Even as it is reportedly set to spend $7 billion on training and inference costs (with an overall $5 billion shortfall), OpenAI is steadfastly seeking to build the world’s first Artificial General Intelligence (AGI).
Project Strawberry is the company’s next step toward that goal, and as of mid September, it’s officially been announced.
What is Project Strawberry?
Project Strawberry is OpenAI’s latest (and potentially greatest) large language model, one that is expected to broadly surpass the capabilities of current state-of-the-art systems with its “human-like reasoning skills” when it rolls out. It just might power the next generation of ChatGPT.
What can Strawberry do?
Project Strawberry will reportedly be a reasoning powerhouse. Using a combination of reinforcement learning and “chain of thought” reasoning, the new model will reportedly be able to solve math problems it has never seen before and act as a high-level agent, creating marketing strategies and autonomously solving complex word puzzles like the NYT’s Connections. It can even “navigate the internet autonomously” to perform “deep research,” according to internal documents viewed by Reuters in July.
During development, o1-preview was given a qualifying exam for the International Mathematics Olympiad, one that the previous GPT-4o model was only able to answer correctly with 13% accuracy. The new model answered 83% of the test’s questions. In terms of coding, o1-preview scored in the 89th percentile in an online Codeforces competition. o1 can even reportedly answer questions that stumped previous models like, “which is bigger, 9.11 or 9.9?” and “how many ‘Rs’ are in the word ‘Strawberry’?”
The Reuters report also notes that Strawberry’s architecture is similar to the Self-Taught Reasoner (STaR) technique. Developed at Stanford in 2022, STaR enables a model to generate training data on which to fine-tune itself, becoming more capable over time.
Why is it called that?
We don’t know the exact reason for the name “Strawberry,” as that’s not something OpenAI has publicly disclosed. It’s a code name chosen for internal reference and to maintain secrecy during development.
This was originally hinted at from a cryptic X post by CEO Sam Altman, which sent the internet into a flurry of rumors and speculation.
i love summer in the garden pic.twitter.com/Ter5Z5nFMc
— Sam Altman (@sama) August 7, 2024
Strawberry did not, however, start out as Strawberry. It reportedly used to be known as Q* (pronounced Q-Star), and when it was, Q* was the linchpin of CEO Sam Altman’s brief ouster last November.
OpenAI researchers specifically cited Q* in their letter to the company’s board decrying the potential risks posed by unregulated advanced AI. After Altman had consolidated power upon returning to OpenAI, Q* was reportedly renamed Strawberry in July 2024.
When will Strawberry be released?
OpenAI launched a preview of Project Strawberry in September 2024, dubbed “o1-preview.” It was accompanied by a smaller, lighter-weight version of the new model, “o1-mini.”
At launch, the new model iteration was made immediately available to ChatGPT Plus and Teams subscribers and arrived for Enterprise and Edu users the following week. OpenAI claims that o1-mini will eventually be made available to free-tier users, though the company did not specify a timeline for that to happen.
Wait, what was that about it being able to autonomously navigate the internet? Isn’t that how we got Ultron?
No, we got Ultron because of comic book legends Roy Thomas and John Buscema, but the fears of an advanced AI growing out of human control are no longer entirely unfounded fantasy. Multiple sets of ex-OpenAI researchers have spoken out against the company’s efforts to develop an AGI, citing its lack of developmental safeguards.
However, those complaints have done little to slow OpenAI’s drive to develop an AGI within the next decade. The company recently released a five-tier scale for measuring the capabilities of the AI systems it builds toward that goal and is currently seeking additional funding (yes, besides Microsoft) in a new round led by Thrive Capitol that values the company in excess of $100 billion.
