A pattern for building personal knowledge bases using LLMs.
This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.
Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.
When using the JBL Flip 3 (or JBL Flip 2 or JBL Flip 4) as a permanent speaker for music or Text-to-speech announcements, connected through the AUX input and powered consantly it still goes into an automatic standby after some time of silence and it's a major problem in this use case. Depending on how long the silence was and how loud the new input audio is, it may take between 3 to 30 seconds in my experience before it resumes playing. For lack of a better place I'm posting here the information I have on solving this.
- Obviously emitting the audio signal at maximum volume from the computer and lowering the volume at the speaker reduces the time for it to wake up. Downsides: less volume control for the computer (especially problematic if it's remotely controlled), still won't work if the audio played is relatively quiet or short (such as a home-automation announcement) so not a definitive solution.
- Mixing unaudible waveforms into the audio output, either low frequency (< 20Hz) or high (> 15kHz). Downs
Due to the massive toxicity of hexavalent chromium and its prohibition in the EU, photographers have searched for a number of years for alternatives to potassium and ammonium dichromates in the gum dichromate and carbon transfer processes. Additionally, in gelatine-based processes like carbon transfer, there is a "dark reaction" where dichromate salts can fog the emulsion.
There are basically three major lines of inquiry: one is the Chiba system and variants, and the latter two are photoreactive azide dyes which act as more or less direct substitutes for dichromate.
The Chiba System uses a ferrous photochemistry to initiate polymerization, exploiting photosensitivity of iron salts used in cyanotype, vandyke, and Pt/Pd processes. It is thorougly documented in a paper available online for carbon and gum printing: