Strawberry- an OpenAI Project
Project Strawberry," previously known as Project Q*, is an ambitious OpenAI effort to really push the reasoning abilities of AI beyond anything previously seen. It's working on AI models that can self-navigate the internet, plan, and conduct what OpenAI calls "deep research." The ultimate goal is developing AI systems that think more like humans, a significant step beyond the current capabilities of most AI models.
According to Bloomberg and Reuters, leaked information has revealed that OpenAI is far more advanced towards meeting this goal. At an internal demo, the company unveiled a model that can reason like humans. Details aren't yet provided, but such properties include complex planning attributes and problem-solving, arguably making it a landmark in AI development.
The most exciting thing about Project Strawberry, however, is its purported success in internal testing. According to some reports, during an AI math skills test, a model achieved a score of 90% on a hard test. While it is not clear whether that is attributed directly to Project Strawberry, it really shows that the project works on improving the mind of AI. Some sources even mentioned attending demos where the AI solved highly complex math and science problems way out of reach for the commercially available AI systems.
How exactly they did this is not entirely public, but presumably, OpenAI fine-tuned its existing large language models. These have been pre-trained with enormous datasets and are fine-tuned using techniques that may borrow from an approach documented by a 2022 Stanford research paper called STaR for Self-Taught Reasoner. STaR trains AI to generate the rationales for its choices, which could be the key to Project Strawberry.
In a recent interview, OpenAI's CTO, Mira Murati, gave hints about what the next generation of AI would bring and said that future AIs are going to have intelligence comparable to a person with a Ph.D. It is then speculated that Project Strawberry may have been the model she was referring to, especially given its focus on improving reasoning abilities.
领英推荐
However, the rate at which such sophisticated AI is being developed also brings along its concerns. Recent departures by key people, including Ilya Sutskever and a part of OpenAI's alignment team, have fanned heated debates on the future course of the organization and the safety of its products. According to some critics, in the race to achieve AGI, OpenAI may end up pushing at an insane pace without being much concerned about the risks.
Project Q*, the precursor of Project Strawberry, was started in early 2022. The large language models were combined with reinforcement learning and search algorithms, mixing a little deep learning with traditional human-coded rules. The architecture impinges on a complex approach to developing AI that can reason complexly, probably due to the use of reinforcement learning in fine-tuning its abilities.
Although it is way too early in the game yet to say that Project Strawberry will run all the way to its taut, towering ambitions, it certainly has something of a sea-change moment about it for AI research. The work, by shifting attention to how best to improve the reasoning powers of AIs, pushes at what is possible with AI toward systems that not just process but think and reason. This can potentially revolutionize AI applications to domains as diverse as scientific research and common problem-solving, ushering in an age of change in the development of artificial intelligence.