OpenAI Project Strawberry: All the Latest Insights and Updates

OpenAI is reportedly investing $7 billion in training and inference costs while facing a significant $5 billion shortfall, as it strives to achieve its ambitious goal of creating the world's first Artificial General Intelligence (AGI).

What is Project Strawberry?

Project Strawberry is OpenAI's latest large language model, anticipated to far exceed the capabilities of existing state-of-the-art systems with its advanced “human-like reasoning skills.” This innovative model could potentially be the driving force behind the next generation of ChatGPT.

What can Strawberry do?

Project Strawberry is designed to be a formidable reasoning engine. By leveraging a mix of reinforcement learning and “chain of thought” reasoning, this new model aims to tackle unprecedented math problems and function as a high-level agent—developing marketing strategies and independently solving intricate puzzles like the NYT's Connections. It can even "autonomously navigate the internet" to conduct "deep research," according to internal documents accessed by Reuters in July.

During its development, the o1-preview model was tested with a qualifying exam for the International Mathematics Olympiad, where it achieved an impressive 83% accuracy—compared to just 13% for its predecessor, GPT-4o. In coding challenges, o1-preview ranked in the 89th percentile in an online Codeforces competition. It is even capable of answering complex questions that had stumped earlier models, such as “which is larger, 9.11 or 9.9?” and “how many ‘Rs’ are in the word ‘Strawberry’?”

Additionally, the architecture of Strawberry is reportedly inspired by the Self-Taught Reasoner (STaR) technique developed at Stanford in 2022, which allows models to generate their own training data for continuous improvement.

Why is it called Strawberry?

The reason behind the name “Strawberry” remains under wraps, as OpenAI has not shared its significance. This code name was selected for internal use to maintain confidentiality during the model's development. The name was first hinted at in a cryptic post by CEO Sam Altman that sparked widespread speculation online.

Interestingly, Strawberry was not always known by this name. Initially, it was referred to as Q (pronounced Q-Star), a key element in the controversies surrounding Altman's removal from the company last November. OpenAI researchers pointed to Q in their appeal to the board over potential risks of unregulated advanced AI systems. After Altman returned to OpenAI and regained influence, Q was renamed Strawberry in July 2024.

When will Strawberry be released?

OpenAI introduced a preview of Project Strawberry in September 2024, called “o1-preview,” alongside a lighter version named “o1-mini.”

Upon launch, this new iteration was made available to ChatGPT Plus and Teams subscribers immediately, with further access for Enterprise and Edu users the subsequent week. OpenAI plans to eventually offer o1-mini to free-tier users; however, a specific timeline has not been provided.

Wait, what was that about it being able to autonomously navigate the internet? Isn’t that how we got Ultron?

Ultron originated from comic book creators Roy Thomas and John Buscema, but concerns about advanced AI crossing into uncontrollable territories are increasingly relevant. Former OpenAI researchers have raised alarms about the company's AGI development, citing insufficient safeguards.

Despite these concerns, OpenAI remains undeterred in its pursuit of AGI within the next decade. Recently, the company introduced a five-tier scale to measure its AI systems’ capabilities as it seeks additional funding—beyond Microsoft—in a new funding round led by Thrive Capital, valuing the company at over $100 billion.