OpenAI’s Strawberry project, builds a new AI model 100 times more powerful than GPT-4

The Information reported on August 26 that Project Strawberry will be mathematically and programming superior to any existing chatbot — the GPT-Next model is 100 times more powerful than GPT-4 and is capable of real-time video analysis, shared “two people involved in the project.”

OpenAI is likely to release its most powerful AI model this fall (September-November), and may integrate it into ChatGPT-5, the new version of its chatbot and virtual assistant that it launched in late 2022.

This secret project, which OpenAI has been working on for a long time, was previously known as Project Q* (Q-star), and is now codenamed Project Strawberry. It is expected to be able to self-study the internet and significantly improve AI reasoning.

This is considered OpenAI’s effort to create Artificial General Intelligence — AI with capabilities similar to the human brain.

On August 7, OpenAI CEO Sam Altman posted an image of strawberries growing in two pots to his X account. The post was seen as confirmation that OpenAI was working on a new, powerful large-scale language model.

The report said OpenAI demonstrated a version of this new model to national security officials, seemingly a statement of its commitment to transparency as the rapid development of AI has raised serious security concerns from national governments.

The integration with ChatGPT will make ChatGPT the most powerful AI chatbot available, the report said. ChatGPT sometimes struggles with math, and experts say these errors may be due to a lack of complete mathematical information in the training data.

A demonstration by the Project Strawberry team showed the new AI model capable of high-level thinking, allowing it to solve puzzles, including The New York Times’ word game ‘Connections’, a particularly difficult puzzle, The Information reports.

Training needs

The Information says Project Strawberry is aimed at raising additional funding, which OpenAI needs for its next model, codenamed Orion.

Creating high-quality training data for Orion is said to be one of the key applications of the Strawberry Project. This is important because most of the training data on the Internet has already been used, and there is currently a lack of freely accessible information for the purpose of training AI models.

As such, OpenAI has recently struck deals with publications to use their content for training.

Project Orion, designed to outperform GPT-4, can use a combination of Project Strawberry and high-quality synthetic data that is likely to reduce errors and illusions compared to previous versions and other AI models.

Generate aggregated data

Altman said that to experiment with different ways to train AI models, OpenAI has been experimenting with generating large amounts of synthetic data. Generative AI models generate synthetic data based on real-world data samples.

Algorithms learn patterns, correlations, and statistical properties of sample data; once trained, the model can generate statistically identical synthetic data.

The large datasets that AI models rely on can be prone to bias and error, or may contain incomplete or inaccurate information — the high-quality synthetic data generated by the Strawberry Project can fill in the gaps in real-world datasets and provide a more comprehensive, inclusive, and balanced training set.

Many believe that using synthetic data can help make future AI models more neutral and fair, and reduce noise and irrelevant information — thus improving both training efficiency and model accuracy.

Based on what is known, Project Strawberry’s improved reasoning, logic, and ability to plan and execute research could allow the model to automatically conduct experiments, analyze data, and generate new hypotheses.

This could lead to scientific breakthroughs, including the discovery of new drugs. These models could also provide personalized education, creating educational content and interactive lessons.