I tested LLMs using Caesar cipher prompts, compiled a list of cheap cloud GPU services like Runpod, and learned how JSR handles package documentation. I also found that averaging embeddings is useful for processing long document inputs.
A few carefully chosen prompts—especially compact logic puzzles and a Caesar-cipher question—can quickly separate ordinary models from the strongest reasoning-capable LLMs.
I discovered E2E for cheap Indian GPU hosting, explored XML tags for better LLM prompting, and tested tools like Jupyter Lite and VoidEditor. I also learned about Ollama's concurrency and animating faces with Segmind's Hallo.
I explored CSS nesting, AWS Lambda performance patterns, and how rain improved WiFi signals. I also learned about AI package hallucinations and summarized Paul Graham’s insights on using low standards and iteration to improve writing quality.
I discovered how LLM planning time rivals massive parameter increases, compared text-to-speech pricing, and tested DuckDB's function chaining. I also explored Deno 2's Node compatibility, Marimo notebooks, and efficient Python Docker builds using uv.
I investigated OpenAI's tactics for system prompt compression and recursive book summarization. I also explored llm-guard for output validation and compared using Google Docs versus email for facilitating collaborative commentary on long-form essays.