Figure drops OpenAI in favor of in-house models

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. Figure drops OpenAI in favor of in-house models

Figure AI, a robotics company aiming to bring a general-purpose humanoid robot into commercial and residential use, has announced it is exiting a deal with OpenAI. The company has opted to focus on in-house AI due to a "major breakthrough."

OpenAI has been a longtime investor in Figure, with the two companies partnering in a deal last year. Figure has raised $1.5 billion from investors. OpenAI has notably hedged its bets in the humanoid space, with the company filing a trademark application with the USPTO involving humanoids.

Read the full story here

2. How AI could make “self-healing” roads a reality

Researchers from King's College London and Swansea University have developed a new type of asphalt that can repair its own cracks over time, eliminating the need for manual maintenance. The innovative material mimics the self-healing abilities of trees and some animals, aiming to create more durable and sustainable roads.

The research, supported by Google Cloud's artificial intelligence, aligns with the UK government's net-zero emissions goals and addresses the financial burden of road maintenance.

Read the full story here

3. OpenAI and SoftBank Group Partner to Develop and Market Advanced Enterprise AI

OpenAI and SoftBank Group have announced a partnership to develop and market Advanced Enterprise AI called "Cristal intelligence". Cristal intelligence will integrate the systems and data of individual enterprises in a customized way, making it the first company in the world to integrate Cristal intelligence at scale.

SoftBank Group Corp. will spend $3 billion US annually to deploy OpenAI's solutions across its group companies, making it the first company in the world to integrate Cristal intelligence at scale. To accelerate the deployment of Cristal intelligence customized for Japan-based companies, OpenAI and the SoftBank Group agreed to establish a joint venture (JV) company called "SB OpenAI Japan." This JV will market Cristal intelligence exclusively to major companies in Japan, enabling clients to securely train their own data and build AI agents fully integrated with their own IT systems.

Read the full story here

4. Constitutional Classifiers: Defending against universal jailbreaks

The Anthropic Safeguards Research Team has developed a method to defend AI models against universal jailbreaks. The method, called Constitutional Classifiers, is based on synthetically generated data that filters the majority of jailbreaks with minimal over-refusals and without incurring a large compute overhead.

The system was tested under human red teaming, with 183 active2 participants spending an estimated >3,000 hours attempting to jailbreak the model. Despite its robustness, the prototype system had some problems, such as refusing too many harmless queries and costing a lot of computational resources. Automated evaluations showed that the Constitutional Classifiers system significantly improved the robustness of the AI model against jailbreaking, with only minimal additional cost. The team is currently working on reducing refusals and compute costs as they refine the technique.

Read the full story here

5. Unsupervised People’s Speech: A Massive Multilingual Audio Dataset

The MLCommons Dataset working group has released the Unsupervised People's Speech dataset, containing over 1 million hours of audio across dozens of languages. The dataset aims to support research and development in various areas of speech technology, including Natural Language Processing research for languages other than English.

The dataset was created through various data pipelines, including speech detection and language identification. Technical challenges include data upload and storage, self-supervised learning potential, and deduplication. The dataset is built from audio data on Archive.org and is open-source to enable community development.

Read the full story here