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·27d ago·5 min read
Grok 4.5
TL;DR
Grok 4.5, SpaceXAI's latest model, was developed to excel in coding and knowledge work. It was trained on extensive datasets using tens of thousands of NVIDIA GB300 GPUs, focusing on intelligent reasoning and efficiency.
✦ Why It Matters
Engineers can adopt Grok 4.5 for complex coding tasks to enhance productivity and accuracy in their projects.
Key Takeaways
- Utilize Grok 4.5 for coding and knowledge work tasks.
- Implement data filtering techniques for model training.
- Explore reinforcement learning for task-specific improvements.
- Leverage high-performance GPUs for large-scale model training.
Full Summary
Grok 4.5 is SpaceXAI's most advanced model, specifically designed for coding, agentic tasks, and knowledge work. It was trained on diverse datasets covering coding, science, engineering, and math, utilizing tens of thousands of NVIDIA GB300 GPUs.
The training process emphasized data quality through deduplication and domain-focused selection, ensuring a high-signal dataset. Reinforcement learning (RL) was scaled to enhance per-token intelligence, with training on hundreds of thousands of tasks related to software engineering.
This approach allowed for asynchronous training, enabling long-duration agentic rollouts while learning continued across GPUs. The outcome is a model that demonstrates superior reasoning capabilities in both engineering and office tasks, achieving the highest scores on benchmarks like Harvey's Legal Agent Benchmark.