Exploring OpenAI's Project Strawberry: A Leap Towards Autonomous AI Research

Last month, Reuters reported that ChatGPT creator OpenAI is working on a project codenamed “Strawberry”. Internal documents reviewed by the publication show that Project Strawberry is about enabling AI to perform deep research autonomously on the web. When asked about the project, OpenAI’s spokesperson said, “these systems will improve in reasoning over time.” Now, what exactly is Project Strawberry and what exactly is OpenAI cooking? Is it the upcoming GPT-5 model, aka ChatGPT 5? Let’s discuss it.

OpenAI’s Project Strawberry Explained

Project Strawberry is said to be related to Q* — the next breakthrough in AI development believed by OpenAI staffers inside the company. There is not much information about Q* in the public domain. However, sources who viewed the demo say that it’s capable of solving science and math-related questions that no other current models can demonstrate.

It has scored over 90% on a MATH dataset, a rigorous benchmark for testing proficiency in math problems. Bloomberg recently reported that OpenAI held an internal demo, showcasing a research project that demonstrated human-like reasoning. OpenAI also presented different levels of AI progress, which we have described in our AGI explainer, and are as follows:

• Level 1: Conversational AI — Can chat like humans in natural language
• Level 2: Reasoning AI — Can reason like humans exhibiting intelligence
• Level 3: Autonomous AI — Perform actions autonomously
• Level 4: Innovating AI — Innovate and improve themselves
• Level 5: Organizational AI — Perform actions and accomplish tasks of an entire organization

Many believe that with Project Strawberry, OpenAI has reached Level 2 where AI systems can reason intelligently like humans. Reuters’ report suggests that Project Strawberry includes a post-training method where the model is trained in a specific way, similar to fine-tuning.

This specific technique for post-training is very similar to Self-Taught Reasoner or STaR, a chain-of-thought method proposed by Stanford researchers in 2022. It aims to improve the model performance on complex reasoning tasks like mathematics or commonsense question-answering by generating step-by-step “chain-of-thought” rationales.

In simple terms, STaR allows the model to improve itself by learning from its own generated reasoning. It seems that OpenAI is likely using the STaR technique for Project Strawberry to mimic human-level reasoning and solve hard math problems.

Project Strawberry is not just limited to advanced reasoning; it can also perform something called Long-horizon Tasks (LHT), according to documents seen by Reuters. Essentially, it can act as an agent that browses the web autonomously, comes up with findings, plans, and performs a series of actions.

The Curious Case of Project Strawberry on X

On August 7, OpenAI CEO Sam Altman shared an image of a strawberry, raising excitement and curiosity among the community. The next day, an X account (@iruletheworldmo), believed to be an OpenAI staffer, tweeted, “welcome to level two. how do you feel? did I make you feel?” Altman commented, “amazing tbh.”

Following this, a flurry of cryptic strawberry-related posts was shared by OpenAI staffers on X. Most of the posts allude to Level 2, aka advanced reasoning capability, and Project Strawberry.

Looking at the development so far, it appears OpenAI is preparing to release the next frontier model pretty soon. Powered by Project Strawberry, the frontier model is likely to bring advanced reasoning capability and unlock agentic workflows. Now whether it will be called GPT-5 or something else, we don’t know. Sam Altman has said before that the successor to the GPT-4 model might not be called GPT-5.

Besides that, it has been over a year since GPT-4 was released. Meanwhile, Anthropic and Google have almost dethroned OpenAI with their iterative updates. In May 2024, OpenAI CTO Mira Murati said the next big thing is coming in 2024 itself. So, the launch seems pretty imminent. Are you ready for that? Let us know in the comments below.