Convo: a conversational programming language

I have made a GPT that can make copy paste GPT. With code analyze and other tools GPT are quite powerful tools for logic. :rabbit: Plain language logic is incredibly powerful especially in GPT builders. Like ChatGPT plus etc.

1 Like

When these people realize the difference between grammar and syntax within the shell of a structured output and the difference between reasoning and rationality versus bulk generation towards statistical higher likelihood of probable correctiveness they are going to bleep their pants.

I am out of love, but yes 100%. We are at a turning point in reality IMO. A paradigm shift. :rabbit::honeybee:

Instead of a paradigm Iā€™d say you know about 600 billion paradigms shifting into opening eyes pocket LOL all for a simulated touring chatbot and itā€™s crazy man.

So this:

was created from this:

implement this using html+js all in a single page:

Application: Component Demo
Create a page that shows a rich set of bootstrap inspired components.

Define the main content area:
    Add a header titled "Component Showcase".
    Below the header, create a scrollable list layout to display various styled page elements and form inputs.

    Add the following elements to the list. Add a title to each section:
        - A section with an "h1" header, "h2" subheader, and a "p" paragraph to demonstrate text styling.
        - A form with text input, email input, password input, a dropdown menu, checkboxes, and radio buttons.
        - A section with primary and secondary buttons styled using bootstrap-inspired classes.
        - An alert box styled as a success message.
        - A progress bar showing 50% completion.
        - A card with a title, text, and a footer with a link button.

Set the sidebar title to "Components" and add a detailed description of what's being shown to the sidebar.
 
Section: Layout
Create a sidebar on the left side of the screen.
Inside the sidebar, place a title aligned to the left and a toggle button aligned to the right.
When the sidebar is closed, hide the title & sidebar contents, and dock the toggle button to the far left of the screen. Leave enough space for the sidebar to be re-opened.

Create a main content area to the right of the sidebar.
Make sure the main content area takes the remaining space on the screen.

Section: Behavior
To toggle the sidebar:
    If the sidebar is open, then close it:
        Hide the title.
        Dock the toggle button to the far left of the screen, pointing right.
    Otherwise, open the sidebar:
        Show the title.
        Move the toggle button to the right side of the sidebar, pointing left.

Section: Style
Apply a "light mode" theme that's inspired by bootstrap.
Use a chevron for the sidebar toggle button.
Style the form inputs, buttons, and other elements with padding, margins, and consistent font styles.
Ensure hover and focus states are visually distinct for interactive elements.

Iā€™m not saying Iā€™m doing anything the LLMā€™s canā€™t already do. Iā€™m trying to work out how you can build up a list of instructions like this to get very consistent code out of the model. We would classically call that ā€œa programā€

UPDATE:
I asked for that as python and it banged out a Flask server implementation

Just asked for C# and it generated the same exact app but using Blazor which is .NETā€™s page syntax.

Yes, that is an emulator like what folks used to make pseudo code for html and old social media platforms. It is a great idea. I love seeing old tech used for cutting edge ideas. Many in this forum are trying to discover standardization in prompts to make repeatable responses .


I made this today.

Iā€™m assuming you used v0 or something? Iā€™m using any chat client or LLM you want.

See making them beginning parts of the absence isnā€™t the hard part of the hard part is going through and hardening the application making it so that it is consistent and able to do everything that itā€™s supposed to do on repetitively inconsistently in different situations. I spent about 100,000 times not making an app in about like a hundred times actually making an app. I think I have a lot of wisdom I could help you with what youā€™re doing if youā€™d like to chat sometime. You are mutual friends with Jochan right?

Oh I was totally Iā€™m just kidding thatā€™s just like a joke that I made about llms reasoning sorryā€¦

1 Like

Iā€™d love to chat and the feedback/discussion is always welcome. Iā€™m banging out some basic UI examples and constructs right now but Iā€™m about to get into using convo to generate server components. I already did a quick prototype of multiple agents interacting with each other using convo:

If you want to get together and code sometime we can get on whatā€™s app or something and we can share some secrets :call_me_hand::sunglasses:.



Iā€™ve been working on a genealogy app today.

To reset just a bitā€¦ There really is nothing overly magic Iā€™m doing here. Iā€™m building on an observation I madeā€¦ LLMā€™s think in terms of concepts and to the LLM JavaScript, Python, Rust, and every other programming language share similar concepts. This alignment of concepts is what makes it easy for an LLM to translate JavaScript to Python. To the LLM it might as well be translating Spanish to French.

The observation I made is that LLMs understand pseudocode and they map pseudocode to the same concepts they have for all other programming languages. The thing thatā€™s different about pseudocode is that pseudocode lets you essentially do ā€œmagicā€. You can gloss over details in pseudocode as TBD. When you ask the model to translate pseudocode to a language like python it decides all of those TBD details on the fly as its generating.

If you use a tool like v0 or even just Copilot youā€™re essentially writing pseudocode to prompt the generations they emit so again thereā€™s nothing magical that Iā€™m doing which those tools dont. What those tools have you do though is to design chunks of your code using generation which you then stitch a bunch of program code around. None of those tools pretend to be capable of just spitting out a whole application.

My goal is to get rid of all that stitching code and just spit out whole applications. I donā€™t know that the current generation of models is capable of that yet but they soon will be. My philosophy is to always skate to where the puck will be, not where it is. So thatā€™s what Iā€™m doingā€¦

2 Likes

Ironically that comes down to the fundamentals of the pursuit of automated code development to that I think most new developers start on like I did myself when I first got under you know Bert models and you know Chat gpt. The whole concept of giving the system that your parameters and what you wanted to make is like the goal of artificial intelligence and the generation of code from thought. Ironically I think what you will find is that the legitimately the best method is to have a direct conversation with a clear understanding of your actual constraints data types and the functionality that you want to have in a back and forth iterative conversation with a model like clod using an IDE like cursor and it is natural easy and ebbs and flows code thatā€™s gold.

1 Like

Check out websim.ai You owe me 1.

Itā€™s very much true that you need a clear definition of your types and functionality but you donā€™t need the back and forth conversation. In fact the longer the back and forth where the discussion is grounded by actual code, the more off the rails your code will go. If anything you want to keep your back and forth conversations at a super high pseudocode level and once youā€™ve ironed all of the details, then you can have the model generate the underlying code. You might need to make a couple of follow-up turns to repair issues but if you need to do major restructuring you should go back to the pseudocode level.

Thatā€™s just based on my experience and what Iā€™ve found works for me. Everyone has to find what works for them. If youā€™re loving cursor then by all means keep using cursor. Iā€™m sure the cursor guys were told more than once that they were dumb for trying to compete with VSCode and Copilot. Every great idea starts off being criticizedā€¦

1 Like

Iā€™ve seen itā€¦ SynthOS is similar to websim in some waysā€¦

(attachments)

1 Like

It is more about structure and consistent language than true code. I can code AI functions in a lot of weird ways like up thread I shared a BASIC AI. But it just shows logic is key, I see it as a cartoon flip book, in that if I can get a GPT to take the next logical step with the most compact structure I can build and in my methods I put it into an instructional GPT so it not only generates it teaches how to use it. Simple language and precise structure imo is key. Out of hearts but this is a wonderful conversation. :heart::rabbit:

2 Likes