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
MIT Technology Review·Jun 19·5 min read
A startup claims it broke through a bottleneck that’s holding back LLMs
TL;DR
Large language models (LLMs) have faced a mathematical bottleneck that limits their efficiency and performance. Subquadratic has developed a new LLM called SubQ, which is faster, cheaper, and more energy-efficient.
✦ Why It Matters
Engineers and researchers can explore SubQ for more efficient processing of large datasets in their AI applications.
Key Takeaways
- Evaluate SubQ for cost-effective LLM tasks.
- Monitor developments in sparse attention techniques.
- Consider the implications of reduced energy usage in AI.
- Stay informed about independent evaluations of new models.
How It Works
SubQ employs sparse attention, which selectively multiplies only certain token values instead of all, drastically reducing computational requirements. This dynamic selection process allows SubQ to maintain performance while processing larger text inputs efficiently.