Thread by @yleflour

🏁 On with Part 4 of my #gpt4 saga

Between pricing (I'm cheap as f😵k) and rate limiting, I'm now spending half my time writing

Truth is, I quite enjoy it so let's gooooo 👇

1/15

So at least the #GPT4's API is available for prompting.

But I don't like the playground

👶🏻 I need pairprog-webview, nextjs, tailwind to chat with your own API
🤖 Sure thing
👵🏻 Wow, this API call looks quite outdated and doesn't work
🤖 ermmm...

How do I solve this? 🤔

2/15

Open up platform.openai.com/docs/api-refer…

Copy, paste into @euangoddard ✨magnificent✨ "Paste as markdown page", copy back, paste to #gpt4

👶🏻 Here's the updated doc
🤖 K boss
✂️ > 📋 > ▶️

✨ It works ✨

3/15

But the #chatgpt site response streaming feature is cool, I want it tooooo 😭

Fortunately, there is an issue's answer on the repo regarding this

So I need to copy, paste as markdown, STOP ✋. It's getting tiring

Wait, wait, wait... Guess it's time for more rambling 😃

4/15

One of the highest costs of training an LM is gathering and formatting the training data

But the web is currently made for humans, not 🤖

So while companies working on large ML have the resource to do so, at a consumer level, we don't

Couple of things could happen 👇

5/15

1️⃣ The web starts to develop a "Prompt" presentation layer

Imagine if every site using @docusaurus had a "copy as prompt" button for each section/page

Now humans can fetch and feed updated doc in #chatgpt in a matter of seconds

Right @sebastienlorber 😉 ?

6/15

Your doc doesn't have it yet? Might not use your tool

Even better, if there's a dedicated endpoint to fetch the doc as prompt formatted text

With the current advent of COSS (Commercial Open Source Software), such accessibility features should become a given

7/15

But fetching is still tiresome, and searching is a time waster (more on this in another part probably)

So let's take a step back and get meta

What is my main tool to get answers right now?

Yep, it starts with "Chat" and ends with "GPT" 😗

8/15

2️⃣ LM to LM interfaces

What's better than a generic #gpt4 LM to answer a precise question about the latest release of a framework?

Another LM trained and updated on this framework's specificities.

What if #chatgpt could ask #chat-nextjs for specific information? 🤯

9/15

This could be the advent of distributed LM and a new version of the web

Responses will always be up to date, with higher quality, and likely to provide references (please do so 🙏)

Want SEO (well PEO)? Be the place to find the answer to a specific question with a prompt

10/15

3️⃣ Knowledge becomes closed source behind prompts 🔒

Smaller LMs would be great, but we still have an issue with large LM (think #gpt)

If large LMs can simply gobble up the knowledge you share openly and replace you, will you still be willing to share it? 🫢

11/15

We already started losing exploration with web 2.0 over personalisation

Do you still go to Youtube in hope of discovering new creators ? If you're leftist leaning, are right leaning news site appearing in your Google results ?

Well, this is worse 😞

12/15

Think Google is a black box ? At least we have a basic understanding of how it works and picks results

In a LLM web you'll never know how you got the info

Actually, most don't care. Reacting to Reddit posts based solely on the headline is already a meme at this point

13/15

And it's already happening, @OpenCage had to put up a blog post to contradict #chatgpt about it's own service. It's own service !

blog.opencagedata.com/post/dont-beli…

Can't fix the result, can't find why #chatgpt thinks this, yet thousands of people take it at face value

14/15

I could go on but this is getting way too convoluted for me

So back to being a simple minded writing a vscode extension

Yay 🥳

15/15

Well what did you expect ?

Of course it's the end of part 4. This extension won't be writing itself while I tweet

If you ❤️ this thread, Like/Retweet the first tweet below if you can

You can also hit @yleflour & 🔔 for more rambling...

... cause there's a lot more coming 😝

https://twitter.com/yleflour/status/1638148557189558273