in language and logic, yet they often face difficulties with basic mathematical problems that require logical thinking.

Murati explains that OpenAI's o1 model employs reinforcement learning. This technique enhances the model's logic by rewarding correct responses and penalizing incorrect ones, allowing it to refine its approach to problem-solving. "This process hones the model's analytical abilities and optimizes the methods it applies to reach conclusions," she states. Thanks to reinforcement learning, machines can now outperform humans in gaming and undertake practical activities, such as creating computer chip designs. This approach is crucial for transforming a large language model (LLM) into an effective and manageable chatbot.

Mark Chen, OpenAI's research vice president, showcased their latest model to WIRED, effectively handling numerous challenges that stumped its predecessor, GPT-4o. Among these were a complex chemistry inquiry and a particularly perplexing math riddle: "Considering a princess's current age equals what the prince's age will be when she is twice as old as he was when her age equaled half their combined ages now, how old are the prince and princess?" (The solution revealed that the prince is 30 years old, while the princess is 40).

"According to Chen, the latest model is developing the ability to independently process information, moving away from merely attempting to replicate human thought processes, which is typical of traditional LLMs."

OpenAI reports that its latest model shows significant improvement across various subject areas such as coding, mathematics, physics, biology, and chemistry. In terms of performance on the American Invitational Mathematics Examination (AIME), which is a challenge for math students, their GPT-4o model was able to correctly solve an average of 12 percent of the questions, whereas the o1 version achieved a correct answer rate of 83 percent, the company states.

The latest version exhibits a slower performance compared to GPT-4o, according to OpenAI. This is partly due to its inability to conduct web searches and its lack of multimodality, as it cannot interpret images or audio, unlike its predecessor.

Enhancing the logical thinking abilities of Large Language Models (LLMs) has become a significant area of interest among researchers for a while now. In fact, competing entities are also exploring comparable avenues of research. In July, Google revealed AlphaProof, an initiative that integrates language models and reinforcement learning to tackle complex mathematical challenges.

AlphaProof mastered the skill of solving mathematical problems by analyzing the right solutions. Expanding this learning approach is difficult because not all scenarios the model faces have definite answers. According to Chen, OpenAI has effectively developed a more versatile reasoning system. "I believe we've achieved significant advancements in this area; it's something that sets us apart," Chen states. "It's quite proficient at logical thinking in various fields."

Stanford professor Noah Goodman, known for his research on enhancing the reasoning skills of Large Language Models (LLMs), suggests that the secret to broader training might lie in employing a "meticulously prompted language model and manually curated data." He also mentions that the ability to exchange result speed for increased precision would represent a significant improvement.

MIT's Assistant Professor Yoon Kim notes that the workings of Large Language Models (LLMs) in problem-solving are still not fully understood, and their methods of step-by-step reasoning could significantly differ from human intellect. This issue gains importance as the technology sees broader application. Kim highlights, "These systems could be making impactful decisions on a vast scale. The critical concern is whether we should fully understand the decision-making process of these computational models."

Today, OpenAI unveiled a method that could play a crucial role in guiding AI models towards more acceptable behavior. According to Murati, this latest development is more adept at preventing the generation of undesirable or dangerous content by contemplating the consequences of its decisions. "Just like children who understand the reasoning behind their actions tend to adopt appropriate norms, behaviors, and values more effectively," she explains.

Oren Etzioni, an esteemed AI authority and professor emeritus at the University of Washington, emphasizes the critical need for Large Language Models (LLMs) to tackle multi-step problem-solving, utilize tools, and address intricate issues. He points out that merely increasing their size won't achieve this. Etzioni also highlights additional obstacles on the horizon, noting that even with advancements in reasoning capabilities, issues such as generating inaccurate information and maintaining factual accuracy remain significant challenges.

OpenAI's Chen mentions that the innovative reasoning method devised by the organization demonstrates that enhancing AI doesn't necessarily require vast amounts of computing resources. "What's thrilling about this new approach is our belief that it will enable us to deliver smarter solutions more affordably," he states, "and I believe that's fundamentally what our company aims to achieve."

Check Out Also …

Delivered directly to your email: A selection of the finest and most unusual tales from the vault of WIRED.

Exploring the Process of Memory Selection by the Brain

Headline News: Introducing Priscila, the Leading Figure of the Rideshare Syndicate

Silicon Valley's wealthy elite show unexpected support for Donald Trump

Occasion: Be part of The Major Interview happening on December 3rd in San Francisco.

Additional Content from WIRED

Evaluations and Instructions

© 2024 Condé Nast. All rights reserved. Purchases made through our site for products linked via our Affiliate Partnerships with retail outlets may generate revenue for WIRED. It is prohibited to reproduce, distribute, transmit, store, or use the content found on this website in any form without obtaining prior written consent from Condé Nast. Choices regarding advertisements are available.

Choose global website