🍓OpenAI's 'Strawberry' Secret Exposed

Decoding project 'Strawberry'—OpenAI's new reasoning breakthrough

Key Points 

• The project aims to enable the company's models to do more than just answer questions; it seeks to allow their AI to autonomously browse the internet and perform what OpenAI calls 'deep research.'

• According to a source, Strawberry involves a specialized approach to fine-tuning AI models after they have been initially trained on vast datasets.

☕News - OpenAI is developing a new approach to its AI models under a project called "Strawberry," according to a source and internal documents seen by Reuters. 

As per the document, the project aims to enable the company's models to do more than just answer questions; it seeks to allow their AI to autonomously browse the internet and perform what OpenAI calls 'deep research.' The source noted that OpenAI plans to significantly improve its AI models' reasoning abilities with Strawberry, explaining that it involves a specialized approach to fine-tuning AI models after they have been initially trained on vast datasets.

🐴From the horse's mouth - When asked about Strawberry and the aforementioned details, an OpenAI spokesperson stated that the company's goal is for their AI models to perceive and understand the world more like humans do. They noted that ongoing research into new AI capabilities is standard in the industry, with the expectation that these systems will improve in reasoning over time, without directly addressing questions about Strawberry.

🍰What's more? Notably, the Strawberry project was previously known as Q*. Earlier this year, two sources witnessed demos of Q* where OpenAI staff claimed it could handle complex science and math questions beyond the capability of current commercial models. Another source familiar with the project mentioned that OpenAI has internally tested AI achieving over 90% accuracy on a competitive math dataset.

Furthermore, Bloomberg reported that during an internal all-hands meeting last Tuesday, OpenAI presented a research project demo that it claimed had new human-like reasoning skills. 

Study: AI boosts individual creativity, reduces collective diversity

Key Points 

• Evaluations showed significant improvements in writing quality and novelty for less creative individuals using AI-generated ideas, compared to those writing independently.

• However, access to generative AI made the resulting stories more similar and less varied within their respective categories. 

• Interestingly, participants with higher creativity metrics did not experience similar improvements with AI assistance. 

🤓News - A new study conducted by researchers from University College London and the University of Exeter has found that while AI can be of great assistance to less naturally creative individuals, this improvement comes at the cost of it dampening overall variety within the group.

The experiment involved hundreds of participants tasked with writing short stories suitable for a broad audience. Different groups were given varying levels of access to GPT-4 for story ideas: one group wrote without AI assistance, another received one AI-generated story idea, and a third could select from up to five AI-generated story starters. Afterward, both the participants and an independent group evaluated the stories, unaware of the AI involvement.

Evaluations showed that individuals with lower creativity metrics received poorer ratings when writing independently. This group significantly improved in writing quality, enjoyability, and novelty with AI-generated story ideas, especially with multiple options. However, participants scoring higher on creativity metrics did not see similar benefits; their writing quality either plateaued or slightly decreased with AI assistance.

💡What's more?  In addition to the evaluations, the researchers conducted their own analyses using OpenAI’s embeddings API. This allowed them to measure how similar each story was to others within its category (human-only, one AI option, or five AI options).

They found that access to generative AI made the resulting stories more similar and less varied within their respective categories. The difference in similarity among the stories ranged from 9% to 10%, indicating some level of uniformity but not complete similarity.

🐙What's the conclusion - The study, albeit focused on short stories, appears to validate a common sentiment: although AI can enhance creativity on a personal level, it may not introduce fundamentally novel elements to creative pursuits. In essence, the findings imply a cautious approach towards widespread adoption of generative AI in creative endeavors, despite its potential to boost individual creativity.