OpenAI is working on a new secret project that is expected to produce a new model that has better-than-human intelligence (? Super-intelligence?). A leaked internal document and information from an insider has revealed that OpenAI is working on a new, secret project, codenamed Project Strawberry, which is expected to produce a model that’s capable of better reasoning, mimicking that of human intelligence.
What is Project Strawberry?
Project Strawberry is a tightly kept secret, even within OpenAI, but the leaked information has revealed that it involves training the new model to “plan ahead enough to navigate the internet autonomously”, without being told to by the user, and undertake “deep research”, demonstrating improved reasoning and human-like intelligence. As it will perform “deep research” it could also mean that the new model will produce fewer hallucinations, which are typical of AI models (like ChatGPT), as they tend to make up information when they lack the training data in a particular subject. None of this is however, new.
Although details are still scarce, it’s thought that Project Strawberry is using a method called Self-Taught Reasoner (STaR), to achieve human-level reasoning, as this technique allows the model to improve, by learning from its own reasoning, and issomething that Open-AI believe will eventually lead to better-than-human-like intelligence. It’s not yet known whether OpenAI plans to release Project Strawberry or whether Strawberry is the highly-anticipated launch of GPT-5, which is expected to be as intelligent as someone with a PhD. Others are speculating that Project Strawberry is actually the infamous Q* model, that was leaked late last year after multiple OpenAI employees wrote a letter to the board to warn them of a new AI breakthrough that “could threaten humanity.”