We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 15·5 min read
Show HN: Can Europe train a frontier AI model on the compute it owns?
TL;DR
Europe faces a challenge in developing a sovereign frontier-class AI model due to delays in establishing new data centers. By utilizing existing public compute resources, such as EuroHPC supercomputers, and employing low-communication (DiLoCo-style) training, Europe can create a competitive AI model.
✦ Why It Matters
Engineers can leverage existing compute resources to accelerate AI model development while awaiting new infrastructure.
Key Takeaways
- Explore federated training methods for AI development.
- Utilize existing public compute resources effectively.
- Anticipate faster AI model deployment timelines.
- Consider the implications of infrastructure delays on AI projects.
Full Summary
Europe is exploring the feasibility of training a frontier-class AI model using its existing public computing resources while waiting for new data centers to come online. The research indicates that Europe has access to tens of exaflops of AI compute power through EuroHPC supercomputers and national AI Factories.
By implementing a federated training approach known as low-communication (DiLoCo-style), which minimizes data transfer between nodes, Europe can effectively utilize its current infrastructure. The findings suggest that this method could produce a frontier-class AI model by 2028, compared to a projected 2033 for new gigawatt data centers.
This strategy not only addresses immediate AI development needs but also leverages existing investments in public computing. The implications for engineers and researchers include the potential for faster AI advancements without waiting for new infrastructure.