Pretty easy to do that, use a fixed point implementation of GPT(N) of whatever size you like and range code your data into the model probabilities. This also w

Pretty easy to do that, use a fixed point implementation of GPT(N) of whatever size you like and range code your data into the model probabilities. This also will achieve a close to rate optimal embedding-- allowing you to embed about as much data as the language model thinks the text has... If you encrypt the data and include a checksum or other identifying bytes in the ciphertext you can even have unwitting human participants in the discussions and if their posts are context your embedded data will be credible replies. You just have to be sure that threading behavior doesn't make it impossible to give the decoder identical context.