August 27, 2024 – OpenAI is reportedly gearing up to launch a groundbreaking AI model named “Strawberry,” which promises to significantly enhance its chatbots’ mathematical and programming capabilities. According to The Information, this advanced model could be unveiled in the coming months, marking a substantial leap forward in AI technology compared to existing chatbot models.
Previously referred to as Q* (Q Star), Strawberry is designed to tackle complex problems that existing models have struggled to address effectively. Internal testing has already showcased Strawberry’s proficiency in solving elementary school-level math problems with high accuracy, marking a significant leap forward as current AI models like ChatGPT still grapple with basic arithmetic.
Strawberry’s capabilities extend beyond mere numerical acumen. The Information’s sources indicate that the AI can also tackle more subjective queries when granted additional time to “ponder.” This feature could substantially expand AI’s potential in addressing intricate challenges.
Furthermore, Strawberry plays a pivotal role in the development of “Orion”, OpenAI’s forthcoming large-scale language model. The larger iteration of Strawberry is being leveraged to generate synthetic data for training new AI models. This approach helps circumvent the hurdles OpenAI faces in acquiring high-quality data from real-world sources. Improvements in data quality could make Orion more reliable and accurate, positioning it as a flagship product in OpenAI’s lineup.
In addition to training other models, Strawberry is being expanded to tackle a range of advanced tasks, including developing market strategies, solving word games like the New York Times’ Connections, and conducting in-depth research into complex topics. This extension significantly broadens AI’s value, moving beyond just business applications to educational realms where deliberation and strategy can be executed with far greater sophistication.
Possible Problems
Although Strawberry brings the promise of improved data quality for AI training, the methods through which this data is being sourced remain shrouded in uncertainty. Recent developments have complicated the landscape for data acquisition, as major content delivery networks (CDNs) such as Cloudflare have introduced options to block AI crawlers at CDN level.
Making the issue worse, there is also the prevalence of AI-generated content already circulating online. The incorporation of such material into AI training datasets can lead to a significant decline in the quality of the resulting models.
The challenge OpenAI faces is now twofold: navigating the new barriers to data collection while also ensuring the quality of the data it does manage to gather, to avoid the pitfalls of AI-generated content contaminating the learning process.
Source: The Information, Tom’s guide
