FIDONet Cybernetic Immortality

FIDONet is a computer network popular in 1990s, years before Internet became widely available. It allowed people all over the world to connect with each other using only modems and telephone lines, which exchanged information nightly.

FIDONet had very distinct culture of communication, which is being missed by the generation of people who widely used it in the past. Nowadays, although technically still working over Internet, FIDONet is largely dead. Group discussions were happening in many topical areas called Echos (Echomail), and messages from its members were disseminated throughout the network in the matter of days.

FIDONet Cybernetic Immortality art installation tries to revive the spirit of FIDONet by showing you endless Echomail conversations generated by a GPT neural network trained on subset of FIDONet archives. This project highlights two topics:

  • The spirit of FIDONet itself as a historic pre-Internet computer network of early computer enthusiasts
  • The concept of Cybernetic Immortality, which is a kind of immortality where an artificial system is created, which mimics the behaviour of the original person.

This project shows that FIDONet spirit is alive again, but FIDONet simulacrum that you see is populated not by the living people, but by their cybernetical immortal digital twins, that talk to each other infinitely in the realms of modern datacenters.

The main difficulty when implementing this project was to obtain FIDONet archives, because largely they are being kept on some media that is not easily accessible now (eg. tapes). Current iteration of the project has been trained on rather small subset of FIDONet from

Exhibited At

Technical Description

The model is based on GPT2-large model, and was fine-tuned for 2 epochs on archives of ExecPC BBS, obtained from here. This process took around 9 hours on NVidia A100 compute in Yandex Datasphere service.

Trained fido-gpt model is available from Huggingface.

Demo that you see on this page shows not a real-time generation, but rather messages picked up from a large dataset of pre-generated content. This is done to minimize the carbon footprint, and to ensure more robust demonstration.