The Job to be Done of Project Strawberry
OpenAI is once again exploring the frontiers of artificial intelligence with a new, project called "Project Strawberry." Previously known as Project Q*, this initiative is set to elevate the reasoning abilities of AI models to new heights.
Recent reports from Bloomberg and Reuters suggest that OpenAI is making significant strides in developing AI that can plan autonomously, navigate the web, and conduct "deep research" akin to human thought. During an internal meeting, a demonstration hinted that these models might soon be able to think more like humans.
The core job-to-be-done (JTBD) for Project Strawberry is to create AI that can reason like humans. Unverified sources indicate that an internally tested model scored an impressive 90% on a complex AI math skills test, though it's unclear if this model is part of Project Strawberry.
What we do know is that Project Strawberry is likely refining OpenAI’s existing large language models, using vast data sets to enhance their reasoning abilities. This approach seems similar to the Self-Taught Reasoner (STaR) technique from a 2022 Stanford study.
The Job to Be Done of Project Strawberry
In a recent Dartmouth Engineering interview, OpenAI’s CTO, Mira Murati, hinted at the next generation of AI, which she compared to someone with a Ph.D. Could this be the model she was referring to?
Sam Altman, OpenAI’s CEO, believes that future breakthroughs will come from improving AI’s reasoning skills. Current AI models excel at generating text and handling specific tasks, but they struggle with complex reasoning and long-term planning.
However, there are concerns. With Ilya Sutskever's departure and key alignment team members leaving, questions arise about the direction of OpenAI’s research and the safety of its products.
A Brief History of Q*
Project Q* began attracting attention in 2023, but its roots trace back to early 2022. The timeline includes:
Early 2022: OpenAI starts improving AI’s reasoning capabilities.
Mid 2022: Ilya Sutskever initiates what becomes Project Q*.
Late 2022: The team successfully gets the model to solve basic math problems.
Early 2023: Internal tensions arise over the project’s direction.
Q* combines large language models, reinforcement learning, and search algorithms, potentially blending Q-learning and A* search methods. The name "Q*" might be a nod to DeepMind’s early work with reinforcement learning.
Final Thoughts on Project Strawberry
While full details of Project Strawberry remain undisclosed, it’s clear that OpenAI is aiming for significant advancements. Whether these developments will lead to Artificial General Intelligence (AGI) or pose a threat to humanity is still up for debate. However, this project could mark a major step toward AI that can reason more generally.
It’s important to recognize that intelligence exists on a spectrum, whether human or artificial. Just as human intelligence varies, AI systems also have different capabilities based on their design and training. In many specialized fields, current AI systems already outperform humans, but humans still excel in general reasoning, creativity, and emotional intelligence.
The future lies not in competing with AI, but in combining human and AI strengths to achieve transformative results. The JTBD for Project Strawberry is clear: to build AI that can reason like humans, pushing the boundaries of what’s possible in the field of artificial intelligence.
