OpenAI is working on new artificial intelligence models in addition to GPT-4o. In fact, the company mentions one of these models, codenamed “Strawberry,” in a document obtained by Reuters. Strawberry is currently being kept under wraps, with little information available about it.
Strawberry, a secret project for now. According to the document, the AI model is part of OpenAI’s internal development and is planned to be incorporated into the update of the controversial Q* algorithm. There’s no detailed information on when the company will make it available, as the intention is to keep it a secret for the time being.
Capable of thoroughly exploring the web. A key feature of Strawberry will reportedly be its capacity to navigate the web. This goes beyond basic Internet access and link reading, something the AI already does effortlessly. This means the model will have the ability to explore the web on its own and utilize this skill to plan complex tasks in advance.
OpenAI aims to enable this AI to access multiple websites and conduct “deep research” based on users’ commands.
A different method of training. OpenAI engineers are allegedly training Strawberry to learn to reason on its own rather than training it on a large dataset.
According to Reuters, the approach is similar to the Self-Taught Reasoner (STaR) method computer scientist Noah Goodman developed in 2022 at Stanford University. This technique involves creating new tasks and generating its own training data as it receives it. In other words, the model is trained based on the data it’s currently reading on a website, like a live update or “post-training,” as experts call it.
Next level. As Bloomberg reports, Strawberry will replace the Q* algorithm, and the first demos show it can surpass the capabilities of any current AI model. According to Bloomberg sources, this new model can answer mathematical and scientific questions that existing AI models have struggled with.
A step closer to Artificial General Intelligence (AGI). According to Bloomberg, OpenAI claims Strawberry and its new “learning” method represent a move from level 1, where AIs can engage in conversations with humans, to level 2, known as “Reasoners.” OpenAI’s internal classification further categorizes future advancements as “Agents” (level 3), “Innovators” (level 4), and “Organizations” (level 5).
Image | Aidin Geranrekab
View 0 comments