Start with a clear question or topic area. Grok searches across the web, X, and its training data to gather relevant sources. You can also paste in specific documents, URLs, or datasets you want included.
Example prompt
Summarize the latest research on transformer architecture scaling laws published in 2025-2026. Focus on compute-optimal training, mixture-of-experts approaches, and inference efficiency.
Grok organizes findings into a structured document with sections, key claims, supporting evidence, and source attribution. You can ask follow-up questions to drill into specific areas, request comparisons between papers, or ask for a different framing.
Example prompt
Compare the scaling recommendations from Chinchilla, the Llama 3 paper, and the DeepSeek MoE work. Where do they agree and where do they diverge?
Refine the synthesis iteratively — expand sections, add counter-arguments, update with new sources. The final output is ready to share with your team, drop into a slide deck, or use as the foundation for a longer report.
Example prompt
Add a section on the practical implications for a team training a 70B model with a fixed compute budget of 10K H100-hours. What does the research suggest about data mix, learning rate schedule, and architecture choices?