What do we know about ‘Project Strawberry,’ OpenAI’s shadowy new AI project?

i love summer in the garden pic.twitter.com/Ter5Z5nFMc

— Sam Altman (@sama) August 7, 2024

Some are taking this to be a reference to Project Strawberry (formerly Q*), the codename for an ongoing effort that’s reportedly taking place within the company. Citing internal documents and an anonymous source close to the matter, a report from Reuters published last month found that the new project revolved around a system capable of reasoning and long-term planning – both of which are considered by many AI experts to be necessary attributes for artificial general intelligence (AGI), an as-yet hypothetical AI system capable of matching the human brain across any cognitive task.

OpenAI states on its website that its “mission is to ensure that artificial general intelligence benefits all of humanity.”

While today’s LLMs can produce text in natural language, companies still haven’t been able to build systems – often called “agents” – capable of acting autonomously across time, using abstract reasoning to identify, plan towards and ultimately achieve a future goal. Humans have a prefrontal cortex, which allows them to consider multiple possibilities and strategize; building similar abilities into machines remains an unsolved problem, and there’s some debate surrounding whether or not it’s even technically feasible.

OpenAI’s reported efforts with Project Strawberry suggest that the company may be making some progress in this regard, but details are extremely scant and hazy. According to Reuters, the company is working on new mechanisms – such as an ability to engage in so-called long-horizon tasks (LHT) – into Strawberry so that the system might be able to formulate its own working models of the world, as well as shape and execute its own plans.

Bloomberg also reported last month that OpenAI had shared a memo with employees outlining the company’s five-step plan for building AGI. The message reportedly claimed that the company was close to reaching level two: systems that can perform reasoning at a PhD level. The next step was designated for agents, which the company defined as systems that could plan and execute tasks over an extended period of time.

It’s worth bearing in mind that Altman seems to have a penchant for sparking rumors through mysterious tweets. In May, he posted the word “her,” which many people interpreted as evidence that OpenAI had used the voice of Scarlett Johansson for its then-newly released model GPT-4o. (Johansson played the voice of an AI companion in Her, a 2013 film also starring Joaquin Phoenix.) OpenAI denied using Johansson’s voice for the model after she threatened legal action against the company. The voice was ultimately deactivated.

OpenAI declined to comment for this story.

For more on the latest happenings in AI, web3 and other cutting-edge technologies, sign up for The Emerging Tech Briefing newsletter.