On March 14, I submitted a hugged face response to a request for information regarding the White House Science and Technology Policy’s White House AI Action Plan. Using this opportunity, we advocated a fundamental role in opening up AI systems, allowing technology to be more performant and efficient, broadly and reliably adopted, and meeting the highest standards of security. This blog post provides a summary of the answers. The full text is available here.
Context: (strongly) Don’t sleep with the open model feature
An open approach to AI development is not only more transparent, adaptive and scientifically sound than (usually) but consistently replicates or outweighs the performance of API only commercial offerings that are widely used in many tasks. And we are doing so more and more with shorter timelines, increasing resource efficiency. Our team’s recent Olympic coders outperform Claude 3.7 for complex coding tasks with 7B parameters and post-open source training recipes, or O1-MINI performance with fully open OLMO 2 models of AI2 (using open training data). These successes demonstrate that robust AI strategies must leverage open and collaboration development to optimally drive technology performance, adoption and security. We will create three main recommendations in this direction:
Recommendation 1: Recognize open source and open science as the basis for AI success
The most advanced AI systems to date can withstand the powerful foundations of open research (attention mechanisms, transformer architectures, cheap post-training algorithms) and open source software (Pytorch, Hugging Face Libraries, Supercomputer Operating Systems). Investment in freely reusable and adaptable systems has also been shown to have a strong economic impact increase effect and to promote a significant proportion of the country’s GDP. As AI systems with open weights and training techniques become increasingly attractive options for developers in both performance and cost terms, prioritizing public research infrastructure and broad access to computing, customizable models, and reliable open data sets are essential for the further technical and economic success of AI technologies.
Recommendation 2: Prioritize efficiency and reliability to unlock a wide range of innovations
Addressing the resource constraints of organizations that adopt and adapt AI technology is essential to fostering spread and innovation from employers across the development chain. Small models (which can also be used with edge devices), techniques to reduce computational requirements in inference, and efforts to promote medium-sized training in organizations with modest to moderate computational resources all support the development of models that meet high-risk settings, such as healthcare, where complete, general models have proven reliability. A more efficient and purpose-built AI system will allow in-context assessment, better resource utilization, and organizations to build technical capabilities at every stage of the AI development chain, allowing all users to take advantage of the system that best suits their needs.
Recommendation 3: Make AI safe through open, traceable, transparent systems
Finally, when decades of information and cybersecurity of open source software are indications, open and transparent AI systems play a fundamental role in ensuring AI development and deployment, especially in the most important settings. A fully transparent model that provides access to training data and procedures can support the widest range of safety certifications. Open infrastructure and open source tools Implementing the latest training technologies allows organizations to train the models they need in a fully controlled environment. Openweight models that can be implemented in air gap environments are a critical component in managing information risk. Prioritizing adoption of the most transparent systems, supporting the development of outlined open resources, and building the ability to leverage them in particular in critical settings for AI adoption is essential to enabling safer AI adoption.
See the full response for more detailed recommendations!
