Mixture of Experts Monster prompt, OpenAI’s business play, nano-banana and US Open experimentations
Aug 29, 2025
Join Aaron Baughman, an IBM Fellow and Master Inventor, and Chris Hay, a distinguished engineer, as they dive into the fascinating world of AI. They discuss KPMG's ambitious 100-page prompt that powers its TaxBot, revealing the complexities behind effective AI prompts. The duo also explores OpenAI's potential move into selling infrastructure. Get ready to giggle with Gemini's quirky 'nano-banana' image generation tech and learn about innovative features at the US Open, like real-time match stats and probabilities, all aimed at enhancing fan engagement.
AI Snips
Chapters
Transcript
Episode notes
KPMG’s 100‑Page TaxBot Prompt
- KPMG built a TaxBot powered by a 100-page prompt to aggregate firm tax expertise and draft advisory opinions.
- Aaron Baughman said he once pasted a 40-page manual into a model to get effective summaries.
Prompt Engineering Isn’t Dead
- Prompt engineering remains important because complex agentic behaviors often need extensive specification.
- Lauren McHugh and others argued fine-tuning and retrieval can reduce prompt complexity over time.
Use Context Window For Reliable Outputs
- Put domain rules and example structure into the context window to get reliably formatted outputs for long documents.
- Prefer context injection for quick iterations instead of expensive fine‑tuning when users need fast feedback.