Rumors are swirling about OpenAI’s new AI system codenamed “Strawberry,” and we might be seeing it as soon as this fall, according to The Information.

The Scoop on Strawberry

According to recent reports from The Information, OpenAI is racing to launch a new AI system that could represent a major leap forward in problem-solving abilities, particularly in math and programming. 

This system, known as “Strawberry,” might be integrated into a chatbot—possibly even ChatGPT—as early as this fall.

But what makes Strawberry so special? Let’s break it down:

1. Math Mastery: Strawberry can reportedly solve math problems it hasn’t seen before, pushing beyond the capabilities of current AI tools.
2. Thinking Time: When given additional time to ponder problems, Strawberry can better perform more subjective tasks, like answering questions about product marketing strategies.
3. Data Generator: A larger version of Strawberry is being used to generate synthetic training data for OpenAI’s upcoming flagship model, codenamed Orion.

Why Strawberry Matters

Strawberry isn’t just another incremental improvement in AI—it represents a significant step towards what OpenAI calls “Level 2” AI: reasoners with human-level problem-solving abilities.

“This is a step toward the artificial general intelligence (AGI) that OpenAI pursues,” Marketing AI Institute founder and CEO Paul Roetzer told me on Episode 113 of The Artificial Intelligence Show.

But the real game-changer here is how Strawberry achieves its impressive capabilities. It all comes down to a concept called “process supervision.”

The Secret Sauce: Process Supervision

In May 2023, OpenAI researchers published a paper titled Let’s Verify Step by Step.

“I think this paper is probably the blueprint to Strawberry,” says Roetzer.

Here’s the key idea: Instead of just rewarding the AI for getting the right final answer (outcome supervision), Strawberry is rewarded for each correct step in its reasoning process (process supervision). It’s like giving the AI a gold star not just for acing the test, but for showing its work every step of the way.

Why is this so important? It allows the AI to:

• Evaluate itself at each step
• Identify and correct errors along the way
• Dramatically reduce hallucinations and mistakes
• Potentially improve itself over time

“Once these models can do this, once they can remove the error rates or dramatically reduce the error rates, they can remove the hallucinations and they can give themselves the rewards along the way,” says Roetzer. “Now they develop the ability to improve themselves.”

“Self-correcting” AI puts us in a situation where AI’s capabilities can take off pretty quickly, says Roetzer.

The Road Ahead

While the potential of Strawberry is exciting, it’s important to note that what we’ll likely see publicly is a scaled-down, “safer” version of the full model, says Roetzer. (OpenAI has a history of releasing more limited versions of their technologies to the public.)

Regardless, expect Strawberry to mark a significant milestone in AI development if it does what we’re speculating it does. By improving an AI’s ability to reason, plan, and self-correct, we’re taking a big step towards more reliable and capable AI systems. While we may not see the full power of Strawberry immediately, its development signals a new era in AI problem-solving capabilities.