Show HN: TweetMirror - Train a LLM on your Tweets

Hi HN,

Tweet Mirror is a toy project I made to learn about Large Language Models. With the speed that the field is advancing, I felt I'd be remiss if I didn't learn more about it.

To use it, you log in with Twitter, pay the amount it costs to train the model (GPUs are expensive otherwise I'd release this for free!) and wait 15m to an hour for fine-tuning to receive tweets generated in your style.

Technically, I fine tune the GPT-J 6B parameter model for a heuristically determined number of epochs with early stopping. I also perform a lot of data preprocessing/sanitization which improves results. In addition, I perform topic modeling using a pipeline similar to BERT-topic and condition some % of your tweets on common topics as well as what is more liked.

With this project, I wanted to engender conversation around a couple of topics that came to mind while building it. Curious for all of your thoughts and feedback at large!

- Economics and defensibility of building AI. I wanted to release this for free but had to charge because it costs a non-trivial amount per job. If fine-tuning models / training models for users will be valuable in many applications, how will this shape the business models of companies that use these technologies? Is there something intrinsic about the cost structure that gives benefit to incumbents?

- Dangers of "loose" AI agents. There is nothing stopping a bad-actor from doing what I have done here and letting it loose on the internet, perhaps even with RL as feedback. Thus creating a seemingly real person that runs on its own. Countries such as Saudi Arabia have already used twitter bots for their agenda, and this technology seems much more powerful than what they used. Should we regulate AI agents being released "into the wild". How do we navigate AI that has convincing agency on the internet, which feels very feasible in the near-term?

- There is something really spooky about the generated tweets because they often share your vernacular and sound quite similar to you. I think this raises interesting philosophical questions. What does it mean to be human in a world where your thought process can be git-forked?

Curious what you all think, and how your Tweet Mirror's make you feel!