GPTweets: Generated tweets like yours with a fine-tuned GPT3 model

I love this, but have found it much too easy to view other people's generated tweets.

Because each generated tweet's URL ID is generated from a numerical value iterated by 1 upon each generation, and because there is no auth in front of tweets (likely to make sharing easier), all generated tweets are trivially browsable at moment of generation via manually iterating the URL's ID number by 1.

I.e. I am able to view with no auth:

https://tweets.gptlabs.us/tweet/1000 https://tweets.gptlabs.us/tweet/1001 https://tweets.gptlabs.us/tweet/1002 https://tweets.gptlabs.us/tweet/1003 https://tweets.gptlabs.us/tweet/1004 https://tweets.gptlabs.us/tweet/1005 https://tweets.gptlabs.us/tweet/XXXX.........

It's therefore trivial to download the entire database of generated tweets via script, including many harmful ones, associated with people who would potentially not consent to ever sharing them.

A trivial fix could be to re-generate all tweet IDs using long, alphanumeric, non-iterable, case-sensitive IDs.

Otherwise, a clear warning of this current default publicity would also be good to input at every point of the journey.

My friend Lewis and I worked on bringing this to life last week. GPTweets works by using the Twitter API to read the last few thousand tweets of yours, and then fine-tunes a Davinci model off of your most successful tweets. The end result is hilarious and also surprisingly insightful.

Training fine tuned Davinci models is expensive ($0.50-4 per user), so we are offering this as a paid service. The cheapest option is a one-time payment of $5.99 for 250 tweets.

Please enjoy!