Can chatgpt-o1 complete a junior frontend developer's task?

October 19, 2024

When I first embarked on my journey as a junior software engineer (about 14 years ago), I remember working on a frontend project that required me to partially fill a "cloud" SVG with a solid color to represent a fancy progress bar.

Being a junior developer in a new position and eager to prove my skills, I remember frantically googling for solutions and only running into examples of filling entire SVG objects. After tinkering with this problem for some time, I came up with a nifty solution that involved manipulating gradient stops

50%
My Solution

The trick was to use a linear gradient with two stops, one with a transparent color and the other with a solid color. By manipulating the offset of these stops, the gradient would shift and gradually fill the cloud with color as the data loads.

Given the lack of online resources about this specific technique, I remember thinking it was a pretty neat trick and wanting to write a blog post about it, something I never got around to doing until now.

My Solution

Here is a modern example of the same technique I used back then.


const CloudFillHuman: React.FC<CloudFillProps> = ({ percentage }) => {
  /* 
   calculate the offset of the gradient stops based on the passed 
   percentage and render the cloud filled with the gradient 
  */
  const clampedPercentage = Math.max(0, Math.min(percentage, 100));
  const offset = 100 - clampedPercentage;

  return (
    <svg 
      xmlns="http://www.w3.org/2000/svg" 
      width="200" 
      height="200" 
      strokeWidth={0.5} 
      viewBox="0 0 24 24" 
      stroke="#fff"
    >
      /* Define the Gradient with two stops */
      <defs>
        <linearGradient id="cloud-gradient" x1="0%" y1="0%" x2="0%" y2="100%">
          <stop offset={`${offset}%`} style={{stopColor: "#ffffff00", stopOpacity: 0}} />
          <stop offset={`${offset}%`} style={{stopColor: "#5cc6fe", stopOpacity: 1}} />
        </linearGradient>
      </defs>

      /* Define the Shape to be filled */
      <path 
        fill="url(#cloud-gradient)" 
        strokeLinecap="round" 
        strokeLinejoin="round" 
        d="M 2.25 15a4.5 4.5 0 0 0 4.5 4.5
           H 18a3.75 3.75 0 0 0 1.332-7.257 
           3 3 0 0 0-3.758-3.848 5.25 5.25 
           0 0 0-10.233 2.33 A 4.502 4.502 
           0 0 0 2.25 15Z"
      />
    </svg>
  )
};

Given the massive amount of online data that LLM models have been trained on, and the time that has passed since I first solved this problem, I was curious to see how ChatGPT-o1 would solve it today. So I decided to put it to the test.

The Prompt

Here is the prompt I used to test ChatGPT-o1. To be fair I gave it pretty clear instructions, something I wish my boss at the time had done.

I have an SVG in the shape of a cloud, I want you to write a typescript function that takes an integer number between 0 and 100 as input and partially fill the cloud with a solid color. For example, if the input value is 50, I expect half the cloud to be filled with the color blue and the other half to be transparent.

here is the SVG shape ...

O1 thought for 30 seconds and then provided the following solution


const CloudFill: React.FC<CloudWithFillProps> = ({ percentage }) => {
  // Ensure percentage is within 0 to 100
  const clampedPercentage = Math.max(0, Math.min(percentage, 100));
  const yPosition = (24 * (100 - clampedPercentage)) / 100; // Calculate Y position based on percentage

  // Generate a unique ID for the clip path to avoid ID conflicts if multiple clouds are rendered
  const clipPathId = React.useMemo(() => `clip-${Math.random().toString(36).substr(2, 9)}`, []);

  return (
    <svg
      xmlns="http://www.w3.org/2000/svg"
      viewBox="0 0 24 24"
      strokeWidth={0.5}
      stroke="currentColor"
      width={200}
      height={200}
    >
      <defs>
        <clipPath id={clipPathId}>
          {/* Rectangle that determines the filled area based on the percentage */}
          <rect x="0" ={yPosition} width="24" height={24 - yPosition} />
        </clipPath>
      </defs>

      {/* Filled cloud shape clipped to the percentage */}
      <path
        fill="#5cc6fe"
        d="M2.25 15a4.5 4.5 0 0 0 4.5 4.5H18a3.75
           3.75 0 0 0 1.332-7.257 3 3 0 0 0-3.758-3.848
           5.25 5.25 0 0 0-10.233
           2.33A4.502 4.502 0 0 0 2.25 15Z"
        clipPath={`url(#${clipPathId})`}
      />

      {/* Cloud outline */}
      <path
        fill="none"
        strokeLinecap="round"
        strokeLinejoin="round"
        d="M2.25 15a4.5 4.5 0 0 0 4.5 4.5H18a3.75
           3.75 0 0 0 1.332-7.257 3 3 0 0 0-3.758-3.848
           5.25 5.25 0 0 0-10.233
           2.33A4.502 4.502 0 0 0 2.25 15Z"
      />
    </svg>
  );
};

GPT o1 used a ClipPath to partially render a filled cloud behind the outline of the cloud.
Honestly the solution is pretty good, and I am impressed by how well it understood the prompt.

Now let's test the two solutions against each other.

Comparing the Solutions

At first glance, the two solutions look pretty similar. The LLM solution appears to be working as intended. But let's see how they perform when we manipulate the percentage value.

If you use the slider below you will see that the LLM solution does not begin to fill the shape until the slider is above 20% and on the upper-bound it fills the entire shape when the percentage is 80% or above.

50%
My Solution
GPT-o1

The Verdict

So where does this leave us? Are LLMs ready to take over moderately complex frontend tasks?

While the technology is undeniably useful, the answer to wether LLMs are ready to take over engineering tasks without any human supervision is still a resounding No, although this does not mean that some companies will not try to go down this path.

It's hard to extrapolate too much from a single example, but one thing is for certain, software engineering has always been about exactness and precision just as much as it is about creativity and innovation.

While LLMs are great at generating content, answering questions, and empowering developers, getting them to produce consistent, bug-free code on their own will take a major breakthrough in the field.

Perhaps multiple agents iterating on the same code could be a step in the right direction? However I beleive this will continue to be a challenge for the foreseeable future, after all, LLMs are trained on billions of lines of "imperfect" code written by humans and we still haven't solved that problem ourselves.