01 / CONTEXT
Conversation-aware carriers
Topics, roles, locations, and phrase choices give each visible turn a conversational setting instead of producing isolated text.
From a broad idea to an inspectable experiment.
The direction changed as each failure became useful evidence: reliability, naturalness, efficiency, and privacy all had to be visible in the same demo.
Create a communication system that uses AI to hide an encrypted message inside an ordinary-looking text message.
The original idea combined a normal conversation, a hidden payload, context-aware language, and recipient-only recovery.
If its not possible to encode the secret message into the carrier, throw an error, dont just append repeated text.
That moved the project away from pretending every carrier was successful and toward explicit capacity checks, validation, and visible failure.
What about using a super lightweight offline LLM that can be configured to do this, then all that's needed is a shared secret such as the seed.
The experiment explored local inference, deterministic generation, probability-based selection, and the practical limits of laptop hardware.
fully implement the phrase bank with openai, I dont want to use the statistical method anymore as its too unreliable
The design then shifted toward structured, inspectable language banks with topic and role control instead of relying on fragile model replay.
Make the messages better by not making it lots of short sentences but instead one legit looking long message?
Recent work focuses on coherent conversation, varied roles, local place choices, server-side state, key handling, and language quality that can be evaluated rather than assumed.
No dont use the API. I want you to generate it and upload it to the server for me.
The phrase-bank path is now designed to be locally reproducible, with the source data and generation process kept available for inspection.