OpenAI's 'Strawberry': Advancing AI Reasoning & Math Problem Solving

OpenAI is reportedly working on a new AI model codenamed "Strawberry," which promises significant improvements in reasoning abilities. According to a report from The Information, this model, previously known as Q*, is designed to tackle complex mathematical problems it hasn't encountered before and programming challenges that current models struggle with. Additionally, Strawberry is said to excel in solving intricate word puzzles and even answering subjective questions, such as those related to marketing strategies. The model is expected to be publicly available as early as this fall.

Strawberry's Role in Developing Orion

In addition to its standalone capabilities, Strawberry is playing a crucial role in the development of OpenAI's next large language model (LLM), called Orion. Strawberry is being utilized to generate synthetic data for training Orion, which could be the successor to GPT-4. Launched in March 2023, GPT-4 was a significant advancement in AI, but in the following months, competing models have caught up, and in some cases, surpassed it.

Advancements by Other AI Giants

Other AI giants are also making headway in mathematical reasoning. Google DeepMind, for instance, has developed AlphaProof and AlphaGeometry 2, two AI systems that have demonstrated advanced mathematical reasoning. These systems recently achieved a silver-medal performance at the 2024 International Mathematical Olympiad, successfully solving four out of six problems.

The Importance of Synthetic Data

A key factor in AlphaGeometry 2's improved performance over its predecessor was its training on a much larger dataset of synthetic data. This allowed the model to handle more complex geometry problems, including those involving object movements and calculations of angles, ratios, or distances.

Orion is rumored to benefit from similar synthetic data, potentially from an enhanced version of the Strawberry model. This data could be pivotal in addressing one of the major limitations of current AI systems: "hallucinations," where models generate plausible-sounding but incorrect information.

The Path to Artificial General Intelligence

The foundation for this project traces back to research by former OpenAI chief scientist Ilya Sutskever and is being compared to Stanford's Self-Taught Reasoner (STaR) method. The STaR approach trains models to reason more effectively by generating explanations for their answers, filtering out incorrect ones, and then fine-tuning the model based on these self-generated rationales. When details about OpenAI’s Q* project first emerged in late 2023, some experts suggested it could be a significant step towards Artificial General Intelligence (AGI), though specific details remain sparse.

Looking Ahead: The Potential Impact of Strawberry

If the Strawberry model meets expectations, it could represent a major advancement in AI reasoning, potentially propelling OpenAI closer to Stage 2 of its five-level roadmap to AGI. With ChatGPT expected to integrate these capabilities this fall, we may be on the brink of the next major wave of AI innovation.