ChatGPT cannot build a simple working site from scratch | Hacker News Reader

ChatGPT cannot build a simple working site from scratch

13 pointshmerchan22 comments

I gave chatGpt a simple task to build a website that takes name of team members and splits them randomly into teams.

It started off well by making me install python and flask and build a skeleton website. But it could not make it work completely. After a few debugging back and forth it went round and round doing and undoing same mistakes, but could not get the task completed..

Am I doing something wrong?

View on HN

You are doing several things wrong.

1: Which ChatGPT are you using? 3 or 4?

2: You are not clearly describing your problem. Teams? How many should it be splitting? Why do you need a backend? Why isn't this just a simple HTML page? Any examples? Have you clearly thought through what you want?

3: You are not sharing the prompts you used. How do you expect us to help you if you don't share what you tried?

In general, people who cannot communicate clearly are going to have a lot of trouble working with something which requires clear communication.

Your hostile / aggressive tone is not ideal either...

Here's an example session I just had with 4, forgive the formatting:

User Please construct a simple webpage which allows me to add an arbitrary amount of team members, and when I am done adding them I should be able to click a button to randomly assign them into teams.

ChatGPT

I can provide you with a simple HTML and JavaScript code to create the webpage you've described. The following code creates a webpage that allows you to add team members, input the number of teams, and click a button to randomly assign team members into teams.

1. Save the code below as an HTML file (e.g., "team_randomizer.html").

2. Open the HTML file in a web browser to use the webpage.

```html

<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Team Randomizer</title> <style> .result { margin-top: 20px; } </style> </head> <body> <h1>Team Randomizer</h1> <div> <label for="memberName">Member Name:</label> <input type="text" id="memberName" /> <button onclick="addMember()">Add Member</button> </div> <div> <label for="numTeams">Number of Teams:</label> <input type="number" id="numTeams" min="1" /> <button onclick="randomizeTeams()">Randomize Teams</button> </div> <div class="result" id="result"></div> <script> const members = [];

        function addMember() {
            const memberName = document.getElementById('memberName').value.trim();
            if (memberName) {
                members.push(memberName);
                document.getElementById('memberName').value = '';
                displayMembers();
            }
        }

        function displayMembers() {
            const result = document.getElementById('result');
            result.innerHTML = `<p>Team Members: ${members.join(', ')}</p>`;
        }

        function randomizeTeams() {
            const numTeams = parseInt(document.getElementById('numTeams').value, 10);
            if (!numTeams || numTeams < 1) return;

            const shuffled = [...members];
            for (let i = shuffled.length - 1; i > 0; i--) {
                const j = Math.floor(Math.random() * (i + 1));
                [shuffled[i], shuffled[j]] = [shuffled[j], shuffled[i]];
            }

            const teams = [];
            for (let i = 0; i < numTeams; i++) {
                teams[i] = [];
            }

            for (let i = 0; i < shuffled.length; i++) {
                teams[i % numTeams].push(shuffled[i]);
            }

            displayResult(teams);
        }

        function displayResult(teams) {
            const result = document.getElementById('result');
            let html = '';
            for (let i = 0; i < teams.length; i++) {
                html += `<p>Team ${i + 1}: ${teams[i].join(', ')}</p>`;
            }
            result.innerHTML = html;
        }
    </script>

</body> </html> ```

This code will create a simple webpage where you can add team members, specify the number of teams, and randomize the teams. The team members and resulting teams will be displayed on the page.

This whole thing feels like "I installed Linux without reading how to do anything and it ate my hard drive and I couldn't fix my computer." just with large language models. But for some reason, ignorance is celebrated if it makes the AI stuff look bad. I get it people. The hype is annoying, but you're not gonna fight the hype by failing to use the tool and then saying the tool just "cannot build a simple webpage"

It's funny because this is exactly my first experience with Linux in the 90s

> people who cannot communicate clearly are going to have a lot of trouble working with something which requires clear communication

Kind of like programming. Framing the solution to a problem in a clear and logical fashion is the hard part, not writing the code.

my coding skill skyrocketed as my ability to ask better phrased questions in Google, the same is true of chat gpt, though now I feel like a 10x programmer.

Communicating like you?

I just used ChatGPT with gpt4, copied your problem from your post, ran it, and the first result worked just fine.

Might be that you’re using gpt3.5(understandable because not everyone has access), or maybe you asked in a different way than in your post.

Plenty of people have been building simple websites with it, so maybe try to figure out what you’re doing wrong instead of declaring the tool a failure.

GPT3.5 easily spat out a working answer in response to "Write an HTTP and JavaScript page, as laconically as possible, that takes a list of newline separated strings from a textbox, and divides them randomly into `n` groups, where `n` is set with a text box."

Using ChatGPT to code effectively requires internalizing the small sliding context window and correct specification of the problem-space. You can get real work done if you thoughtfully split your project into parts, and you can save time by jumping back to a contextual fork where your program was fully specified.

To get useful responses to prompts:

* start with a clear objective

* include context

* describe criteria to meet expectations

* iterate providing updated context

I used chatGPT-3 yesterday to write a Ruby client that runs code generated from prompts. (Was inspired by the gepl project seen here https://github.com/lmortimer/gepl)

https://github.com/evalgpt/evalgpt

Almost every demo of ChatGPT or other generative AI that goes viral on Twitter nowadays is cherry-picked and/or does not include other relevant attributes that are critical to solving the problem, such as unique prompt engineering.

Thanks for the feedback, please see my prompt snippets here, https://drive.google.com/file/d/1vBbbP6bDkkR4R2PMcgkWN2eCgOO...

basically the chat went on for quite long and the solution kept generating some errors and it came to the point where it was un-doing the fixes it did the day before.

It's (GPT-4) tuned to give stackoverflow-like answers and respond that it can't provide you full working solutions if you ask for something complete. You have to play to its strengths and adapt your workflow to whatever its capabilities are today, for your specific project.

I've tried to get it to do 100% of my coding for smallish projects, and it's pretty frustrating. I'd only do that as a challenge.

I feel like this is like saying I don't use autocomplete because I still have to type...

Until GPT can download and install itself and other DL projects from GitHub without problems on any OS, I'm not going to expect much else from it.

are you trying to get it to spit out the entire solution in one chat? It's easier for answering one-off questions when you are stuck.

I'm in the middle of converting some ruby code to javascript, so asking it something like how to split a string in nodejs into directory and file components results in a quick and correct answer.

Who told you ChatGPT can do this? ChatGPT can assist you in doing this by (potentially) making the process faster (I say potentially because you typically have to know what to ask and how to ask). But you generally have to know what you're doing otherwise you'll be stuck in the hallucination wonderland and dumb q/a loops the result of which will range from bad (code doesn't work) to terrible (code works, but not really the way you intended or has silent bugs).

Tldr: LLMs are like fire. A good servant but a bad master.

IIRC, when GPT4 was released OpenAI demoed the AI taking a drawing of a website and building a working one.

3.5? 4? Is your prompt detailed enough?

In other breaking news, water is wet.

Reply on news.ycombinator.com