Rewriting your emails with AI to be
 

Transparency - Our Costs

PolitePost was never intended to be profitable, but I also didn't expect it to become so popular and cost so much. We ask for coffee donations to keep the site running. We're very thankful to everyone who donates, and your generosity has kept my personal contributions within reason. What we offer in return is transparency.

PolitePost uses Together AI's inference to run. Their pricing is located at: https://www.together.ai/pricing. Our web hosting is negligible. Since PolitePost started in December 2022, our monthly costs with Together (and historically Anyscale, OpenAI, and Microsoft) are as follows:

The charge per email depends on the selected model and API. We currently use Together.ai because cheaper AI models have finally reached a point where they are good enough and are substantially cheaper than the GPT models we used before. In late March 2024, we decided to start using a mix of several models on Anyscale's platform, which is roughly a sixth of the cost. The models we are using are meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo, Llama-3-8b-chat-hf, zephyr-7b-beta, and NeuralHermes-2.5-Mistral-7B. In June 2024, we had to abruptly change to Together.ai because Anyscale's platform was undergoing unexpected changes. We leveraged the same models and made a fairly painless change. Prior March 2024, we used the gpt-3.5-turbo model paired with Microsoft's Chat API to produce rewrites. Open source and lower models did not generate good results at the time even if they were substantially cheaper and acceptable for some use cases. Prior to March 2023, we were using the da-vinci-003 model on the OpenAI Completion API. It was 20x our current price but performs very similarly in our tests. Between March and July 2023, we also opened an additional page to use the GPT-4 model. GPT-4 is a substantial 30x our current costs, and frankly, did not perform very well in comparison. We stopped using GPT-4 entirely in July 2023.

To add to the confusion, AI providers charge per "token". The word count of your original email and the rewrite are added together and converted to tokens. It is roughly 4 characters a token, or about 75 words for every 100 tokens. The longer your email, the higher the charge. We've limited your submissions to 2,000 characters to keep prices low and ensure everyone has a fair chance to use PolitePost.

Transparency - Privacy and Your Data

PolitePost has ZERO interest in your data, including your email content. However, we use Anyscale's Endpoint APIs to run our service. We have a pretty short privacy policy available for those looking for some assurance. But we want to just share our practices here in plain English.

Our Tech Stack

PolitePost essentially is a front end webpage that sends API calls directly from your computer to Anyscale. We run it through our own backend to prevent abuse. We do not cache or store any of your data, we simply act as a middle man and pass it through. Our backend is only there to filter some obvious hate messages, trim down on longer emails, and throttle usage if there are too many requests going at a time.

Microsoft

We take your privacy seriously. PolitePost switched from providers from OpenAI LLC to Microsoft Corporation on August 1, 2023 and from Microsoft Corporation to Anyscale, Inc. on March 20, 2024. We took the time to review the Terms of Use and Privacy Policy offered by Anyscale, Inc. and found similar guarantees as Microsoft, which are better than OpenAI. Your data will NOT be used to train or fine-tune models. While we still recommend you do not use PolitePost for top secret, highly sensitive information, we have done everything we can as a small site to ensure your data is private.

Web Analytics

During your use of PolitePost, we don't require any logins or credentials. We don't keep track of any of your identifiers, however, we do use analytics on our platform. We specifically chose Cloudflare Web Analytics to give us very basic data from a vendor who stresses privacy-first. They claim that their model is not driven by ad revenue, and that they don't track anything to do with you such as IP address or other identifiers. From what we've seen, we're basically able to see the number of views and what country the user is from.