OpenAI's Strawberry Model Autumn Reveal🍂🍓

Hey Everyone,

What does it mean when OpenAI is showing off its latest tech to the Pentagon, before the general public? OpenAI has apparently showcased its latest AI breakthrough, codenamed “Strawberry,” to U.S. national security officials, according to a new report by The Information.

This model forms the "secret sauce" of their new Orion model. We still don't know a ton about it all. The story goes that OpenAI is high likely to release a chatbot version of Strawberry as early as this fall which like the Q* rumor is really good at Math and doing some tasks more independently.

While many are awaiting Claude 3.5 Opus by Anthropic, Strawberry will be its approximate competition it seems. The product, called “Strawberry,” will be able to solve problems and tasks that are beyond the capabilities of current AI models though not many details have been given.

OpenAI's product-marketing and branding have gone a bit off the rails recently as Google and Anthropic have not only caught up with them, but surpassed them in many ways and according to ML engineers and their consensus. Strawberry, previously known as Q* has been a matter of speculation for months, while most believe GPT-5 will be available sometime next year in 2025. OpenAI has teased various products like Sora, SearchGPT, Orion model and others but with limited access or information.

Meanwhile, Claude and Perplexity are now growing faster with ChatGPT euphoria to have peaked or plateaued. So it's a dilemma for OpenAI, how do they keep enthusiasm high for their product that has been somewhat left behind in recent months? Meanwhile there has been a talent exodus at the company.

I speculate here a bit, OpenAI is using Strawberry to generate high-quality training data for its next flagship language model, codenamed "Orion." This is a bid not only to reduce errors and "hallucinations" in AI responses by providing more accurate examples of complex reasoning but heading into Agentic AI and the ability to chain tasks together like many other AI coding Startups are now undertaking.

OpenAI Pivoting to more Applied AI

Strawberry is designed to solve complex math and programming problems more effectively than current systems. It aims to generate high-quality synthetic training data for Orion, OpenAI's next flagship large language model (LLM). Despite its rapid revenue growth, OpenAI is facing high costs, particularly for cloud servers, and is looking for ways to reduce losses and raise more capital. Losing $3 Billion a year is likely not sustainable especially as many users switch from ChatGPT to Claude.

I'm not sure the Strawberry marketing is reassuring potential investors much! How do you evaluate something like this? Strawberry can tackle previously unseen math problems and optimize programming tasks. It has demonstrated the ability to solve the New York Times word puzzle 'Connections' and scored over 90% on the MATH benchmark.

The company is also exploring a smaller, distilled version of Strawberry for potential integration into chat-based applications like ChatGPT. This model could enhance reasoning capabilities in scenarios where users require more thoughtful, detailed answers rather than quick responses. I suppose this is to improve their Enterprise AI offering where the likes of Anthropic, Cohere, Databricks, Aleph Alpha and many others are rapidly advancing.

What's clear is the window of OpenAI's first mover advantage is rapidly closing and how they build products isn't always aligned with their customers. The company also launched a new SearchGPT tool July 25, making the prototype available to a limited test group. This as Perplexity adoption is now growing fairly fast as they continue to improve the product.

OpenAI has shown Strawberry to American national security officials and appear to be increasing their potential association with the Pentagon. OpenAI has kept many details about Q* under wraps to prevent misuse, which has sparked debates on transparency and safety in AI development. However, OpenAI have so many products in the pipeline they appear scattered and slow to execute anything well since the original ChatGPT and GPT-4 moment. Their path to profitability looks fairly poor when you consider the progress competitors like Google and Anthropic have made in 2024.

Still if you follow Sam Altman, you can still believe in the Strawberry. 🍓... We'll know more this Autumn.