In the November 2023 OpenAI crisis times, the existence of Q*, a secret AI model with the potential for super-intelligence, became public. Nowadays, this model is going by the name “Strawberry”. Its launch seems imminent.
A new article from The Information characterizes Strawberry not as a “super-intelligence,” but rather as an impressive step forward for AI. For example, it is a math prodigy compared to it’s predecessor GPT-4. On the popular MATH benchmark, it achieves a score of 90 percent, while GPT-4 scored 53 percent and its upgraded cousingot to 76.6 percent.
The name Strawberry seems a cheeky one. A familiar way of pointing out GenAI’s limitations is to refer to the fact that ChatGPT regularly miscounts the number of R’s in the word “strawberry.” Since the new model would complete complex tasks with multiple reasoning steps, adding up some vowels in a word might be child’s play.
Problems all around
To an outsider, Strawberry simply seems like a logical evolution of GPT-3 (2020) and GPT-4 (2023). Perhaps it will also simply be called GPT-5, too. Yet behind the scenes, much has gone wrong within OpenAI. First, in the months following the board conflict in November, there was an exodus of OpenAI veterans, including the “genius” Ilya Sutskever, as CEO Sam Altman described him.
Sutskever is now trying to get ahead of OpenAI with human-level AI thinking with the new AI company Safe Superintelligence Inc (SSI). Meanwhile, his old employer is having mounting financial issues. For that reason, it is not so surprising that the trajectory toward an eventual launch of Strawberry this fall has been described as a “race.” After all, if it is lost, the ChatGPT maker might face existential problems.
Read more: OpenAI’s business model isn’t working as bankruptcy looms