Exciting Launch: OpenAI's New Strawberry Model Set to Go Live!

On the evening of August 27, 2024, The Information reported that OpenAI is showcasing its “Strawberry” model to attract additional funding. This model not only improves its capabilities but also serves as training data for the upcoming “Orion” model, internally codenamed Orion.

Strawberry boasts enhanced reasoning skills and exceptional mathematical abilities, excelling in solving complex mathematical problems and optimizing coding tasks. It scored over 90% on the MATH benchmark, demonstrating its capacity to invest more time in thoughtful analysis. In July, Reuters reported that OpenAI showcased this technology to U.S. national security officials. Due to safety considerations, the model is slated for release after the U.S. elections. OpenAI has decided not to offer this powerful model directly to the public, to prevent its misuse by nations restricted by U.S. regulations who might employ it to train more advanced models.

OpenAI's CEO, Sam Altman, shared a photo of a strawberry plant on social media, hinting at the ongoing development of the Strawberry model. Additionally, OpenAI plans to provide a smaller, distilled version of Strawberry for public access, expected to debut around October's DevDay, which will also lower deployment costs. The full Strawberry model will be utilized internally to generate high-quality synthetic data for training Orion, which aims to surpass GPT-4.

This initiative is focused on reasoning and is seen as a significant step towards achieving artificial general intelligence (AGI)—the Holy Grail of AI research, where machines can think, reason, and act like humans. Will the Strawberry model meet expectations? Can it push the boundaries of current artificial intelligence? For now, we can only speculate and eagerly await the experiences it will offer, hoping they are as delightful as strawberries themselves. Do you think OpenAI's latest project will change the game?