OpenAI has a new model making waves behind the scenes and Project Strawberry could be the start of thinking, reasoning ChatGPT.
The Q Model was leaked last year and that was thought to be the start of a higher level ChatGPT. Project Strawberry seems to have grown out of that early model. As well as evaluating existing information to answer queries more sepcifically, it can go online to gather more information to come up with a deeper answer to your questions.
ChatGPT has come a long way in a short time, but it is still limited to pretty generic filler content. It can often repeat itself and seemingly get lost in small details. That’s because an LLM is basically a really advanced next token predictor and so you finish with an average of everything that’s already out there.
Interpretation and extrapolation are kind of possible, but you still need a human touch to really get the most out of the limited analysis and reasoning power that ChatGPT-4o demonstrates. OpenAI CTO Mira Murati recently said that the next generation would have the intelligence of a PhD graduate.
This could also be a firm step up to Level 2 on the way to AGI.
What do we know about Project Strawberry?
A Reuters report, based on internal documents and insiders say this is a major upgrade. According to the source, Strawberry is a “tightly kept secret” within OpenAI. While it may become part of ChatGPT’s main architecture, its primary purpose is to perform “deep research” beyond simple user queries.
Strawberry’s capabilities include autonomous and reliable internet navigation, without explicit user prompts. ChatGPT 4o uses up-to-date data and can access the internet. But it needs direction.
Now this might be a step too far, but recently there were several high-profile departures from people openly criticizing the safety protocols at OpenAI. Sam Altman was even ousted from the board for a matter of hours before he stamped his authority all over them and returned to the helm.
Do Ilya Sutskever and Helen Toner think this next generation of AI is going to be the one to go all Skynet? Or is it just weird timing?
When will we see Project Strawberry in ChatGPT?
Any improvements in reasoning from Strawberry are likely to be integrated into the training for GPT-5. Strawberry might represent a new model that is inherently a reasoner, built from the ground up without needing fine-tuning or additional training. Although details are scarce, OpenAI hinted to Reuters that human-like AI understanding is a crucial goal.
That might be closer than we thought.