šŸ”€ OpenAI's Strategy Shift Amid GPT AI Improvement Slowdown

OpenAI explores new strategies to enhance AI models as performance improvements slow, signaling a shift in the AI development landscape.

Author

Jack Lajoie
November 11, 2024 ā€¢ read time ~ 8 minutes

The Daily Current āš”ļø

Welcome to the creatives, builders, pioneers, and thought leaders ever driving further into the liminal space.

OpenAI's strategic shift amid slowing GPT improvements highlights a broader trend towards efficiency and specialized AI solutions. The era of ā€œquality dataā€ may be upon us. Meanwhile, the education sector, exemplified by Chegg's decline due to ChatGPT, continues to face pivotal changes and innovations such as LucidSim's zero-shot transfer for robotics and The Beatles' AI-assisted Grammy nomination underscore the transformative impact of AI across industries.

 šŸ”Œ Plug Into These Headlines:

OpenAIā€™s upcoming model Orion is showing signs of slowing improvement rates compared to previous iterations. While still outperforming existing models, the performance gap between Orion and its predecessors is smaller than the leap observed between GPT-3 and GPT-4. In response, OpenAI is reportedly exploring new strategies to enhance model performance, including post-training improvements and the use of AI-generated training data.

  • OpenAI has created a foundations team to address the scarcity of high-quality training data.

  • Orion was partially trained on AI-generated data, which may lead to resemblance to older models.

  • The industry appears to be shifting efforts to improving models after their initial training, potentially yielding a different type of scaling law.

šŸ’° As the cost-benefit ratio of training increasingly large models becomes less favorable, OpenAIā€™s strategy shift may foreshadow a broader industry trend towards efficiency and specialized AI solutions.

The rise of ChatGPT exposed the vulnerability of Cheggā€™s subscription-based homework help service. As students discovered they could obtain similar assistance for free from the AI chatbot, Cheggā€™s market value dropped from $12 billion to under $2 billion in two years. This dramatic shift illustrates how quickly AI can reshape established business models in the education sector, forcing companies to rapidly adapt or face obsolescence.

  • Cheggā€™s stock price fell by 50% in a single day in May 2023 following the announcement of declining subscribers due to ChatGPTā€™s impact.

  • The companyā€™s market value dropped from $12 billion to under $2 billion within two years (2021-2023).

  • Cheggā€™s subscriber base decreased from 5.4 million to 4.4 million in one year (2022-2023).

  • ChatGPT demonstrated the ability to solve complex math problems and write essays, directly competing with Cheggā€™s services.

  • Chegg attempted to integrate AI into its platform, launching CheggMate, an AI-powered tool in May 2023.

šŸ”„ As Chegg struggles to reinvent itself, the education sector faces a pivotal moment where the line between technological assistance and academic integrity becomes increasingly blurred.

By leveraging generative models to create physically accurate, diverse image sequences, the researchers trained a quadruped robot to perform complex locomotion tasks. The key innovation lies in the robotā€™s ability to transfer its learned skills from simulated environments to real-world scenarios without any prior real-world data exposure. This zero-shot transfer capability demonstrates the potential of using generated images for robust sim-to-real robotics applications.

  • The research, titled ā€œLucidSim: Learning Agile Visual Locomotion from Generated Images,ā€ will be presented at the 8th Annual Conference on Robot Learning in 2024.

  • The robot demonstrates zero-shot transfer to RGB-only observations in the real world.

  • Fast and accurate physics simulation is crucial for robot learning, allowing exploration of failure scenarios.

  • The robot relies solely on RGB camera input, without depth sensors, making it a cost-effective solution.

  • A set of visually complex evaluation environments was created using 3D Gaussian Splatting to test the policyā€™s generalization capabilities.

šŸ§  The success of zero-shot transfer in this study suggests that AI-generated environments may become a crucial tool for developing adaptable and robust robotic systems.

The Beatlesā€™ song ā€œNow and Thenā€ has received a Grammy nomination for Best Pop Duo/Group Performance, becoming the first AI-assisted track to achieve this recognition. The song features John Lennonā€™s vocals extracted from a 1970s demo using AI technology, completed by Paul McCartney and Ringo Starr with contributions from George Harrison. This nomination highlights the evolving intersection of artificial intelligence and music creation, sparking discussions about AIā€™s role in the industry and its eligibility for awards.

  • The Recording Academy has stated that AI-assisted works can be nominated if they meet certain human authorship criteria.

  • The use of AI in ā€œNow and Thenā€ was primarily for audio separation, not for generating new content.

  • The song debuted at No. 1 on the UK Singles Chart, making it The Beatlesā€™ 18th chart-topper.

  • This nomination comes amid ongoing discussions about AIā€™s impact on the music industry and copyright laws.

šŸŖ² Leave it to The Beatles to, even about 70 years now after they became famous, continued pushing barriers and innovating in the realm of music.

As of November 11, 2024, Suno AI has released up to version 3.5. This version allows users to create full four-minute songs, even with a free account. The company has announced that version 4 (v4) is under development, aiming to introduce improvements in vocal quality, prompt adherence, and overall coherence of the generated music.