I discovered Hermes 3's special reasoning tokens, Lumentis for transcript documentation, and the 27% reuse threshold for Anthropic’s prompt caching. I also noted that LLMs write better code in Markdown than JSON and explored Copilot's system prompt.
Hermes 3 has special tokens like <SCRATCHPAD>, <RESTATEMENT>, <THOUGHT_*>, <PYDANTIC_SCHEMAS>, <SCHEMA_*>, <REASONING>, <INNER_MONOLOGUE>, <PLAN>, <EXECUTION>, <REFLECTION>, <THINKING>, <SOLUTION>, <EXPLANATION>, <UNIT_TEST>, etc. This extends the capability dramatically.
I discovered how double-checking LLM outputs can slash error rates and compared Anthropic's new search tool pricing. I also found snapdom for element capture, explored Gemini's prompt caching, and documented some prompt evaluation frameworks.
I share my strategies for better LLM usage, focusing on voice interaction and 'impossibility lists.' I also cover Luis Alvarez’s diverse scientific discoveries, compare AI coding assistants, and break down common patterns found in system prompts.
I explored using LLMs for educational games and market research, compared log storage costs across cloud providers, and tested document conversion tools. I also found that NumPy often outperforms HNSW indexing for similarity searches on datasets under 1M vectors.
I explored techniques for managing LLM coding agents, Anthropic's multi-agent architecture, and persona vectors. I also found handy tools like gitingest for repo ingestion, the O*NET database for job analysis, and modern browser APIs for file access.
I explored Agentic RAG for complex retrieval, fine-tuning with LoRAX, and practical LLM strategies. Key takeaways include using N-shot prompting before scaling models, automating workflows via disposable apps, and leveraging context caching to significantly reduce inference costs.