Report: OpenAI Could Release New 'Strawberry' AI Model This Fall

0:00

/1861

1×

Facts

• Media reports have revealed new details regarding OpenAI's forthcoming model, codenamed Strawberry, and its next large-language model (LLM) Orion. Strawberry — capable of more complex reasoning — is reportedly being pushed for ChatGPT integration this fall.^[1]
• The existence of Strawberry, formerly known as Q*, was first reported by Reuters this July, with unnamed OpenAI sources saying it could handle complex math and science problems beyond what their current products are able to solve. Being able to plan ahead and browse the internet autonomously are reportedly key targets for Strawberry.^[2]
• In addition, Strawberry is said to be used to generate training data for Orion, the company's next planned LLM, a class of AI products that includes ChatGPT. High-quality synthetic training data, which is not made by humans, would help rectify LLM 'hallucinations,' which occur when a model generates false information.^[3][1]
• In the weeks since the first leak, OpenAI chief Sam Altman has posted images of strawberries on his social media in a seeming reference to the project. Reports claim that Strawberry would be a 'level two' product out of the five steps OpenAI believes are needed to reach artificial general intelligence (AGI), which would have human-like reasoning.^[4][5]
• Most recently, a report from The Information claims that OpenAI has demonstrated Strawberry's capabilities to US national security officials. Internally, Strawberry is also reportedly able to solve word puzzles, generate subjective answers to questions, and solve logic and math problems that it hasn't yet encountered.^[1]
• Strawberry has reportedly scored 90% on a math problem dataset, whereas GPT-4 scored only 53% before upgrades. OpenAI may include a simplified version of Strawberry in ChatGPT in the near future, while Altman said in May that they have enough data to produce their next model.^[6][1]

Sources: ^[1]Lesswrong, ^[2]Reuters, ^[3]Research & Development World, ^[4]Tom's Guide, ^[5]X and ^[6]Techzine Global.

Narratives

• Narrative A, as provided by Windows Central. The news of Strawberry's accelerated development, coupled with the precarious financial situation of OpenAI, gives us much reason to worry. In order to secure funding, the company may be rushing a problem without proper safeguards, and could unleash the most powerful AI model the world has yet seen with little warning. We must scrutinize OpenAI's practices closely, as even former employees say that safety has taken a backseat to profit.
• Narrative B, as provided by New York Times. AI is facing a reckoning, though one regarding its deficiencies as opposed to its potentially life-altering abilities. The truth is that AI has become very good at producing garbage, and significant investment is needed to get it past this hump, according to projections. The claims of AGI posing an existential risk are much exaggerated, and we need to keep pushing forward with new products, like Strawberry, that get us closer to AI being useful for humanity at large.

Predictions