Twitter Grok AI Large Language Model Released for Free!

---

---

AI Summary

• Introduction to Grok One
  • Open-source language model by X (formerly Twitter)
  • 314 billion parameters, mixture of experts model
  • Not fine-tuned for instruction following
  • Released under Apache 2 license
• Performance Comparison
  • Outperforms GPT-3.5
  • Falls short of GPT-4 and Cloe 2
  • Based on benchmarks
• Grok Chat and Usage
  • Powers the Grok chat on Twitter
  • Available on GitHub and Hugging Face
  • Requires multiple GPUs due to size
• Technical Specifications
  • 64 layers, 48 attention heads for queries, 8 for key/values
  • Maximum sequence length of 892 tokens
  • 8 experts, 2 active per response
• Coding Ability Test
  • Instruction fine-tuned version tested
  • Successfully completed tasks up to “very hard” level
  • Encountered issues with “expert level” challenge
• Model Comparison on Datasets
  • Grok One excels in logical reasoning and math
  • Surpasses LLaMA 270B, Inflection 1, and GPT-3.5
• Final Thoughts
  • Excitement about open-source model capabilities
  • Plans to create more videos testing the model
  • Encouragement to like, share, and subscribe