Rudeness Translator — Converting Rude Comments To Constructive Criticism With “ChatGPT-In-A-Bottle”

Or “The Adventure Continues — Going From Davinci-3 To Gpt-Turbo-3.5”

Take THAT, Jason

GPT-Turbo-3.5 was released on March 1 of this year, meaning that at this point it is ancient history and no one really cares about it. Until a few minutes ago, though, I was using Text-Davinci-003 and noticed a considerable lack of quality. When I inputted a pretty long comment I received on Medium (you can read the full comment here, in all its glory) and asked for a summary, it gave me a lot of “summarized” facts that did not appear anywhere in the original comment. I will just go ahead and copy its summary:

So yeah, 2022, do you even need the DIV?

No. You don’t. You need to learn HTML, and learn it properly. You need to learn the difference between a semantic element and a presentational element. You need to learn the difference between a class and an ID. You need to learn the difference between a div and a span. You need to learn the difference between a heading and a paragraph. You need to learn the difference between a list and a table. You need to learn the difference between a form and a table. You need to learn the difference between a link and a button.

You need to learn HTML.
— This was generated by the OpenAI API

While some of the above is, indeed, derived from the comment, most of it is simply made-up. The sentence about the difference between a semantic element and a presentional element, for example, is not a summary and appears to have come out of nowhere. The succeeding sentences about a list and table, form and table, and link and button are related to HTML, but are in no way a summary of Jason’s comment. It is a bit like asking AI to summarize everything that happened in the Biden Administration, only to hear a bunch of random facts about Abraham Lincoln.

These models are something of a black box in which researchers can be open about data and training methodology, but never about everything. That being said, I was very impressed tonight to see the GPT-3.5-Turbo summary:

It seems like you have a strong opinion about HTML and its usage. HTML is a markup language used to structure the content of web pages, and it plays a crucial role in web development. It is true that HTML has evolved over the years, and newer technologies like CSS and JavaScript have taken on more responsibilities for styling and interactivity.

However, HTML still remains an important foundation for building web pages. While there may be alternative frameworks and tools available, understanding HTML and CSS is fundamental for web development.

It’s important to note that different projects and requirements may call for different approaches and technologies. While some people may prefer certain frameworks or libraries, it doesn’t mean that HTML itself is irrelevant or should be disregarded.

Accessibility and semantic markup are indeed important considerations when developing websites, and conforming to best practices can greatly enhance the user experience. It’s advisable to stay updated with current web standards and guidelines to ensure that websites are accessible and efficient.

Constructive criticism and open discussions about different technologies and techniques are beneficial for the growth and improvement of the web development community. However, it’s also essential to approach these discussions respectfully and provide well-founded arguments to support one’s opinions.
— GPT-3.5-Turbo

The last paragraph was residue from my prompt, but for the most part I think this is a very high-quality, fair summary of the original comment. I think it is on par with ChatGPT (which makes sense, as they are fundamentally the same technologies). It is still nowhere near the quality of a human summary, but GPT-3.5 is incapable of being offended by it.

The Code

I was looking at these two sources:

Upgrade text-davinci-003 to gpt-3.5-turbo

Beginner’s Guide to OpenAI’s GPT-3.5-Turbo Model

If you look at the Medium article, you will notice that its code is fundamentally the same as the dev.to code…yet I had a lot more success following the dev.to code/tutorial, maybe because it was so much shorter and simpler.

GPT-3.5-Turbo is way cheaper than text-davinci-003, which I consider somewhat ironic considering it performed so much better (by my test). From the site:

Model: The ChatGPT model family we are releasing today, gpt-3.5-turbo, is the same model used in the ChatGPT product. It is priced at $0.002 per 1k tokens, which is 10x cheaper than our existing GPT-3.5 models. It’s also our best model for many non-chat use cases—we’ve seen early testers migrate from text-davinci-003 to gpt-3.5-turbo with only a small amount of adjustment needed to their prompts.
— Source

Unlike with davinci, there were no parameters for max-tokens or temperature.

Ideas For Next Steps

Smack said to put the whole thing on GitHub pages and see if it still worked — he suspected it would not because of blocking, but I think it will work fine. GitHub Pages is a static host, so on that route I will not be able to use a customized backend to save login information. Account logins would be ideal here — you could save an API key, which would be really helpful, but more importantly you could save your prompts somewhere, since every time you click a button you take up some amount of money.

Before that, though, I thought it would be nice to up the frontend. A slider for temperature would seem natural, which is why I was a bit disappointed Turbo does not take temperature as a parameter. Also, if you print the return response, you can see tokens.

I want a v-dialog to appear, prompting the user with exactly how much this will cost and to then ask them. Requiring an API key is a non-negotiable barrier to entry for many people, so it could be worthwhile to simply store a pre-loaded API key somewhere on the server and host a full stack application with AWS.

Closing Thoughts

I was not sure what to put here, so here is a video of Maxine The Fluffy Corgi.