The developers of ChatGPT, OpenAI, are working on a top-secret AI project called “Strawberry” that is aimed at greatly improving AI reasoning capabilities and could result in advances in autonomous research and problem-solving.
Although specifics are being kept under wraps, Reuters has discovered through internal documents and sources that Strawberry uses a unique method of processing and training AI models to accomplish tasks that have eluded previous systems.
In contrast to existing models that mostly concentrate on producing text-based responses, Strawberry seeks to give AI the capacity to “plan ” and independently traverse the internet in order to carry out “deep research,” as described by OpenAI. Given that it necessitates a deeper comprehension of context, reasoning, and multi-step problem-solving, this is a big advancement.
The industry is focused on developing AI that can think at the human level; Google, Microsoft, and Meta are just a few of the companies investigating different approaches. Achieving this breakthrough, according to experts, might open up new possibilities for AI, including the ability to drive scientific discoveries, create intricate software, and take on problems that presently require human intuition and forethought.
“We want our AI models to see and understand the world more like we do,” an OpenAI spokeswoman told Reuters, despite the company not having publicly verified any details regarding Strawberry. In the industry, it is customary to conduct ongoing research into new AI capabilities, with the shared expectation that these systems will eventually become more intelligent.”
Strawberry seems to be a development of an earlier OpenAI project called Q*, which caused enthusiasm within because of its capacity for sophisticated reasoning. Witnesses to Q* demonstrations claimed that the technology could tackle intricate maths and science issues better than any commercial AI currently on the market.
Although the precise workings are yet unknown, insiders speculate that Strawberry uses a unique type of “post-training,” which is the process of fine-tuning AI models after they have been trained on enormous datasets. Finding the AI’s reasoning skills requires this post-training step, which may involve methods like “fine-tuning” and self-generated training data.
