Yet, AI sometimes stumbles over the simplest of tasks—like counting the letter "r" in the word "strawberry." Several top-of-the-line AI models like GPT-4o and Claude, have claimed that there are only two "r"s in "strawberry," when there are actually three.

Apart from this embarrassing error, the red colored fruit has been making headlines for more than just a humorous mishap.

OpenAI Uses Strawberry To Develop Orion

OpenAI is developing a new model called "Strawberry" which would prove to be training wheels for the company’s next big AI model called ‘Orion’--touted as the successor to GPT-4o (the company’s most advanced model yet), as per a report by The Information.

Orion will be based on the Strawberry technology (also being called GPT-5). But what sets Strawberry apart? Its key feature is its ability to create high-quality synthetic data, which helps tackle the issue of limited training data available online.

Data scraping, and the scarcity of accessible data, has become a major challenge for AI companies attempting to train their models from the internet and other sources.

Afterall, generative AI models need scores and scores of data to generate meaningful and accurate outputs.

OpenAI’s AI models have come under fire for scraping data off the internet without permission. Author George R. R. Martin and comedian-actress Sarah Silverman have sued the company for copyright infringement in the past.

So, the goal here is to not only avoid lawsuits by using the Strawberry synthetic data, but to also cut down on the mistakes or "hallucinations" that often happen with AI.

It is reportedly believed by the team at OpenAI that this should improve Orion's performance across fields like healthcare, banking, and scientific research.

There is however a problem with using synthetic data to train AI models: it may lead to a ‘model collapse,’ which happens when AI repeatedly learns from its own outputs, leading to degraded quality, distorted images, and misinformation.

To avoid this, AI needs fresh, real data to prevent drifting away from reality.

OpenAI’s CEO Sam Altman revealed they generate 100 billion words daily, some of which end up online.

From Project Q Star to Strawberry

The story gets interesting when you look at the backstory of "Strawberry." While the name sounds innocent, this AI model was initially known as Q* (pronounced Q-Star) and played a big role in some chaos at OpenAI last year—when Altman was ousted and then quickly brought back.

In an interview with podcaster Lex Fridman, when asked about Q Star, Altman had said that there is nothing to talk about.

But, a couple of days ago, he hinted at Strawberry's existence with a cryptic tweet showing a pot of strawberries, adding to the mystery and speculation surrounding the project.

However, the development of Strawberry has been somewhat secretive, with much of the information coming from internal leaks and hints. 

Reportedly, Altman’s ouster was partly due to concerns over a major AI breakthrough involving Q*. Some researchers at OpenAI even wrote to the board, warning about the risks of advanced AI and specifically this model, which is now called Strawberry.

There were claims that Q* might be a big step toward artificial general intelligence (AGI)—an AI that can think, learn, and understand like a human across many different tasks. The worry with AGI is that it could potentially get out of control, develop its own goals that don’t align with human values, and cause serious problems.

OpenAI is tracking its progress toward artificial general intelligence (AGI) using a five-level scale. Right now, they believe they’re getting close to level two–Reasoning AI.

OpenAI Shows Strawberry To US Govt

Given the atmosphere in the US around bringing a stringent AI regulation law, OpenAI has open agreements with US federal agencies to provide early access to upcoming AI models, including Strawberry, reported The Information. 

Confirming the same, Altman posted on X:

This move is part of OpenAI's strategy to work closely with government bodies, like the US AI Safety Institute, to ensure AI advancements align with national security and safety standards. 

After facing criticism for neglecting safety research, OpenAI has strengthened its focus on AI safety by forming a Safety and Security Committee and bringing in experts to guide these efforts, dedicating significant resources to responsible AI development.

This is going to be an important year for OpenAI, which is also reportedly in talks to raise a big chunk of cash, led by Thrive Capital, pushing its valuation over $100 billion—way up from its previous $86 billion. 

Thrive might chip in around $1 billion, and apart from the usual suspect Microsoft, it has been suggested that Nvidia and Apple might also join the round. 

This would be OpenAI’s biggest cash boost since Microsoft’s $10 billion investment in early 2023. 

Despite generating over $3.4 billion in annual revenue, OpenAI is expected to lose nearly $5 billion by the end of this year and has already burned through $8.5 billion on AI training and staffing.

These financial strains reflect the high costs of developing cutting-edge AI technologies, as seen in OpenAI's rapid rollout of increasingly advanced models.

They released GPT-3 in 2020, GPT-3.5 in late 2022, GPT-4 in early 2023, and GPT-4o model this year. Both Strawberry and Orion are expected to be launched sometime in the fall or autmn as we Indians call the season.